diff options
Diffstat (limited to 'pkg')
528 files changed, 32432 insertions, 9398 deletions
diff --git a/pkg/abi/linux/BUILD b/pkg/abi/linux/BUILD index 1f3c0c687..322d1ccc4 100644 --- a/pkg/abi/linux/BUILD +++ b/pkg/abi/linux/BUILD @@ -17,6 +17,8 @@ go_library( "dev.go", "elf.go", "epoll.go", + "epoll_amd64.go", + "epoll_arm64.go", "errors.go", "eventfd.go", "exec.go", @@ -28,6 +30,7 @@ go_library( "futex.go", "inotify.go", "ioctl.go", + "ioctl_tun.go", "ip.go", "ipc.go", "limits.go", @@ -59,6 +62,7 @@ go_library( "wait.go", "xattr.go", ], + marshal = True, visibility = ["//visibility:public"], deps = [ "//pkg/abi", diff --git a/pkg/abi/linux/elf.go b/pkg/abi/linux/elf.go index 40f0459a0..7c9a02f20 100644 --- a/pkg/abi/linux/elf.go +++ b/pkg/abi/linux/elf.go @@ -102,4 +102,7 @@ const ( // NT_X86_XSTATE is for x86 extended state using xsave. NT_X86_XSTATE = 0x202 + + // NT_ARM_TLS is for ARM TLS register. + NT_ARM_TLS = 0x401 ) diff --git a/pkg/abi/linux/epoll.go b/pkg/abi/linux/epoll.go index 0e881aa3c..1121a1a92 100644 --- a/pkg/abi/linux/epoll.go +++ b/pkg/abi/linux/epoll.go @@ -14,12 +14,9 @@ package linux -// EpollEvent is equivalent to struct epoll_event from epoll(2). -type EpollEvent struct { - Events uint32 - Fd int32 - Data int32 -} +import ( + "gvisor.dev/gvisor/pkg/binary" +) // Event masks. const ( @@ -60,3 +57,6 @@ const ( EPOLL_CTL_DEL = 0x2 EPOLL_CTL_MOD = 0x3 ) + +// SizeOfEpollEvent is the size of EpollEvent struct. +var SizeOfEpollEvent = int(binary.Size(EpollEvent{})) diff --git a/pkg/usermem/usermem_unsafe.go b/pkg/abi/linux/epoll_amd64.go index 876783e78..34ff18009 100644 --- a/pkg/usermem/usermem_unsafe.go +++ b/pkg/abi/linux/epoll_amd64.go @@ -12,16 +12,16 @@ // See the License for the specific language governing permissions and // limitations under the License. -package usermem +package linux -import ( - "unsafe" -) - -// stringFromImmutableBytes is equivalent to string(bs), except that it never -// copies even if escape analysis can't prove that bs does not escape. This is -// only valid if bs is never mutated after stringFromImmutableBytes returns. -func stringFromImmutableBytes(bs []byte) string { - // Compare strings.Builder.String(). - return *(*string)(unsafe.Pointer(&bs)) +// EpollEvent is equivalent to struct epoll_event from epoll(2). +// +// +marshal +type EpollEvent struct { + Events uint32 + // Linux makes struct epoll_event::data a __u64. We represent it as + // [2]int32 because, on amd64, Linux also makes struct epoll_event + // __attribute__((packed)), such that there is no padding between Events + // and Data. + Data [2]int32 } diff --git a/pkg/abi/linux/epoll_arm64.go b/pkg/abi/linux/epoll_arm64.go new file mode 100644 index 000000000..f86c35329 --- /dev/null +++ b/pkg/abi/linux/epoll_arm64.go @@ -0,0 +1,26 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package linux + +// EpollEvent is equivalent to struct epoll_event from epoll(2). +// +// +marshal +type EpollEvent struct { + Events uint32 + // Linux makes struct epoll_event a __u64, necessitating 4 bytes of padding + // here. + _ int32 + Data [2]int32 +} diff --git a/pkg/abi/linux/file.go b/pkg/abi/linux/file.go index c3ab15a4f..055ac1d7c 100644 --- a/pkg/abi/linux/file.go +++ b/pkg/abi/linux/file.go @@ -241,6 +241,8 @@ const ( ) // Statx represents struct statx. +// +// +marshal type Statx struct { Mask uint32 Blksize uint32 @@ -264,6 +266,9 @@ type Statx struct { DevMinor uint32 } +// SizeOfStatx is the size of a Statx struct. +var SizeOfStatx = binary.Size(Statx{}) + // FileMode represents a mode_t. type FileMode uint16 @@ -282,6 +287,11 @@ func (m FileMode) ExtraBits() FileMode { return m &^ (PermissionsMask | FileTypeMask) } +// IsDir returns true if file type represents a directory. +func (m FileMode) IsDir() bool { + return m.FileType() == S_IFDIR +} + // String returns a string representation of m. func (m FileMode) String() string { var s []string diff --git a/pkg/abi/linux/file_amd64.go b/pkg/abi/linux/file_amd64.go index 9d307e840..6b72364ea 100644 --- a/pkg/abi/linux/file_amd64.go +++ b/pkg/abi/linux/file_amd64.go @@ -12,6 +12,8 @@ // See the License for the specific language governing permissions and // limitations under the License. +// +build amd64 + package linux // Constants for open(2). @@ -23,6 +25,8 @@ const ( ) // Stat represents struct stat. +// +// +marshal type Stat struct { Dev uint64 Ino uint64 diff --git a/pkg/abi/linux/file_arm64.go b/pkg/abi/linux/file_arm64.go index 26a54f416..6492c9038 100644 --- a/pkg/abi/linux/file_arm64.go +++ b/pkg/abi/linux/file_arm64.go @@ -12,6 +12,8 @@ // See the License for the specific language governing permissions and // limitations under the License. +// +build arm64 + package linux // Constants for open(2). @@ -23,6 +25,8 @@ const ( ) // Stat represents struct stat. +// +// +marshal type Stat struct { Dev uint64 Ino uint64 diff --git a/pkg/abi/linux/fs.go b/pkg/abi/linux/fs.go index 2c652baa2..158d2db5b 100644 --- a/pkg/abi/linux/fs.go +++ b/pkg/abi/linux/fs.go @@ -38,6 +38,8 @@ const ( ) // Statfs is struct statfs, from uapi/asm-generic/statfs.h. +// +// +marshal type Statfs struct { // Type is one of the filesystem magic values, defined above. Type uint64 diff --git a/pkg/abi/linux/ioctl.go b/pkg/abi/linux/ioctl.go index 0e18db9ef..2062e6a4b 100644 --- a/pkg/abi/linux/ioctl.go +++ b/pkg/abi/linux/ioctl.go @@ -72,3 +72,29 @@ const ( SIOCGMIIPHY = 0x8947 SIOCGMIIREG = 0x8948 ) + +// ioctl(2) directions. Used to calculate requests number. +// Constants from asm-generic/ioctl.h. +const ( + _IOC_NONE = 0 + _IOC_WRITE = 1 + _IOC_READ = 2 +) + +// Constants from asm-generic/ioctl.h. +const ( + _IOC_NRBITS = 8 + _IOC_TYPEBITS = 8 + _IOC_SIZEBITS = 14 + _IOC_DIRBITS = 2 + + _IOC_NRSHIFT = 0 + _IOC_TYPESHIFT = _IOC_NRSHIFT + _IOC_NRBITS + _IOC_SIZESHIFT = _IOC_TYPESHIFT + _IOC_TYPEBITS + _IOC_DIRSHIFT = _IOC_SIZESHIFT + _IOC_SIZEBITS +) + +// IOC outputs the result of _IOC macro in asm-generic/ioctl.h. +func IOC(dir, typ, nr, size uint32) uint32 { + return uint32(dir)<<_IOC_DIRSHIFT | typ<<_IOC_TYPESHIFT | nr<<_IOC_NRSHIFT | size<<_IOC_SIZESHIFT +} diff --git a/pkg/abi/linux/ioctl_tun.go b/pkg/abi/linux/ioctl_tun.go new file mode 100644 index 000000000..c59c9c136 --- /dev/null +++ b/pkg/abi/linux/ioctl_tun.go @@ -0,0 +1,29 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package linux + +// ioctl(2) request numbers from linux/if_tun.h +var ( + TUNSETIFF = IOC(_IOC_WRITE, 'T', 202, 4) + TUNGETIFF = IOC(_IOC_READ, 'T', 210, 4) +) + +// Flags from net/if_tun.h +const ( + IFF_TUN = 0x0001 + IFF_TAP = 0x0002 + IFF_NO_PI = 0x1000 + IFF_NOFILTER = 0x1000 +) diff --git a/pkg/abi/linux/netfilter.go b/pkg/abi/linux/netfilter.go index bbc4df74c..a8d4f9d69 100644 --- a/pkg/abi/linux/netfilter.go +++ b/pkg/abi/linux/netfilter.go @@ -158,10 +158,32 @@ type IPTIP struct { // Flags define matching behavior for the IP header. Flags uint8 - // InverseFlags invert the meaning of fields in struct IPTIP. + // InverseFlags invert the meaning of fields in struct IPTIP. See the + // IPT_INV_* flags. InverseFlags uint8 } +// Flags in IPTIP.InverseFlags. Corresponding constants are in +// include/uapi/linux/netfilter_ipv4/ip_tables.h. +const ( + // Invert the meaning of InputInterface. + IPT_INV_VIA_IN = 0x01 + // Invert the meaning of OutputInterface. + IPT_INV_VIA_OUT = 0x02 + // Unclear what this is, as no references to it exist in the kernel. + IPT_INV_TOS = 0x04 + // Invert the meaning of Src. + IPT_INV_SRCIP = 0x08 + // Invert the meaning of Dst. + IPT_INV_DSTIP = 0x10 + // Invert the meaning of the IPT_F_FRAG flag. + IPT_INV_FRAG = 0x20 + // Invert the meaning of the Protocol field. + IPT_INV_PROTO = 0x40 + // Enable all flags. + IPT_INV_MASK = 0x7F +) + // SizeOfIPTIP is the size of an IPTIP. const SizeOfIPTIP = 84 @@ -225,11 +247,14 @@ type XTEntryTarget struct { // SizeOfXTEntryTarget is the size of an XTEntryTarget. const SizeOfXTEntryTarget = 32 -// XTStandardTarget is a builtin target, one of ACCEPT, DROP, JUMP, QUEUE, or -// RETURN. It corresponds to struct xt_standard_target in +// XTStandardTarget is a built-in target, one of ACCEPT, DROP, JUMP, QUEUE, +// RETURN, or jump. It corresponds to struct xt_standard_target in // include/uapi/linux/netfilter/x_tables.h. type XTStandardTarget struct { - Target XTEntryTarget + Target XTEntryTarget + // A positive verdict indicates a jump, and is the offset from the + // start of the table to jump to. A negative value means one of the + // other built-in targets. Verdict int32 _ [4]byte } @@ -250,6 +275,50 @@ type XTErrorTarget struct { // SizeOfXTErrorTarget is the size of an XTErrorTarget. const SizeOfXTErrorTarget = 64 +// Flag values for NfNATIPV4Range. The values indicate whether to map +// protocol specific part(ports) or IPs. It corresponds to values in +// include/uapi/linux/netfilter/nf_nat.h. +const ( + NF_NAT_RANGE_MAP_IPS = 1 << 0 + NF_NAT_RANGE_PROTO_SPECIFIED = 1 << 1 + NF_NAT_RANGE_PROTO_RANDOM = 1 << 2 + NF_NAT_RANGE_PERSISTENT = 1 << 3 + NF_NAT_RANGE_PROTO_RANDOM_FULLY = 1 << 4 + NF_NAT_RANGE_PROTO_RANDOM_ALL = (NF_NAT_RANGE_PROTO_RANDOM | NF_NAT_RANGE_PROTO_RANDOM_FULLY) + NF_NAT_RANGE_MASK = (NF_NAT_RANGE_MAP_IPS | + NF_NAT_RANGE_PROTO_SPECIFIED | NF_NAT_RANGE_PROTO_RANDOM | + NF_NAT_RANGE_PERSISTENT | NF_NAT_RANGE_PROTO_RANDOM_FULLY) +) + +// NfNATIPV4Range corresponds to struct nf_nat_ipv4_range +// in include/uapi/linux/netfilter/nf_nat.h. The fields are in +// network byte order. +type NfNATIPV4Range struct { + Flags uint32 + MinIP [4]byte + MaxIP [4]byte + MinPort uint16 + MaxPort uint16 +} + +// NfNATIPV4MultiRangeCompat corresponds to struct +// nf_nat_ipv4_multi_range_compat in include/uapi/linux/netfilter/nf_nat.h. +type NfNATIPV4MultiRangeCompat struct { + RangeSize uint32 + RangeIPV4 NfNATIPV4Range +} + +// XTRedirectTarget triggers a redirect when reached. +// Adding 4 bytes of padding to make the struct 8 byte aligned. +type XTRedirectTarget struct { + Target XTEntryTarget + NfRange NfNATIPV4MultiRangeCompat + _ [4]byte +} + +// SizeOfXTRedirectTarget is the size of an XTRedirectTarget. +const SizeOfXTRedirectTarget = 56 + // IPTGetinfo is the argument for the IPT_SO_GET_INFO sockopt. It corresponds // to struct ipt_getinfo in include/uapi/linux/netfilter_ipv4/ip_tables.h. type IPTGetinfo struct { @@ -440,3 +509,44 @@ const ( // Enable all flags. XT_UDP_INV_MASK = 0x03 ) + +// IPTOwnerInfo holds data for matching packets with owner. It corresponds +// to struct ipt_owner_info in libxt_owner.c of iptables binary. +type IPTOwnerInfo struct { + // UID is user id which created the packet. + UID uint32 + + // GID is group id which created the packet. + GID uint32 + + // PID is process id of the process which created the packet. + PID uint32 + + // SID is session id which created the packet. + SID uint32 + + // Comm is the command name which created the packet. + Comm [16]byte + + // Match is used to match UID/GID of the socket. See the + // XT_OWNER_* flags below. + Match uint8 + + // Invert flips the meaning of Match field. + Invert uint8 +} + +// SizeOfIPTOwnerInfo is the size of an XTOwnerMatchInfo. +const SizeOfIPTOwnerInfo = 34 + +// Flags in IPTOwnerInfo.Match. Corresponding constants are in +// include/uapi/linux/netfilter/xt_owner.h. +const ( + // Match the UID of the packet. + XT_OWNER_UID = 1 << 0 + // Match the GID of the packet. + XT_OWNER_GID = 1 << 1 + // Match if the socket exists for the packet. Forwarded + // packets do not have an associated socket. + XT_OWNER_SOCKET = 1 << 2 +) diff --git a/pkg/abi/linux/netfilter_test.go b/pkg/abi/linux/netfilter_test.go index 21e237f92..565dd550e 100644 --- a/pkg/abi/linux/netfilter_test.go +++ b/pkg/abi/linux/netfilter_test.go @@ -29,6 +29,7 @@ func TestSizes(t *testing.T) { {IPTGetEntries{}, SizeOfIPTGetEntries}, {IPTGetinfo{}, SizeOfIPTGetinfo}, {IPTIP{}, SizeOfIPTIP}, + {IPTOwnerInfo{}, SizeOfIPTOwnerInfo}, {IPTReplace{}, SizeOfIPTReplace}, {XTCounters{}, SizeOfXTCounters}, {XTEntryMatch{}, SizeOfXTEntryMatch}, diff --git a/pkg/abi/linux/signal.go b/pkg/abi/linux/signal.go index c69b04ea9..1c330e763 100644 --- a/pkg/abi/linux/signal.go +++ b/pkg/abi/linux/signal.go @@ -115,6 +115,8 @@ const ( ) // SignalSet is a signal mask with a bit corresponding to each signal. +// +// +marshal type SignalSet uint64 // SignalSetSize is the size in bytes of a SignalSet. diff --git a/pkg/abi/linux/socket.go b/pkg/abi/linux/socket.go index 766ee4014..4a14ef691 100644 --- a/pkg/abi/linux/socket.go +++ b/pkg/abi/linux/socket.go @@ -411,6 +411,15 @@ type ControlMessageCredentials struct { GID uint32 } +// A ControlMessageIPPacketInfo is IP_PKTINFO socket control message. +// +// ControlMessageIPPacketInfo represents struct in_pktinfo from linux/in.h. +type ControlMessageIPPacketInfo struct { + NIC int32 + LocalAddr InetAddr + DestinationAddr InetAddr +} + // SizeOfControlMessageCredentials is the binary size of a // ControlMessageCredentials struct. var SizeOfControlMessageCredentials = int(binary.Size(ControlMessageCredentials{})) @@ -431,6 +440,10 @@ const SizeOfControlMessageTOS = 1 // SizeOfControlMessageTClass is the size of an IPV6_TCLASS control message. const SizeOfControlMessageTClass = 4 +// SizeOfControlMessageIPPacketInfo is the size of an IP_PKTINFO +// control message. +const SizeOfControlMessageIPPacketInfo = 12 + // SCM_MAX_FD is the maximum number of FDs accepted in a single sendmsg call. // From net/scm.h. const SCM_MAX_FD = 253 diff --git a/pkg/abi/linux/time.go b/pkg/abi/linux/time.go index 5c5a58cd4..e6860ed49 100644 --- a/pkg/abi/linux/time.go +++ b/pkg/abi/linux/time.go @@ -101,6 +101,8 @@ func NsecToTimeT(nsec int64) TimeT { } // Timespec represents struct timespec in <time.h>. +// +// +marshal type Timespec struct { Sec int64 Nsec int64 @@ -155,6 +157,8 @@ func DurationToTimespec(dur time.Duration) Timespec { const SizeOfTimeval = 16 // Timeval represents struct timeval in <time.h>. +// +// +marshal type Timeval struct { Sec int64 Usec int64 @@ -228,6 +232,8 @@ type Tms struct { type TimerID int32 // StatxTimestamp represents struct statx_timestamp. +// +// +marshal type StatxTimestamp struct { Sec int64 Nsec uint32 @@ -256,6 +262,8 @@ func NsecToStatxTimestamp(nsec int64) (ts StatxTimestamp) { } // Utime represents struct utimbuf used by utimes(2). +// +// +marshal type Utime struct { Actime int64 Modtime int64 diff --git a/pkg/abi/linux/xattr.go b/pkg/abi/linux/xattr.go index a3b6406fa..99180b208 100644 --- a/pkg/abi/linux/xattr.go +++ b/pkg/abi/linux/xattr.go @@ -18,6 +18,7 @@ package linux const ( XATTR_NAME_MAX = 255 XATTR_SIZE_MAX = 65536 + XATTR_LIST_MAX = 65536 XATTR_CREATE = 1 XATTR_REPLACE = 2 diff --git a/pkg/atomicbitops/BUILD b/pkg/atomicbitops/BUILD index 3948074ba..1a30f6967 100644 --- a/pkg/atomicbitops/BUILD +++ b/pkg/atomicbitops/BUILD @@ -5,10 +5,10 @@ package(licenses = ["notice"]) go_library( name = "atomicbitops", srcs = [ - "atomic_bitops.go", - "atomic_bitops_amd64.s", - "atomic_bitops_arm64.s", - "atomic_bitops_common.go", + "atomicbitops.go", + "atomicbitops_amd64.s", + "atomicbitops_arm64.s", + "atomicbitops_noasm.go", ], visibility = ["//:sandbox"], ) @@ -16,7 +16,7 @@ go_library( go_test( name = "atomicbitops_test", size = "small", - srcs = ["atomic_bitops_test.go"], + srcs = ["atomicbitops_test.go"], library = ":atomicbitops", deps = ["//pkg/sync"], ) diff --git a/pkg/atomicbitops/atomic_bitops.go b/pkg/atomicbitops/atomicbitops.go index fcc41a9ea..1be081719 100644 --- a/pkg/atomicbitops/atomic_bitops.go +++ b/pkg/atomicbitops/atomicbitops.go @@ -14,47 +14,34 @@ // +build amd64 arm64 -// Package atomicbitops provides basic bitwise operations in an atomic way. -// The implementation on amd64 leverages the LOCK prefix directly instead of -// relying on the generic cas primitives, and the arm64 leverages the LDAXR -// and STLXR pair primitives. +// Package atomicbitops provides extensions to the sync/atomic package. // -// WARNING: the bitwise ops provided in this package doesn't imply any memory -// ordering. Using them to construct locks must employ proper memory barriers. +// All read-modify-write operations implemented by this package have +// acquire-release memory ordering (like sync/atomic). package atomicbitops -// AndUint32 atomically applies bitwise and operation to *addr with val. +// AndUint32 atomically applies bitwise AND operation to *addr with val. func AndUint32(addr *uint32, val uint32) -// OrUint32 atomically applies bitwise or operation to *addr with val. +// OrUint32 atomically applies bitwise OR operation to *addr with val. func OrUint32(addr *uint32, val uint32) -// XorUint32 atomically applies bitwise xor operation to *addr with val. +// XorUint32 atomically applies bitwise XOR operation to *addr with val. func XorUint32(addr *uint32, val uint32) // CompareAndSwapUint32 is like sync/atomic.CompareAndSwapUint32, but returns // the value previously stored at addr. func CompareAndSwapUint32(addr *uint32, old, new uint32) uint32 -// AndUint64 atomically applies bitwise and operation to *addr with val. +// AndUint64 atomically applies bitwise AND operation to *addr with val. func AndUint64(addr *uint64, val uint64) -// OrUint64 atomically applies bitwise or operation to *addr with val. +// OrUint64 atomically applies bitwise OR operation to *addr with val. func OrUint64(addr *uint64, val uint64) -// XorUint64 atomically applies bitwise xor operation to *addr with val. +// XorUint64 atomically applies bitwise XOR operation to *addr with val. func XorUint64(addr *uint64, val uint64) // CompareAndSwapUint64 is like sync/atomic.CompareAndSwapUint64, but returns // the value previously stored at addr. func CompareAndSwapUint64(addr *uint64, old, new uint64) uint64 - -// IncUnlessZeroInt32 increments the value stored at the given address and -// returns true; unless the value stored in the pointer is zero, in which case -// it is left unmodified and false is returned. -func IncUnlessZeroInt32(addr *int32) bool - -// DecUnlessOneInt32 decrements the value stored at the given address and -// returns true; unless the value stored in the pointer is 1, in which case it -// is left unmodified and false is returned. -func DecUnlessOneInt32(addr *int32) bool diff --git a/pkg/atomicbitops/atomic_bitops_amd64.s b/pkg/atomicbitops/atomicbitops_amd64.s index db0972001..54c887ee5 100644 --- a/pkg/atomicbitops/atomic_bitops_amd64.s +++ b/pkg/atomicbitops/atomicbitops_amd64.s @@ -75,41 +75,3 @@ TEXT ·CompareAndSwapUint64(SB),$0-32 CMPXCHGQ DX, 0(DI) MOVQ AX, ret+24(FP) RET - -TEXT ·IncUnlessZeroInt32(SB),NOSPLIT,$0-9 - MOVQ addr+0(FP), DI - MOVL 0(DI), AX - -retry: - TESTL AX, AX - JZ fail - LEAL 1(AX), DX - LOCK - CMPXCHGL DX, 0(DI) - JNZ retry - - SETEQ ret+8(FP) - RET - -fail: - MOVB AX, ret+8(FP) - RET - -TEXT ·DecUnlessOneInt32(SB),NOSPLIT,$0-9 - MOVQ addr+0(FP), DI - MOVL 0(DI), AX - -retry: - LEAL -1(AX), DX - TESTL DX, DX - JZ fail - LOCK - CMPXCHGL DX, 0(DI) - JNZ retry - - SETEQ ret+8(FP) - RET - -fail: - MOVB DX, ret+8(FP) - RET diff --git a/pkg/atomicbitops/atomic_bitops_arm64.s b/pkg/atomicbitops/atomicbitops_arm64.s index 97f8808c1..5c780851b 100644 --- a/pkg/atomicbitops/atomic_bitops_arm64.s +++ b/pkg/atomicbitops/atomicbitops_arm64.s @@ -50,7 +50,6 @@ TEXT ·CompareAndSwapUint32(SB),$0-20 MOVD addr+0(FP), R0 MOVW old+8(FP), R1 MOVW new+12(FP), R2 - again: LDAXRW (R0), R3 CMPW R1, R3 @@ -95,7 +94,6 @@ TEXT ·CompareAndSwapUint64(SB),$0-32 MOVD addr+0(FP), R0 MOVD old+8(FP), R1 MOVD new+16(FP), R2 - again: LDAXR (R0), R3 CMP R1, R3 @@ -105,35 +103,3 @@ again: done: MOVD R3, prev+24(FP) RET - -TEXT ·IncUnlessZeroInt32(SB),NOSPLIT,$0-9 - MOVD addr+0(FP), R0 - -again: - LDAXRW (R0), R1 - CBZ R1, fail - ADDW $1, R1 - STLXRW R1, (R0), R2 - CBNZ R2, again - MOVW $1, R2 - MOVB R2, ret+8(FP) - RET -fail: - MOVB ZR, ret+8(FP) - RET - -TEXT ·DecUnlessOneInt32(SB),NOSPLIT,$0-9 - MOVD addr+0(FP), R0 - -again: - LDAXRW (R0), R1 - SUBSW $1, R1, R1 - BEQ fail - STLXRW R1, (R0), R2 - CBNZ R2, again - MOVW $1, R2 - MOVB R2, ret+8(FP) - RET -fail: - MOVB ZR, ret+8(FP) - RET diff --git a/pkg/atomicbitops/atomic_bitops_common.go b/pkg/atomicbitops/atomicbitops_noasm.go index 85163ad62..3b2898256 100644 --- a/pkg/atomicbitops/atomic_bitops_common.go +++ b/pkg/atomicbitops/atomicbitops_noasm.go @@ -20,7 +20,6 @@ import ( "sync/atomic" ) -// AndUint32 atomically applies bitwise and operation to *addr with val. func AndUint32(addr *uint32, val uint32) { for { o := atomic.LoadUint32(addr) @@ -31,7 +30,6 @@ func AndUint32(addr *uint32, val uint32) { } } -// OrUint32 atomically applies bitwise or operation to *addr with val. func OrUint32(addr *uint32, val uint32) { for { o := atomic.LoadUint32(addr) @@ -42,7 +40,6 @@ func OrUint32(addr *uint32, val uint32) { } } -// XorUint32 atomically applies bitwise xor operation to *addr with val. func XorUint32(addr *uint32, val uint32) { for { o := atomic.LoadUint32(addr) @@ -53,8 +50,6 @@ func XorUint32(addr *uint32, val uint32) { } } -// CompareAndSwapUint32 is like sync/atomic.CompareAndSwapUint32, but returns -// the value previously stored at addr. func CompareAndSwapUint32(addr *uint32, old, new uint32) (prev uint32) { for { prev = atomic.LoadUint32(addr) @@ -67,7 +62,6 @@ func CompareAndSwapUint32(addr *uint32, old, new uint32) (prev uint32) { } } -// AndUint64 atomically applies bitwise and operation to *addr with val. func AndUint64(addr *uint64, val uint64) { for { o := atomic.LoadUint64(addr) @@ -78,7 +72,6 @@ func AndUint64(addr *uint64, val uint64) { } } -// OrUint64 atomically applies bitwise or operation to *addr with val. func OrUint64(addr *uint64, val uint64) { for { o := atomic.LoadUint64(addr) @@ -89,7 +82,6 @@ func OrUint64(addr *uint64, val uint64) { } } -// XorUint64 atomically applies bitwise xor operation to *addr with val. func XorUint64(addr *uint64, val uint64) { for { o := atomic.LoadUint64(addr) @@ -100,8 +92,6 @@ func XorUint64(addr *uint64, val uint64) { } } -// CompareAndSwapUint64 is like sync/atomic.CompareAndSwapUint64, but returns -// the value previously stored at addr. func CompareAndSwapUint64(addr *uint64, old, new uint64) (prev uint64) { for { prev = atomic.LoadUint64(addr) @@ -113,35 +103,3 @@ func CompareAndSwapUint64(addr *uint64, old, new uint64) (prev uint64) { } } } - -// IncUnlessZeroInt32 increments the value stored at the given address and -// returns true; unless the value stored in the pointer is zero, in which case -// it is left unmodified and false is returned. -func IncUnlessZeroInt32(addr *int32) bool { - for { - v := atomic.LoadInt32(addr) - if v == 0 { - return false - } - - if atomic.CompareAndSwapInt32(addr, v, v+1) { - return true - } - } -} - -// DecUnlessOneInt32 decrements the value stored at the given address and -// returns true; unless the value stored in the pointer is 1, in which case it -// is left unmodified and false is returned. -func DecUnlessOneInt32(addr *int32) bool { - for { - v := atomic.LoadInt32(addr) - if v == 1 { - return false - } - - if atomic.CompareAndSwapInt32(addr, v, v-1) { - return true - } - } -} diff --git a/pkg/atomicbitops/atomic_bitops_test.go b/pkg/atomicbitops/atomicbitops_test.go index 9466d3e23..73af71bb4 100644 --- a/pkg/atomicbitops/atomic_bitops_test.go +++ b/pkg/atomicbitops/atomicbitops_test.go @@ -196,67 +196,3 @@ func TestCompareAndSwapUint64(t *testing.T) { } } } - -func TestIncUnlessZeroInt32(t *testing.T) { - for _, test := range []struct { - initial int32 - final int32 - ret bool - }{ - { - initial: 0, - final: 0, - ret: false, - }, - { - initial: 1, - final: 2, - ret: true, - }, - { - initial: 2, - final: 3, - ret: true, - }, - } { - val := test.initial - if got, want := IncUnlessZeroInt32(&val), test.ret; got != want { - t.Errorf("For initial value of %d: incorrect return value: got %v, wanted %v", test.initial, got, want) - } - if got, want := val, test.final; got != want { - t.Errorf("For initial value of %d: incorrect final value: got %d, wanted %d", test.initial, got, want) - } - } -} - -func TestDecUnlessOneInt32(t *testing.T) { - for _, test := range []struct { - initial int32 - final int32 - ret bool - }{ - { - initial: 0, - final: -1, - ret: true, - }, - { - initial: 1, - final: 1, - ret: false, - }, - { - initial: 2, - final: 1, - ret: true, - }, - } { - val := test.initial - if got, want := DecUnlessOneInt32(&val), test.ret; got != want { - t.Errorf("For initial value of %d: incorrect return value: got %v, wanted %v", test.initial, got, want) - } - if got, want := val, test.final; got != want { - t.Errorf("For initial value of %d: incorrect final value: got %d, wanted %d", test.initial, got, want) - } - } -} diff --git a/pkg/binary/binary.go b/pkg/binary/binary.go index 631785f7b..25065aef9 100644 --- a/pkg/binary/binary.go +++ b/pkg/binary/binary.go @@ -254,3 +254,13 @@ func WriteUint64(w io.Writer, order binary.ByteOrder, num uint64) error { _, err := w.Write(buf) return err } + +// AlignUp rounds a length up to an alignment. align must be a power of 2. +func AlignUp(length int, align uint) int { + return (length + int(align) - 1) & ^(int(align) - 1) +} + +// AlignDown rounds a length down to an alignment. align must be a power of 2. +func AlignDown(length int, align uint) int { + return length & ^(int(align) - 1) +} diff --git a/pkg/bits/bits_template.go b/pkg/bits/bits_template.go index 93a435b80..998645388 100644 --- a/pkg/bits/bits_template.go +++ b/pkg/bits/bits_template.go @@ -42,3 +42,11 @@ func Mask(is ...int) T { func MaskOf(i int) T { return T(1) << T(i) } + +// IsPowerOfTwo returns true if v is power of 2. +func IsPowerOfTwo(v T) bool { + if v == 0 { + return false + } + return v&(v-1) == 0 +} diff --git a/pkg/bits/uint64_test.go b/pkg/bits/uint64_test.go index 1b018d808..193d1ebcd 100644 --- a/pkg/bits/uint64_test.go +++ b/pkg/bits/uint64_test.go @@ -114,3 +114,21 @@ func TestIsOn(t *testing.T) { } } } + +func TestIsPowerOfTwo(t *testing.T) { + for _, tc := range []struct { + v uint64 + want bool + }{ + {v: 0, want: false}, + {v: 1, want: true}, + {v: 2, want: true}, + {v: 3, want: false}, + {v: 4, want: true}, + {v: 5, want: false}, + } { + if got := IsPowerOfTwo64(tc.v); got != tc.want { + t.Errorf("IsPowerOfTwo(%d) = %t, want: %t", tc.v, got, tc.want) + } + } +} diff --git a/pkg/buffer/BUILD b/pkg/buffer/BUILD new file mode 100644 index 000000000..dcd086298 --- /dev/null +++ b/pkg/buffer/BUILD @@ -0,0 +1,43 @@ +load("//tools:defs.bzl", "go_library", "go_test") +load("//tools/go_generics:defs.bzl", "go_template_instance") + +package(licenses = ["notice"]) + +go_template_instance( + name = "buffer_list", + out = "buffer_list.go", + package = "buffer", + prefix = "buffer", + template = "//pkg/ilist:generic_list", + types = { + "Element": "*buffer", + "Linker": "*buffer", + }, +) + +go_library( + name = "buffer", + srcs = [ + "buffer.go", + "buffer_list.go", + "safemem.go", + "view.go", + "view_unsafe.go", + ], + visibility = ["//visibility:public"], + deps = [ + "//pkg/log", + "//pkg/safemem", + ], +) + +go_test( + name = "buffer_test", + size = "small", + srcs = [ + "safemem_test.go", + "view_test.go", + ], + library = ":buffer", + deps = ["//pkg/safemem"], +) diff --git a/pkg/buffer/buffer.go b/pkg/buffer/buffer.go new file mode 100644 index 000000000..c6d089fd9 --- /dev/null +++ b/pkg/buffer/buffer.go @@ -0,0 +1,94 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// Package buffer provides the implementation of a buffer view. +// +// A view is an flexible buffer, backed by a pool, supporting the safecopy +// operations natively as well as the ability to grow via either prepend or +// append, as well as shrink. +package buffer + +import ( + "sync" +) + +const bufferSize = 8144 // See below. + +// buffer encapsulates a queueable byte buffer. +// +// Note that the total size is slightly less than two pages. This is done +// intentionally to ensure that the buffer object aligns with runtime +// internals. We have no hard size or alignment requirements. This two page +// size will effectively minimize internal fragmentation, but still have a +// large enough chunk to limit excessive segmentation. +// +// +stateify savable +type buffer struct { + data [bufferSize]byte + read int + write int + bufferEntry +} + +// reset resets internal data. +// +// This must be called before returning the buffer to the pool. +func (b *buffer) Reset() { + b.read = 0 + b.write = 0 +} + +// Full indicates the buffer is full. +// +// This indicates there is no capacity left to write. +func (b *buffer) Full() bool { + return b.write == len(b.data) +} + +// ReadSize returns the number of bytes available for reading. +func (b *buffer) ReadSize() int { + return b.write - b.read +} + +// ReadMove advances the read index by the given amount. +func (b *buffer) ReadMove(n int) { + b.read += n +} + +// ReadSlice returns the read slice for this buffer. +func (b *buffer) ReadSlice() []byte { + return b.data[b.read:b.write] +} + +// WriteSize returns the number of bytes available for writing. +func (b *buffer) WriteSize() int { + return len(b.data) - b.write +} + +// WriteMove advances the write index by the given amount. +func (b *buffer) WriteMove(n int) { + b.write += n +} + +// WriteSlice returns the write slice for this buffer. +func (b *buffer) WriteSlice() []byte { + return b.data[b.write:] +} + +// bufferPool is a pool for buffers. +var bufferPool = sync.Pool{ + New: func() interface{} { + return new(buffer) + }, +} diff --git a/pkg/buffer/safemem.go b/pkg/buffer/safemem.go new file mode 100644 index 000000000..0e5b86344 --- /dev/null +++ b/pkg/buffer/safemem.go @@ -0,0 +1,129 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package buffer + +import ( + "gvisor.dev/gvisor/pkg/safemem" +) + +// WriteBlock returns this buffer as a write Block. +func (b *buffer) WriteBlock() safemem.Block { + return safemem.BlockFromSafeSlice(b.WriteSlice()) +} + +// ReadBlock returns this buffer as a read Block. +func (b *buffer) ReadBlock() safemem.Block { + return safemem.BlockFromSafeSlice(b.ReadSlice()) +} + +// WriteFromBlocks implements safemem.Writer.WriteFromBlocks. +// +// This will advance the write index. +func (v *View) WriteFromBlocks(srcs safemem.BlockSeq) (uint64, error) { + need := int(srcs.NumBytes()) + if need == 0 { + return 0, nil + } + + var ( + dst safemem.BlockSeq + blocks []safemem.Block + ) + + // Need at least one buffer. + firstBuf := v.data.Back() + if firstBuf == nil { + firstBuf = bufferPool.Get().(*buffer) + v.data.PushBack(firstBuf) + } + + // Does the last block have sufficient capacity alone? + if l := firstBuf.WriteSize(); l >= need { + dst = safemem.BlockSeqOf(firstBuf.WriteBlock()) + } else { + // Append blocks until sufficient. + need -= l + blocks = append(blocks, firstBuf.WriteBlock()) + for need > 0 { + emptyBuf := bufferPool.Get().(*buffer) + v.data.PushBack(emptyBuf) + need -= emptyBuf.WriteSize() + blocks = append(blocks, emptyBuf.WriteBlock()) + } + dst = safemem.BlockSeqFromSlice(blocks) + } + + // Perform the copy. + n, err := safemem.CopySeq(dst, srcs) + v.size += int64(n) + + // Update all indices. + for left := int(n); left > 0; firstBuf = firstBuf.Next() { + if l := firstBuf.WriteSize(); left >= l { + firstBuf.WriteMove(l) // Whole block. + left -= l + } else { + firstBuf.WriteMove(left) // Partial block. + left = 0 + } + } + + return n, err +} + +// ReadToBlocks implements safemem.Reader.ReadToBlocks. +// +// This will not advance the read index; the caller should follow +// this call with a call to TrimFront in order to remove the read +// data from the buffer. This is done to support pipe sematics. +func (v *View) ReadToBlocks(dsts safemem.BlockSeq) (uint64, error) { + need := int(dsts.NumBytes()) + if need == 0 { + return 0, nil + } + + var ( + src safemem.BlockSeq + blocks []safemem.Block + ) + + firstBuf := v.data.Front() + if firstBuf == nil { + return 0, nil // No EOF. + } + + // Is all the data in a single block? + if l := firstBuf.ReadSize(); l >= need { + src = safemem.BlockSeqOf(firstBuf.ReadBlock()) + } else { + // Build a list of all the buffers. + need -= l + blocks = append(blocks, firstBuf.ReadBlock()) + for buf := firstBuf.Next(); buf != nil && need > 0; buf = buf.Next() { + need -= buf.ReadSize() + blocks = append(blocks, buf.ReadBlock()) + } + src = safemem.BlockSeqFromSlice(blocks) + } + + // Perform the copy. + n, err := safemem.CopySeq(dsts, src) + + // See above: we would normally advance the read index here, but we + // don't do that in order to support pipe semantics. We rely on a + // separate call to TrimFront() in this case. + + return n, err +} diff --git a/pkg/buffer/safemem_test.go b/pkg/buffer/safemem_test.go new file mode 100644 index 000000000..47f357e0c --- /dev/null +++ b/pkg/buffer/safemem_test.go @@ -0,0 +1,170 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package buffer + +import ( + "bytes" + "strings" + "testing" + + "gvisor.dev/gvisor/pkg/safemem" +) + +func TestSafemem(t *testing.T) { + testCases := []struct { + name string + input string + output string + readLen int + op func(*View) + }{ + // Basic coverage. + { + name: "short", + input: "010", + output: "010", + }, + { + name: "long", + input: "0" + strings.Repeat("1", bufferSize) + "0", + output: "0" + strings.Repeat("1", bufferSize) + "0", + }, + { + name: "short-read", + input: "0", + readLen: 100, // > size. + output: "0", + }, + { + name: "zero-read", + input: "0", + output: "", + }, + { + name: "read-empty", + input: "", + readLen: 1, // > size. + output: "", + }, + + // Ensure offsets work. + { + name: "offsets-short", + input: "012", + output: "2", + op: func(v *View) { + v.TrimFront(2) + }, + }, + { + name: "offsets-long0", + input: "0" + strings.Repeat("1", bufferSize) + "0", + output: strings.Repeat("1", bufferSize) + "0", + op: func(v *View) { + v.TrimFront(1) + }, + }, + { + name: "offsets-long1", + input: "0" + strings.Repeat("1", bufferSize) + "0", + output: strings.Repeat("1", bufferSize-1) + "0", + op: func(v *View) { + v.TrimFront(2) + }, + }, + { + name: "offsets-long2", + input: "0" + strings.Repeat("1", bufferSize) + "0", + output: "10", + op: func(v *View) { + v.TrimFront(bufferSize) + }, + }, + + // Ensure truncation works. + { + name: "truncate-short", + input: "012", + output: "01", + op: func(v *View) { + v.Truncate(2) + }, + }, + { + name: "truncate-long0", + input: "0" + strings.Repeat("1", bufferSize) + "0", + output: "0" + strings.Repeat("1", bufferSize), + op: func(v *View) { + v.Truncate(bufferSize + 1) + }, + }, + { + name: "truncate-long1", + input: "0" + strings.Repeat("1", bufferSize) + "0", + output: "0" + strings.Repeat("1", bufferSize-1), + op: func(v *View) { + v.Truncate(bufferSize) + }, + }, + { + name: "truncate-long2", + input: "0" + strings.Repeat("1", bufferSize) + "0", + output: "01", + op: func(v *View) { + v.Truncate(2) + }, + }, + } + + for _, tc := range testCases { + t.Run(tc.name, func(t *testing.T) { + // Construct the new view. + var view View + bs := safemem.BlockSeqOf(safemem.BlockFromSafeSlice([]byte(tc.input))) + n, err := view.WriteFromBlocks(bs) + if err != nil { + t.Errorf("expected err nil, got %v", err) + } + if n != uint64(len(tc.input)) { + t.Errorf("expected %d bytes, got %d", len(tc.input), n) + } + + // Run the operation. + if tc.op != nil { + tc.op(&view) + } + + // Read and validate. + readLen := tc.readLen + if readLen == 0 { + readLen = len(tc.output) // Default. + } + out := make([]byte, readLen) + bs = safemem.BlockSeqOf(safemem.BlockFromSafeSlice(out)) + n, err = view.ReadToBlocks(bs) + if err != nil { + t.Errorf("expected nil, got %v", err) + } + if n != uint64(len(tc.output)) { + t.Errorf("expected %d bytes, got %d", len(tc.output), n) + } + + // Ensure the contents are correct. + if !bytes.Equal(out[:n], []byte(tc.output[:n])) { + t.Errorf("contents are wrong: expected %q, got %q", tc.output, string(out)) + } + }) + } +} diff --git a/pkg/buffer/view.go b/pkg/buffer/view.go new file mode 100644 index 000000000..e6901eadb --- /dev/null +++ b/pkg/buffer/view.go @@ -0,0 +1,390 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package buffer + +import ( + "fmt" + "io" +) + +// View is a non-linear buffer. +// +// All methods are thread compatible. +// +// +stateify savable +type View struct { + data bufferList + size int64 +} + +// TrimFront removes the first count bytes from the buffer. +func (v *View) TrimFront(count int64) { + if count >= v.size { + v.advanceRead(v.size) + } else { + v.advanceRead(count) + } +} + +// ReadAt implements io.ReaderAt.ReadAt. +func (v *View) ReadAt(p []byte, offset int64) (int, error) { + var ( + skipped int64 + done int64 + ) + for buf := v.data.Front(); buf != nil && done < int64(len(p)); buf = buf.Next() { + needToSkip := int(offset - skipped) + if sz := buf.ReadSize(); sz <= needToSkip { + skipped += int64(sz) + continue + } + + // Actually read data. + n := copy(p[done:], buf.ReadSlice()[needToSkip:]) + skipped += int64(needToSkip) + done += int64(n) + } + if int(done) < len(p) || offset+done == v.size { + return int(done), io.EOF + } + return int(done), nil +} + +// advanceRead advances the view's read index. +// +// Precondition: there must be sufficient bytes in the buffer. +func (v *View) advanceRead(count int64) { + for buf := v.data.Front(); buf != nil && count > 0; { + sz := int64(buf.ReadSize()) + if sz > count { + // There is still data for reading. + buf.ReadMove(int(count)) + v.size -= count + count = 0 + break + } + + // Consume the whole buffer. + oldBuf := buf + buf = buf.Next() // Iterate. + v.data.Remove(oldBuf) + oldBuf.Reset() + bufferPool.Put(oldBuf) + + // Update counts. + count -= sz + v.size -= sz + } + if count > 0 { + panic(fmt.Sprintf("advanceRead still has %d bytes remaining", count)) + } +} + +// Truncate truncates the view to the given bytes. +// +// This will not grow the view, only shrink it. If a length is passed that is +// greater than the current size of the view, then nothing will happen. +// +// Precondition: length must be >= 0. +func (v *View) Truncate(length int64) { + if length < 0 { + panic("negative length provided") + } + if length >= v.size { + return // Nothing to do. + } + for buf := v.data.Back(); buf != nil && v.size > length; buf = v.data.Back() { + sz := int64(buf.ReadSize()) + if after := v.size - sz; after < length { + // Truncate the buffer locally. + left := (length - after) + buf.write = buf.read + int(left) + v.size = length + break + } + + // Drop the buffer completely; see above. + v.data.Remove(buf) + buf.Reset() + bufferPool.Put(buf) + v.size -= sz + } +} + +// Grow grows the given view to the number of bytes, which will be appended. If +// zero is true, all these bytes will be zero. If zero is false, then this is +// the caller's responsibility. +// +// Precondition: length must be >= 0. +func (v *View) Grow(length int64, zero bool) { + if length < 0 { + panic("negative length provided") + } + for v.size < length { + buf := v.data.Back() + + // Is there some space in the last buffer? + if buf == nil || buf.Full() { + buf = bufferPool.Get().(*buffer) + v.data.PushBack(buf) + } + + // Write up to length bytes. + sz := buf.WriteSize() + if int64(sz) > length-v.size { + sz = int(length - v.size) + } + + // Zero the written section; note that this pattern is + // specifically recognized and optimized by the compiler. + if zero { + for i := buf.write; i < buf.write+sz; i++ { + buf.data[i] = 0 + } + } + + // Advance the index. + buf.WriteMove(sz) + v.size += int64(sz) + } +} + +// Prepend prepends the given data. +func (v *View) Prepend(data []byte) { + // Is there any space in the first buffer? + if buf := v.data.Front(); buf != nil && buf.read > 0 { + // Fill up before the first write. + avail := buf.read + bStart := 0 + dStart := len(data) - avail + if avail > len(data) { + bStart = avail - len(data) + dStart = 0 + } + n := copy(buf.data[bStart:], data[dStart:]) + data = data[:dStart] + v.size += int64(n) + buf.read -= n + } + + for len(data) > 0 { + // Do we need an empty buffer? + buf := bufferPool.Get().(*buffer) + v.data.PushFront(buf) + + // The buffer is empty; copy last chunk. + avail := len(buf.data) + bStart := 0 + dStart := len(data) - avail + if avail > len(data) { + bStart = avail - len(data) + dStart = 0 + } + + // We have to put the data at the end of the current + // buffer in order to ensure that the next prepend will + // correctly fill up the beginning of this buffer. + n := copy(buf.data[bStart:], data[dStart:]) + data = data[:dStart] + v.size += int64(n) + buf.read = len(buf.data) - n + buf.write = len(buf.data) + } +} + +// Append appends the given data. +func (v *View) Append(data []byte) { + for done := 0; done < len(data); { + buf := v.data.Back() + + // Ensure there's a buffer with space. + if buf == nil || buf.Full() { + buf = bufferPool.Get().(*buffer) + v.data.PushBack(buf) + } + + // Copy in to the given buffer. + n := copy(buf.WriteSlice(), data[done:]) + done += n + buf.WriteMove(n) + v.size += int64(n) + } +} + +// Flatten returns a flattened copy of this data. +// +// This method should not be used in any performance-sensitive paths. It may +// allocate a fresh byte slice sufficiently large to contain all the data in +// the buffer. This is principally for debugging. +// +// N.B. Tee data still belongs to this view, as if there is a single buffer +// present, then it will be returned directly. This should be used for +// temporary use only, and a reference to the given slice should not be held. +func (v *View) Flatten() []byte { + if buf := v.data.Front(); buf == nil { + return nil // No data at all. + } else if buf.Next() == nil { + return buf.ReadSlice() // Only one buffer. + } + data := make([]byte, 0, v.size) // Need to flatten. + for buf := v.data.Front(); buf != nil; buf = buf.Next() { + // Copy to the allocated slice. + data = append(data, buf.ReadSlice()...) + } + return data +} + +// Size indicates the total amount of data available in this view. +func (v *View) Size() int64 { + return v.size +} + +// Copy makes a strict copy of this view. +func (v *View) Copy() (other View) { + for buf := v.data.Front(); buf != nil; buf = buf.Next() { + other.Append(buf.ReadSlice()) + } + return +} + +// Apply applies the given function across all valid data. +func (v *View) Apply(fn func([]byte)) { + for buf := v.data.Front(); buf != nil; buf = buf.Next() { + fn(buf.ReadSlice()) + } +} + +// Merge merges the provided View with this one. +// +// The other view will be appended to v, and other will be empty after this +// operation completes. +func (v *View) Merge(other *View) { + // Copy over all buffers. + for buf := other.data.Front(); buf != nil; buf = other.data.Front() { + other.data.Remove(buf) + v.data.PushBack(buf) + } + + // Adjust sizes. + v.size += other.size + other.size = 0 +} + +// WriteFromReader writes to the buffer from an io.Reader. +// +// A minimum read size equal to unsafe.Sizeof(unintptr) is enforced, +// provided that count is greater than or equal to unsafe.Sizeof(uintptr). +func (v *View) WriteFromReader(r io.Reader, count int64) (int64, error) { + var ( + done int64 + n int + err error + ) + for done < count { + buf := v.data.Back() + + // Ensure we have an empty buffer. + if buf == nil || buf.Full() { + buf = bufferPool.Get().(*buffer) + v.data.PushBack(buf) + } + + // Is this less than the minimum batch? + if buf.WriteSize() < minBatch && (count-done) >= int64(minBatch) { + tmp := make([]byte, minBatch) + n, err = r.Read(tmp) + v.Append(tmp[:n]) + done += int64(n) + if err != nil { + break + } + continue + } + + // Limit the read, if necessary. + sz := buf.WriteSize() + if left := count - done; int64(sz) > left { + sz = int(left) + } + + // Pass the relevant portion of the buffer. + n, err = r.Read(buf.WriteSlice()[:sz]) + buf.WriteMove(n) + done += int64(n) + v.size += int64(n) + if err == io.EOF { + err = nil // Short write allowed. + break + } else if err != nil { + break + } + } + return done, err +} + +// ReadToWriter reads from the buffer into an io.Writer. +// +// N.B. This does not consume the bytes read. TrimFront should +// be called appropriately after this call in order to do so. +// +// A minimum write size equal to unsafe.Sizeof(unintptr) is enforced, +// provided that count is greater than or equal to unsafe.Sizeof(uintptr). +func (v *View) ReadToWriter(w io.Writer, count int64) (int64, error) { + var ( + done int64 + n int + err error + ) + offset := 0 // Spill-over for batching. + for buf := v.data.Front(); buf != nil && done < count; buf = buf.Next() { + // Has this been consumed? Skip it. + sz := buf.ReadSize() + if sz <= offset { + offset -= sz + continue + } + sz -= offset + + // Is this less than the minimum batch? + left := count - done + if sz < minBatch && left >= int64(minBatch) && (v.size-done) >= int64(minBatch) { + tmp := make([]byte, minBatch) + n, err = v.ReadAt(tmp, done) + w.Write(tmp[:n]) + done += int64(n) + offset = n - sz // Reset below. + if err != nil { + break + } + continue + } + + // Limit the write if necessary. + if int64(sz) >= left { + sz = int(left) + } + + // Perform the actual write. + n, err = w.Write(buf.ReadSlice()[offset : offset+sz]) + done += int64(n) + if err != nil { + break + } + + // Reset spill-over. + offset = 0 + } + return done, err +} diff --git a/pkg/buffer/view_test.go b/pkg/buffer/view_test.go new file mode 100644 index 000000000..3db1bc6ee --- /dev/null +++ b/pkg/buffer/view_test.go @@ -0,0 +1,467 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package buffer + +import ( + "bytes" + "io" + "strings" + "testing" +) + +func fillAppend(v *View, data []byte) { + v.Append(data) +} + +func fillAppendEnd(v *View, data []byte) { + v.Grow(bufferSize-1, false) + v.Append(data) + v.TrimFront(bufferSize - 1) +} + +func fillWriteFromReader(v *View, data []byte) { + b := bytes.NewBuffer(data) + v.WriteFromReader(b, int64(len(data))) +} + +func fillWriteFromReaderEnd(v *View, data []byte) { + v.Grow(bufferSize-1, false) + b := bytes.NewBuffer(data) + v.WriteFromReader(b, int64(len(data))) + v.TrimFront(bufferSize - 1) +} + +var fillFuncs = map[string]func(*View, []byte){ + "append": fillAppend, + "appendEnd": fillAppendEnd, + "writeFromReader": fillWriteFromReader, + "writeFromReaderEnd": fillWriteFromReaderEnd, +} + +func testReadAt(t *testing.T, v *View, offset int64, n int, wantStr string, wantErr error) { + t.Helper() + d := make([]byte, n) + n, err := v.ReadAt(d, offset) + if n != len(wantStr) { + t.Errorf("got %d, want %d", n, len(wantStr)) + } + if err != wantErr { + t.Errorf("got err %v, want %v", err, wantErr) + } + if !bytes.Equal(d[:n], []byte(wantStr)) { + t.Errorf("got %q, want %q", string(d[:n]), wantStr) + } +} + +func TestView(t *testing.T) { + testCases := []struct { + name string + input string + output string + op func(*testing.T, *View) + }{ + // Preconditions. + { + name: "truncate-check", + input: "hello", + output: "hello", // Not touched. + op: func(t *testing.T, v *View) { + defer func() { + if r := recover(); r == nil { + t.Errorf("Truncate(-1) did not panic") + } + }() + v.Truncate(-1) + }, + }, + { + name: "grow-check", + input: "hello", + output: "hello", // Not touched. + op: func(t *testing.T, v *View) { + defer func() { + if r := recover(); r == nil { + t.Errorf("Grow(-1) did not panic") + } + }() + v.Grow(-1, false) + }, + }, + { + name: "advance-check", + input: "hello", + output: "", // Consumed. + op: func(t *testing.T, v *View) { + defer func() { + if r := recover(); r == nil { + t.Errorf("advanceRead(Size()+1) did not panic") + } + }() + v.advanceRead(v.Size() + 1) + }, + }, + + // Prepend. + { + name: "prepend", + input: "world", + output: "hello world", + op: func(t *testing.T, v *View) { + v.Prepend([]byte("hello ")) + }, + }, + { + name: "prepend-backfill-full", + input: "hello world", + output: "jello world", + op: func(t *testing.T, v *View) { + v.TrimFront(1) + v.Prepend([]byte("j")) + }, + }, + { + name: "prepend-backfill-under", + input: "hello world", + output: "hola world", + op: func(t *testing.T, v *View) { + v.TrimFront(5) + v.Prepend([]byte("hola")) + }, + }, + { + name: "prepend-backfill-over", + input: "hello world", + output: "smello world", + op: func(t *testing.T, v *View) { + v.TrimFront(1) + v.Prepend([]byte("sm")) + }, + }, + { + name: "prepend-fill", + input: strings.Repeat("1", bufferSize-1), + output: "0" + strings.Repeat("1", bufferSize-1), + op: func(t *testing.T, v *View) { + v.Prepend([]byte("0")) + }, + }, + { + name: "prepend-overflow", + input: strings.Repeat("1", bufferSize), + output: "0" + strings.Repeat("1", bufferSize), + op: func(t *testing.T, v *View) { + v.Prepend([]byte("0")) + }, + }, + { + name: "prepend-multiple-buffers", + input: strings.Repeat("1", bufferSize-1), + output: strings.Repeat("0", bufferSize*3) + strings.Repeat("1", bufferSize-1), + op: func(t *testing.T, v *View) { + v.Prepend([]byte(strings.Repeat("0", bufferSize*3))) + }, + }, + + // Append and write. + { + name: "append", + input: "hello", + output: "hello world", + op: func(t *testing.T, v *View) { + v.Append([]byte(" world")) + }, + }, + { + name: "append-fill", + input: strings.Repeat("1", bufferSize-1), + output: strings.Repeat("1", bufferSize-1) + "0", + op: func(t *testing.T, v *View) { + v.Append([]byte("0")) + }, + }, + { + name: "append-overflow", + input: strings.Repeat("1", bufferSize), + output: strings.Repeat("1", bufferSize) + "0", + op: func(t *testing.T, v *View) { + v.Append([]byte("0")) + }, + }, + { + name: "append-multiple-buffers", + input: strings.Repeat("1", bufferSize-1), + output: strings.Repeat("1", bufferSize-1) + strings.Repeat("0", bufferSize*3), + op: func(t *testing.T, v *View) { + v.Append([]byte(strings.Repeat("0", bufferSize*3))) + }, + }, + + // Truncate. + { + name: "truncate", + input: "hello world", + output: "hello", + op: func(t *testing.T, v *View) { + v.Truncate(5) + }, + }, + { + name: "truncate-noop", + input: "hello world", + output: "hello world", + op: func(t *testing.T, v *View) { + v.Truncate(v.Size() + 1) + }, + }, + { + name: "truncate-multiple-buffers", + input: strings.Repeat("1", bufferSize*2), + output: strings.Repeat("1", bufferSize*2-1), + op: func(t *testing.T, v *View) { + v.Truncate(bufferSize*2 - 1) + }, + }, + { + name: "truncate-multiple-buffers-to-one", + input: strings.Repeat("1", bufferSize*2), + output: "11111", + op: func(t *testing.T, v *View) { + v.Truncate(5) + }, + }, + + // TrimFront. + { + name: "trim", + input: "hello world", + output: "world", + op: func(t *testing.T, v *View) { + v.TrimFront(6) + }, + }, + { + name: "trim-too-large", + input: "hello world", + output: "", + op: func(t *testing.T, v *View) { + v.TrimFront(v.Size() + 1) + }, + }, + { + name: "trim-multiple-buffers", + input: strings.Repeat("1", bufferSize*2), + output: strings.Repeat("1", bufferSize*2-1), + op: func(t *testing.T, v *View) { + v.TrimFront(1) + }, + }, + { + name: "trim-multiple-buffers-to-one-buffer", + input: strings.Repeat("1", bufferSize*2), + output: "1", + op: func(t *testing.T, v *View) { + v.TrimFront(bufferSize*2 - 1) + }, + }, + + // Grow. + { + name: "grow", + input: "hello world", + output: "hello world", + op: func(t *testing.T, v *View) { + v.Grow(1, true) + }, + }, + { + name: "grow-from-zero", + output: strings.Repeat("\x00", 1024), + op: func(t *testing.T, v *View) { + v.Grow(1024, true) + }, + }, + { + name: "grow-from-non-zero", + input: strings.Repeat("1", bufferSize), + output: strings.Repeat("1", bufferSize) + strings.Repeat("\x00", bufferSize), + op: func(t *testing.T, v *View) { + v.Grow(bufferSize*2, true) + }, + }, + + // Copy. + { + name: "copy", + input: "hello", + output: "hello", + op: func(t *testing.T, v *View) { + other := v.Copy() + bs := other.Flatten() + want := []byte("hello") + if !bytes.Equal(bs, want) { + t.Errorf("expected %v, got %v", want, bs) + } + }, + }, + { + name: "copy-large", + input: strings.Repeat("1", bufferSize+1), + output: strings.Repeat("1", bufferSize+1), + op: func(t *testing.T, v *View) { + other := v.Copy() + bs := other.Flatten() + want := []byte(strings.Repeat("1", bufferSize+1)) + if !bytes.Equal(bs, want) { + t.Errorf("expected %v, got %v", want, bs) + } + }, + }, + + // Merge. + { + name: "merge", + input: "hello", + output: "hello world", + op: func(t *testing.T, v *View) { + var other View + other.Append([]byte(" world")) + v.Merge(&other) + if sz := other.Size(); sz != 0 { + t.Errorf("expected 0, got %d", sz) + } + }, + }, + { + name: "merge-large", + input: strings.Repeat("1", bufferSize+1), + output: strings.Repeat("1", bufferSize+1) + strings.Repeat("0", bufferSize+1), + op: func(t *testing.T, v *View) { + var other View + other.Append([]byte(strings.Repeat("0", bufferSize+1))) + v.Merge(&other) + if sz := other.Size(); sz != 0 { + t.Errorf("expected 0, got %d", sz) + } + }, + }, + + // ReadAt. + { + name: "readat", + input: "hello", + output: "hello", + op: func(t *testing.T, v *View) { testReadAt(t, v, 0, 6, "hello", io.EOF) }, + }, + { + name: "readat-long", + input: "hello", + output: "hello", + op: func(t *testing.T, v *View) { testReadAt(t, v, 0, 8, "hello", io.EOF) }, + }, + { + name: "readat-short", + input: "hello", + output: "hello", + op: func(t *testing.T, v *View) { testReadAt(t, v, 0, 3, "hel", nil) }, + }, + { + name: "readat-offset", + input: "hello", + output: "hello", + op: func(t *testing.T, v *View) { testReadAt(t, v, 2, 3, "llo", io.EOF) }, + }, + { + name: "readat-long-offset", + input: "hello", + output: "hello", + op: func(t *testing.T, v *View) { testReadAt(t, v, 2, 8, "llo", io.EOF) }, + }, + { + name: "readat-short-offset", + input: "hello", + output: "hello", + op: func(t *testing.T, v *View) { testReadAt(t, v, 2, 2, "ll", nil) }, + }, + { + name: "readat-skip-all", + input: "hello", + output: "hello", + op: func(t *testing.T, v *View) { testReadAt(t, v, bufferSize+1, 1, "", io.EOF) }, + }, + { + name: "readat-second-buffer", + input: strings.Repeat("0", bufferSize+1) + "12", + output: strings.Repeat("0", bufferSize+1) + "12", + op: func(t *testing.T, v *View) { testReadAt(t, v, bufferSize+1, 1, "1", nil) }, + }, + { + name: "readat-second-buffer-end", + input: strings.Repeat("0", bufferSize+1) + "12", + output: strings.Repeat("0", bufferSize+1) + "12", + op: func(t *testing.T, v *View) { testReadAt(t, v, bufferSize+1, 2, "12", io.EOF) }, + }, + } + + for _, tc := range testCases { + for fillName, fn := range fillFuncs { + t.Run(fillName+"/"+tc.name, func(t *testing.T) { + // Construct & fill the view. + var view View + fn(&view, []byte(tc.input)) + + // Run the operation. + if tc.op != nil { + tc.op(t, &view) + } + + // Flatten and validate. + out := view.Flatten() + if !bytes.Equal([]byte(tc.output), out) { + t.Errorf("expected %q, got %q", tc.output, string(out)) + } + + // Ensure the size is correct. + if len(out) != int(view.Size()) { + t.Errorf("size is wrong: expected %d, got %d", len(out), view.Size()) + } + + // Calculate contents via apply. + var appliedOut []byte + view.Apply(func(b []byte) { + appliedOut = append(appliedOut, b...) + }) + if len(appliedOut) != len(out) { + t.Errorf("expected %d, got %d", len(out), len(appliedOut)) + } + if !bytes.Equal(appliedOut, out) { + t.Errorf("expected %v, got %v", out, appliedOut) + } + + // Calculate contents via ReadToWriter. + var b bytes.Buffer + n, err := view.ReadToWriter(&b, int64(len(out))) + if n != int64(len(out)) { + t.Errorf("expected %d, got %d", len(out), n) + } + if err != nil { + t.Errorf("expected nil, got %v", err) + } + if !bytes.Equal(b.Bytes(), out) { + t.Errorf("expected %v, got %v", out, b.Bytes()) + } + }) + } + } +} diff --git a/pkg/fspath/builder_unsafe.go b/pkg/buffer/view_unsafe.go index 75606808d..d1ef39b26 100644 --- a/pkg/fspath/builder_unsafe.go +++ b/pkg/buffer/view_unsafe.go @@ -1,4 +1,4 @@ -// Copyright 2019 The gVisor Authors. +// Copyright 2020 The gVisor Authors. // // Licensed under the Apache License, Version 2.0 (the "License"); // you may not use this file except in compliance with the License. @@ -12,16 +12,14 @@ // See the License for the specific language governing permissions and // limitations under the License. -package fspath +package buffer import ( "unsafe" ) -// String returns the accumulated string. No other methods should be called -// after String. -func (b *Builder) String() string { - bs := b.buf[b.start:] - // Compare strings.Builder.String(). - return *(*string)(unsafe.Pointer(&bs)) -} +// minBatch is the smallest Read or Write operation that the +// WriteFromReader and ReadToWriter functions will use. +// +// This is defined as the size of a native pointer. +const minBatch = int(unsafe.Sizeof(uintptr(0))) diff --git a/pkg/context/context.go b/pkg/context/context.go index 23e009ef3..5319b6d8d 100644 --- a/pkg/context/context.go +++ b/pkg/context/context.go @@ -127,10 +127,6 @@ func (logContext) Value(key interface{}) interface{} { var bgContext = &logContext{Logger: log.Log()} // Background returns an empty context using the default logger. -// -// Users should be wary of using a Background context. Please tag any use with -// FIXME(b/38173783) and a note to remove this use. -// // Generally, one should use the Task as their context when available, or avoid // having to use a context in places where a Task is unavailable. // diff --git a/pkg/cpuid/cpuid_parse_x86_test.go b/pkg/cpuid/cpuid_parse_x86_test.go index d48418e69..c9bd40e1b 100644 --- a/pkg/cpuid/cpuid_parse_x86_test.go +++ b/pkg/cpuid/cpuid_parse_x86_test.go @@ -12,7 +12,7 @@ // See the License for the specific language governing permissions and // limitations under the License. -// +build i386 amd64 +// +build 386 amd64 package cpuid diff --git a/pkg/cpuid/cpuid_x86.go b/pkg/cpuid/cpuid_x86.go index 333ca0a04..562f8f405 100644 --- a/pkg/cpuid/cpuid_x86.go +++ b/pkg/cpuid/cpuid_x86.go @@ -12,7 +12,7 @@ // See the License for the specific language governing permissions and // limitations under the License. -// +build i386 amd64 +// +build 386 amd64 package cpuid @@ -235,7 +235,9 @@ const ( X86FeaturePERFCTR_TSC X86FeaturePERFCTR_LLC X86FeatureMWAITX - // ECX[31:30] are reserved. + // TODO(b/152776797): Some CPUs set this but it is not documented anywhere. + X86FeatureBlock5Bit30 + _ // ecx bit 31 is reserved. ) // Block 6 constants are the extended feature bits in @@ -438,6 +440,9 @@ var x86FeatureParseOnlyStrings = map[Feature]string{ // Block 3. X86FeaturePREFETCHWT1: "prefetchwt1", + + // Block 5. + X86FeatureBlock5Bit30: "block5_bit30", } // intelCacheDescriptors describe the caches and TLBs on the system. They are @@ -725,6 +730,18 @@ func vendorIDFromRegs(bx, cx, dx uint32) string { return string(bytes) } +var maxXsaveSize = func() uint32 { + // Leaf 0 of xsaveinfo function returns the size for currently + // enabled xsave features in ebx, the maximum size if all valid + // features are saved with xsave in ecx, and valid XCR0 bits in + // edx:eax. + // + // If xSaveInfo isn't supported, cpuid will not fault but will + // return bogus values. + _, _, maxXsaveSize, _ := HostID(uint32(xSaveInfo), 0) + return maxXsaveSize +}() + // ExtendedStateSize returns the number of bytes needed to save the "extended // state" for this processor and the boundary it must be aligned to. Extended // state includes floating point registers, and other cpu state that's not @@ -736,12 +753,7 @@ func vendorIDFromRegs(bx, cx, dx uint32) string { // about 2.5K worst case, with avx512). func (fs *FeatureSet) ExtendedStateSize() (size, align uint) { if fs.UseXsave() { - // Leaf 0 of xsaveinfo function returns the size for currently - // enabled xsave features in ebx, the maximum size if all valid - // features are saved with xsave in ecx, and valid XCR0 bits in - // edx:eax. - _, _, maxSize, _ := HostID(uint32(xSaveInfo), 0) - return uint(maxSize), 64 + return uint(maxXsaveSize), 64 } // If we don't support xsave, we fall back to fxsave, which requires diff --git a/pkg/cpuid/cpuid_x86_test.go b/pkg/cpuid/cpuid_x86_test.go index 0fe20c213..bacf345c8 100644 --- a/pkg/cpuid/cpuid_x86_test.go +++ b/pkg/cpuid/cpuid_x86_test.go @@ -12,7 +12,7 @@ // See the License for the specific language governing permissions and // limitations under the License. -// +build i386 amd64 +// +build 386 amd64 package cpuid diff --git a/pkg/eventchannel/event_test.go b/pkg/eventchannel/event_test.go index 7f41b4a27..43750360b 100644 --- a/pkg/eventchannel/event_test.go +++ b/pkg/eventchannel/event_test.go @@ -78,7 +78,7 @@ func TestMultiEmitter(t *testing.T) { for _, name := range names { m := testMessage{name: name} if _, err := me.Emit(m); err != nil { - t.Fatal("me.Emit(%v) failed: %v", m, err) + t.Fatalf("me.Emit(%v) failed: %v", m, err) } } @@ -96,7 +96,7 @@ func TestMultiEmitter(t *testing.T) { // Close multiEmitter. if err := me.Close(); err != nil { - t.Fatal("me.Close() failed: %v", err) + t.Fatalf("me.Close() failed: %v", err) } // All testEmitters should be closed. diff --git a/pkg/fspath/BUILD b/pkg/fspath/BUILD index ee84471b2..67dd1e225 100644 --- a/pkg/fspath/BUILD +++ b/pkg/fspath/BUILD @@ -8,9 +8,11 @@ go_library( name = "fspath", srcs = [ "builder.go", - "builder_unsafe.go", "fspath.go", ], + deps = [ + "//pkg/gohacks", + ], ) go_test( diff --git a/pkg/fspath/builder.go b/pkg/fspath/builder.go index 7ddb36826..6318d3874 100644 --- a/pkg/fspath/builder.go +++ b/pkg/fspath/builder.go @@ -16,6 +16,8 @@ package fspath import ( "fmt" + + "gvisor.dev/gvisor/pkg/gohacks" ) // Builder is similar to strings.Builder, but is used to produce pathnames @@ -102,3 +104,9 @@ func (b *Builder) AppendString(str string) { copy(b.buf[b.start:], b.buf[oldStart:]) copy(b.buf[len(b.buf)-len(str):], str) } + +// String returns the accumulated string. No other methods should be called +// after String. +func (b *Builder) String() string { + return gohacks.StringFromImmutableBytes(b.buf[b.start:]) +} diff --git a/pkg/fspath/fspath.go b/pkg/fspath/fspath.go index 9fb3fee24..4c983d5fd 100644 --- a/pkg/fspath/fspath.go +++ b/pkg/fspath/fspath.go @@ -67,7 +67,8 @@ func Parse(pathname string) Path { // Path contains the information contained in a pathname string. // -// Path is copyable by value. +// Path is copyable by value. The zero value for Path is equivalent to +// fspath.Parse(""), i.e. the empty path. type Path struct { // Begin is an iterator to the first path component in the relative part of // the path. diff --git a/pkg/gate/gate_test.go b/pkg/gate/gate_test.go index 850693df8..316015e06 100644 --- a/pkg/gate/gate_test.go +++ b/pkg/gate/gate_test.go @@ -15,6 +15,7 @@ package gate_test import ( + "runtime" "testing" "time" @@ -165,6 +166,8 @@ func worker(g *gate.Gate, done *sync.WaitGroup) { if !g.Enter() { break } + // Golang before v1.14 doesn't preempt busyloops. + runtime.Gosched() g.Leave() } done.Done() diff --git a/pkg/gohacks/BUILD b/pkg/gohacks/BUILD new file mode 100644 index 000000000..798a65eca --- /dev/null +++ b/pkg/gohacks/BUILD @@ -0,0 +1,11 @@ +load("//tools:defs.bzl", "go_library") + +package(licenses = ["notice"]) + +go_library( + name = "gohacks", + srcs = [ + "gohacks_unsafe.go", + ], + visibility = ["//:sandbox"], +) diff --git a/pkg/gohacks/gohacks_unsafe.go b/pkg/gohacks/gohacks_unsafe.go new file mode 100644 index 000000000..aad675172 --- /dev/null +++ b/pkg/gohacks/gohacks_unsafe.go @@ -0,0 +1,57 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// Package gohacks contains utilities for subverting the Go compiler. +package gohacks + +import ( + "reflect" + "unsafe" +) + +// Noescape hides a pointer from escape analysis. Noescape is the identity +// function but escape analysis doesn't think the output depends on the input. +// Noescape is inlined and currently compiles down to zero instructions. +// USE CAREFULLY! +// +// (Noescape is copy/pasted from Go's runtime/stubs.go:noescape().) +// +//go:nosplit +func Noescape(p unsafe.Pointer) unsafe.Pointer { + x := uintptr(p) + return unsafe.Pointer(x ^ 0) +} + +// ImmutableBytesFromString is equivalent to []byte(s), except that it uses the +// same memory backing s instead of making a heap-allocated copy. This is only +// valid if the returned slice is never mutated. +func ImmutableBytesFromString(s string) []byte { + shdr := (*reflect.StringHeader)(unsafe.Pointer(&s)) + var bs []byte + bshdr := (*reflect.SliceHeader)(unsafe.Pointer(&bs)) + bshdr.Data = shdr.Data + bshdr.Len = shdr.Len + bshdr.Cap = shdr.Len + return bs +} + +// StringFromImmutableBytes is equivalent to string(bs), except that it uses +// the same memory backing bs instead of making a heap-allocated copy. This is +// only valid if bs is never mutated after StringFromImmutableBytes returns. +func StringFromImmutableBytes(bs []byte) string { + // This is cheaper than messing with reflect.StringHeader and + // reflect.SliceHeader, which as of this writing produces many dead stores + // of zeroes. Compare strings.Builder.String(). + return *(*string)(unsafe.Pointer(&bs)) +} diff --git a/pkg/ilist/list.go b/pkg/ilist/list.go index 019caadca..0d07da3b1 100644 --- a/pkg/ilist/list.go +++ b/pkg/ilist/list.go @@ -86,11 +86,21 @@ func (l *List) Back() Element { return l.tail } +// Len returns the number of elements in the list. +// +// NOTE: This is an O(n) operation. +func (l *List) Len() (count int) { + for e := l.Front(); e != nil; e = e.Next() { + count++ + } + return count +} + // PushFront inserts the element e at the front of list l. func (l *List) PushFront(e Element) { - ElementMapper{}.linkerFor(e).SetNext(l.head) - ElementMapper{}.linkerFor(e).SetPrev(nil) - + linker := ElementMapper{}.linkerFor(e) + linker.SetNext(l.head) + linker.SetPrev(nil) if l.head != nil { ElementMapper{}.linkerFor(l.head).SetPrev(e) } else { @@ -102,9 +112,9 @@ func (l *List) PushFront(e Element) { // PushBack inserts the element e at the back of list l. func (l *List) PushBack(e Element) { - ElementMapper{}.linkerFor(e).SetNext(nil) - ElementMapper{}.linkerFor(e).SetPrev(l.tail) - + linker := ElementMapper{}.linkerFor(e) + linker.SetNext(nil) + linker.SetPrev(l.tail) if l.tail != nil { ElementMapper{}.linkerFor(l.tail).SetNext(e) } else { @@ -125,17 +135,20 @@ func (l *List) PushBackList(m *List) { l.tail = m.tail } - m.head = nil m.tail = nil } // InsertAfter inserts e after b. func (l *List) InsertAfter(b, e Element) { - a := ElementMapper{}.linkerFor(b).Next() - ElementMapper{}.linkerFor(e).SetNext(a) - ElementMapper{}.linkerFor(e).SetPrev(b) - ElementMapper{}.linkerFor(b).SetNext(e) + bLinker := ElementMapper{}.linkerFor(b) + eLinker := ElementMapper{}.linkerFor(e) + + a := bLinker.Next() + + eLinker.SetNext(a) + eLinker.SetPrev(b) + bLinker.SetNext(e) if a != nil { ElementMapper{}.linkerFor(a).SetPrev(e) @@ -146,10 +159,13 @@ func (l *List) InsertAfter(b, e Element) { // InsertBefore inserts e before a. func (l *List) InsertBefore(a, e Element) { - b := ElementMapper{}.linkerFor(a).Prev() - ElementMapper{}.linkerFor(e).SetNext(a) - ElementMapper{}.linkerFor(e).SetPrev(b) - ElementMapper{}.linkerFor(a).SetPrev(e) + aLinker := ElementMapper{}.linkerFor(a) + eLinker := ElementMapper{}.linkerFor(e) + + b := aLinker.Prev() + eLinker.SetNext(a) + eLinker.SetPrev(b) + aLinker.SetPrev(e) if b != nil { ElementMapper{}.linkerFor(b).SetNext(e) @@ -160,8 +176,9 @@ func (l *List) InsertBefore(a, e Element) { // Remove removes e from l. func (l *List) Remove(e Element) { - prev := ElementMapper{}.linkerFor(e).Prev() - next := ElementMapper{}.linkerFor(e).Next() + linker := ElementMapper{}.linkerFor(e) + prev := linker.Prev() + next := linker.Next() if prev != nil { ElementMapper{}.linkerFor(prev).SetNext(next) @@ -174,6 +191,9 @@ func (l *List) Remove(e Element) { } else { l.tail = prev } + + linker.SetNext(nil) + linker.SetPrev(nil) } // Entry is a default implementation of Linker. Users can add anonymous fields diff --git a/pkg/log/glog.go b/pkg/log/glog.go index cab5fae55..f57c4427b 100644 --- a/pkg/log/glog.go +++ b/pkg/log/glog.go @@ -25,7 +25,7 @@ import ( // GoogleEmitter is a wrapper that emits logs in a format compatible with // package github.com/golang/glog. type GoogleEmitter struct { - Writer + *Writer } // pid is used for the threadid component of the header. @@ -46,7 +46,7 @@ var pid = os.Getpid() // line The line number // msg The user-supplied message // -func (g *GoogleEmitter) Emit(level Level, timestamp time.Time, format string, args ...interface{}) { +func (g GoogleEmitter) Emit(depth int, level Level, timestamp time.Time, format string, args ...interface{}) { // Log level. prefix := byte('?') switch level { @@ -64,9 +64,7 @@ func (g *GoogleEmitter) Emit(level Level, timestamp time.Time, format string, ar microsecond := int(timestamp.Nanosecond() / 1000) // 0 = this frame. - // 1 = Debugf, etc. - // 2 = Caller. - _, file, line, ok := runtime.Caller(2) + _, file, line, ok := runtime.Caller(depth + 1) if ok { // Trim any directory path from the file. slash := strings.LastIndexByte(file, byte('/')) @@ -83,5 +81,5 @@ func (g *GoogleEmitter) Emit(level Level, timestamp time.Time, format string, ar message := fmt.Sprintf(format, args...) // Emit the formatted result. - fmt.Fprintf(&g.Writer, "%c%02d%02d %02d:%02d:%02d.%06d % 7d %s:%d] %s\n", prefix, int(month), day, hour, minute, second, microsecond, pid, file, line, message) + fmt.Fprintf(g.Writer, "%c%02d%02d %02d:%02d:%02d.%06d % 7d %s:%d] %s\n", prefix, int(month), day, hour, minute, second, microsecond, pid, file, line, message) } diff --git a/pkg/log/json.go b/pkg/log/json.go index a278c8fc8..bdf9d691e 100644 --- a/pkg/log/json.go +++ b/pkg/log/json.go @@ -58,11 +58,11 @@ func (lv *Level) UnmarshalJSON(b []byte) error { // JSONEmitter logs messages in json format. type JSONEmitter struct { - Writer + *Writer } // Emit implements Emitter.Emit. -func (e JSONEmitter) Emit(level Level, timestamp time.Time, format string, v ...interface{}) { +func (e JSONEmitter) Emit(_ int, level Level, timestamp time.Time, format string, v ...interface{}) { j := jsonLog{ Msg: fmt.Sprintf(format, v...), Level: level, diff --git a/pkg/log/json_k8s.go b/pkg/log/json_k8s.go index cee6eb514..5883e95e1 100644 --- a/pkg/log/json_k8s.go +++ b/pkg/log/json_k8s.go @@ -29,11 +29,11 @@ type k8sJSONLog struct { // K8sJSONEmitter logs messages in json format that is compatible with // Kubernetes fluent configuration. type K8sJSONEmitter struct { - Writer + *Writer } // Emit implements Emitter.Emit. -func (e *K8sJSONEmitter) Emit(level Level, timestamp time.Time, format string, v ...interface{}) { +func (e K8sJSONEmitter) Emit(_ int, level Level, timestamp time.Time, format string, v ...interface{}) { j := k8sJSONLog{ Log: fmt.Sprintf(format, v...), Level: level, diff --git a/pkg/log/log.go b/pkg/log/log.go index 5056f17e6..37e0605ad 100644 --- a/pkg/log/log.go +++ b/pkg/log/log.go @@ -79,7 +79,7 @@ func (l Level) String() string { type Emitter interface { // Emit emits the given log statement. This allows for control over the // timestamp used for logging. - Emit(level Level, timestamp time.Time, format string, v ...interface{}) + Emit(depth int, level Level, timestamp time.Time, format string, v ...interface{}) } // Writer writes the output to the given writer. @@ -142,7 +142,7 @@ func (l *Writer) Write(data []byte) (int, error) { } // Emit emits the message. -func (l *Writer) Emit(level Level, timestamp time.Time, format string, args ...interface{}) { +func (l *Writer) Emit(_ int, _ Level, _ time.Time, format string, args ...interface{}) { fmt.Fprintf(l, format, args...) } @@ -150,9 +150,9 @@ func (l *Writer) Emit(level Level, timestamp time.Time, format string, args ...i type MultiEmitter []Emitter // Emit emits to all emitters. -func (m *MultiEmitter) Emit(level Level, timestamp time.Time, format string, v ...interface{}) { +func (m *MultiEmitter) Emit(depth int, level Level, timestamp time.Time, format string, v ...interface{}) { for _, e := range *m { - e.Emit(level, timestamp, format, v...) + e.Emit(1+depth, level, timestamp, format, v...) } } @@ -167,7 +167,7 @@ type TestEmitter struct { } // Emit emits to the TestLogger. -func (t *TestEmitter) Emit(level Level, timestamp time.Time, format string, v ...interface{}) { +func (t *TestEmitter) Emit(_ int, level Level, timestamp time.Time, format string, v ...interface{}) { t.Logf(format, v...) } @@ -198,22 +198,37 @@ type BasicLogger struct { // Debugf implements logger.Debugf. func (l *BasicLogger) Debugf(format string, v ...interface{}) { - if l.IsLogging(Debug) { - l.Emit(Debug, time.Now(), format, v...) - } + l.DebugfAtDepth(1, format, v...) } // Infof implements logger.Infof. func (l *BasicLogger) Infof(format string, v ...interface{}) { - if l.IsLogging(Info) { - l.Emit(Info, time.Now(), format, v...) - } + l.InfofAtDepth(1, format, v...) } // Warningf implements logger.Warningf. func (l *BasicLogger) Warningf(format string, v ...interface{}) { + l.WarningfAtDepth(1, format, v...) +} + +// DebugfAtDepth logs at a specific depth. +func (l *BasicLogger) DebugfAtDepth(depth int, format string, v ...interface{}) { + if l.IsLogging(Debug) { + l.Emit(1+depth, Debug, time.Now(), format, v...) + } +} + +// InfofAtDepth logs at a specific depth. +func (l *BasicLogger) InfofAtDepth(depth int, format string, v ...interface{}) { + if l.IsLogging(Info) { + l.Emit(1+depth, Info, time.Now(), format, v...) + } +} + +// WarningfAtDepth logs at a specific depth. +func (l *BasicLogger) WarningfAtDepth(depth int, format string, v ...interface{}) { if l.IsLogging(Warning) { - l.Emit(Warning, time.Now(), format, v...) + l.Emit(1+depth, Warning, time.Now(), format, v...) } } @@ -257,17 +272,32 @@ func SetLevel(newLevel Level) { // Debugf logs to the global logger. func Debugf(format string, v ...interface{}) { - Log().Debugf(format, v...) + Log().DebugfAtDepth(1, format, v...) } // Infof logs to the global logger. func Infof(format string, v ...interface{}) { - Log().Infof(format, v...) + Log().InfofAtDepth(1, format, v...) } // Warningf logs to the global logger. func Warningf(format string, v ...interface{}) { - Log().Warningf(format, v...) + Log().WarningfAtDepth(1, format, v...) +} + +// DebugfAtDepth logs to the global logger. +func DebugfAtDepth(depth int, format string, v ...interface{}) { + Log().DebugfAtDepth(1+depth, format, v...) +} + +// InfofAtDepth logs to the global logger. +func InfofAtDepth(depth int, format string, v ...interface{}) { + Log().InfofAtDepth(1+depth, format, v...) +} + +// WarningfAtDepth logs to the global logger. +func WarningfAtDepth(depth int, format string, v ...interface{}) { + Log().WarningfAtDepth(1+depth, format, v...) } // defaultStackSize is the default buffer size to allocate for stack traces. @@ -344,5 +374,5 @@ func CopyStandardLogTo(l Level) error { func init() { // Store the initial value for the log. - log.Store(&BasicLogger{Level: Info, Emitter: &GoogleEmitter{Writer{Next: os.Stderr}}}) + log.Store(&BasicLogger{Level: Info, Emitter: GoogleEmitter{&Writer{Next: os.Stderr}}}) } diff --git a/pkg/log/log_test.go b/pkg/log/log_test.go index 402cc29ae..9ff18559b 100644 --- a/pkg/log/log_test.go +++ b/pkg/log/log_test.go @@ -52,7 +52,7 @@ func TestDropMessages(t *testing.T) { t.Fatalf("Write should have failed") } - fmt.Printf("writer: %+v\n", w) + fmt.Printf("writer: %#v\n", &w) tw.fail = false if _, err := w.Write([]byte("line 2\n")); err != nil { @@ -76,7 +76,7 @@ func TestDropMessages(t *testing.T) { func TestCaller(t *testing.T) { tw := &testWriter{} - e := &GoogleEmitter{Writer: Writer{Next: tw}} + e := GoogleEmitter{Writer: &Writer{Next: tw}} bl := &BasicLogger{ Emitter: e, Level: Debug, @@ -94,7 +94,7 @@ func BenchmarkGoogleLogging(b *testing.B) { tw := &testWriter{ limit: 1, // Only record one message. } - e := &GoogleEmitter{Writer: Writer{Next: tw}} + e := GoogleEmitter{Writer: &Writer{Next: tw}} bl := &BasicLogger{ Emitter: e, Level: Debug, diff --git a/pkg/metric/metric.go b/pkg/metric/metric.go index 006fcd9ab..64aa365ce 100644 --- a/pkg/metric/metric.go +++ b/pkg/metric/metric.go @@ -39,16 +39,11 @@ var ( // Uint64Metric encapsulates a uint64 that represents some kind of metric to be // monitored. // -// All metrics must be cumulative, meaning that their values will only increase -// over time. -// // Metrics are not saved across save/restore and thus reset to zero on restore. // -// TODO(b/67298402): Support non-cumulative metrics. // TODO(b/67298427): Support metric fields. type Uint64Metric struct { - // value is the actual value of the metric. It must be accessed - // atomically. + // value is the actual value of the metric. It must be accessed atomically. value uint64 } @@ -110,13 +105,10 @@ type customUint64Metric struct { // Register must only be called at init and will return and error if called // after Initialized. // -// All metrics must be cumulative, meaning that the return values of value must -// only increase over time. -// // Preconditions: // * name must be globally unique. // * Initialize/Disable have not been called. -func RegisterCustomUint64Metric(name string, sync bool, description string, value func() uint64) error { +func RegisterCustomUint64Metric(name string, cumulative, sync bool, units pb.MetricMetadata_Units, description string, value func() uint64) error { if initialized { return ErrInitializationDone } @@ -129,9 +121,10 @@ func RegisterCustomUint64Metric(name string, sync bool, description string, valu metadata: &pb.MetricMetadata{ Name: name, Description: description, - Cumulative: true, + Cumulative: cumulative, Sync: sync, - Type: pb.MetricMetadata_UINT64, + Type: pb.MetricMetadata_TYPE_UINT64, + Units: units, }, value: value, } @@ -140,24 +133,32 @@ func RegisterCustomUint64Metric(name string, sync bool, description string, valu // MustRegisterCustomUint64Metric calls RegisterCustomUint64Metric and panics // if it returns an error. -func MustRegisterCustomUint64Metric(name string, sync bool, description string, value func() uint64) { - if err := RegisterCustomUint64Metric(name, sync, description, value); err != nil { +func MustRegisterCustomUint64Metric(name string, cumulative, sync bool, description string, value func() uint64) { + if err := RegisterCustomUint64Metric(name, cumulative, sync, pb.MetricMetadata_UNITS_NONE, description, value); err != nil { panic(fmt.Sprintf("Unable to register metric %q: %v", name, err)) } } -// NewUint64Metric creates and registers a new metric with the given name. +// NewUint64Metric creates and registers a new cumulative metric with the given name. // // Metrics must be statically defined (i.e., at init). -func NewUint64Metric(name string, sync bool, description string) (*Uint64Metric, error) { +func NewUint64Metric(name string, sync bool, units pb.MetricMetadata_Units, description string) (*Uint64Metric, error) { var m Uint64Metric - return &m, RegisterCustomUint64Metric(name, sync, description, m.Value) + return &m, RegisterCustomUint64Metric(name, true /* cumulative */, sync, units, description, m.Value) } -// MustCreateNewUint64Metric calls NewUint64Metric and panics if it returns an -// error. +// MustCreateNewUint64Metric calls NewUint64Metric and panics if it returns an error. func MustCreateNewUint64Metric(name string, sync bool, description string) *Uint64Metric { - m, err := NewUint64Metric(name, sync, description) + m, err := NewUint64Metric(name, sync, pb.MetricMetadata_UNITS_NONE, description) + if err != nil { + panic(fmt.Sprintf("Unable to create metric %q: %v", name, err)) + } + return m +} + +// MustCreateNewUint64NanosecondsMetric calls NewUint64Metric and panics if it returns an error. +func MustCreateNewUint64NanosecondsMetric(name string, sync bool, description string) *Uint64Metric { + m, err := NewUint64Metric(name, sync, pb.MetricMetadata_UNITS_NANOSECONDS, description) if err != nil { panic(fmt.Sprintf("Unable to create metric %q: %v", name, err)) } @@ -244,6 +245,6 @@ func EmitMetricUpdate() { return } - log.Debugf("Emitting metrics: %v", m) + log.Debugf("Emitting metrics: %v", &m) eventchannel.Emit(&m) } diff --git a/pkg/metric/metric.proto b/pkg/metric/metric.proto index a2c2bd1ba..3cc89047d 100644 --- a/pkg/metric/metric.proto +++ b/pkg/metric/metric.proto @@ -36,10 +36,18 @@ message MetricMetadata { // the monitoring system. bool sync = 4; - enum Type { UINT64 = 0; } + enum Type { TYPE_UINT64 = 0; } // type is the type of the metric value. Type type = 5; + + enum Units { + UNITS_NONE = 0; + UNITS_NANOSECONDS = 1; + } + + // units is the units of the metric value. + Units units = 6; } // MetricRegistration contains the metadata for all metrics that will be in diff --git a/pkg/metric/metric_test.go b/pkg/metric/metric_test.go index 34969385a..c425ea532 100644 --- a/pkg/metric/metric_test.go +++ b/pkg/metric/metric_test.go @@ -66,12 +66,12 @@ const ( func TestInitialize(t *testing.T) { defer reset() - _, err := NewUint64Metric("/foo", false, fooDescription) + _, err := NewUint64Metric("/foo", false, pb.MetricMetadata_UNITS_NONE, fooDescription) if err != nil { t.Fatalf("NewUint64Metric got err %v want nil", err) } - _, err = NewUint64Metric("/bar", true, barDescription) + _, err = NewUint64Metric("/bar", true, pb.MetricMetadata_UNITS_NANOSECONDS, barDescription) if err != nil { t.Fatalf("NewUint64Metric got err %v want nil", err) } @@ -94,8 +94,8 @@ func TestInitialize(t *testing.T) { foundFoo := false foundBar := false for _, m := range mr.Metrics { - if m.Type != pb.MetricMetadata_UINT64 { - t.Errorf("Metadata %+v Type got %v want %v", m, m.Type, pb.MetricMetadata_UINT64) + if m.Type != pb.MetricMetadata_TYPE_UINT64 { + t.Errorf("Metadata %+v Type got %v want %v", m, m.Type, pb.MetricMetadata_TYPE_UINT64) } if !m.Cumulative { t.Errorf("Metadata %+v Cumulative got false want true", m) @@ -110,6 +110,9 @@ func TestInitialize(t *testing.T) { if m.Sync { t.Errorf("/foo %+v Sync got true want false", m) } + if m.Units != pb.MetricMetadata_UNITS_NONE { + t.Errorf("/foo %+v Units got %v want %v", m, m.Units, pb.MetricMetadata_UNITS_NONE) + } case "/bar": foundBar = true if m.Description != barDescription { @@ -118,6 +121,9 @@ func TestInitialize(t *testing.T) { if !m.Sync { t.Errorf("/bar %+v Sync got true want false", m) } + if m.Units != pb.MetricMetadata_UNITS_NANOSECONDS { + t.Errorf("/bar %+v Units got %v want %v", m, m.Units, pb.MetricMetadata_UNITS_NANOSECONDS) + } } } @@ -132,12 +138,12 @@ func TestInitialize(t *testing.T) { func TestDisable(t *testing.T) { defer reset() - _, err := NewUint64Metric("/foo", false, fooDescription) + _, err := NewUint64Metric("/foo", false, pb.MetricMetadata_UNITS_NONE, fooDescription) if err != nil { t.Fatalf("NewUint64Metric got err %v want nil", err) } - _, err = NewUint64Metric("/bar", true, barDescription) + _, err = NewUint64Metric("/bar", true, pb.MetricMetadata_UNITS_NONE, barDescription) if err != nil { t.Fatalf("NewUint64Metric got err %v want nil", err) } @@ -161,12 +167,12 @@ func TestDisable(t *testing.T) { func TestEmitMetricUpdate(t *testing.T) { defer reset() - foo, err := NewUint64Metric("/foo", false, fooDescription) + foo, err := NewUint64Metric("/foo", false, pb.MetricMetadata_UNITS_NONE, fooDescription) if err != nil { t.Fatalf("NewUint64Metric got err %v want nil", err) } - _, err = NewUint64Metric("/bar", true, barDescription) + _, err = NewUint64Metric("/bar", true, pb.MetricMetadata_UNITS_NONE, barDescription) if err != nil { t.Fatalf("NewUint64Metric got err %v want nil", err) } diff --git a/pkg/p9/buffer.go b/pkg/p9/buffer.go index 249536d8a..6a4951821 100644 --- a/pkg/p9/buffer.go +++ b/pkg/p9/buffer.go @@ -20,16 +20,16 @@ import ( // encoder is used for messages and 9P primitives. type encoder interface { - // Decode decodes from the given buffer. Decode may be called more than once + // decode decodes from the given buffer. decode may be called more than once // to reuse the instance. It must clear any previous state. // // This may not fail, exhaustion will be recorded in the buffer. - Decode(b *buffer) + decode(b *buffer) - // Encode encodes to the given buffer. + // encode encodes to the given buffer. // // This may not fail. - Encode(b *buffer) + encode(b *buffer) } // order is the byte order used for encoding. @@ -39,7 +39,7 @@ var order = binary.LittleEndian // // This is passed to the encoder methods. type buffer struct { - // data is the underlying data. This may grow during Encode. + // data is the underlying data. This may grow during encode. data []byte // overflow indicates whether an overflow has occurred. diff --git a/pkg/p9/client.go b/pkg/p9/client.go index a6f493b82..71e944c30 100644 --- a/pkg/p9/client.go +++ b/pkg/p9/client.go @@ -174,7 +174,7 @@ func NewClient(socket *unet.Socket, messageSize uint32, version string) (*Client // our sendRecv function to use that functionality. Otherwise, // we stick to sendRecvLegacy. rversion := Rversion{} - err := c.sendRecvLegacy(&Tversion{ + _, err := c.sendRecvLegacy(&Tversion{ Version: versionString(requested), MSize: messageSize, }, &rversion) @@ -219,11 +219,11 @@ func NewClient(socket *unet.Socket, messageSize uint32, version string) (*Client c.sendRecv = c.sendRecvChannel } else { // Channel setup failed; fallback. - c.sendRecv = c.sendRecvLegacy + c.sendRecv = c.sendRecvLegacySyscallErr } } else { // No channels available: use the legacy mechanism. - c.sendRecv = c.sendRecvLegacy + c.sendRecv = c.sendRecvLegacySyscallErr } // Ensure that the socket and channels are closed when the socket is shut @@ -305,7 +305,7 @@ func (c *Client) openChannel(id int) error { ) // Open the data channel. - if err := c.sendRecvLegacy(&Tchannel{ + if _, err := c.sendRecvLegacy(&Tchannel{ ID: uint32(id), Control: 0, }, &rchannel0); err != nil { @@ -319,7 +319,7 @@ func (c *Client) openChannel(id int) error { defer rchannel0.FilePayload().Close() // Open the channel for file descriptors. - if err := c.sendRecvLegacy(&Tchannel{ + if _, err := c.sendRecvLegacy(&Tchannel{ ID: uint32(id), Control: 1, }, &rchannel1); err != nil { @@ -431,13 +431,28 @@ func (c *Client) waitAndRecv(done chan error) error { } } +// sendRecvLegacySyscallErr is a wrapper for sendRecvLegacy that converts all +// non-syscall errors to EIO. +func (c *Client) sendRecvLegacySyscallErr(t message, r message) error { + received, err := c.sendRecvLegacy(t, r) + if !received { + log.Warningf("p9.Client.sendRecvChannel: %v", err) + return syscall.EIO + } + return err +} + // sendRecvLegacy performs a roundtrip message exchange. // +// sendRecvLegacy returns true if a message was received. This allows us to +// differentiate between failed receives and successful receives where the +// response was an error message. +// // This is called by internal functions. -func (c *Client) sendRecvLegacy(t message, r message) error { +func (c *Client) sendRecvLegacy(t message, r message) (bool, error) { tag, ok := c.tagPool.Get() if !ok { - return ErrOutOfTags + return false, ErrOutOfTags } defer c.tagPool.Put(tag) @@ -457,12 +472,12 @@ func (c *Client) sendRecvLegacy(t message, r message) error { err := send(c.socket, Tag(tag), t) c.sendMu.Unlock() if err != nil { - return err + return false, err } // Co-ordinate with other receivers. if err := c.waitAndRecv(resp.done); err != nil { - return err + return false, err } // Is it an error message? @@ -470,14 +485,14 @@ func (c *Client) sendRecvLegacy(t message, r message) error { // For convenience, we transform these directly // into errors. Handlers need not handle this case. if rlerr, ok := resp.r.(*Rlerror); ok { - return syscall.Errno(rlerr.Error) + return true, syscall.Errno(rlerr.Error) } // At this point, we know it matches. // // Per recv call above, we will only allow a type // match (and give our r) or an instance of Rlerror. - return nil + return true, nil } // sendRecvChannel uses channels to send a message. @@ -486,7 +501,7 @@ func (c *Client) sendRecvChannel(t message, r message) error { c.channelsMu.Lock() if len(c.availableChannels) == 0 { c.channelsMu.Unlock() - return c.sendRecvLegacy(t, r) + return c.sendRecvLegacySyscallErr(t, r) } idx := len(c.availableChannels) - 1 ch := c.availableChannels[idx] @@ -526,7 +541,11 @@ func (c *Client) sendRecvChannel(t message, r message) error { } // Parse the server's response. - _, retErr := ch.recv(r, rsz) + resp, retErr := ch.recv(r, rsz) + if resp == nil { + log.Warningf("p9.Client.sendRecvChannel: p9.channel.recv: %v", retErr) + retErr = syscall.EIO + } // Release the channel. c.channelsMu.Lock() diff --git a/pkg/p9/client_test.go b/pkg/p9/client_test.go index 29a0afadf..c757583e0 100644 --- a/pkg/p9/client_test.go +++ b/pkg/p9/client_test.go @@ -96,7 +96,12 @@ func benchmarkSendRecv(b *testing.B, fn func(c *Client) func(message, message) e } func BenchmarkSendRecvLegacy(b *testing.B) { - benchmarkSendRecv(b, func(c *Client) func(message, message) error { return c.sendRecvLegacy }) + benchmarkSendRecv(b, func(c *Client) func(message, message) error { + return func(t message, r message) error { + _, err := c.sendRecvLegacy(t, r) + return err + } + }) } func BenchmarkSendRecvChannel(b *testing.B) { diff --git a/pkg/p9/file.go b/pkg/p9/file.go index d4ffbc8e3..cab35896f 100644 --- a/pkg/p9/file.go +++ b/pkg/p9/file.go @@ -97,12 +97,12 @@ type File interface { // free to ignore the hint entirely (i.e. the value returned may be larger // than size). All size checking is done independently at the syscall layer. // - // TODO(b/127675828): Determine concurrency guarantees once implemented. + // On the server, GetXattr has a read concurrency guarantee. GetXattr(name string, size uint64) (string, error) // SetXattr sets extended attributes on this node. // - // TODO(b/127675828): Determine concurrency guarantees once implemented. + // On the server, SetXattr has a write concurrency guarantee. SetXattr(name, value string, flags uint32) error // ListXattr lists the names of the extended attributes on this node. @@ -113,12 +113,12 @@ type File interface { // free to ignore the hint entirely (i.e. the value returned may be larger // than size). All size checking is done independently at the syscall layer. // - // TODO(b/148303075): Determine concurrency guarantees once implemented. + // On the server, ListXattr has a read concurrency guarantee. ListXattr(size uint64) (map[string]struct{}, error) // RemoveXattr removes extended attributes on this node. // - // TODO(b/148303075): Determine concurrency guarantees once implemented. + // On the server, RemoveXattr has a write concurrency guarantee. RemoveXattr(name string) error // Allocate allows the caller to directly manipulate the allocated disk space diff --git a/pkg/p9/handlers.go b/pkg/p9/handlers.go index 2ac45eb80..1db5797dd 100644 --- a/pkg/p9/handlers.go +++ b/pkg/p9/handlers.go @@ -48,6 +48,8 @@ func ExtractErrno(err error) syscall.Errno { return ExtractErrno(e.Err) case *os.SyscallError: return ExtractErrno(e.Err) + case *os.LinkError: + return ExtractErrno(e.Err) } // Default case. @@ -920,8 +922,15 @@ func (t *Tgetxattr) handle(cs *connState) message { } defer ref.DecRef() - val, err := ref.file.GetXattr(t.Name, t.Size) - if err != nil { + var val string + if err := ref.safelyRead(func() (err error) { + // Don't allow getxattr on files that have been deleted. + if ref.isDeleted() { + return syscall.EINVAL + } + val, err = ref.file.GetXattr(t.Name, t.Size) + return err + }); err != nil { return newErr(err) } return &Rgetxattr{Value: val} @@ -935,7 +944,13 @@ func (t *Tsetxattr) handle(cs *connState) message { } defer ref.DecRef() - if err := ref.file.SetXattr(t.Name, t.Value, t.Flags); err != nil { + if err := ref.safelyWrite(func() error { + // Don't allow setxattr on files that have been deleted. + if ref.isDeleted() { + return syscall.EINVAL + } + return ref.file.SetXattr(t.Name, t.Value, t.Flags) + }); err != nil { return newErr(err) } return &Rsetxattr{} @@ -949,10 +964,18 @@ func (t *Tlistxattr) handle(cs *connState) message { } defer ref.DecRef() - xattrs, err := ref.file.ListXattr(t.Size) - if err != nil { + var xattrs map[string]struct{} + if err := ref.safelyRead(func() (err error) { + // Don't allow listxattr on files that have been deleted. + if ref.isDeleted() { + return syscall.EINVAL + } + xattrs, err = ref.file.ListXattr(t.Size) + return err + }); err != nil { return newErr(err) } + xattrList := make([]string, 0, len(xattrs)) for x := range xattrs { xattrList = append(xattrList, x) @@ -968,7 +991,13 @@ func (t *Tremovexattr) handle(cs *connState) message { } defer ref.DecRef() - if err := ref.file.RemoveXattr(t.Name); err != nil { + if err := ref.safelyWrite(func() error { + // Don't allow removexattr on files that have been deleted. + if ref.isDeleted() { + return syscall.EINVAL + } + return ref.file.RemoveXattr(t.Name) + }); err != nil { return newErr(err) } return &Rremovexattr{} diff --git a/pkg/p9/messages.go b/pkg/p9/messages.go index b1cede5f5..57b89ad7d 100644 --- a/pkg/p9/messages.go +++ b/pkg/p9/messages.go @@ -51,7 +51,7 @@ type payloader interface { // SetPayload returns the decoded message. // // This is going to be total message size - FixedSize. But this should - // be validated during Decode, which will be called after SetPayload. + // be validated during decode, which will be called after SetPayload. SetPayload([]byte) } @@ -90,14 +90,14 @@ type Tversion struct { Version string } -// Decode implements encoder.Decode. -func (t *Tversion) Decode(b *buffer) { +// decode implements encoder.decode. +func (t *Tversion) decode(b *buffer) { t.MSize = b.Read32() t.Version = b.ReadString() } -// Encode implements encoder.Encode. -func (t *Tversion) Encode(b *buffer) { +// encode implements encoder.encode. +func (t *Tversion) encode(b *buffer) { b.Write32(t.MSize) b.WriteString(t.Version) } @@ -121,14 +121,14 @@ type Rversion struct { Version string } -// Decode implements encoder.Decode. -func (r *Rversion) Decode(b *buffer) { +// decode implements encoder.decode. +func (r *Rversion) decode(b *buffer) { r.MSize = b.Read32() r.Version = b.ReadString() } -// Encode implements encoder.Encode. -func (r *Rversion) Encode(b *buffer) { +// encode implements encoder.encode. +func (r *Rversion) encode(b *buffer) { b.Write32(r.MSize) b.WriteString(r.Version) } @@ -149,13 +149,13 @@ type Tflush struct { OldTag Tag } -// Decode implements encoder.Decode. -func (t *Tflush) Decode(b *buffer) { +// decode implements encoder.decode. +func (t *Tflush) decode(b *buffer) { t.OldTag = b.ReadTag() } -// Encode implements encoder.Encode. -func (t *Tflush) Encode(b *buffer) { +// encode implements encoder.encode. +func (t *Tflush) encode(b *buffer) { b.WriteTag(t.OldTag) } @@ -173,12 +173,12 @@ func (t *Tflush) String() string { type Rflush struct { } -// Decode implements encoder.Decode. -func (*Rflush) Decode(*buffer) { +// decode implements encoder.decode. +func (*Rflush) decode(*buffer) { } -// Encode implements encoder.Encode. -func (*Rflush) Encode(*buffer) { +// encode implements encoder.encode. +func (*Rflush) encode(*buffer) { } // Type implements message.Type. @@ -203,8 +203,8 @@ type Twalk struct { Names []string } -// Decode implements encoder.Decode. -func (t *Twalk) Decode(b *buffer) { +// decode implements encoder.decode. +func (t *Twalk) decode(b *buffer) { t.FID = b.ReadFID() t.NewFID = b.ReadFID() n := b.Read16() @@ -214,8 +214,8 @@ func (t *Twalk) Decode(b *buffer) { } } -// Encode implements encoder.Encode. -func (t *Twalk) Encode(b *buffer) { +// encode implements encoder.encode. +func (t *Twalk) encode(b *buffer) { b.WriteFID(t.FID) b.WriteFID(t.NewFID) b.Write16(uint16(len(t.Names))) @@ -240,22 +240,22 @@ type Rwalk struct { QIDs []QID } -// Decode implements encoder.Decode. -func (r *Rwalk) Decode(b *buffer) { +// decode implements encoder.decode. +func (r *Rwalk) decode(b *buffer) { n := b.Read16() r.QIDs = r.QIDs[:0] for i := 0; i < int(n); i++ { var q QID - q.Decode(b) + q.decode(b) r.QIDs = append(r.QIDs, q) } } -// Encode implements encoder.Encode. -func (r *Rwalk) Encode(b *buffer) { +// encode implements encoder.encode. +func (r *Rwalk) encode(b *buffer) { b.Write16(uint16(len(r.QIDs))) for _, q := range r.QIDs { - q.Encode(b) + q.encode(b) } } @@ -275,13 +275,13 @@ type Tclunk struct { FID FID } -// Decode implements encoder.Decode. -func (t *Tclunk) Decode(b *buffer) { +// decode implements encoder.decode. +func (t *Tclunk) decode(b *buffer) { t.FID = b.ReadFID() } -// Encode implements encoder.Encode. -func (t *Tclunk) Encode(b *buffer) { +// encode implements encoder.encode. +func (t *Tclunk) encode(b *buffer) { b.WriteFID(t.FID) } @@ -299,12 +299,12 @@ func (t *Tclunk) String() string { type Rclunk struct { } -// Decode implements encoder.Decode. -func (*Rclunk) Decode(*buffer) { +// decode implements encoder.decode. +func (*Rclunk) decode(*buffer) { } -// Encode implements encoder.Encode. -func (*Rclunk) Encode(*buffer) { +// encode implements encoder.encode. +func (*Rclunk) encode(*buffer) { } // Type implements message.Type. @@ -325,13 +325,13 @@ type Tremove struct { FID FID } -// Decode implements encoder.Decode. -func (t *Tremove) Decode(b *buffer) { +// decode implements encoder.decode. +func (t *Tremove) decode(b *buffer) { t.FID = b.ReadFID() } -// Encode implements encoder.Encode. -func (t *Tremove) Encode(b *buffer) { +// encode implements encoder.encode. +func (t *Tremove) encode(b *buffer) { b.WriteFID(t.FID) } @@ -349,12 +349,12 @@ func (t *Tremove) String() string { type Rremove struct { } -// Decode implements encoder.Decode. -func (*Rremove) Decode(*buffer) { +// decode implements encoder.decode. +func (*Rremove) decode(*buffer) { } -// Encode implements encoder.Encode. -func (*Rremove) Encode(*buffer) { +// encode implements encoder.encode. +func (*Rremove) encode(*buffer) { } // Type implements message.Type. @@ -374,13 +374,13 @@ type Rlerror struct { Error uint32 } -// Decode implements encoder.Decode. -func (r *Rlerror) Decode(b *buffer) { +// decode implements encoder.decode. +func (r *Rlerror) decode(b *buffer) { r.Error = b.Read32() } -// Encode implements encoder.Encode. -func (r *Rlerror) Encode(b *buffer) { +// encode implements encoder.encode. +func (r *Rlerror) encode(b *buffer) { b.Write32(r.Error) } @@ -409,16 +409,16 @@ type Tauth struct { UID UID } -// Decode implements encoder.Decode. -func (t *Tauth) Decode(b *buffer) { +// decode implements encoder.decode. +func (t *Tauth) decode(b *buffer) { t.AuthenticationFID = b.ReadFID() t.UserName = b.ReadString() t.AttachName = b.ReadString() t.UID = b.ReadUID() } -// Encode implements encoder.Encode. -func (t *Tauth) Encode(b *buffer) { +// encode implements encoder.encode. +func (t *Tauth) encode(b *buffer) { b.WriteFID(t.AuthenticationFID) b.WriteString(t.UserName) b.WriteString(t.AttachName) @@ -437,7 +437,7 @@ func (t *Tauth) String() string { // Rauth is an authentication response. // -// Encode, Decode and Length are inherited directly from QID. +// encode and decode are inherited directly from QID. type Rauth struct { QID } @@ -463,16 +463,16 @@ type Tattach struct { Auth Tauth } -// Decode implements encoder.Decode. -func (t *Tattach) Decode(b *buffer) { +// decode implements encoder.decode. +func (t *Tattach) decode(b *buffer) { t.FID = b.ReadFID() - t.Auth.Decode(b) + t.Auth.decode(b) } -// Encode implements encoder.Encode. -func (t *Tattach) Encode(b *buffer) { +// encode implements encoder.encode. +func (t *Tattach) encode(b *buffer) { b.WriteFID(t.FID) - t.Auth.Encode(b) + t.Auth.encode(b) } // Type implements message.Type. @@ -509,14 +509,14 @@ type Tlopen struct { Flags OpenFlags } -// Decode implements encoder.Decode. -func (t *Tlopen) Decode(b *buffer) { +// decode implements encoder.decode. +func (t *Tlopen) decode(b *buffer) { t.FID = b.ReadFID() t.Flags = b.ReadOpenFlags() } -// Encode implements encoder.Encode. -func (t *Tlopen) Encode(b *buffer) { +// encode implements encoder.encode. +func (t *Tlopen) encode(b *buffer) { b.WriteFID(t.FID) b.WriteOpenFlags(t.Flags) } @@ -542,15 +542,15 @@ type Rlopen struct { filePayload } -// Decode implements encoder.Decode. -func (r *Rlopen) Decode(b *buffer) { - r.QID.Decode(b) +// decode implements encoder.decode. +func (r *Rlopen) decode(b *buffer) { + r.QID.decode(b) r.IoUnit = b.Read32() } -// Encode implements encoder.Encode. -func (r *Rlopen) Encode(b *buffer) { - r.QID.Encode(b) +// encode implements encoder.encode. +func (r *Rlopen) encode(b *buffer) { + r.QID.encode(b) b.Write32(r.IoUnit) } @@ -587,8 +587,8 @@ type Tlcreate struct { GID GID } -// Decode implements encoder.Decode. -func (t *Tlcreate) Decode(b *buffer) { +// decode implements encoder.decode. +func (t *Tlcreate) decode(b *buffer) { t.FID = b.ReadFID() t.Name = b.ReadString() t.OpenFlags = b.ReadOpenFlags() @@ -596,8 +596,8 @@ func (t *Tlcreate) Decode(b *buffer) { t.GID = b.ReadGID() } -// Encode implements encoder.Encode. -func (t *Tlcreate) Encode(b *buffer) { +// encode implements encoder.encode. +func (t *Tlcreate) encode(b *buffer) { b.WriteFID(t.FID) b.WriteString(t.Name) b.WriteOpenFlags(t.OpenFlags) @@ -617,7 +617,7 @@ func (t *Tlcreate) String() string { // Rlcreate is a create response. // -// The Encode, Decode, etc. methods are inherited from Rlopen. +// The encode, decode, etc. methods are inherited from Rlopen. type Rlcreate struct { Rlopen } @@ -647,16 +647,16 @@ type Tsymlink struct { GID GID } -// Decode implements encoder.Decode. -func (t *Tsymlink) Decode(b *buffer) { +// decode implements encoder.decode. +func (t *Tsymlink) decode(b *buffer) { t.Directory = b.ReadFID() t.Name = b.ReadString() t.Target = b.ReadString() t.GID = b.ReadGID() } -// Encode implements encoder.Encode. -func (t *Tsymlink) Encode(b *buffer) { +// encode implements encoder.encode. +func (t *Tsymlink) encode(b *buffer) { b.WriteFID(t.Directory) b.WriteString(t.Name) b.WriteString(t.Target) @@ -679,14 +679,14 @@ type Rsymlink struct { QID QID } -// Decode implements encoder.Decode. -func (r *Rsymlink) Decode(b *buffer) { - r.QID.Decode(b) +// decode implements encoder.decode. +func (r *Rsymlink) decode(b *buffer) { + r.QID.decode(b) } -// Encode implements encoder.Encode. -func (r *Rsymlink) Encode(b *buffer) { - r.QID.Encode(b) +// encode implements encoder.encode. +func (r *Rsymlink) encode(b *buffer) { + r.QID.encode(b) } // Type implements message.Type. @@ -711,15 +711,15 @@ type Tlink struct { Name string } -// Decode implements encoder.Decode. -func (t *Tlink) Decode(b *buffer) { +// decode implements encoder.decode. +func (t *Tlink) decode(b *buffer) { t.Directory = b.ReadFID() t.Target = b.ReadFID() t.Name = b.ReadString() } -// Encode implements encoder.Encode. -func (t *Tlink) Encode(b *buffer) { +// encode implements encoder.encode. +func (t *Tlink) encode(b *buffer) { b.WriteFID(t.Directory) b.WriteFID(t.Target) b.WriteString(t.Name) @@ -744,12 +744,12 @@ func (*Rlink) Type() MsgType { return MsgRlink } -// Decode implements encoder.Decode. -func (*Rlink) Decode(*buffer) { +// decode implements encoder.decode. +func (*Rlink) decode(*buffer) { } -// Encode implements encoder.Encode. -func (*Rlink) Encode(*buffer) { +// encode implements encoder.encode. +func (*Rlink) encode(*buffer) { } // String implements fmt.Stringer. @@ -772,16 +772,16 @@ type Trenameat struct { NewName string } -// Decode implements encoder.Decode. -func (t *Trenameat) Decode(b *buffer) { +// decode implements encoder.decode. +func (t *Trenameat) decode(b *buffer) { t.OldDirectory = b.ReadFID() t.OldName = b.ReadString() t.NewDirectory = b.ReadFID() t.NewName = b.ReadString() } -// Encode implements encoder.Encode. -func (t *Trenameat) Encode(b *buffer) { +// encode implements encoder.encode. +func (t *Trenameat) encode(b *buffer) { b.WriteFID(t.OldDirectory) b.WriteString(t.OldName) b.WriteFID(t.NewDirectory) @@ -802,12 +802,12 @@ func (t *Trenameat) String() string { type Rrenameat struct { } -// Decode implements encoder.Decode. -func (*Rrenameat) Decode(*buffer) { +// decode implements encoder.decode. +func (*Rrenameat) decode(*buffer) { } -// Encode implements encoder.Encode. -func (*Rrenameat) Encode(*buffer) { +// encode implements encoder.encode. +func (*Rrenameat) encode(*buffer) { } // Type implements message.Type. @@ -832,15 +832,15 @@ type Tunlinkat struct { Flags uint32 } -// Decode implements encoder.Decode. -func (t *Tunlinkat) Decode(b *buffer) { +// decode implements encoder.decode. +func (t *Tunlinkat) decode(b *buffer) { t.Directory = b.ReadFID() t.Name = b.ReadString() t.Flags = b.Read32() } -// Encode implements encoder.Encode. -func (t *Tunlinkat) Encode(b *buffer) { +// encode implements encoder.encode. +func (t *Tunlinkat) encode(b *buffer) { b.WriteFID(t.Directory) b.WriteString(t.Name) b.Write32(t.Flags) @@ -860,12 +860,12 @@ func (t *Tunlinkat) String() string { type Runlinkat struct { } -// Decode implements encoder.Decode. -func (*Runlinkat) Decode(*buffer) { +// decode implements encoder.decode. +func (*Runlinkat) decode(*buffer) { } -// Encode implements encoder.Encode. -func (*Runlinkat) Encode(*buffer) { +// encode implements encoder.encode. +func (*Runlinkat) encode(*buffer) { } // Type implements message.Type. @@ -893,15 +893,15 @@ type Trename struct { Name string } -// Decode implements encoder.Decode. -func (t *Trename) Decode(b *buffer) { +// decode implements encoder.decode. +func (t *Trename) decode(b *buffer) { t.FID = b.ReadFID() t.Directory = b.ReadFID() t.Name = b.ReadString() } -// Encode implements encoder.Encode. -func (t *Trename) Encode(b *buffer) { +// encode implements encoder.encode. +func (t *Trename) encode(b *buffer) { b.WriteFID(t.FID) b.WriteFID(t.Directory) b.WriteString(t.Name) @@ -921,12 +921,12 @@ func (t *Trename) String() string { type Rrename struct { } -// Decode implements encoder.Decode. -func (*Rrename) Decode(*buffer) { +// decode implements encoder.decode. +func (*Rrename) decode(*buffer) { } -// Encode implements encoder.Encode. -func (*Rrename) Encode(*buffer) { +// encode implements encoder.encode. +func (*Rrename) encode(*buffer) { } // Type implements message.Type. @@ -945,13 +945,13 @@ type Treadlink struct { FID FID } -// Decode implements encoder.Decode. -func (t *Treadlink) Decode(b *buffer) { +// decode implements encoder.decode. +func (t *Treadlink) decode(b *buffer) { t.FID = b.ReadFID() } -// Encode implements encoder.Encode. -func (t *Treadlink) Encode(b *buffer) { +// encode implements encoder.encode. +func (t *Treadlink) encode(b *buffer) { b.WriteFID(t.FID) } @@ -971,13 +971,13 @@ type Rreadlink struct { Target string } -// Decode implements encoder.Decode. -func (r *Rreadlink) Decode(b *buffer) { +// decode implements encoder.decode. +func (r *Rreadlink) decode(b *buffer) { r.Target = b.ReadString() } -// Encode implements encoder.Encode. -func (r *Rreadlink) Encode(b *buffer) { +// encode implements encoder.encode. +func (r *Rreadlink) encode(b *buffer) { b.WriteString(r.Target) } @@ -1003,15 +1003,15 @@ type Tread struct { Count uint32 } -// Decode implements encoder.Decode. -func (t *Tread) Decode(b *buffer) { +// decode implements encoder.decode. +func (t *Tread) decode(b *buffer) { t.FID = b.ReadFID() t.Offset = b.Read64() t.Count = b.Read32() } -// Encode implements encoder.Encode. -func (t *Tread) Encode(b *buffer) { +// encode implements encoder.encode. +func (t *Tread) encode(b *buffer) { b.WriteFID(t.FID) b.Write64(t.Offset) b.Write32(t.Count) @@ -1033,20 +1033,20 @@ type Rread struct { Data []byte } -// Decode implements encoder.Decode. +// decode implements encoder.decode. // // Data is automatically decoded via Payload. -func (r *Rread) Decode(b *buffer) { +func (r *Rread) decode(b *buffer) { count := b.Read32() if count != uint32(len(r.Data)) { b.markOverrun() } } -// Encode implements encoder.Encode. +// encode implements encoder.encode. // // Data is automatically encoded via Payload. -func (r *Rread) Encode(b *buffer) { +func (r *Rread) encode(b *buffer) { b.Write32(uint32(len(r.Data))) } @@ -1087,8 +1087,8 @@ type Twrite struct { Data []byte } -// Decode implements encoder.Decode. -func (t *Twrite) Decode(b *buffer) { +// decode implements encoder.decode. +func (t *Twrite) decode(b *buffer) { t.FID = b.ReadFID() t.Offset = b.Read64() count := b.Read32() @@ -1097,10 +1097,10 @@ func (t *Twrite) Decode(b *buffer) { } } -// Encode implements encoder.Encode. +// encode implements encoder.encode. // // This uses the buffer payload to avoid a copy. -func (t *Twrite) Encode(b *buffer) { +func (t *Twrite) encode(b *buffer) { b.WriteFID(t.FID) b.Write64(t.Offset) b.Write32(uint32(len(t.Data))) @@ -1137,13 +1137,13 @@ type Rwrite struct { Count uint32 } -// Decode implements encoder.Decode. -func (r *Rwrite) Decode(b *buffer) { +// decode implements encoder.decode. +func (r *Rwrite) decode(b *buffer) { r.Count = b.Read32() } -// Encode implements encoder.Encode. -func (r *Rwrite) Encode(b *buffer) { +// encode implements encoder.encode. +func (r *Rwrite) encode(b *buffer) { b.Write32(r.Count) } @@ -1178,8 +1178,8 @@ type Tmknod struct { GID GID } -// Decode implements encoder.Decode. -func (t *Tmknod) Decode(b *buffer) { +// decode implements encoder.decode. +func (t *Tmknod) decode(b *buffer) { t.Directory = b.ReadFID() t.Name = b.ReadString() t.Mode = b.ReadFileMode() @@ -1188,8 +1188,8 @@ func (t *Tmknod) Decode(b *buffer) { t.GID = b.ReadGID() } -// Encode implements encoder.Encode. -func (t *Tmknod) Encode(b *buffer) { +// encode implements encoder.encode. +func (t *Tmknod) encode(b *buffer) { b.WriteFID(t.Directory) b.WriteString(t.Name) b.WriteFileMode(t.Mode) @@ -1214,14 +1214,14 @@ type Rmknod struct { QID QID } -// Decode implements encoder.Decode. -func (r *Rmknod) Decode(b *buffer) { - r.QID.Decode(b) +// decode implements encoder.decode. +func (r *Rmknod) decode(b *buffer) { + r.QID.decode(b) } -// Encode implements encoder.Encode. -func (r *Rmknod) Encode(b *buffer) { - r.QID.Encode(b) +// encode implements encoder.encode. +func (r *Rmknod) encode(b *buffer) { + r.QID.encode(b) } // Type implements message.Type. @@ -1249,16 +1249,16 @@ type Tmkdir struct { GID GID } -// Decode implements encoder.Decode. -func (t *Tmkdir) Decode(b *buffer) { +// decode implements encoder.decode. +func (t *Tmkdir) decode(b *buffer) { t.Directory = b.ReadFID() t.Name = b.ReadString() t.Permissions = b.ReadPermissions() t.GID = b.ReadGID() } -// Encode implements encoder.Encode. -func (t *Tmkdir) Encode(b *buffer) { +// encode implements encoder.encode. +func (t *Tmkdir) encode(b *buffer) { b.WriteFID(t.Directory) b.WriteString(t.Name) b.WritePermissions(t.Permissions) @@ -1281,14 +1281,14 @@ type Rmkdir struct { QID QID } -// Decode implements encoder.Decode. -func (r *Rmkdir) Decode(b *buffer) { - r.QID.Decode(b) +// decode implements encoder.decode. +func (r *Rmkdir) decode(b *buffer) { + r.QID.decode(b) } -// Encode implements encoder.Encode. -func (r *Rmkdir) Encode(b *buffer) { - r.QID.Encode(b) +// encode implements encoder.encode. +func (r *Rmkdir) encode(b *buffer) { + r.QID.encode(b) } // Type implements message.Type. @@ -1310,16 +1310,16 @@ type Tgetattr struct { AttrMask AttrMask } -// Decode implements encoder.Decode. -func (t *Tgetattr) Decode(b *buffer) { +// decode implements encoder.decode. +func (t *Tgetattr) decode(b *buffer) { t.FID = b.ReadFID() - t.AttrMask.Decode(b) + t.AttrMask.decode(b) } -// Encode implements encoder.Encode. -func (t *Tgetattr) Encode(b *buffer) { +// encode implements encoder.encode. +func (t *Tgetattr) encode(b *buffer) { b.WriteFID(t.FID) - t.AttrMask.Encode(b) + t.AttrMask.encode(b) } // Type implements message.Type. @@ -1344,18 +1344,18 @@ type Rgetattr struct { Attr Attr } -// Decode implements encoder.Decode. -func (r *Rgetattr) Decode(b *buffer) { - r.Valid.Decode(b) - r.QID.Decode(b) - r.Attr.Decode(b) +// decode implements encoder.decode. +func (r *Rgetattr) decode(b *buffer) { + r.Valid.decode(b) + r.QID.decode(b) + r.Attr.decode(b) } -// Encode implements encoder.Encode. -func (r *Rgetattr) Encode(b *buffer) { - r.Valid.Encode(b) - r.QID.Encode(b) - r.Attr.Encode(b) +// encode implements encoder.encode. +func (r *Rgetattr) encode(b *buffer) { + r.Valid.encode(b) + r.QID.encode(b) + r.Attr.encode(b) } // Type implements message.Type. @@ -1380,18 +1380,18 @@ type Tsetattr struct { SetAttr SetAttr } -// Decode implements encoder.Decode. -func (t *Tsetattr) Decode(b *buffer) { +// decode implements encoder.decode. +func (t *Tsetattr) decode(b *buffer) { t.FID = b.ReadFID() - t.Valid.Decode(b) - t.SetAttr.Decode(b) + t.Valid.decode(b) + t.SetAttr.decode(b) } -// Encode implements encoder.Encode. -func (t *Tsetattr) Encode(b *buffer) { +// encode implements encoder.encode. +func (t *Tsetattr) encode(b *buffer) { b.WriteFID(t.FID) - t.Valid.Encode(b) - t.SetAttr.Encode(b) + t.Valid.encode(b) + t.SetAttr.encode(b) } // Type implements message.Type. @@ -1408,12 +1408,12 @@ func (t *Tsetattr) String() string { type Rsetattr struct { } -// Decode implements encoder.Decode. -func (*Rsetattr) Decode(*buffer) { +// decode implements encoder.decode. +func (*Rsetattr) decode(*buffer) { } -// Encode implements encoder.Encode. -func (*Rsetattr) Encode(*buffer) { +// encode implements encoder.encode. +func (*Rsetattr) encode(*buffer) { } // Type implements message.Type. @@ -1435,18 +1435,18 @@ type Tallocate struct { Length uint64 } -// Decode implements encoder.Decode. -func (t *Tallocate) Decode(b *buffer) { +// decode implements encoder.decode. +func (t *Tallocate) decode(b *buffer) { t.FID = b.ReadFID() - t.Mode.Decode(b) + t.Mode.decode(b) t.Offset = b.Read64() t.Length = b.Read64() } -// Encode implements encoder.Encode. -func (t *Tallocate) Encode(b *buffer) { +// encode implements encoder.encode. +func (t *Tallocate) encode(b *buffer) { b.WriteFID(t.FID) - t.Mode.Encode(b) + t.Mode.encode(b) b.Write64(t.Offset) b.Write64(t.Length) } @@ -1465,12 +1465,12 @@ func (t *Tallocate) String() string { type Rallocate struct { } -// Decode implements encoder.Decode. -func (*Rallocate) Decode(*buffer) { +// decode implements encoder.decode. +func (*Rallocate) decode(*buffer) { } -// Encode implements encoder.Encode. -func (*Rallocate) Encode(*buffer) { +// encode implements encoder.encode. +func (*Rallocate) encode(*buffer) { } // Type implements message.Type. @@ -1492,14 +1492,14 @@ type Tlistxattr struct { Size uint64 } -// Decode implements encoder.Decode. -func (t *Tlistxattr) Decode(b *buffer) { +// decode implements encoder.decode. +func (t *Tlistxattr) decode(b *buffer) { t.FID = b.ReadFID() t.Size = b.Read64() } -// Encode implements encoder.Encode. -func (t *Tlistxattr) Encode(b *buffer) { +// encode implements encoder.encode. +func (t *Tlistxattr) encode(b *buffer) { b.WriteFID(t.FID) b.Write64(t.Size) } @@ -1520,8 +1520,8 @@ type Rlistxattr struct { Xattrs []string } -// Decode implements encoder.Decode. -func (r *Rlistxattr) Decode(b *buffer) { +// decode implements encoder.decode. +func (r *Rlistxattr) decode(b *buffer) { n := b.Read16() r.Xattrs = r.Xattrs[:0] for i := 0; i < int(n); i++ { @@ -1529,8 +1529,8 @@ func (r *Rlistxattr) Decode(b *buffer) { } } -// Encode implements encoder.Encode. -func (r *Rlistxattr) Encode(b *buffer) { +// encode implements encoder.encode. +func (r *Rlistxattr) encode(b *buffer) { b.Write16(uint16(len(r.Xattrs))) for _, x := range r.Xattrs { b.WriteString(x) @@ -1559,15 +1559,15 @@ type Txattrwalk struct { Name string } -// Decode implements encoder.Decode. -func (t *Txattrwalk) Decode(b *buffer) { +// decode implements encoder.decode. +func (t *Txattrwalk) decode(b *buffer) { t.FID = b.ReadFID() t.NewFID = b.ReadFID() t.Name = b.ReadString() } -// Encode implements encoder.Encode. -func (t *Txattrwalk) Encode(b *buffer) { +// encode implements encoder.encode. +func (t *Txattrwalk) encode(b *buffer) { b.WriteFID(t.FID) b.WriteFID(t.NewFID) b.WriteString(t.Name) @@ -1589,13 +1589,13 @@ type Rxattrwalk struct { Size uint64 } -// Decode implements encoder.Decode. -func (r *Rxattrwalk) Decode(b *buffer) { +// decode implements encoder.decode. +func (r *Rxattrwalk) decode(b *buffer) { r.Size = b.Read64() } -// Encode implements encoder.Encode. -func (r *Rxattrwalk) Encode(b *buffer) { +// encode implements encoder.encode. +func (r *Rxattrwalk) encode(b *buffer) { b.Write64(r.Size) } @@ -1627,16 +1627,16 @@ type Txattrcreate struct { Flags uint32 } -// Decode implements encoder.Decode. -func (t *Txattrcreate) Decode(b *buffer) { +// decode implements encoder.decode. +func (t *Txattrcreate) decode(b *buffer) { t.FID = b.ReadFID() t.Name = b.ReadString() t.AttrSize = b.Read64() t.Flags = b.Read32() } -// Encode implements encoder.Encode. -func (t *Txattrcreate) Encode(b *buffer) { +// encode implements encoder.encode. +func (t *Txattrcreate) encode(b *buffer) { b.WriteFID(t.FID) b.WriteString(t.Name) b.Write64(t.AttrSize) @@ -1657,12 +1657,12 @@ func (t *Txattrcreate) String() string { type Rxattrcreate struct { } -// Decode implements encoder.Decode. -func (r *Rxattrcreate) Decode(*buffer) { +// decode implements encoder.decode. +func (r *Rxattrcreate) decode(*buffer) { } -// Encode implements encoder.Encode. -func (r *Rxattrcreate) Encode(*buffer) { +// encode implements encoder.encode. +func (r *Rxattrcreate) encode(*buffer) { } // Type implements message.Type. @@ -1687,15 +1687,15 @@ type Tgetxattr struct { Size uint64 } -// Decode implements encoder.Decode. -func (t *Tgetxattr) Decode(b *buffer) { +// decode implements encoder.decode. +func (t *Tgetxattr) decode(b *buffer) { t.FID = b.ReadFID() t.Name = b.ReadString() t.Size = b.Read64() } -// Encode implements encoder.Encode. -func (t *Tgetxattr) Encode(b *buffer) { +// encode implements encoder.encode. +func (t *Tgetxattr) encode(b *buffer) { b.WriteFID(t.FID) b.WriteString(t.Name) b.Write64(t.Size) @@ -1717,13 +1717,13 @@ type Rgetxattr struct { Value string } -// Decode implements encoder.Decode. -func (r *Rgetxattr) Decode(b *buffer) { +// decode implements encoder.decode. +func (r *Rgetxattr) decode(b *buffer) { r.Value = b.ReadString() } -// Encode implements encoder.Encode. -func (r *Rgetxattr) Encode(b *buffer) { +// encode implements encoder.encode. +func (r *Rgetxattr) encode(b *buffer) { b.WriteString(r.Value) } @@ -1752,16 +1752,16 @@ type Tsetxattr struct { Flags uint32 } -// Decode implements encoder.Decode. -func (t *Tsetxattr) Decode(b *buffer) { +// decode implements encoder.decode. +func (t *Tsetxattr) decode(b *buffer) { t.FID = b.ReadFID() t.Name = b.ReadString() t.Value = b.ReadString() t.Flags = b.Read32() } -// Encode implements encoder.Encode. -func (t *Tsetxattr) Encode(b *buffer) { +// encode implements encoder.encode. +func (t *Tsetxattr) encode(b *buffer) { b.WriteFID(t.FID) b.WriteString(t.Name) b.WriteString(t.Value) @@ -1782,12 +1782,12 @@ func (t *Tsetxattr) String() string { type Rsetxattr struct { } -// Decode implements encoder.Decode. -func (r *Rsetxattr) Decode(*buffer) { +// decode implements encoder.decode. +func (r *Rsetxattr) decode(*buffer) { } -// Encode implements encoder.Encode. -func (r *Rsetxattr) Encode(*buffer) { +// encode implements encoder.encode. +func (r *Rsetxattr) encode(*buffer) { } // Type implements message.Type. @@ -1809,14 +1809,14 @@ type Tremovexattr struct { Name string } -// Decode implements encoder.Decode. -func (t *Tremovexattr) Decode(b *buffer) { +// decode implements encoder.decode. +func (t *Tremovexattr) decode(b *buffer) { t.FID = b.ReadFID() t.Name = b.ReadString() } -// Encode implements encoder.Encode. -func (t *Tremovexattr) Encode(b *buffer) { +// encode implements encoder.encode. +func (t *Tremovexattr) encode(b *buffer) { b.WriteFID(t.FID) b.WriteString(t.Name) } @@ -1835,12 +1835,12 @@ func (t *Tremovexattr) String() string { type Rremovexattr struct { } -// Decode implements encoder.Decode. -func (r *Rremovexattr) Decode(*buffer) { +// decode implements encoder.decode. +func (r *Rremovexattr) decode(*buffer) { } -// Encode implements encoder.Encode. -func (r *Rremovexattr) Encode(*buffer) { +// encode implements encoder.encode. +func (r *Rremovexattr) encode(*buffer) { } // Type implements message.Type. @@ -1865,15 +1865,15 @@ type Treaddir struct { Count uint32 } -// Decode implements encoder.Decode. -func (t *Treaddir) Decode(b *buffer) { +// decode implements encoder.decode. +func (t *Treaddir) decode(b *buffer) { t.Directory = b.ReadFID() t.Offset = b.Read64() t.Count = b.Read32() } -// Encode implements encoder.Encode. -func (t *Treaddir) Encode(b *buffer) { +// encode implements encoder.encode. +func (t *Treaddir) encode(b *buffer) { b.WriteFID(t.Directory) b.Write64(t.Offset) b.Write32(t.Count) @@ -1907,14 +1907,14 @@ type Rreaddir struct { payload []byte } -// Decode implements encoder.Decode. -func (r *Rreaddir) Decode(b *buffer) { +// decode implements encoder.decode. +func (r *Rreaddir) decode(b *buffer) { r.Count = b.Read32() entriesBuf := buffer{data: r.payload} r.Entries = r.Entries[:0] for { var d Dirent - d.Decode(&entriesBuf) + d.decode(&entriesBuf) if entriesBuf.isOverrun() { // Couldn't decode a complete entry. break @@ -1923,22 +1923,20 @@ func (r *Rreaddir) Decode(b *buffer) { } } -// Encode implements encoder.Encode. -func (r *Rreaddir) Encode(b *buffer) { +// encode implements encoder.encode. +func (r *Rreaddir) encode(b *buffer) { entriesBuf := buffer{} + payloadSize := 0 for _, d := range r.Entries { - d.Encode(&entriesBuf) - if len(entriesBuf.data) >= int(r.Count) { + d.encode(&entriesBuf) + if len(entriesBuf.data) > int(r.Count) { break } + payloadSize = len(entriesBuf.data) } - if len(entriesBuf.data) < int(r.Count) { - r.Count = uint32(len(entriesBuf.data)) - r.payload = entriesBuf.data - } else { - r.payload = entriesBuf.data[:r.Count] - } - b.Write32(uint32(r.Count)) + r.Count = uint32(payloadSize) + r.payload = entriesBuf.data[:payloadSize] + b.Write32(r.Count) } // Type implements message.Type. @@ -1972,13 +1970,13 @@ type Tfsync struct { FID FID } -// Decode implements encoder.Decode. -func (t *Tfsync) Decode(b *buffer) { +// decode implements encoder.decode. +func (t *Tfsync) decode(b *buffer) { t.FID = b.ReadFID() } -// Encode implements encoder.Encode. -func (t *Tfsync) Encode(b *buffer) { +// encode implements encoder.encode. +func (t *Tfsync) encode(b *buffer) { b.WriteFID(t.FID) } @@ -1996,12 +1994,12 @@ func (t *Tfsync) String() string { type Rfsync struct { } -// Decode implements encoder.Decode. -func (*Rfsync) Decode(*buffer) { +// decode implements encoder.decode. +func (*Rfsync) decode(*buffer) { } -// Encode implements encoder.Encode. -func (*Rfsync) Encode(*buffer) { +// encode implements encoder.encode. +func (*Rfsync) encode(*buffer) { } // Type implements message.Type. @@ -2020,13 +2018,13 @@ type Tstatfs struct { FID FID } -// Decode implements encoder.Decode. -func (t *Tstatfs) Decode(b *buffer) { +// decode implements encoder.decode. +func (t *Tstatfs) decode(b *buffer) { t.FID = b.ReadFID() } -// Encode implements encoder.Encode. -func (t *Tstatfs) Encode(b *buffer) { +// encode implements encoder.encode. +func (t *Tstatfs) encode(b *buffer) { b.WriteFID(t.FID) } @@ -2046,14 +2044,14 @@ type Rstatfs struct { FSStat FSStat } -// Decode implements encoder.Decode. -func (r *Rstatfs) Decode(b *buffer) { - r.FSStat.Decode(b) +// decode implements encoder.decode. +func (r *Rstatfs) decode(b *buffer) { + r.FSStat.decode(b) } -// Encode implements encoder.Encode. -func (r *Rstatfs) Encode(b *buffer) { - r.FSStat.Encode(b) +// encode implements encoder.encode. +func (r *Rstatfs) encode(b *buffer) { + r.FSStat.encode(b) } // Type implements message.Type. @@ -2072,13 +2070,13 @@ type Tflushf struct { FID FID } -// Decode implements encoder.Decode. -func (t *Tflushf) Decode(b *buffer) { +// decode implements encoder.decode. +func (t *Tflushf) decode(b *buffer) { t.FID = b.ReadFID() } -// Encode implements encoder.Encode. -func (t *Tflushf) Encode(b *buffer) { +// encode implements encoder.encode. +func (t *Tflushf) encode(b *buffer) { b.WriteFID(t.FID) } @@ -2096,12 +2094,12 @@ func (t *Tflushf) String() string { type Rflushf struct { } -// Decode implements encoder.Decode. -func (*Rflushf) Decode(*buffer) { +// decode implements encoder.decode. +func (*Rflushf) decode(*buffer) { } -// Encode implements encoder.Encode. -func (*Rflushf) Encode(*buffer) { +// encode implements encoder.encode. +func (*Rflushf) encode(*buffer) { } // Type implements message.Type. @@ -2126,8 +2124,8 @@ type Twalkgetattr struct { Names []string } -// Decode implements encoder.Decode. -func (t *Twalkgetattr) Decode(b *buffer) { +// decode implements encoder.decode. +func (t *Twalkgetattr) decode(b *buffer) { t.FID = b.ReadFID() t.NewFID = b.ReadFID() n := b.Read16() @@ -2137,8 +2135,8 @@ func (t *Twalkgetattr) Decode(b *buffer) { } } -// Encode implements encoder.Encode. -func (t *Twalkgetattr) Encode(b *buffer) { +// encode implements encoder.encode. +func (t *Twalkgetattr) encode(b *buffer) { b.WriteFID(t.FID) b.WriteFID(t.NewFID) b.Write16(uint16(len(t.Names))) @@ -2169,26 +2167,26 @@ type Rwalkgetattr struct { QIDs []QID } -// Decode implements encoder.Decode. -func (r *Rwalkgetattr) Decode(b *buffer) { - r.Valid.Decode(b) - r.Attr.Decode(b) +// decode implements encoder.decode. +func (r *Rwalkgetattr) decode(b *buffer) { + r.Valid.decode(b) + r.Attr.decode(b) n := b.Read16() r.QIDs = r.QIDs[:0] for i := 0; i < int(n); i++ { var q QID - q.Decode(b) + q.decode(b) r.QIDs = append(r.QIDs, q) } } -// Encode implements encoder.Encode. -func (r *Rwalkgetattr) Encode(b *buffer) { - r.Valid.Encode(b) - r.Attr.Encode(b) +// encode implements encoder.encode. +func (r *Rwalkgetattr) encode(b *buffer) { + r.Valid.encode(b) + r.Attr.encode(b) b.Write16(uint16(len(r.QIDs))) for _, q := range r.QIDs { - q.Encode(b) + q.encode(b) } } @@ -2210,15 +2208,15 @@ type Tucreate struct { UID UID } -// Decode implements encoder.Decode. -func (t *Tucreate) Decode(b *buffer) { - t.Tlcreate.Decode(b) +// decode implements encoder.decode. +func (t *Tucreate) decode(b *buffer) { + t.Tlcreate.decode(b) t.UID = b.ReadUID() } -// Encode implements encoder.Encode. -func (t *Tucreate) Encode(b *buffer) { - t.Tlcreate.Encode(b) +// encode implements encoder.encode. +func (t *Tucreate) encode(b *buffer) { + t.Tlcreate.encode(b) b.WriteUID(t.UID) } @@ -2255,15 +2253,15 @@ type Tumkdir struct { UID UID } -// Decode implements encoder.Decode. -func (t *Tumkdir) Decode(b *buffer) { - t.Tmkdir.Decode(b) +// decode implements encoder.decode. +func (t *Tumkdir) decode(b *buffer) { + t.Tmkdir.decode(b) t.UID = b.ReadUID() } -// Encode implements encoder.Encode. -func (t *Tumkdir) Encode(b *buffer) { - t.Tmkdir.Encode(b) +// encode implements encoder.encode. +func (t *Tumkdir) encode(b *buffer) { + t.Tmkdir.encode(b) b.WriteUID(t.UID) } @@ -2300,15 +2298,15 @@ type Tumknod struct { UID UID } -// Decode implements encoder.Decode. -func (t *Tumknod) Decode(b *buffer) { - t.Tmknod.Decode(b) +// decode implements encoder.decode. +func (t *Tumknod) decode(b *buffer) { + t.Tmknod.decode(b) t.UID = b.ReadUID() } -// Encode implements encoder.Encode. -func (t *Tumknod) Encode(b *buffer) { - t.Tmknod.Encode(b) +// encode implements encoder.encode. +func (t *Tumknod) encode(b *buffer) { + t.Tmknod.encode(b) b.WriteUID(t.UID) } @@ -2345,15 +2343,15 @@ type Tusymlink struct { UID UID } -// Decode implements encoder.Decode. -func (t *Tusymlink) Decode(b *buffer) { - t.Tsymlink.Decode(b) +// decode implements encoder.decode. +func (t *Tusymlink) decode(b *buffer) { + t.Tsymlink.decode(b) t.UID = b.ReadUID() } -// Encode implements encoder.Encode. -func (t *Tusymlink) Encode(b *buffer) { - t.Tsymlink.Encode(b) +// encode implements encoder.encode. +func (t *Tusymlink) encode(b *buffer) { + t.Tsymlink.encode(b) b.WriteUID(t.UID) } @@ -2391,14 +2389,14 @@ type Tlconnect struct { Flags ConnectFlags } -// Decode implements encoder.Decode. -func (t *Tlconnect) Decode(b *buffer) { +// decode implements encoder.decode. +func (t *Tlconnect) decode(b *buffer) { t.FID = b.ReadFID() t.Flags = b.ReadConnectFlags() } -// Encode implements encoder.Encode. -func (t *Tlconnect) Encode(b *buffer) { +// encode implements encoder.encode. +func (t *Tlconnect) encode(b *buffer) { b.WriteFID(t.FID) b.WriteConnectFlags(t.Flags) } @@ -2418,11 +2416,11 @@ type Rlconnect struct { filePayload } -// Decode implements encoder.Decode. -func (r *Rlconnect) Decode(*buffer) {} +// decode implements encoder.decode. +func (r *Rlconnect) decode(*buffer) {} -// Encode implements encoder.Encode. -func (r *Rlconnect) Encode(*buffer) {} +// encode implements encoder.encode. +func (r *Rlconnect) encode(*buffer) {} // Type implements message.Type. func (*Rlconnect) Type() MsgType { @@ -2445,14 +2443,14 @@ type Tchannel struct { Control uint32 } -// Decode implements encoder.Decode. -func (t *Tchannel) Decode(b *buffer) { +// decode implements encoder.decode. +func (t *Tchannel) decode(b *buffer) { t.ID = b.Read32() t.Control = b.Read32() } -// Encode implements encoder.Encode. -func (t *Tchannel) Encode(b *buffer) { +// encode implements encoder.encode. +func (t *Tchannel) encode(b *buffer) { b.Write32(t.ID) b.Write32(t.Control) } @@ -2474,14 +2472,14 @@ type Rchannel struct { filePayload } -// Decode implements encoder.Decode. -func (r *Rchannel) Decode(b *buffer) { +// decode implements encoder.decode. +func (r *Rchannel) decode(b *buffer) { r.Offset = b.Read64() r.Length = b.Read64() } -// Encode implements encoder.Encode. -func (r *Rchannel) Encode(b *buffer) { +// encode implements encoder.encode. +func (r *Rchannel) encode(b *buffer) { b.Write64(r.Offset) b.Write64(r.Length) } @@ -2577,7 +2575,7 @@ func calculateSize(m message) uint32 { return p.FixedSize() } var dataBuf buffer - m.Encode(&dataBuf) + m.encode(&dataBuf) return uint32(len(dataBuf.data)) } diff --git a/pkg/p9/messages_test.go b/pkg/p9/messages_test.go index 825c939da..7facc9f5e 100644 --- a/pkg/p9/messages_test.go +++ b/pkg/p9/messages_test.go @@ -216,7 +216,7 @@ func TestEncodeDecode(t *testing.T) { }, &Rreaddir{ // Count must be sufficient to encode a dirent. - Count: 0x18, + Count: 0x1a, Entries: []Dirent{{QID: QID{Type: 2}}}, }, &Tfsync{ @@ -382,7 +382,7 @@ func TestEncodeDecode(t *testing.T) { // Encode the original. data := make([]byte, initialBufferLength) buf := buffer{data: data[:0]} - enc.Encode(&buf) + enc.encode(&buf) // Create a new object, same as the first. enc2 := reflect.New(reflect.ValueOf(enc).Elem().Type()).Interface().(encoder) @@ -399,7 +399,7 @@ func TestEncodeDecode(t *testing.T) { } // Mark sure it was okay. - enc2.Decode(&buf2) + enc2.decode(&buf2) if buf2.isOverrun() { t.Errorf("object %#v->%#v got overrun on decode", enc, enc2) continue diff --git a/pkg/p9/p9.go b/pkg/p9/p9.go index 20ab31f7a..28d851ff5 100644 --- a/pkg/p9/p9.go +++ b/pkg/p9/p9.go @@ -450,15 +450,15 @@ func (q QID) String() string { return fmt.Sprintf("QID{Type: %d, Version: %d, Path: %d}", q.Type, q.Version, q.Path) } -// Decode implements encoder.Decode. -func (q *QID) Decode(b *buffer) { +// decode implements encoder.decode. +func (q *QID) decode(b *buffer) { q.Type = b.ReadQIDType() q.Version = b.Read32() q.Path = b.Read64() } -// Encode implements encoder.Encode. -func (q *QID) Encode(b *buffer) { +// encode implements encoder.encode. +func (q *QID) encode(b *buffer) { b.WriteQIDType(q.Type) b.Write32(q.Version) b.Write64(q.Path) @@ -515,8 +515,8 @@ type FSStat struct { NameLength uint32 } -// Decode implements encoder.Decode. -func (f *FSStat) Decode(b *buffer) { +// decode implements encoder.decode. +func (f *FSStat) decode(b *buffer) { f.Type = b.Read32() f.BlockSize = b.Read32() f.Blocks = b.Read64() @@ -528,8 +528,8 @@ func (f *FSStat) Decode(b *buffer) { f.NameLength = b.Read32() } -// Encode implements encoder.Encode. -func (f *FSStat) Encode(b *buffer) { +// encode implements encoder.encode. +func (f *FSStat) encode(b *buffer) { b.Write32(f.Type) b.Write32(f.BlockSize) b.Write64(f.Blocks) @@ -679,8 +679,8 @@ func (a AttrMask) String() string { return fmt.Sprintf("AttrMask{with: %s}", strings.Join(masks, " ")) } -// Decode implements encoder.Decode. -func (a *AttrMask) Decode(b *buffer) { +// decode implements encoder.decode. +func (a *AttrMask) decode(b *buffer) { mask := b.Read64() a.Mode = mask&0x00000001 != 0 a.NLink = mask&0x00000002 != 0 @@ -698,8 +698,8 @@ func (a *AttrMask) Decode(b *buffer) { a.DataVersion = mask&0x00002000 != 0 } -// Encode implements encoder.Encode. -func (a *AttrMask) Encode(b *buffer) { +// encode implements encoder.encode. +func (a *AttrMask) encode(b *buffer) { var mask uint64 if a.Mode { mask |= 0x00000001 @@ -774,8 +774,8 @@ func (a Attr) String() string { a.Mode, a.UID, a.GID, a.NLink, a.RDev, a.Size, a.BlockSize, a.Blocks, a.ATimeSeconds, a.ATimeNanoSeconds, a.MTimeSeconds, a.MTimeNanoSeconds, a.CTimeSeconds, a.CTimeNanoSeconds, a.BTimeSeconds, a.BTimeNanoSeconds, a.Gen, a.DataVersion) } -// Encode implements encoder.Encode. -func (a *Attr) Encode(b *buffer) { +// encode implements encoder.encode. +func (a *Attr) encode(b *buffer) { b.WriteFileMode(a.Mode) b.WriteUID(a.UID) b.WriteGID(a.GID) @@ -796,8 +796,8 @@ func (a *Attr) Encode(b *buffer) { b.Write64(a.DataVersion) } -// Decode implements encoder.Decode. -func (a *Attr) Decode(b *buffer) { +// decode implements encoder.decode. +func (a *Attr) decode(b *buffer) { a.Mode = b.ReadFileMode() a.UID = b.ReadUID() a.GID = b.ReadGID() @@ -926,8 +926,8 @@ func (s SetAttrMask) Empty() bool { return !s.Permissions && !s.UID && !s.GID && !s.Size && !s.ATime && !s.MTime && !s.CTime && !s.ATimeNotSystemTime && !s.MTimeNotSystemTime } -// Decode implements encoder.Decode. -func (s *SetAttrMask) Decode(b *buffer) { +// decode implements encoder.decode. +func (s *SetAttrMask) decode(b *buffer) { mask := b.Read32() s.Permissions = mask&0x00000001 != 0 s.UID = mask&0x00000002 != 0 @@ -972,8 +972,8 @@ func (s SetAttrMask) bitmask() uint32 { return mask } -// Encode implements encoder.Encode. -func (s *SetAttrMask) Encode(b *buffer) { +// encode implements encoder.encode. +func (s *SetAttrMask) encode(b *buffer) { b.Write32(s.bitmask()) } @@ -994,8 +994,8 @@ func (s SetAttr) String() string { return fmt.Sprintf("SetAttr{Permissions: 0o%o, UID: %d, GID: %d, Size: %d, ATime: {Sec: %d, NanoSec: %d}, MTime: {Sec: %d, NanoSec: %d}}", s.Permissions, s.UID, s.GID, s.Size, s.ATimeSeconds, s.ATimeNanoSeconds, s.MTimeSeconds, s.MTimeNanoSeconds) } -// Decode implements encoder.Decode. -func (s *SetAttr) Decode(b *buffer) { +// decode implements encoder.decode. +func (s *SetAttr) decode(b *buffer) { s.Permissions = b.ReadPermissions() s.UID = b.ReadUID() s.GID = b.ReadGID() @@ -1006,8 +1006,8 @@ func (s *SetAttr) Decode(b *buffer) { s.MTimeNanoSeconds = b.Read64() } -// Encode implements encoder.Encode. -func (s *SetAttr) Encode(b *buffer) { +// encode implements encoder.encode. +func (s *SetAttr) encode(b *buffer) { b.WritePermissions(s.Permissions) b.WriteUID(s.UID) b.WriteGID(s.GID) @@ -1064,17 +1064,17 @@ func (d Dirent) String() string { return fmt.Sprintf("Dirent{QID: %d, Offset: %d, Type: 0x%X, Name: %s}", d.QID, d.Offset, d.Type, d.Name) } -// Decode implements encoder.Decode. -func (d *Dirent) Decode(b *buffer) { - d.QID.Decode(b) +// decode implements encoder.decode. +func (d *Dirent) decode(b *buffer) { + d.QID.decode(b) d.Offset = b.Read64() d.Type = b.ReadQIDType() d.Name = b.ReadString() } -// Encode implements encoder.Encode. -func (d *Dirent) Encode(b *buffer) { - d.QID.Encode(b) +// encode implements encoder.encode. +func (d *Dirent) encode(b *buffer) { + d.QID.encode(b) b.Write64(d.Offset) b.WriteQIDType(d.Type) b.WriteString(d.Name) @@ -1118,8 +1118,8 @@ func (a *AllocateMode) ToLinux() uint32 { return rv } -// Decode implements encoder.Decode. -func (a *AllocateMode) Decode(b *buffer) { +// decode implements encoder.decode. +func (a *AllocateMode) decode(b *buffer) { mask := b.Read32() a.KeepSize = mask&0x01 != 0 a.PunchHole = mask&0x02 != 0 @@ -1130,8 +1130,8 @@ func (a *AllocateMode) Decode(b *buffer) { a.Unshare = mask&0x40 != 0 } -// Encode implements encoder.Encode. -func (a *AllocateMode) Encode(b *buffer) { +// encode implements encoder.encode. +func (a *AllocateMode) encode(b *buffer) { mask := uint32(0) if a.KeepSize { mask |= 0x01 diff --git a/pkg/p9/transport.go b/pkg/p9/transport.go index 9c11e28ce..7cec0e86d 100644 --- a/pkg/p9/transport.go +++ b/pkg/p9/transport.go @@ -80,7 +80,7 @@ func send(s *unet.Socket, tag Tag, m message) error { } // Encode the message. The buffer will grow automatically. - m.Encode(&dataBuf) + m.encode(&dataBuf) // Get our vectors to send. var hdr [headerLength]byte @@ -316,7 +316,7 @@ func recv(s *unet.Socket, msize uint32, lookup lookupTagAndType) (Tag, message, } // Decode the message data. - m.Decode(&dataBuf) + m.decode(&dataBuf) if dataBuf.isOverrun() { // No need to drain the socket. return NoTag, nil, ErrNoValidMessage diff --git a/pkg/p9/transport_flipcall.go b/pkg/p9/transport_flipcall.go index 233f825e3..38038abdf 100644 --- a/pkg/p9/transport_flipcall.go +++ b/pkg/p9/transport_flipcall.go @@ -151,7 +151,7 @@ func (ch *channel) send(m message) (uint32, error) { } else { ch.buf.Write8(0) // No incoming FD. } - m.Encode(&ch.buf) + m.encode(&ch.buf) ssz := uint32(len(ch.buf.data)) // Updated below. // Is there a payload? @@ -205,7 +205,7 @@ func (ch *channel) recv(r message, rsz uint32) (message, error) { ch.buf.data = ch.buf.data[:fs] } - r.Decode(&ch.buf) + r.decode(&ch.buf) if ch.buf.isOverrun() { // Nothing valid was available. log.Debugf("recv [got %d bytes, needed more]", rsz) @@ -236,7 +236,7 @@ func (ch *channel) recv(r message, rsz uint32) (message, error) { // Convert errors appropriately; see above. if rlerr, ok := r.(*Rlerror); ok { - return nil, syscall.Errno(rlerr.Error) + return r, syscall.Errno(rlerr.Error) } return r, nil diff --git a/pkg/p9/transport_test.go b/pkg/p9/transport_test.go index 2f50ff3ea..3668fcad7 100644 --- a/pkg/p9/transport_test.go +++ b/pkg/p9/transport_test.go @@ -56,8 +56,8 @@ func TestSendRecv(t *testing.T) { // badDecode overruns on decode. type badDecode struct{} -func (*badDecode) Decode(b *buffer) { b.markOverrun() } -func (*badDecode) Encode(b *buffer) {} +func (*badDecode) decode(b *buffer) { b.markOverrun() } +func (*badDecode) encode(b *buffer) {} func (*badDecode) Type() MsgType { return MsgTypeBadDecode } func (*badDecode) String() string { return "badDecode{}" } @@ -81,8 +81,8 @@ func TestRecvOverrun(t *testing.T) { // unregistered is not registered on decode. type unregistered struct{} -func (*unregistered) Decode(b *buffer) {} -func (*unregistered) Encode(b *buffer) {} +func (*unregistered) decode(b *buffer) {} +func (*unregistered) encode(b *buffer) {} func (*unregistered) Type() MsgType { return MsgTypeUnregistered } func (*unregistered) String() string { return "unregistered{}" } diff --git a/pkg/rand/rand_linux.go b/pkg/rand/rand_linux.go index 0bdad5fad..fa6a21026 100644 --- a/pkg/rand/rand_linux.go +++ b/pkg/rand/rand_linux.go @@ -17,6 +17,7 @@ package rand import ( + "bufio" "crypto/rand" "io" @@ -45,8 +46,22 @@ func (r *reader) Read(p []byte) (int, error) { return rand.Read(p) } +// bufferedReader implements a threadsafe buffered io.Reader. +type bufferedReader struct { + mu sync.Mutex + r *bufio.Reader +} + +// Read implements io.Reader.Read. +func (b *bufferedReader) Read(p []byte) (int, error) { + b.mu.Lock() + n, err := b.r.Read(p) + b.mu.Unlock() + return n, err +} + // Reader is the default reader. -var Reader io.Reader = &reader{} +var Reader io.Reader = &bufferedReader{r: bufio.NewReader(&reader{})} // Read reads from the default reader. func Read(b []byte) (int, error) { diff --git a/pkg/safecopy/memcpy_amd64.s b/pkg/safecopy/memcpy_amd64.s index 129691d68..00b46c18f 100644 --- a/pkg/safecopy/memcpy_amd64.s +++ b/pkg/safecopy/memcpy_amd64.s @@ -55,15 +55,9 @@ TEXT ·memcpy(SB), NOSPLIT, $0-36 MOVQ from+8(FP), SI MOVQ n+16(FP), BX - // REP instructions have a high startup cost, so we handle small sizes - // with some straightline code. The REP MOVSQ instruction is really fast - // for large sizes. The cutover is approximately 2K. tail: - // move_129through256 or smaller work whether or not the source and the - // destination memory regions overlap because they load all data into - // registers before writing it back. move_256through2048 on the other - // hand can be used only when the memory regions don't overlap or the copy - // direction is forward. + // BSR+branch table make almost all memmove/memclr benchmarks worse. Not + // worth doing. TESTQ BX, BX JEQ move_0 CMPQ BX, $2 @@ -83,31 +77,45 @@ tail: JBE move_65through128 CMPQ BX, $256 JBE move_129through256 - // TODO: use branch table and BSR to make this just a single dispatch -/* - * forward copy loop - */ - CMPQ BX, $2048 - JLS move_256through2048 - - // Check alignment - MOVL SI, AX - ORL DI, AX - TESTL $7, AX - JEQ fwdBy8 - - // Do 1 byte at a time - MOVQ BX, CX - REP; MOVSB - RET - -fwdBy8: - // Do 8 bytes at a time - MOVQ BX, CX - SHRQ $3, CX - ANDQ $7, BX - REP; MOVSQ +move_257plus: + SUBQ $256, BX + MOVOU (SI), X0 + MOVOU X0, (DI) + MOVOU 16(SI), X1 + MOVOU X1, 16(DI) + MOVOU 32(SI), X2 + MOVOU X2, 32(DI) + MOVOU 48(SI), X3 + MOVOU X3, 48(DI) + MOVOU 64(SI), X4 + MOVOU X4, 64(DI) + MOVOU 80(SI), X5 + MOVOU X5, 80(DI) + MOVOU 96(SI), X6 + MOVOU X6, 96(DI) + MOVOU 112(SI), X7 + MOVOU X7, 112(DI) + MOVOU 128(SI), X8 + MOVOU X8, 128(DI) + MOVOU 144(SI), X9 + MOVOU X9, 144(DI) + MOVOU 160(SI), X10 + MOVOU X10, 160(DI) + MOVOU 176(SI), X11 + MOVOU X11, 176(DI) + MOVOU 192(SI), X12 + MOVOU X12, 192(DI) + MOVOU 208(SI), X13 + MOVOU X13, 208(DI) + MOVOU 224(SI), X14 + MOVOU X14, 224(DI) + MOVOU 240(SI), X15 + MOVOU X15, 240(DI) + CMPQ BX, $256 + LEAQ 256(SI), SI + LEAQ 256(DI), DI + JGE move_257plus JMP tail move_1or2: @@ -209,42 +217,3 @@ move_129through256: MOVOU -16(SI)(BX*1), X15 MOVOU X15, -16(DI)(BX*1) RET -move_256through2048: - SUBQ $256, BX - MOVOU (SI), X0 - MOVOU X0, (DI) - MOVOU 16(SI), X1 - MOVOU X1, 16(DI) - MOVOU 32(SI), X2 - MOVOU X2, 32(DI) - MOVOU 48(SI), X3 - MOVOU X3, 48(DI) - MOVOU 64(SI), X4 - MOVOU X4, 64(DI) - MOVOU 80(SI), X5 - MOVOU X5, 80(DI) - MOVOU 96(SI), X6 - MOVOU X6, 96(DI) - MOVOU 112(SI), X7 - MOVOU X7, 112(DI) - MOVOU 128(SI), X8 - MOVOU X8, 128(DI) - MOVOU 144(SI), X9 - MOVOU X9, 144(DI) - MOVOU 160(SI), X10 - MOVOU X10, 160(DI) - MOVOU 176(SI), X11 - MOVOU X11, 176(DI) - MOVOU 192(SI), X12 - MOVOU X12, 192(DI) - MOVOU 208(SI), X13 - MOVOU X13, 208(DI) - MOVOU 224(SI), X14 - MOVOU X14, 224(DI) - MOVOU 240(SI), X15 - MOVOU X15, 240(DI) - CMPQ BX, $256 - LEAQ 256(SI), SI - LEAQ 256(DI), DI - JGE move_256through2048 - JMP tail diff --git a/pkg/safemem/seq_test.go b/pkg/safemem/seq_test.go index eba4bb535..de34005e9 100644 --- a/pkg/safemem/seq_test.go +++ b/pkg/safemem/seq_test.go @@ -20,6 +20,27 @@ import ( "testing" ) +func TestBlockSeqOfEmptyBlock(t *testing.T) { + bs := BlockSeqOf(Block{}) + if !bs.IsEmpty() { + t.Errorf("BlockSeqOf(Block{}).IsEmpty(): got false, wanted true; BlockSeq is %v", bs) + } +} + +func TestBlockSeqOfNonemptyBlock(t *testing.T) { + b := BlockFromSafeSlice(make([]byte, 1)) + bs := BlockSeqOf(b) + if bs.IsEmpty() { + t.Fatalf("BlockSeqOf(non-empty Block).IsEmpty(): got true, wanted false; BlockSeq is %v", bs) + } + if head := bs.Head(); head != b { + t.Fatalf("BlockSeqOf(non-empty Block).Head(): got %v, wanted %v", head, b) + } + if tail := bs.Tail(); !tail.IsEmpty() { + t.Fatalf("BlockSeqOf(non-empty Block).Tail().IsEmpty(): got false, wanted true: tail is %v", tail) + } +} + type blockSeqTest struct { desc string diff --git a/pkg/safemem/seq_unsafe.go b/pkg/safemem/seq_unsafe.go index dcdfc9600..f5f0574f8 100644 --- a/pkg/safemem/seq_unsafe.go +++ b/pkg/safemem/seq_unsafe.go @@ -56,6 +56,9 @@ type BlockSeq struct { // BlockSeqOf returns a BlockSeq representing the single Block b. func BlockSeqOf(b Block) BlockSeq { + if b.length == 0 { + return BlockSeq{} + } bs := BlockSeq{ data: b.start, length: -1, diff --git a/pkg/seccomp/seccomp_test.go b/pkg/seccomp/seccomp_test.go index da5a5e4b2..88766f33b 100644 --- a/pkg/seccomp/seccomp_test.go +++ b/pkg/seccomp/seccomp_test.go @@ -451,7 +451,7 @@ func TestRandom(t *testing.T) { } } - fmt.Printf("Testing filters: %v", syscallRules) + t.Logf("Testing filters: %v", syscallRules) instrs, err := BuildProgram([]RuleSet{ RuleSet{ Rules: syscallRules, diff --git a/pkg/segment/test/segment_test.go b/pkg/segment/test/segment_test.go index f19a005f3..97b16c158 100644 --- a/pkg/segment/test/segment_test.go +++ b/pkg/segment/test/segment_test.go @@ -63,7 +63,7 @@ func checkSet(s *Set, expectedSegments int) error { return fmt.Errorf("incorrect order: key %d (segment %d) >= key %d (segment %d)", prev, nrSegments-1, next, nrSegments) } if got, want := seg.Value(), seg.Start()+valueOffset; got != want { - return fmt.Errorf("segment %d has key %d, value %d (expected %d)", nrSegments, seg.Start, got, want) + return fmt.Errorf("segment %d has key %d, value %d (expected %d)", nrSegments, seg.Start(), got, want) } prev = next havePrev = true diff --git a/pkg/sentry/arch/arch.go b/pkg/sentry/arch/arch.go index 1d11cc472..a903d031c 100644 --- a/pkg/sentry/arch/arch.go +++ b/pkg/sentry/arch/arch.go @@ -88,6 +88,9 @@ type Context interface { // SyscallNo returns the syscall number. SyscallNo() uintptr + // SyscallSaveOrig save orignal register value. + SyscallSaveOrig() + // SyscallArgs returns the syscall arguments in an array. SyscallArgs() SyscallArguments diff --git a/pkg/sentry/arch/arch_aarch64.go b/pkg/sentry/arch/arch_aarch64.go index 3b6987665..c29e1b841 100644 --- a/pkg/sentry/arch/arch_aarch64.go +++ b/pkg/sentry/arch/arch_aarch64.go @@ -32,29 +32,38 @@ import ( const ( // SyscallWidth is the width of insturctions. SyscallWidth = 4 + + // fpsimdMagic is the magic number which is used in fpsimd_context. + fpsimdMagic = 0x46508001 + + // fpsimdContextSize is the size of fpsimd_context. + fpsimdContextSize = 0x210 ) +// ARMTrapFlag is the mask for the trap flag. +const ARMTrapFlag = uint64(1) << 21 + // aarch64FPState is aarch64 floating point state. type aarch64FPState []byte -// initAarch64FPState (defined in asm files) sets up initial state. -func initAarch64FPState(data *FloatingPointData) { - // TODO(gvisor.dev/issue/1238): floating-point is not supported. +// initAarch64FPState sets up initial state. +func initAarch64FPState(data aarch64FPState) { + binary.LittleEndian.PutUint32(data, fpsimdMagic) + binary.LittleEndian.PutUint32(data[4:], fpsimdContextSize) } func newAarch64FPStateSlice() []byte { - return alignedBytes(4096, 32)[:4096] + return alignedBytes(4096, 16)[:fpsimdContextSize] } // newAarch64FPState returns an initialized floating point state. // // The returned state is large enough to store all floating point state // supported by host, even if the app won't use much of it due to a restricted -// FeatureSet. Since they may still be able to see state not advertised by -// CPUID we must ensure it does not contain any sentry state. +// FeatureSet. func newAarch64FPState() aarch64FPState { f := aarch64FPState(newAarch64FPStateSlice()) - initAarch64FPState(f.FloatingPointData()) + initAarch64FPState(f) return f } @@ -86,8 +95,14 @@ type State struct { // Our floating point state. aarch64FPState `state:"wait"` + // TLS pointer + TPValue uint64 + // FeatureSet is a pointer to the currently active feature set. FeatureSet *cpuid.FeatureSet + + // OrigR0 stores the value of register R0. + OrigR0 uint64 } // Proto returns a protobuf representation of the system registers in State. @@ -133,10 +148,12 @@ func (s State) Proto() *rpb.Registers { // Fork creates and returns an identical copy of the state. func (s *State) Fork() State { - // TODO(gvisor.dev/issue/1238): floating-point is not supported. return State{ - Regs: s.Regs, - FeatureSet: s.FeatureSet, + Regs: s.Regs, + aarch64FPState: s.aarch64FPState.fork(), + TPValue: s.TPValue, + FeatureSet: s.FeatureSet, + OrigR0: s.OrigR0, } } @@ -246,6 +263,7 @@ func (s *State) PtraceSetFPRegs(src io.Reader) (int, error) { const ( _NT_PRSTATUS = 1 _NT_PRFPREG = 2 + _NT_ARM_TLS = 0x401 ) // PtraceGetRegSet implements Context.PtraceGetRegSet. @@ -285,8 +303,10 @@ func New(arch Arch, fs *cpuid.FeatureSet) Context { case ARM64: return &context64{ State{ - FeatureSet: fs, + aarch64FPState: newAarch64FPState(), + FeatureSet: fs, }, + []aarch64FPState(nil), } } panic(fmt.Sprintf("unknown architecture %v", arch)) diff --git a/pkg/sentry/arch/arch_arm64.go b/pkg/sentry/arch/arch_arm64.go index ac98897b5..db99c5acb 100644 --- a/pkg/sentry/arch/arch_arm64.go +++ b/pkg/sentry/arch/arch_arm64.go @@ -53,6 +53,11 @@ const ( preferredPIELoadAddr usermem.Addr = maxAddr64 / 6 * 5 ) +var ( + // CPUIDInstruction doesn't exist on ARM64. + CPUIDInstruction = []byte{} +) + // These constants are selected as heuristics to help make the Platform's // potentially limited address space conform as closely to Linux as possible. const ( @@ -68,6 +73,7 @@ const ( // context64 represents an ARM64 context. type context64 struct { State + sigFPState []aarch64FPState // fpstate to be restored on sigreturn. } // Arch implements Context.Arch. @@ -75,10 +81,19 @@ func (c *context64) Arch() Arch { return ARM64 } +func (c *context64) copySigFPState() []aarch64FPState { + var sigfps []aarch64FPState + for _, s := range c.sigFPState { + sigfps = append(sigfps, s.fork()) + } + return sigfps +} + // Fork returns an exact copy of this context. func (c *context64) Fork() Context { return &context64{ - State: c.State.Fork(), + State: c.State.Fork(), + sigFPState: c.copySigFPState(), } } @@ -125,16 +140,17 @@ func (c *context64) SetStack(value uintptr) { // TLS returns the current TLS pointer. func (c *context64) TLS() uintptr { - // TODO(gvisor.dev/issue/1238): TLS is not supported. - // MRS_TPIDR_EL0 - return 0 + return uintptr(c.TPValue) } // SetTLS sets the current TLS pointer. Returns false if value is invalid. func (c *context64) SetTLS(value uintptr) bool { - // TODO(gvisor.dev/issue/1238): TLS is not supported. - // MSR_TPIDR_EL0 - return false + if value >= uintptr(maxAddr64) { + return false + } + + c.TPValue = uint64(value) + return true } // SetOldRSeqInterruptedIP implements Context.SetOldRSeqInterruptedIP. diff --git a/pkg/sentry/arch/arch_state_x86.go b/pkg/sentry/arch/arch_state_x86.go index e35c9214a..aa31169e0 100644 --- a/pkg/sentry/arch/arch_state_x86.go +++ b/pkg/sentry/arch/arch_state_x86.go @@ -12,7 +12,7 @@ // See the License for the specific language governing permissions and // limitations under the License. -// +build amd64 i386 +// +build amd64 386 package arch diff --git a/pkg/sentry/arch/arch_x86.go b/pkg/sentry/arch/arch_x86.go index 88b40a9d1..7fc4c0473 100644 --- a/pkg/sentry/arch/arch_x86.go +++ b/pkg/sentry/arch/arch_x86.go @@ -12,7 +12,7 @@ // See the License for the specific language governing permissions and // limitations under the License. -// +build amd64 i386 +// +build amd64 386 package arch diff --git a/pkg/sentry/arch/arch_x86_impl.go b/pkg/sentry/arch/arch_x86_impl.go index 04ac283c6..3edf40764 100644 --- a/pkg/sentry/arch/arch_x86_impl.go +++ b/pkg/sentry/arch/arch_x86_impl.go @@ -12,7 +12,7 @@ // See the License for the specific language governing permissions and // limitations under the License. -// +build amd64 i386 +// +build amd64 386 package arch diff --git a/pkg/sentry/arch/signal_arm64.go b/pkg/sentry/arch/signal_arm64.go index 4f4cc46a8..0c1db4b13 100644 --- a/pkg/sentry/arch/signal_arm64.go +++ b/pkg/sentry/arch/signal_arm64.go @@ -30,14 +30,29 @@ type SignalContext64 struct { Sp uint64 Pc uint64 Pstate uint64 - _pad [8]byte // __attribute__((__aligned__(16))) - Reserved [4096]uint8 + _pad [8]byte // __attribute__((__aligned__(16))) + Fpsimd64 FpsimdContext // size = 528 + Reserved [3568]uint8 +} + +type aarch64Ctx struct { + Magic uint32 + Size uint32 +} + +// FpsimdContext is equivalent to struct fpsimd_context on arm64 +// (arch/arm64/include/uapi/asm/sigcontext.h). +type FpsimdContext struct { + Head aarch64Ctx + Fpsr uint32 + Fpcr uint32 + Vregs [64]uint64 // actually [32]uint128 } // UContext64 is equivalent to ucontext on arm64(arch/arm64/include/uapi/asm/ucontext.h). type UContext64 struct { Flags uint64 - Link *UContext64 + Link uint64 Stack SignalStack Sigset linux.SignalSet // glibc uses a 1024-bit sigset_t diff --git a/pkg/sentry/arch/signal_stack.go b/pkg/sentry/arch/signal_stack.go index 1a6056171..e58f055c7 100644 --- a/pkg/sentry/arch/signal_stack.go +++ b/pkg/sentry/arch/signal_stack.go @@ -12,7 +12,7 @@ // See the License for the specific language governing permissions and // limitations under the License. -// +build i386 amd64 arm64 +// +build 386 amd64 arm64 package arch diff --git a/pkg/sentry/arch/stack.go b/pkg/sentry/arch/stack.go index 09bceabc9..1108fa0bd 100644 --- a/pkg/sentry/arch/stack.go +++ b/pkg/sentry/arch/stack.go @@ -97,7 +97,6 @@ func (s *Stack) Push(vals ...interface{}) (usermem.Addr, error) { if c < 0 { return 0, fmt.Errorf("bad binary.Size for %T", v) } - // TODO(b/38173783): Use a real context.Context. n, err := usermem.CopyObjectOut(context.Background(), s.IO, s.Bottom-usermem.Addr(c), norm, usermem.IOOpts{}) if err != nil || c != n { return 0, err @@ -121,11 +120,9 @@ func (s *Stack) Pop(vals ...interface{}) (usermem.Addr, error) { var err error if isVaddr { value := s.Arch.Native(uintptr(0)) - // TODO(b/38173783): Use a real context.Context. n, err = usermem.CopyObjectIn(context.Background(), s.IO, s.Bottom, value, usermem.IOOpts{}) *vaddr = usermem.Addr(s.Arch.Value(value)) } else { - // TODO(b/38173783): Use a real context.Context. n, err = usermem.CopyObjectIn(context.Background(), s.IO, s.Bottom, v, usermem.IOOpts{}) } if err != nil { diff --git a/pkg/sentry/arch/syscalls_amd64.go b/pkg/sentry/arch/syscalls_amd64.go index 8b4f23007..3859f41ee 100644 --- a/pkg/sentry/arch/syscalls_amd64.go +++ b/pkg/sentry/arch/syscalls_amd64.go @@ -18,6 +18,13 @@ package arch const restartSyscallNr = uintptr(219) +// SyscallSaveOrig save the value of the register which is clobbered in +// syscall handler(doSyscall()). +// +// Noop on x86. +func (c *context64) SyscallSaveOrig() { +} + // SyscallNo returns the syscall number according to the 64-bit convention. func (c *context64) SyscallNo() uintptr { return uintptr(c.Regs.Orig_rax) diff --git a/pkg/sentry/arch/syscalls_arm64.go b/pkg/sentry/arch/syscalls_arm64.go index 00d5ef461..92d062513 100644 --- a/pkg/sentry/arch/syscalls_arm64.go +++ b/pkg/sentry/arch/syscalls_arm64.go @@ -18,6 +18,17 @@ package arch const restartSyscallNr = uintptr(128) +// SyscallSaveOrig save the value of the register R0 which is clobbered in +// syscall handler(doSyscall()). +// +// In linux, at the entry of the syscall handler(el0_svc_common()), value of R0 +// is saved to the pt_regs.orig_x0 in kernel code. But currently, the orig_x0 +// was not accessible to the user space application, so we have to do the same +// operation in the sentry code to save the R0 value into the App context. +func (c *context64) SyscallSaveOrig() { + c.OrigR0 = c.Regs.Regs[0] +} + // SyscallNo returns the syscall number according to the 64-bit convention. func (c *context64) SyscallNo() uintptr { return uintptr(c.Regs.Regs[8]) @@ -40,7 +51,7 @@ func (c *context64) SyscallNo() uintptr { // R30: the link register. func (c *context64) SyscallArgs() SyscallArguments { return SyscallArguments{ - SyscallArgument{Value: uintptr(c.Regs.Regs[0])}, + SyscallArgument{Value: uintptr(c.OrigR0)}, SyscallArgument{Value: uintptr(c.Regs.Regs[1])}, SyscallArgument{Value: uintptr(c.Regs.Regs[2])}, SyscallArgument{Value: uintptr(c.Regs.Regs[3])}, @@ -50,13 +61,21 @@ func (c *context64) SyscallArgs() SyscallArguments { } // RestartSyscall implements Context.RestartSyscall. +// Prepare for system call restart, OrigR0 will be restored to R0. +// Please see the linux code as reference: +// arch/arm64/kernel/signal.c:do_signal() func (c *context64) RestartSyscall() { c.Regs.Pc -= SyscallWidth - c.Regs.Regs[8] = uint64(restartSyscallNr) + // R0 will be backed up into OrigR0 when entering doSyscall(). + // Please see the linux code as reference: + // arch/arm64/kernel/syscall.c:el0_svc_common(). + // Here we restore it back. + c.Regs.Regs[0] = uint64(c.OrigR0) } // RestartSyscallWithRestartBlock implements Context.RestartSyscallWithRestartBlock. func (c *context64) RestartSyscallWithRestartBlock() { c.Regs.Pc -= SyscallWidth + c.Regs.Regs[0] = uint64(c.OrigR0) c.Regs.Regs[8] = uint64(restartSyscallNr) } diff --git a/pkg/sentry/contexttest/contexttest.go b/pkg/sentry/contexttest/contexttest.go index 031fc64ec..8e5658c7a 100644 --- a/pkg/sentry/contexttest/contexttest.go +++ b/pkg/sentry/contexttest/contexttest.go @@ -97,7 +97,7 @@ type hostClock struct { } // Now implements ktime.Clock.Now. -func (hostClock) Now() ktime.Time { +func (*hostClock) Now() ktime.Time { return ktime.FromNanoseconds(time.Now().UnixNano()) } @@ -127,7 +127,7 @@ func (t *TestContext) Value(key interface{}) interface{} { case uniqueid.CtxInotifyCookie: return atomic.AddUint32(&lastInotifyCookie, 1) case ktime.CtxRealtimeClock: - return hostClock{} + return &hostClock{} default: if val, ok := t.otherValues[key]; ok { return val diff --git a/pkg/sentry/control/BUILD b/pkg/sentry/control/BUILD index e69496477..d16d78aa5 100644 --- a/pkg/sentry/control/BUILD +++ b/pkg/sentry/control/BUILD @@ -16,10 +16,13 @@ go_library( ], deps = [ "//pkg/abi/linux", + "//pkg/context", "//pkg/fd", + "//pkg/fspath", "//pkg/log", "//pkg/sentry/fs", "//pkg/sentry/fs/host", + "//pkg/sentry/fsbridge", "//pkg/sentry/kernel", "//pkg/sentry/kernel/auth", "//pkg/sentry/kernel/time", @@ -27,8 +30,10 @@ go_library( "//pkg/sentry/state", "//pkg/sentry/strace", "//pkg/sentry/usage", + "//pkg/sentry/vfs", "//pkg/sentry/watchdog", "//pkg/sync", + "//pkg/syserror", "//pkg/tcpip/link/sniffer", "//pkg/urpc", ], diff --git a/pkg/sentry/control/pprof.go b/pkg/sentry/control/pprof.go index 151808911..663e51989 100644 --- a/pkg/sentry/control/pprof.go +++ b/pkg/sentry/control/pprof.go @@ -117,9 +117,9 @@ func (p *Profile) HeapProfile(o *ProfileOpts, _ *struct{}) error { return nil } -// Goroutine is an RPC stub which dumps out the stack trace for all running -// goroutines. -func (p *Profile) Goroutine(o *ProfileOpts, _ *struct{}) error { +// GoroutineProfile is an RPC stub which dumps out the stack trace for all +// running goroutines. +func (p *Profile) GoroutineProfile(o *ProfileOpts, _ *struct{}) error { if len(o.FilePayload.Files) < 1 { return errNoOutput } @@ -131,6 +131,34 @@ func (p *Profile) Goroutine(o *ProfileOpts, _ *struct{}) error { return nil } +// BlockProfile is an RPC stub which dumps out the stack trace that led to +// blocking on synchronization primitives. +func (p *Profile) BlockProfile(o *ProfileOpts, _ *struct{}) error { + if len(o.FilePayload.Files) < 1 { + return errNoOutput + } + output := o.FilePayload.Files[0] + defer output.Close() + if err := pprof.Lookup("block").WriteTo(output, 0); err != nil { + return err + } + return nil +} + +// MutexProfile is an RPC stub which dumps out the stack trace of holders of +// contended mutexes. +func (p *Profile) MutexProfile(o *ProfileOpts, _ *struct{}) error { + if len(o.FilePayload.Files) < 1 { + return errNoOutput + } + output := o.FilePayload.Files[0] + defer output.Close() + if err := pprof.Lookup("mutex").WriteTo(output, 0); err != nil { + return err + } + return nil +} + // StartTrace is an RPC stub which starts collection of an execution trace. func (p *Profile) StartTrace(o *ProfileOpts, _ *struct{}) error { if len(o.FilePayload.Files) < 1 { diff --git a/pkg/sentry/control/proc.go b/pkg/sentry/control/proc.go index ced51c66c..b51fb3959 100644 --- a/pkg/sentry/control/proc.go +++ b/pkg/sentry/control/proc.go @@ -18,19 +18,26 @@ import ( "bytes" "encoding/json" "fmt" + "path" "sort" "strings" "text/tabwriter" "time" "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/log" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fs/host" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" "gvisor.dev/gvisor/pkg/sentry/limits" "gvisor.dev/gvisor/pkg/sentry/usage" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/urpc" ) @@ -60,6 +67,12 @@ type ExecArgs struct { // process's MountNamespace. MountNamespace *fs.MountNamespace + // MountNamespaceVFS2 is the mount namespace to execute the new process in. + // A reference on MountNamespace must be held for the lifetime of the + // ExecArgs. If MountNamespace is nil, it will default to the init + // process's MountNamespace. + MountNamespaceVFS2 *vfs.MountNamespace + // WorkingDirectory defines the working directory for the new process. WorkingDirectory string `json:"wd"` @@ -150,6 +163,7 @@ func (proc *Proc) execAsync(args *ExecArgs) (*kernel.ThreadGroup, kernel.ThreadI Envv: args.Envv, WorkingDirectory: args.WorkingDirectory, MountNamespace: args.MountNamespace, + MountNamespaceVFS2: args.MountNamespaceVFS2, Credentials: creds, FDTable: fdTable, Umask: 0022, @@ -166,24 +180,51 @@ func (proc *Proc) execAsync(args *ExecArgs) (*kernel.ThreadGroup, kernel.ThreadI // be donated to the new process in CreateProcess. initArgs.MountNamespace.IncRef() } + if initArgs.MountNamespaceVFS2 != nil { + // initArgs must hold a reference on MountNamespaceVFS2, which will + // be donated to the new process in CreateProcess. + initArgs.MountNamespaceVFS2.IncRef() + } ctx := initArgs.NewContext(proc.Kernel) if initArgs.Filename == "" { - // Get the full path to the filename from the PATH env variable. - paths := fs.GetPath(initArgs.Envv) - mns := initArgs.MountNamespace - if mns == nil { - mns = proc.Kernel.GlobalInit().Leader().MountNamespace() - } - f, err := mns.ResolveExecutablePath(ctx, initArgs.WorkingDirectory, initArgs.Argv[0], paths) - if err != nil { - return nil, 0, nil, fmt.Errorf("error finding executable %q in PATH %v: %v", initArgs.Argv[0], paths, err) + if kernel.VFS2Enabled { + // Get the full path to the filename from the PATH env variable. + if initArgs.MountNamespaceVFS2 == nil { + // Set initArgs so that 'ctx' returns the namespace. + // + // MountNamespaceVFS2 adds a reference to the namespace, which is + // transferred to the new process. + initArgs.MountNamespaceVFS2 = proc.Kernel.GlobalInit().Leader().MountNamespaceVFS2() + } + + paths := fs.GetPath(initArgs.Envv) + vfsObj := proc.Kernel.VFS() + file, err := ResolveExecutablePath(ctx, vfsObj, initArgs.WorkingDirectory, initArgs.Argv[0], paths) + if err != nil { + return nil, 0, nil, fmt.Errorf("error finding executable %q in PATH %v: %v", initArgs.Argv[0], paths, err) + } + initArgs.File = fsbridge.NewVFSFile(file) + } else { + // Get the full path to the filename from the PATH env variable. + paths := fs.GetPath(initArgs.Envv) + if initArgs.MountNamespace == nil { + // Set initArgs so that 'ctx' returns the namespace. + initArgs.MountNamespace = proc.Kernel.GlobalInit().Leader().MountNamespace() + + // initArgs must hold a reference on MountNamespace, which will + // be donated to the new process in CreateProcess. + initArgs.MountNamespaceVFS2.IncRef() + } + f, err := initArgs.MountNamespace.ResolveExecutablePath(ctx, initArgs.WorkingDirectory, initArgs.Argv[0], paths) + if err != nil { + return nil, 0, nil, fmt.Errorf("error finding executable %q in PATH %v: %v", initArgs.Argv[0], paths, err) + } + initArgs.Filename = f } - initArgs.Filename = f } - mounter := fs.FileOwnerFromContext(ctx) - + // TODO(gvisor.dev/issue/1623): Use host FD when supported in VFS2. var ttyFile *fs.File for appFD, hostFile := range args.FilePayload.Files { var appFile *fs.File @@ -192,7 +233,7 @@ func (proc *Proc) execAsync(args *ExecArgs) (*kernel.ThreadGroup, kernel.ThreadI // Import the file as a host TTY file. if ttyFile == nil { var err error - appFile, err = host.ImportFile(ctx, int(hostFile.Fd()), mounter, true /* isTTY */) + appFile, err = host.ImportFile(ctx, int(hostFile.Fd()), true /* isTTY */) if err != nil { return nil, 0, nil, err } @@ -211,7 +252,7 @@ func (proc *Proc) execAsync(args *ExecArgs) (*kernel.ThreadGroup, kernel.ThreadI } else { // Import the file as a regular host file. var err error - appFile, err = host.ImportFile(ctx, int(hostFile.Fd()), mounter, false /* isTTY */) + appFile, err = host.ImportFile(ctx, int(hostFile.Fd()), false /* isTTY */) if err != nil { return nil, 0, nil, err } @@ -411,3 +452,67 @@ func ttyName(tty *kernel.TTY) string { } return fmt.Sprintf("pts/%d", tty.Index) } + +// ResolveExecutablePath resolves the given executable name given a set of +// paths that might contain it. +func ResolveExecutablePath(ctx context.Context, vfsObj *vfs.VirtualFilesystem, wd, name string, paths []string) (*vfs.FileDescription, error) { + root := vfs.RootFromContext(ctx) + defer root.DecRef() + creds := auth.CredentialsFromContext(ctx) + + // Absolute paths can be used directly. + if path.IsAbs(name) { + return openExecutable(ctx, vfsObj, creds, root, name) + } + + // Paths with '/' in them should be joined to the working directory, or + // to the root if working directory is not set. + if strings.IndexByte(name, '/') > 0 { + if len(wd) == 0 { + wd = "/" + } + if !path.IsAbs(wd) { + return nil, fmt.Errorf("working directory %q must be absolute", wd) + } + return openExecutable(ctx, vfsObj, creds, root, path.Join(wd, name)) + } + + // Otherwise, we must lookup the name in the paths, starting from the + // calling context's root directory. + for _, p := range paths { + if !path.IsAbs(p) { + // Relative paths aren't safe, no one should be using them. + log.Warningf("Skipping relative path %q in $PATH", p) + continue + } + + binPath := path.Join(p, name) + f, err := openExecutable(ctx, vfsObj, creds, root, binPath) + if err != nil { + return nil, err + } + if f == nil { + continue // Not found/no access. + } + return f, nil + } + return nil, syserror.ENOENT +} + +func openExecutable(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, root vfs.VirtualDentry, path string) (*vfs.FileDescription, error) { + pop := vfs.PathOperation{ + Root: root, + Start: root, // binPath is absolute, Start can be anything. + Path: fspath.Parse(path), + FollowFinalSymlink: true, + } + opts := &vfs.OpenOptions{ + Flags: linux.O_RDONLY, + FileExec: true, + } + f, err := vfsObj.OpenAt(ctx, creds, &pop, opts) + if err == syserror.ENOENT || err == syserror.EACCES { + return nil, nil + } + return f, err +} diff --git a/pkg/sentry/fs/copy_up.go b/pkg/sentry/fs/copy_up.go index b060a12ff..ab1424c95 100644 --- a/pkg/sentry/fs/copy_up.go +++ b/pkg/sentry/fs/copy_up.go @@ -222,8 +222,8 @@ func copyUpLocked(ctx context.Context, parent *Dirent, next *Dirent) error { } childUpper, err := parentUpper.Lookup(ctx, next.name) if err != nil { - log.Warningf("copy up failed to lookup directory: %v", err) - cleanupUpper(ctx, parentUpper, next.name) + werr := fmt.Errorf("copy up failed to lookup directory: %v", err) + cleanupUpper(ctx, parentUpper, next.name, werr) return syserror.EIO } defer childUpper.DecRef() @@ -242,8 +242,8 @@ func copyUpLocked(ctx context.Context, parent *Dirent, next *Dirent) error { } childUpper, err := parentUpper.Lookup(ctx, next.name) if err != nil { - log.Warningf("copy up failed to lookup symlink: %v", err) - cleanupUpper(ctx, parentUpper, next.name) + werr := fmt.Errorf("copy up failed to lookup symlink: %v", err) + cleanupUpper(ctx, parentUpper, next.name, werr) return syserror.EIO } defer childUpper.DecRef() @@ -256,23 +256,23 @@ func copyUpLocked(ctx context.Context, parent *Dirent, next *Dirent) error { // Bring file attributes up to date. This does not include size, which will be // brought up to date with copyContentsLocked. if err := copyAttributesLocked(ctx, childUpperInode, next.Inode.overlay.lower); err != nil { - log.Warningf("copy up failed to copy up attributes: %v", err) - cleanupUpper(ctx, parentUpper, next.name) + werr := fmt.Errorf("copy up failed to copy up attributes: %v", err) + cleanupUpper(ctx, parentUpper, next.name, werr) return syserror.EIO } // Copy the entire file. if err := copyContentsLocked(ctx, childUpperInode, next.Inode.overlay.lower, attrs.Size); err != nil { - log.Warningf("copy up failed to copy up contents: %v", err) - cleanupUpper(ctx, parentUpper, next.name) + werr := fmt.Errorf("copy up failed to copy up contents: %v", err) + cleanupUpper(ctx, parentUpper, next.name, werr) return syserror.EIO } lowerMappable := next.Inode.overlay.lower.Mappable() upperMappable := childUpperInode.Mappable() if lowerMappable != nil && upperMappable == nil { - log.Warningf("copy up failed: cannot ensure memory mapping coherence") - cleanupUpper(ctx, parentUpper, next.name) + werr := fmt.Errorf("copy up failed: cannot ensure memory mapping coherence") + cleanupUpper(ctx, parentUpper, next.name, werr) return syserror.EIO } @@ -324,12 +324,14 @@ func copyUpLocked(ctx context.Context, parent *Dirent, next *Dirent) error { return nil } -// cleanupUpper removes name from parent, and panics if it is unsuccessful. -func cleanupUpper(ctx context.Context, parent *Inode, name string) { +// cleanupUpper is called when copy-up fails. It logs the copy-up error and +// attempts to remove name from parent. If that fails, then it panics. +func cleanupUpper(ctx context.Context, parent *Inode, name string, copyUpErr error) { + log.Warningf(copyUpErr.Error()) if err := parent.InodeOperations.Remove(ctx, parent, name); err != nil { // Unfortunately we don't have much choice. We shouldn't // willingly give the caller access to a nonsense filesystem. - panic(fmt.Sprintf("overlay filesystem is in an inconsistent state: failed to remove %q from upper filesystem: %v", name, err)) + panic(fmt.Sprintf("overlay filesystem is in an inconsistent state: copyUp got error: %v; then cleanup failed to remove %q from upper filesystem: %v.", copyUpErr, name, err)) } } diff --git a/pkg/sentry/fs/dev/BUILD b/pkg/sentry/fs/dev/BUILD index 4c4b7d5cc..9379a4d7b 100644 --- a/pkg/sentry/fs/dev/BUILD +++ b/pkg/sentry/fs/dev/BUILD @@ -9,6 +9,7 @@ go_library( "device.go", "fs.go", "full.go", + "net_tun.go", "null.go", "random.go", "tty.go", @@ -19,15 +20,20 @@ go_library( "//pkg/context", "//pkg/rand", "//pkg/safemem", + "//pkg/sentry/arch", "//pkg/sentry/device", "//pkg/sentry/fs", "//pkg/sentry/fs/fsutil", "//pkg/sentry/fs/ramfs", "//pkg/sentry/fs/tmpfs", + "//pkg/sentry/inet", + "//pkg/sentry/kernel", "//pkg/sentry/memmap", "//pkg/sentry/mm", "//pkg/sentry/pgalloc", + "//pkg/sentry/socket/netstack", "//pkg/syserror", + "//pkg/tcpip/link/tun", "//pkg/usermem", "//pkg/waiter", ], diff --git a/pkg/sentry/fs/dev/dev.go b/pkg/sentry/fs/dev/dev.go index 35bd23991..acbd401a0 100644 --- a/pkg/sentry/fs/dev/dev.go +++ b/pkg/sentry/fs/dev/dev.go @@ -22,6 +22,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fs/ramfs" "gvisor.dev/gvisor/pkg/sentry/fs/tmpfs" + "gvisor.dev/gvisor/pkg/sentry/inet" "gvisor.dev/gvisor/pkg/usermem" ) @@ -66,8 +67,8 @@ func newMemDevice(ctx context.Context, iops fs.InodeOperations, msrc *fs.MountSo }) } -func newDirectory(ctx context.Context, msrc *fs.MountSource) *fs.Inode { - iops := ramfs.NewDir(ctx, nil, fs.RootOwner, fs.FilePermsFromMode(0555)) +func newDirectory(ctx context.Context, contents map[string]*fs.Inode, msrc *fs.MountSource) *fs.Inode { + iops := ramfs.NewDir(ctx, contents, fs.RootOwner, fs.FilePermsFromMode(0555)) return fs.NewInode(ctx, iops, msrc, fs.StableAttr{ DeviceID: devDevice.DeviceID(), InodeID: devDevice.NextIno(), @@ -111,7 +112,7 @@ func New(ctx context.Context, msrc *fs.MountSource) *fs.Inode { // A devpts is typically mounted at /dev/pts to provide // pseudoterminal support. Place an empty directory there for // the devpts to be mounted over. - "pts": newDirectory(ctx, msrc), + "pts": newDirectory(ctx, nil, msrc), // Similarly, applications expect a ptmx device at /dev/ptmx // connected to the terminals provided by /dev/pts/. Rather // than creating a device directly (which requires a hairy @@ -126,6 +127,12 @@ func New(ctx context.Context, msrc *fs.MountSource) *fs.Inode { "tty": newCharacterDevice(ctx, newTTYDevice(ctx, fs.RootOwner, 0666), msrc, ttyDevMajor, ttyDevMinor), } + if isNetTunSupported(inet.StackFromContext(ctx)) { + contents["net"] = newDirectory(ctx, map[string]*fs.Inode{ + "tun": newCharacterDevice(ctx, newNetTunDevice(ctx, fs.RootOwner, 0666), msrc, netTunDevMajor, netTunDevMinor), + }, msrc) + } + iops := ramfs.NewDir(ctx, contents, fs.RootOwner, fs.FilePermsFromMode(0555)) return fs.NewInode(ctx, iops, msrc, fs.StableAttr{ DeviceID: devDevice.DeviceID(), diff --git a/pkg/sentry/fs/dev/net_tun.go b/pkg/sentry/fs/dev/net_tun.go new file mode 100644 index 000000000..dc7ad075a --- /dev/null +++ b/pkg/sentry/fs/dev/net_tun.go @@ -0,0 +1,177 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package dev + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/fs/fsutil" + "gvisor.dev/gvisor/pkg/sentry/inet" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/socket/netstack" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/tcpip/link/tun" + "gvisor.dev/gvisor/pkg/usermem" + "gvisor.dev/gvisor/pkg/waiter" +) + +const ( + netTunDevMajor = 10 + netTunDevMinor = 200 +) + +// +stateify savable +type netTunInodeOperations struct { + fsutil.InodeGenericChecker `state:"nosave"` + fsutil.InodeNoExtendedAttributes `state:"nosave"` + fsutil.InodeNoopAllocate `state:"nosave"` + fsutil.InodeNoopRelease `state:"nosave"` + fsutil.InodeNoopTruncate `state:"nosave"` + fsutil.InodeNoopWriteOut `state:"nosave"` + fsutil.InodeNotDirectory `state:"nosave"` + fsutil.InodeNotMappable `state:"nosave"` + fsutil.InodeNotSocket `state:"nosave"` + fsutil.InodeNotSymlink `state:"nosave"` + fsutil.InodeVirtual `state:"nosave"` + + fsutil.InodeSimpleAttributes +} + +var _ fs.InodeOperations = (*netTunInodeOperations)(nil) + +func newNetTunDevice(ctx context.Context, owner fs.FileOwner, mode linux.FileMode) *netTunInodeOperations { + return &netTunInodeOperations{ + InodeSimpleAttributes: fsutil.NewInodeSimpleAttributes(ctx, owner, fs.FilePermsFromMode(mode), linux.TMPFS_MAGIC), + } +} + +// GetFile implements fs.InodeOperations.GetFile. +func (iops *netTunInodeOperations) GetFile(ctx context.Context, d *fs.Dirent, flags fs.FileFlags) (*fs.File, error) { + return fs.NewFile(ctx, d, flags, &netTunFileOperations{}), nil +} + +// +stateify savable +type netTunFileOperations struct { + fsutil.FileNoSeek `state:"nosave"` + fsutil.FileNoMMap `state:"nosave"` + fsutil.FileNoSplice `state:"nosave"` + fsutil.FileNoopFlush `state:"nosave"` + fsutil.FileNoopFsync `state:"nosave"` + fsutil.FileNotDirReaddir `state:"nosave"` + fsutil.FileUseInodeUnstableAttr `state:"nosave"` + + device tun.Device +} + +var _ fs.FileOperations = (*netTunFileOperations)(nil) + +// Release implements fs.FileOperations.Release. +func (fops *netTunFileOperations) Release() { + fops.device.Release() +} + +// Ioctl implements fs.FileOperations.Ioctl. +func (fops *netTunFileOperations) Ioctl(ctx context.Context, file *fs.File, io usermem.IO, args arch.SyscallArguments) (uintptr, error) { + request := args[1].Uint() + data := args[2].Pointer() + + switch request { + case linux.TUNSETIFF: + t := kernel.TaskFromContext(ctx) + if t == nil { + panic("Ioctl should be called from a task context") + } + if !t.HasCapability(linux.CAP_NET_ADMIN) { + return 0, syserror.EPERM + } + stack, ok := t.NetworkContext().(*netstack.Stack) + if !ok { + return 0, syserror.EINVAL + } + + var req linux.IFReq + if _, err := usermem.CopyObjectIn(ctx, io, data, &req, usermem.IOOpts{ + AddressSpaceActive: true, + }); err != nil { + return 0, err + } + flags := usermem.ByteOrder.Uint16(req.Data[:]) + return 0, fops.device.SetIff(stack.Stack, req.Name(), flags) + + case linux.TUNGETIFF: + var req linux.IFReq + + copy(req.IFName[:], fops.device.Name()) + + // Linux adds IFF_NOFILTER (the same value as IFF_NO_PI unfortunately) when + // there is no sk_filter. See __tun_chr_ioctl() in net/drivers/tun.c. + flags := fops.device.Flags() | linux.IFF_NOFILTER + usermem.ByteOrder.PutUint16(req.Data[:], flags) + + _, err := usermem.CopyObjectOut(ctx, io, data, &req, usermem.IOOpts{ + AddressSpaceActive: true, + }) + return 0, err + + default: + return 0, syserror.ENOTTY + } +} + +// Write implements fs.FileOperations.Write. +func (fops *netTunFileOperations) Write(ctx context.Context, file *fs.File, src usermem.IOSequence, offset int64) (int64, error) { + data := make([]byte, src.NumBytes()) + if _, err := src.CopyIn(ctx, data); err != nil { + return 0, err + } + return fops.device.Write(data) +} + +// Read implements fs.FileOperations.Read. +func (fops *netTunFileOperations) Read(ctx context.Context, file *fs.File, dst usermem.IOSequence, offset int64) (int64, error) { + data, err := fops.device.Read() + if err != nil { + return 0, err + } + n, err := dst.CopyOut(ctx, data) + if n > 0 && n < len(data) { + // Not an error for partial copying. Packet truncated. + err = nil + } + return int64(n), err +} + +// Readiness implements watier.Waitable.Readiness. +func (fops *netTunFileOperations) Readiness(mask waiter.EventMask) waiter.EventMask { + return fops.device.Readiness(mask) +} + +// EventRegister implements watier.Waitable.EventRegister. +func (fops *netTunFileOperations) EventRegister(e *waiter.Entry, mask waiter.EventMask) { + fops.device.EventRegister(e, mask) +} + +// EventUnregister implements watier.Waitable.EventUnregister. +func (fops *netTunFileOperations) EventUnregister(e *waiter.Entry) { + fops.device.EventUnregister(e) +} + +// isNetTunSupported returns whether /dev/net/tun device is supported for s. +func isNetTunSupported(s inet.Stack) bool { + _, ok := s.(*netstack.Stack) + return ok +} diff --git a/pkg/sentry/fs/dirent.go b/pkg/sentry/fs/dirent.go index acab0411a..65be12175 100644 --- a/pkg/sentry/fs/dirent.go +++ b/pkg/sentry/fs/dirent.go @@ -17,7 +17,6 @@ package fs import ( "fmt" "path" - "sort" "sync/atomic" "syscall" @@ -121,9 +120,6 @@ type Dirent struct { // deleted may be set atomically when removed. deleted int32 - // frozen indicates this entry can't walk to unknown nodes. - frozen bool - // mounted is true if Dirent is a mount point, similar to include/linux/dcache.h:DCACHE_MOUNTED. mounted bool @@ -253,8 +249,7 @@ func (d *Dirent) IsNegative() bool { return d.Inode == nil } -// hashChild will hash child into the children list of its new parent d, carrying over -// any "frozen" state from d. +// hashChild will hash child into the children list of its new parent d. // // Returns (*WeakRef, true) if hashing child caused a Dirent to be unhashed. The caller must // validate the returned unhashed weak reference. Common cases: @@ -282,9 +277,6 @@ func (d *Dirent) hashChild(child *Dirent) (*refs.WeakRef, bool) { d.IncRef() } - // Carry over parent's frozen state. - child.frozen = d.frozen - return d.hashChildParentSet(child) } @@ -320,9 +312,9 @@ func (d *Dirent) SyncAll(ctx context.Context) { // There is nothing to sync for a read-only filesystem. if !d.Inode.MountSource.Flags.ReadOnly { - // FIXME(b/34856369): This should be a mount traversal, not a - // Dirent traversal, because some Inodes that need to be synced - // may no longer be reachable by name (after sys_unlink). + // NOTE(b/34856369): This should be a mount traversal, not a Dirent + // traversal, because some Inodes that need to be synced may no longer + // be reachable by name (after sys_unlink). // // Write out metadata, dirty page cached pages, and sync disk/remote // caches. @@ -400,38 +392,6 @@ func (d *Dirent) MountRoot() *Dirent { return mountRoot } -// Freeze prevents this dirent from walking to more nodes. Freeze is applied -// recursively to all children. -// -// If this particular Dirent represents a Virtual node, then Walks and Creates -// may proceed as before. -// -// Freeze can only be called before the application starts running, otherwise -// the root it might be out of sync with the application root if modified by -// sys_chroot. -func (d *Dirent) Freeze() { - d.mu.Lock() - defer d.mu.Unlock() - if d.frozen { - // Already frozen. - return - } - d.frozen = true - - // Take a reference when freezing. - for _, w := range d.children { - if child := w.Get(); child != nil { - // NOTE: We would normally drop the reference here. But - // instead we're hanging on to it. - ch := child.(*Dirent) - ch.Freeze() - } - } - - // Drop all expired weak references. - d.flush() -} - // descendantOf returns true if the receiver dirent is equal to, or a // descendant of, the argument dirent. // @@ -524,11 +484,6 @@ func (d *Dirent) walk(ctx context.Context, root *Dirent, name string, walkMayUnl w.Drop() } - // Are we allowed to do the lookup? - if d.frozen && !d.Inode.IsVirtual() { - return nil, syscall.ENOENT - } - // Slow path: load the InodeOperations into memory. Since this is a hot path and the lookup may be // expensive, if possible release the lock and re-acquire it. if walkMayUnlock { @@ -659,11 +614,6 @@ func (d *Dirent) Create(ctx context.Context, root *Dirent, name string, flags Fi return nil, syscall.EEXIST } - // Are we frozen? - if d.frozen && !d.Inode.IsVirtual() { - return nil, syscall.ENOENT - } - // Try the create. We need to trust the file system to return EEXIST (or something // that will translate to EEXIST) if name already exists. file, err := d.Inode.Create(ctx, d, name, flags, perms) @@ -727,11 +677,6 @@ func (d *Dirent) genericCreate(ctx context.Context, root *Dirent, name string, c return syscall.EEXIST } - // Are we frozen? - if d.frozen && !d.Inode.IsVirtual() { - return syscall.ENOENT - } - // Remove any negative Dirent. We've already asserted above with d.exists // that the only thing remaining here can be a negative Dirent. if w, ok := d.children[name]; ok { @@ -862,49 +807,6 @@ func (d *Dirent) GetDotAttrs(root *Dirent) (DentAttr, DentAttr) { return dot, dot } -// readdirFrozen returns readdir results based solely on the frozen children. -func (d *Dirent) readdirFrozen(root *Dirent, offset int64, dirCtx *DirCtx) (int64, error) { - // Collect attrs for "." and "..". - attrs := make(map[string]DentAttr) - names := []string{".", ".."} - attrs["."], attrs[".."] = d.GetDotAttrs(root) - - // Get info from all children. - d.mu.Lock() - defer d.mu.Unlock() - for name, w := range d.children { - if child := w.Get(); child != nil { - defer child.DecRef() - - // Skip negative children. - if child.(*Dirent).IsNegative() { - continue - } - - sattr := child.(*Dirent).Inode.StableAttr - attrs[name] = DentAttr{ - Type: sattr.Type, - InodeID: sattr.InodeID, - } - names = append(names, name) - } - } - - sort.Strings(names) - - if int(offset) >= len(names) { - return offset, nil - } - names = names[int(offset):] - for _, name := range names { - if err := dirCtx.DirEmit(name, attrs[name]); err != nil { - return offset, err - } - offset++ - } - return offset, nil -} - // DirIterator is an open directory containing directory entries that can be read. type DirIterator interface { // IterateDir emits directory entries by calling dirCtx.EmitDir, beginning @@ -964,10 +866,6 @@ func direntReaddir(ctx context.Context, d *Dirent, it DirIterator, root *Dirent, return offset, nil } - if d.frozen { - return d.readdirFrozen(root, offset, dirCtx) - } - // Collect attrs for "." and "..". dot, dotdot := d.GetDotAttrs(root) @@ -1068,11 +966,6 @@ func (d *Dirent) mount(ctx context.Context, inode *Inode) (newChild *Dirent, err return nil, syserror.EINVAL } - // Are we frozen? - if d.parent.frozen && !d.parent.Inode.IsVirtual() { - return nil, syserror.ENOENT - } - // Dirent that'll replace d. // // Note that NewDirent returns with one reference taken; the reference @@ -1101,11 +994,6 @@ func (d *Dirent) unmount(ctx context.Context, replacement *Dirent) error { return syserror.ENOENT } - // Are we frozen? - if d.parent.frozen && !d.parent.Inode.IsVirtual() { - return syserror.ENOENT - } - // Remount our former child in its place. // // As replacement used to be our child, it must already have the right @@ -1135,11 +1023,6 @@ func (d *Dirent) Remove(ctx context.Context, root *Dirent, name string, dirPath unlock := d.lockDirectory() defer unlock() - // Are we frozen? - if d.frozen && !d.Inode.IsVirtual() { - return syscall.ENOENT - } - // Try to walk to the node. child, err := d.walk(ctx, root, name, false /* may unlock */) if err != nil { @@ -1201,11 +1084,6 @@ func (d *Dirent) RemoveDirectory(ctx context.Context, root *Dirent, name string) unlock := d.lockDirectory() defer unlock() - // Are we frozen? - if d.frozen && !d.Inode.IsVirtual() { - return syscall.ENOENT - } - // Check for dots. if name == "." { // Rejected as the last component by rmdir(2). @@ -1438,8 +1316,8 @@ func lockForRename(oldParent *Dirent, oldName string, newParent *Dirent, newName }, nil } -func checkSticky(ctx context.Context, dir *Dirent, victim *Dirent) error { - uattr, err := dir.Inode.UnstableAttr(ctx) +func (d *Dirent) checkSticky(ctx context.Context, victim *Dirent) error { + uattr, err := d.Inode.UnstableAttr(ctx) if err != nil { return syserror.EPERM } @@ -1465,30 +1343,33 @@ func checkSticky(ctx context.Context, dir *Dirent, victim *Dirent) error { return syserror.EPERM } -// MayDelete determines whether `name`, a child of `dir`, can be deleted or +// MayDelete determines whether `name`, a child of `d`, can be deleted or // renamed by `ctx`. // // Compare Linux kernel fs/namei.c:may_delete. -func MayDelete(ctx context.Context, root, dir *Dirent, name string) error { - if err := dir.Inode.CheckPermission(ctx, PermMask{Write: true, Execute: true}); err != nil { +func (d *Dirent) MayDelete(ctx context.Context, root *Dirent, name string) error { + if err := d.Inode.CheckPermission(ctx, PermMask{Write: true, Execute: true}); err != nil { return err } - victim, err := dir.Walk(ctx, root, name) + unlock := d.lockDirectory() + defer unlock() + + victim, err := d.walk(ctx, root, name, true /* may unlock */) if err != nil { return err } defer victim.DecRef() - return mayDelete(ctx, dir, victim) + return d.mayDelete(ctx, victim) } // mayDelete determines whether `victim`, a child of `dir`, can be deleted or // renamed by `ctx`. // // Preconditions: `dir` is writable and executable by `ctx`. -func mayDelete(ctx context.Context, dir, victim *Dirent) error { - if err := checkSticky(ctx, dir, victim); err != nil { +func (d *Dirent) mayDelete(ctx context.Context, victim *Dirent) error { + if err := d.checkSticky(ctx, victim); err != nil { return err } @@ -1516,15 +1397,6 @@ func Rename(ctx context.Context, root *Dirent, oldParent *Dirent, oldName string return err } - // Are we frozen? - // TODO(jamieliu): Is this the right errno? - if oldParent.frozen && !oldParent.Inode.IsVirtual() { - return syscall.ENOENT - } - if newParent.frozen && !newParent.Inode.IsVirtual() { - return syscall.ENOENT - } - // Do we have general permission to remove from oldParent and // create/replace in newParent? if err := oldParent.Inode.CheckPermission(ctx, PermMask{Write: true, Execute: true}); err != nil { @@ -1542,7 +1414,7 @@ func Rename(ctx context.Context, root *Dirent, oldParent *Dirent, oldName string defer renamed.DecRef() // Check that the renamed dirent is deletable. - if err := mayDelete(ctx, oldParent, renamed); err != nil { + if err := oldParent.mayDelete(ctx, renamed); err != nil { return err } @@ -1580,7 +1452,7 @@ func Rename(ctx context.Context, root *Dirent, oldParent *Dirent, oldName string // across the Rename, so must call DecRef manually (no defer). // Check that we can delete replaced. - if err := mayDelete(ctx, newParent, replaced); err != nil { + if err := newParent.mayDelete(ctx, replaced); err != nil { replaced.DecRef() return err } diff --git a/pkg/sentry/fs/dirent_cache.go b/pkg/sentry/fs/dirent_cache.go index 25514ace4..33de32c69 100644 --- a/pkg/sentry/fs/dirent_cache.go +++ b/pkg/sentry/fs/dirent_cache.go @@ -101,8 +101,6 @@ func (c *DirentCache) remove(d *Dirent) { panic(fmt.Sprintf("trying to remove %v, which is not in the dirent cache", d)) } c.list.Remove(d) - d.SetPrev(nil) - d.SetNext(nil) d.DecRef() c.currentSize-- if c.limit != nil { diff --git a/pkg/sentry/fs/fdpipe/pipe_test.go b/pkg/sentry/fs/fdpipe/pipe_test.go index 5aff0cc95..a0082ecca 100644 --- a/pkg/sentry/fs/fdpipe/pipe_test.go +++ b/pkg/sentry/fs/fdpipe/pipe_test.go @@ -119,7 +119,7 @@ func TestNewPipe(t *testing.T) { continue } if flags := p.flags; test.flags != flags { - t.Errorf("%s: got file flags %s, want %s", test.desc, flags, test.flags) + t.Errorf("%s: got file flags %v, want %v", test.desc, flags, test.flags) continue } if len(test.readAheadBuffer) != len(p.readAheadBuffer) { @@ -136,7 +136,7 @@ func TestNewPipe(t *testing.T) { continue } if !fdnotifier.HasFD(int32(f.FD())) { - t.Errorf("%s: pipe fd %d is not registered for events", test.desc, f.FD) + t.Errorf("%s: pipe fd %d is not registered for events", test.desc, f.FD()) } } } diff --git a/pkg/sentry/fs/file.go b/pkg/sentry/fs/file.go index 78100e448..846252c89 100644 --- a/pkg/sentry/fs/file.go +++ b/pkg/sentry/fs/file.go @@ -44,7 +44,7 @@ var ( RecordWaitTime = false reads = metric.MustCreateNewUint64Metric("/fs/reads", false /* sync */, "Number of file reads.") - readWait = metric.MustCreateNewUint64Metric("/fs/read_wait", false /* sync */, "Time waiting on file reads, in nanoseconds.") + readWait = metric.MustCreateNewUint64NanosecondsMetric("/fs/read_wait", false /* sync */, "Time waiting on file reads, in nanoseconds.") ) // IncrementWait increments the given wait time metric, if enabled. diff --git a/pkg/sentry/fs/file_overlay_test.go b/pkg/sentry/fs/file_overlay_test.go index a76d87e3a..1971cc680 100644 --- a/pkg/sentry/fs/file_overlay_test.go +++ b/pkg/sentry/fs/file_overlay_test.go @@ -175,90 +175,6 @@ func TestReaddirRevalidation(t *testing.T) { } } -// TestReaddirOverlayFrozen tests that calling Readdir on an overlay file with -// a frozen dirent tree does not make Readdir calls to the underlying files. -// This is a regression test for b/114808269. -func TestReaddirOverlayFrozen(t *testing.T) { - ctx := contexttest.Context(t) - - // Create an overlay with two directories, each with two files. - upper := newTestRamfsDir(ctx, []dirContent{{name: "upper-file1"}, {name: "upper-file2"}}, nil) - lower := newTestRamfsDir(ctx, []dirContent{{name: "lower-file1"}, {name: "lower-file2"}}, nil) - overlayInode := fs.NewTestOverlayDir(ctx, upper, lower, false) - - // Set that overlay as the root. - root := fs.NewDirent(ctx, overlayInode, "root") - ctx = &rootContext{ - Context: ctx, - root: root, - } - - // Check that calling Readdir on the root now returns all 4 files (2 - // from each layer in the overlay). - rootFile, err := root.Inode.GetFile(ctx, root, fs.FileFlags{Read: true}) - if err != nil { - t.Fatalf("root.Inode.GetFile failed: %v", err) - } - defer rootFile.DecRef() - ser := &fs.CollectEntriesSerializer{} - if err := rootFile.Readdir(ctx, ser); err != nil { - t.Fatalf("rootFile.Readdir failed: %v", err) - } - if got, want := ser.Order, []string{".", "..", "lower-file1", "lower-file2", "upper-file1", "upper-file2"}; !reflect.DeepEqual(got, want) { - t.Errorf("Readdir got names %v, want %v", got, want) - } - - // Readdir should have been called on upper and lower. - upperDir := upper.InodeOperations.(*dir) - lowerDir := lower.InodeOperations.(*dir) - if !upperDir.ReaddirCalled { - t.Errorf("upperDir.ReaddirCalled got %v, want true", upperDir.ReaddirCalled) - } - if !lowerDir.ReaddirCalled { - t.Errorf("lowerDir.ReaddirCalled got %v, want true", lowerDir.ReaddirCalled) - } - - // Reset. - upperDir.ReaddirCalled = false - lowerDir.ReaddirCalled = false - - // Take references on "upper-file1" and "lower-file1", pinning them in - // the dirent tree. - for _, name := range []string{"upper-file1", "lower-file1"} { - if _, err := root.Walk(ctx, root, name); err != nil { - t.Fatalf("root.Walk(%q) failed: %v", name, err) - } - // Don't drop a reference on the returned dirent so that it - // will stay in the tree. - } - - // Freeze the dirent tree. - root.Freeze() - - // Seek back to the beginning of the file. - if _, err := rootFile.Seek(ctx, fs.SeekSet, 0); err != nil { - t.Fatalf("error seeking to beginning of directory: %v", err) - } - - // Calling Readdir on the root now will return only the pinned - // children. - ser = &fs.CollectEntriesSerializer{} - if err := rootFile.Readdir(ctx, ser); err != nil { - t.Fatalf("rootFile.Readdir failed: %v", err) - } - if got, want := ser.Order, []string{".", "..", "lower-file1", "upper-file1"}; !reflect.DeepEqual(got, want) { - t.Errorf("Readdir got names %v, want %v", got, want) - } - - // Readdir should NOT have been called on upper or lower. - if upperDir.ReaddirCalled { - t.Errorf("upperDir.ReaddirCalled got %v, want false", upperDir.ReaddirCalled) - } - if lowerDir.ReaddirCalled { - t.Errorf("lowerDir.ReaddirCalled got %v, want false", lowerDir.ReaddirCalled) - } -} - type rootContext struct { context.Context root *fs.Dirent diff --git a/pkg/sentry/fs/fsutil/host_file_mapper.go b/pkg/sentry/fs/fsutil/host_file_mapper.go index 67278aa86..e82afd112 100644 --- a/pkg/sentry/fs/fsutil/host_file_mapper.go +++ b/pkg/sentry/fs/fsutil/host_file_mapper.go @@ -65,13 +65,18 @@ type mapping struct { writable bool } -// NewHostFileMapper returns a HostFileMapper with no references or cached -// mappings. +// Init must be called on zero-value HostFileMappers before first use. +func (f *HostFileMapper) Init() { + f.refs = make(map[uint64]int32) + f.mappings = make(map[uint64]mapping) +} + +// NewHostFileMapper returns an initialized HostFileMapper allocated on the +// heap with no references or cached mappings. func NewHostFileMapper() *HostFileMapper { - return &HostFileMapper{ - refs: make(map[uint64]int32), - mappings: make(map[uint64]mapping), - } + f := &HostFileMapper{} + f.Init() + return f } // IncRefOn increments the reference count on all offsets in mr. diff --git a/pkg/sentry/fs/fsutil/inode.go b/pkg/sentry/fs/fsutil/inode.go index daecc4ffe..1922ff08c 100644 --- a/pkg/sentry/fs/fsutil/inode.go +++ b/pkg/sentry/fs/fsutil/inode.go @@ -259,8 +259,8 @@ func (i *InodeSimpleExtendedAttributes) ListXattr(context.Context, *fs.Inode, ui // RemoveXattr implements fs.InodeOperations.RemoveXattr. func (i *InodeSimpleExtendedAttributes) RemoveXattr(_ context.Context, _ *fs.Inode, name string) error { - i.mu.RLock() - defer i.mu.RUnlock() + i.mu.Lock() + defer i.mu.Unlock() if _, ok := i.xattrs[name]; ok { delete(i.xattrs, name) return nil diff --git a/pkg/sentry/fs/gofer/file.go b/pkg/sentry/fs/gofer/file.go index 23296f246..b2fcab127 100644 --- a/pkg/sentry/fs/gofer/file.go +++ b/pkg/sentry/fs/gofer/file.go @@ -37,9 +37,9 @@ var ( opens9P = metric.MustCreateNewUint64Metric("/gofer/opens_9p", false /* sync */, "Number of times a 9P file was opened from a gofer.") opensHost = metric.MustCreateNewUint64Metric("/gofer/opens_host", false /* sync */, "Number of times a host file was opened from a gofer.") reads9P = metric.MustCreateNewUint64Metric("/gofer/reads_9p", false /* sync */, "Number of 9P file reads from a gofer.") - readWait9P = metric.MustCreateNewUint64Metric("/gofer/read_wait_9p", false /* sync */, "Time waiting on 9P file reads from a gofer, in nanoseconds.") + readWait9P = metric.MustCreateNewUint64NanosecondsMetric("/gofer/read_wait_9p", false /* sync */, "Time waiting on 9P file reads from a gofer, in nanoseconds.") readsHost = metric.MustCreateNewUint64Metric("/gofer/reads_host", false /* sync */, "Number of host file reads from a gofer.") - readWaitHost = metric.MustCreateNewUint64Metric("/gofer/read_wait_host", false /* sync */, "Time waiting on host file reads from a gofer, in nanoseconds.") + readWaitHost = metric.MustCreateNewUint64NanosecondsMetric("/gofer/read_wait_host", false /* sync */, "Time waiting on host file reads from a gofer, in nanoseconds.") ) // fileOperations implements fs.FileOperations for a remote file system. diff --git a/pkg/sentry/fs/gofer/file_state.go b/pkg/sentry/fs/gofer/file_state.go index ff96b28ba..edd6576aa 100644 --- a/pkg/sentry/fs/gofer/file_state.go +++ b/pkg/sentry/fs/gofer/file_state.go @@ -34,7 +34,6 @@ func (f *fileOperations) afterLoad() { flags := f.flags flags.Truncate = false - // TODO(b/38173783): Context is not plumbed to save/restore. f.handles, err = f.inodeOperations.fileState.getHandles(context.Background(), flags, f.inodeOperations.cachingInodeOps) if err != nil { return fmt.Errorf("failed to re-open handle: %v", err) diff --git a/pkg/sentry/fs/gofer/handles.go b/pkg/sentry/fs/gofer/handles.go index 9f7c3e89f..fc14249be 100644 --- a/pkg/sentry/fs/gofer/handles.go +++ b/pkg/sentry/fs/gofer/handles.go @@ -57,7 +57,6 @@ func (h *handles) DecRef() { } } } - // FIXME(b/38173783): Context is not plumbed here. if err := h.File.close(context.Background()); err != nil { log.Warningf("error closing p9 file: %v", err) } diff --git a/pkg/sentry/fs/gofer/inode.go b/pkg/sentry/fs/gofer/inode.go index 1c934981b..a016c896e 100644 --- a/pkg/sentry/fs/gofer/inode.go +++ b/pkg/sentry/fs/gofer/inode.go @@ -273,7 +273,7 @@ func (i *inodeFileState) recreateReadHandles(ctx context.Context, writer *handle // operations on the old will see the new data. Then, make the new handle take // ownereship of the old FD and mark the old readHandle to not close the FD // when done. - if err := syscall.Dup3(h.Host.FD(), i.readHandles.Host.FD(), 0); err != nil { + if err := syscall.Dup3(h.Host.FD(), i.readHandles.Host.FD(), syscall.O_CLOEXEC); err != nil { return err } @@ -710,13 +710,10 @@ func init() { } // AddLink implements InodeOperations.AddLink, but is currently a noop. -// FIXME(b/63117438): Remove this from InodeOperations altogether. func (*inodeOperations) AddLink() {} // DropLink implements InodeOperations.DropLink, but is currently a noop. -// FIXME(b/63117438): Remove this from InodeOperations altogether. func (*inodeOperations) DropLink() {} // NotifyStatusChange implements fs.InodeOperations.NotifyStatusChange. -// FIXME(b/63117438): Remove this from InodeOperations altogether. func (i *inodeOperations) NotifyStatusChange(ctx context.Context) {} diff --git a/pkg/sentry/fs/gofer/inode_state.go b/pkg/sentry/fs/gofer/inode_state.go index 238f7804c..a3402e343 100644 --- a/pkg/sentry/fs/gofer/inode_state.go +++ b/pkg/sentry/fs/gofer/inode_state.go @@ -123,7 +123,6 @@ func (i *inodeFileState) afterLoad() { // beforeSave. return fmt.Errorf("failed to find path for inode number %d. Device %s contains %s", i.sattr.InodeID, i.s.connID, fs.InodeMappings(i.s.inodeMappings)) } - // TODO(b/38173783): Context is not plumbed to save/restore. ctx := &dummyClockContext{context.Background()} _, i.file, err = i.s.attach.walk(ctx, splitAbsolutePath(name)) diff --git a/pkg/sentry/fs/gofer/session_state.go b/pkg/sentry/fs/gofer/session_state.go index 111da59f9..2d398b753 100644 --- a/pkg/sentry/fs/gofer/session_state.go +++ b/pkg/sentry/fs/gofer/session_state.go @@ -104,7 +104,6 @@ func (s *session) afterLoad() { // If private unix sockets are enabled, create and fill the session's endpoint // maps. if opts.privateunixsocket { - // TODO(b/38173783): Context is not plumbed to save/restore. ctx := &dummyClockContext{context.Background()} if err = s.restoreEndpointMaps(ctx); err != nil { diff --git a/pkg/sentry/fs/gofer/util.go b/pkg/sentry/fs/gofer/util.go index 2d8d3a2ea..47a6c69bf 100644 --- a/pkg/sentry/fs/gofer/util.go +++ b/pkg/sentry/fs/gofer/util.go @@ -20,17 +20,29 @@ import ( "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/p9" "gvisor.dev/gvisor/pkg/sentry/fs" + ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" ) func utimes(ctx context.Context, file contextFile, ts fs.TimeSpec) error { if ts.ATimeOmit && ts.MTimeOmit { return nil } + + // Replace requests to use the "system time" with the current time to + // ensure that timestamps remain consistent with the remote + // filesystem. + now := ktime.NowFromContext(ctx) + if ts.ATimeSetSystemTime { + ts.ATime = now + } + if ts.MTimeSetSystemTime { + ts.MTime = now + } mask := p9.SetAttrMask{ ATime: !ts.ATimeOmit, - ATimeNotSystemTime: !ts.ATimeSetSystemTime, + ATimeNotSystemTime: true, MTime: !ts.MTimeOmit, - MTimeNotSystemTime: !ts.MTimeSetSystemTime, + MTimeNotSystemTime: true, } as, ans := ts.ATime.Unix() ms, mns := ts.MTime.Unix() diff --git a/pkg/sentry/fs/host/BUILD b/pkg/sentry/fs/host/BUILD index 21003ea45..aabce6cc9 100644 --- a/pkg/sentry/fs/host/BUILD +++ b/pkg/sentry/fs/host/BUILD @@ -10,7 +10,7 @@ go_library( "descriptor_state.go", "device.go", "file.go", - "fs.go", + "host.go", "inode.go", "inode_state.go", "ioctl_unsafe.go", @@ -62,18 +62,15 @@ go_test( size = "small", srcs = [ "descriptor_test.go", - "fs_test.go", "inode_test.go", "socket_test.go", "wait_test.go", ], library = ":host", deps = [ - "//pkg/context", "//pkg/fd", "//pkg/fdnotifier", "//pkg/sentry/contexttest", - "//pkg/sentry/fs", "//pkg/sentry/kernel/time", "//pkg/sentry/socket", "//pkg/sentry/socket/unix/transport", diff --git a/pkg/sentry/fs/host/control.go b/pkg/sentry/fs/host/control.go index 1658979fc..52c0504b6 100644 --- a/pkg/sentry/fs/host/control.go +++ b/pkg/sentry/fs/host/control.go @@ -32,6 +32,8 @@ func newSCMRights(fds []int) control.SCMRights { } // Files implements control.SCMRights.Files. +// +// TODO(gvisor.dev/issue/2017): Port to VFS2. func (c *scmRights) Files(ctx context.Context, max int) (control.RightsFiles, bool) { n := max var trunc bool @@ -76,7 +78,7 @@ func fdsToFiles(ctx context.Context, fds []int) []*fs.File { } // Create the file backed by hostFD. - file, err := NewFile(ctx, fd, fs.FileOwnerFromContext(ctx)) + file, err := NewFile(ctx, fd) if err != nil { ctx.Warningf("Error creating file from host FD: %v", err) break diff --git a/pkg/sentry/fs/host/descriptor.go b/pkg/sentry/fs/host/descriptor.go index 2a4d1b291..cfdce6a74 100644 --- a/pkg/sentry/fs/host/descriptor.go +++ b/pkg/sentry/fs/host/descriptor.go @@ -16,7 +16,6 @@ package host import ( "fmt" - "path" "syscall" "gvisor.dev/gvisor/pkg/fdnotifier" @@ -28,12 +27,9 @@ import ( // // +stateify savable type descriptor struct { - // donated is true if the host fd was donated by another process. - donated bool - // If origFD >= 0, it is the host fd that this file was originally created // from, which must be available at time of restore. The FD can be closed - // after descriptor is created. Only set if donated is true. + // after descriptor is created. origFD int // wouldBlock is true if value (below) points to a file that can @@ -41,15 +37,13 @@ type descriptor struct { wouldBlock bool // value is the wrapped host fd. It is never saved or restored - // directly. How it is restored depends on whether it was - // donated and the fs.MountSource it was originally - // opened/created from. + // directly. value int `state:"nosave"` } // newDescriptor returns a wrapped host file descriptor. On success, // the descriptor is registered for event notifications with queue. -func newDescriptor(fd int, donated bool, saveable bool, wouldBlock bool, queue *waiter.Queue) (*descriptor, error) { +func newDescriptor(fd int, saveable bool, wouldBlock bool, queue *waiter.Queue) (*descriptor, error) { ownedFD := fd origFD := -1 if saveable { @@ -69,7 +63,6 @@ func newDescriptor(fd int, donated bool, saveable bool, wouldBlock bool, queue * } } return &descriptor{ - donated: donated, origFD: origFD, wouldBlock: wouldBlock, value: ownedFD, @@ -77,25 +70,11 @@ func newDescriptor(fd int, donated bool, saveable bool, wouldBlock bool, queue * } // initAfterLoad initializes the value of the descriptor after Load. -func (d *descriptor) initAfterLoad(mo *superOperations, id uint64, queue *waiter.Queue) error { - if d.donated { - var err error - d.value, err = syscall.Dup(d.origFD) - if err != nil { - return fmt.Errorf("failed to dup restored fd %d: %v", d.origFD, err) - } - } else { - name, ok := mo.inodeMappings[id] - if !ok { - return fmt.Errorf("failed to find path for inode number %d", id) - } - fullpath := path.Join(mo.root, name) - - var err error - d.value, err = open(nil, fullpath) - if err != nil { - return fmt.Errorf("failed to open %q: %v", fullpath, err) - } +func (d *descriptor) initAfterLoad(id uint64, queue *waiter.Queue) error { + var err error + d.value, err = syscall.Dup(d.origFD) + if err != nil { + return fmt.Errorf("failed to dup restored fd %d: %v", d.origFD, err) } if d.wouldBlock { if err := syscall.SetNonblock(d.value, true); err != nil { diff --git a/pkg/sentry/fs/host/descriptor_state.go b/pkg/sentry/fs/host/descriptor_state.go index 8167390a9..e880582ab 100644 --- a/pkg/sentry/fs/host/descriptor_state.go +++ b/pkg/sentry/fs/host/descriptor_state.go @@ -16,7 +16,7 @@ package host // beforeSave is invoked by stateify. func (d *descriptor) beforeSave() { - if d.donated && d.origFD < 0 { + if d.origFD < 0 { panic("donated file descriptor cannot be saved") } } diff --git a/pkg/sentry/fs/host/descriptor_test.go b/pkg/sentry/fs/host/descriptor_test.go index 4205981f5..d8e4605b6 100644 --- a/pkg/sentry/fs/host/descriptor_test.go +++ b/pkg/sentry/fs/host/descriptor_test.go @@ -47,10 +47,10 @@ func TestDescriptorRelease(t *testing.T) { // FD ownership is transferred to the descritor. queue := &waiter.Queue{} - d, err := newDescriptor(fd, false /* donated*/, tc.saveable, tc.wouldBlock, queue) + d, err := newDescriptor(fd, tc.saveable, tc.wouldBlock, queue) if err != nil { syscall.Close(fd) - t.Fatalf("newDescriptor(%d, %t, false, %t, queue) failed, err: %v", fd, tc.saveable, tc.wouldBlock, err) + t.Fatalf("newDescriptor(%d, %t, %t, queue) failed, err: %v", fd, tc.saveable, tc.wouldBlock, err) } if tc.saveable { if d.origFD < 0 { diff --git a/pkg/sentry/fs/host/file.go b/pkg/sentry/fs/host/file.go index e08f56d04..3e48b8b2c 100644 --- a/pkg/sentry/fs/host/file.go +++ b/pkg/sentry/fs/host/file.go @@ -60,8 +60,8 @@ var _ fs.FileOperations = (*fileOperations)(nil) // The returned File cannot be saved, since there is no guarantee that the same // FD will exist or represent the same file at time of restore. If such a // guarantee does exist, use ImportFile instead. -func NewFile(ctx context.Context, fd int, mounter fs.FileOwner) (*fs.File, error) { - return newFileFromDonatedFD(ctx, fd, mounter, false, false) +func NewFile(ctx context.Context, fd int) (*fs.File, error) { + return newFileFromDonatedFD(ctx, fd, false, false) } // ImportFile creates a new File backed by the provided host file descriptor. @@ -71,13 +71,13 @@ func NewFile(ctx context.Context, fd int, mounter fs.FileOwner) (*fs.File, error // If the returned file is saved, it will be restored by re-importing the FD // originally passed to ImportFile. It is the restorer's responsibility to // ensure that the FD represents the same file. -func ImportFile(ctx context.Context, fd int, mounter fs.FileOwner, isTTY bool) (*fs.File, error) { - return newFileFromDonatedFD(ctx, fd, mounter, true, isTTY) +func ImportFile(ctx context.Context, fd int, isTTY bool) (*fs.File, error) { + return newFileFromDonatedFD(ctx, fd, true, isTTY) } // newFileFromDonatedFD returns an fs.File from a donated FD. If the FD is // saveable, then saveable is true. -func newFileFromDonatedFD(ctx context.Context, donated int, mounter fs.FileOwner, saveable, isTTY bool) (*fs.File, error) { +func newFileFromDonatedFD(ctx context.Context, donated int, saveable, isTTY bool) (*fs.File, error) { var s syscall.Stat_t if err := syscall.Fstat(donated, &s); err != nil { return nil, err @@ -101,8 +101,8 @@ func newFileFromDonatedFD(ctx context.Context, donated int, mounter fs.FileOwner }) return s, nil default: - msrc := newMountSource(ctx, "/", mounter, &Filesystem{}, fs.MountSourceFlags{}, false /* dontTranslateOwnership */) - inode, err := newInode(ctx, msrc, donated, saveable, true /* donated */) + msrc := fs.NewNonCachingMountSource(ctx, &filesystem{}, fs.MountSourceFlags{}) + inode, err := newInode(ctx, msrc, donated, saveable) if err != nil { return nil, err } diff --git a/pkg/sentry/fs/host/fs.go b/pkg/sentry/fs/host/fs.go deleted file mode 100644 index d3e8e3a36..000000000 --- a/pkg/sentry/fs/host/fs.go +++ /dev/null @@ -1,339 +0,0 @@ -// Copyright 2018 The gVisor Authors. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -// Package host implements an fs.Filesystem for files backed by host -// file descriptors. -package host - -import ( - "fmt" - "path" - "path/filepath" - "strconv" - "strings" - - "gvisor.dev/gvisor/pkg/context" - "gvisor.dev/gvisor/pkg/log" - "gvisor.dev/gvisor/pkg/sentry/fs" -) - -// FilesystemName is the name under which Filesystem is registered. -const FilesystemName = "whitelistfs" - -const ( - // whitelistKey is the mount option containing a comma-separated list - // of host paths to whitelist. - whitelistKey = "whitelist" - - // rootPathKey is the mount option containing the root path of the - // mount. - rootPathKey = "root" - - // dontTranslateOwnershipKey is the key to superOperations.dontTranslateOwnership. - dontTranslateOwnershipKey = "dont_translate_ownership" -) - -// maxTraversals determines link traversals in building the whitelist. -const maxTraversals = 10 - -// Filesystem is a pseudo file system that is only available during the setup -// to lock down the configurations. This filesystem should only be mounted at root. -// -// Think twice before exposing this to applications. -// -// +stateify savable -type Filesystem struct { - // whitelist is a set of host paths to whitelist. - paths []string -} - -var _ fs.Filesystem = (*Filesystem)(nil) - -// Name is the identifier of this file system. -func (*Filesystem) Name() string { - return FilesystemName -} - -// AllowUserMount prohibits users from using mount(2) with this file system. -func (*Filesystem) AllowUserMount() bool { - return false -} - -// AllowUserList allows this filesystem to be listed in /proc/filesystems. -func (*Filesystem) AllowUserList() bool { - return true -} - -// Flags returns that there is nothing special about this file system. -func (*Filesystem) Flags() fs.FilesystemFlags { - return 0 -} - -// Mount returns an fs.Inode exposing the host file system. It is intended to be locked -// down in PreExec below. -func (f *Filesystem) Mount(ctx context.Context, _ string, flags fs.MountSourceFlags, data string, _ interface{}) (*fs.Inode, error) { - // Parse generic comma-separated key=value options. - options := fs.GenericMountSourceOptions(data) - - // Grab the whitelist if one was specified. - // TODO(edahlgren/mpratt/hzy): require another option "testonly" in order to allow - // no whitelist. - if wl, ok := options[whitelistKey]; ok { - f.paths = strings.Split(wl, "|") - delete(options, whitelistKey) - } - - // If the rootPath was set, use it. Othewise default to the root of the - // host fs. - rootPath := "/" - if rp, ok := options[rootPathKey]; ok { - rootPath = rp - delete(options, rootPathKey) - - // We must relativize the whitelisted paths to the new root. - for i, p := range f.paths { - rel, err := filepath.Rel(rootPath, p) - if err != nil { - return nil, fmt.Errorf("whitelist path %q must be a child of root path %q", p, rootPath) - } - f.paths[i] = path.Join("/", rel) - } - } - fd, err := open(nil, rootPath) - if err != nil { - return nil, fmt.Errorf("failed to find root: %v", err) - } - - var dontTranslateOwnership bool - if v, ok := options[dontTranslateOwnershipKey]; ok { - b, err := strconv.ParseBool(v) - if err != nil { - return nil, fmt.Errorf("invalid value for %q: %v", dontTranslateOwnershipKey, err) - } - dontTranslateOwnership = b - delete(options, dontTranslateOwnershipKey) - } - - // Fail if the caller passed us more options than we know about. - if len(options) > 0 { - return nil, fmt.Errorf("unsupported mount options: %v", options) - } - - // The mounting EUID/EGID will be cached by this file system. This will - // be used to assign ownership to files that we own. - owner := fs.FileOwnerFromContext(ctx) - - // Construct the host file system mount and inode. - msrc := newMountSource(ctx, rootPath, owner, f, flags, dontTranslateOwnership) - return newInode(ctx, msrc, fd, false /* saveable */, false /* donated */) -} - -// InstallWhitelist locks down the MountNamespace to only the currently installed -// Dirents and the given paths. -func (f *Filesystem) InstallWhitelist(ctx context.Context, m *fs.MountNamespace) error { - return installWhitelist(ctx, m, f.paths) -} - -func installWhitelist(ctx context.Context, m *fs.MountNamespace, paths []string) error { - if len(paths) == 0 || (len(paths) == 1 && paths[0] == "") { - // Warning will be logged during filter installation if the empty - // whitelist matters (allows for host file access). - return nil - } - - // Done tracks entries already added. - done := make(map[string]bool) - root := m.Root() - defer root.DecRef() - - for i := 0; i < len(paths); i++ { - // Make sure the path is absolute. This is a sanity check. - if !path.IsAbs(paths[i]) { - return fmt.Errorf("path %q is not absolute", paths[i]) - } - - // We need to add all the intermediate paths, in case one of - // them is a symlink that needs to be resolved. - for j := 1; j <= len(paths[i]); j++ { - if j < len(paths[i]) && paths[i][j] != '/' { - continue - } - current := paths[i][:j] - - // Lookup the given component in the tree. - remainingTraversals := uint(maxTraversals) - d, err := m.FindLink(ctx, root, nil, current, &remainingTraversals) - if err != nil { - log.Warningf("populate failed for %q: %v", current, err) - continue - } - - // It's critical that this DecRef happens after the - // freeze below. This ensures that the dentry is in - // place to be frozen. Otherwise, we freeze without - // these entries. - defer d.DecRef() - - // Expand the last component if necessary. - if current == paths[i] { - // Is it a directory or symlink? - sattr := d.Inode.StableAttr - if fs.IsDir(sattr) { - for name := range childDentAttrs(ctx, d) { - paths = append(paths, path.Join(current, name)) - } - } - if fs.IsSymlink(sattr) { - // Only expand symlinks once. The - // folder structure may contain - // recursive symlinks and we don't want - // to end up infinitely expanding this - // symlink. This is safe because this - // is the last component. If a later - // path wants to symlink something - // beneath this symlink that will still - // be handled by the FindLink above. - if done[current] { - continue - } - - s, err := d.Inode.Readlink(ctx) - if err != nil { - log.Warningf("readlink failed for %q: %v", current, err) - continue - } - if path.IsAbs(s) { - paths = append(paths, s) - } else { - target := path.Join(path.Dir(current), s) - paths = append(paths, target) - } - } - } - - // Only report this one once even though we may look - // it up more than once. If we whitelist /a/b,/a then - // /a will be "done" when it is looked up for /a/b, - // however we still need to expand all of its contents - // when whitelisting /a. - if !done[current] { - log.Debugf("whitelisted: %s", current) - } - done[current] = true - } - } - - // Freeze the mount tree in place. This prevents any new paths from - // being opened and any old ones from being removed. If we do provide - // tmpfs mounts, we'll want to freeze/thaw those separately. - m.Freeze() - return nil -} - -func childDentAttrs(ctx context.Context, d *fs.Dirent) map[string]fs.DentAttr { - dirname, _ := d.FullName(nil /* root */) - dir, err := d.Inode.GetFile(ctx, d, fs.FileFlags{Read: true}) - if err != nil { - log.Warningf("failed to open directory %q: %v", dirname, err) - return nil - } - dir.DecRef() - var stubSerializer fs.CollectEntriesSerializer - if err := dir.Readdir(ctx, &stubSerializer); err != nil { - log.Warningf("failed to iterate on host directory %q: %v", dirname, err) - return nil - } - delete(stubSerializer.Entries, ".") - delete(stubSerializer.Entries, "..") - return stubSerializer.Entries -} - -// newMountSource constructs a new host fs.MountSource -// relative to a root path. The root should match the mount point. -func newMountSource(ctx context.Context, root string, mounter fs.FileOwner, filesystem fs.Filesystem, flags fs.MountSourceFlags, dontTranslateOwnership bool) *fs.MountSource { - return fs.NewMountSource(ctx, &superOperations{ - root: root, - inodeMappings: make(map[uint64]string), - mounter: mounter, - dontTranslateOwnership: dontTranslateOwnership, - }, filesystem, flags) -} - -// superOperations implements fs.MountSourceOperations. -// -// +stateify savable -type superOperations struct { - fs.SimpleMountSourceOperations - - // root is the path of the mount point. All inode mappings - // are relative to this root. - root string - - // inodeMappings contains mappings of fs.Inodes associated - // with this MountSource to paths under root. - inodeMappings map[uint64]string - - // mounter is the cached EUID/EGID that mounted this file system. - mounter fs.FileOwner - - // dontTranslateOwnership indicates whether to not translate file - // ownership. - // - // By default, files/directories owned by the sandbox uses UID/GID - // of the mounter. For files/directories that are not owned by the - // sandbox, file UID/GID is translated to a UID/GID which cannot - // be mapped in the sandboxed application's user namespace. The - // UID/GID will look like the nobody UID/GID (65534) but is not - // strictly owned by the user "nobody". - // - // If whitelistfs is a lower filesystem in an overlay, set - // dont_translate_ownership=true in mount options. - dontTranslateOwnership bool -} - -var _ fs.MountSourceOperations = (*superOperations)(nil) - -// ResetInodeMappings implements fs.MountSourceOperations.ResetInodeMappings. -func (m *superOperations) ResetInodeMappings() { - m.inodeMappings = make(map[uint64]string) -} - -// SaveInodeMapping implements fs.MountSourceOperations.SaveInodeMapping. -func (m *superOperations) SaveInodeMapping(inode *fs.Inode, path string) { - // This is very unintuitive. We *CANNOT* trust the inode's StableAttrs, - // because overlay copyUp may have changed them out from under us. - // So much for "immutable". - sattr := inode.InodeOperations.(*inodeOperations).fileState.sattr - m.inodeMappings[sattr.InodeID] = path -} - -// Keep implements fs.MountSourceOperations.Keep. -// -// TODO(b/72455313,b/77596690): It is possible to change the permissions on a -// host file while it is in the dirent cache (say from RO to RW), but it is not -// possible to re-open the file with more relaxed permissions, since the host -// FD is already open and stored in the inode. -// -// Using the dirent LRU cache increases the odds that this bug is encountered. -// Since host file access is relatively fast anyways, we disable the LRU cache -// for host fs files. Once we can properly deal with permissions changes and -// re-opening host files, we should revisit whether or not to make use of the -// LRU cache. -func (*superOperations) Keep(*fs.Dirent) bool { - return false -} - -func init() { - fs.RegisterFilesystem(&Filesystem{}) -} diff --git a/pkg/sentry/fs/host/fs_test.go b/pkg/sentry/fs/host/fs_test.go deleted file mode 100644 index 3111d2df9..000000000 --- a/pkg/sentry/fs/host/fs_test.go +++ /dev/null @@ -1,380 +0,0 @@ -// Copyright 2018 The gVisor Authors. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -package host - -import ( - "fmt" - "io/ioutil" - "os" - "path" - "reflect" - "sort" - "testing" - - "gvisor.dev/gvisor/pkg/context" - "gvisor.dev/gvisor/pkg/sentry/contexttest" - "gvisor.dev/gvisor/pkg/sentry/fs" -) - -// newTestMountNamespace creates a MountNamespace with a ramfs root. -// It returns the host folder created, which should be removed when done. -func newTestMountNamespace(t *testing.T) (*fs.MountNamespace, string, error) { - p, err := ioutil.TempDir("", "root") - if err != nil { - return nil, "", err - } - - fd, err := open(nil, p) - if err != nil { - os.RemoveAll(p) - return nil, "", err - } - ctx := contexttest.Context(t) - root, err := newInode(ctx, newMountSource(ctx, p, fs.RootOwner, &Filesystem{}, fs.MountSourceFlags{}, false), fd, false, false) - if err != nil { - os.RemoveAll(p) - return nil, "", err - } - mm, err := fs.NewMountNamespace(ctx, root) - if err != nil { - os.RemoveAll(p) - return nil, "", err - } - return mm, p, nil -} - -// createTestDirs populates the root with some test files and directories. -// /a/a1.txt -// /a/a2.txt -// /b/b1.txt -// /b/c/c1.txt -// /symlinks/normal.txt -// /symlinks/to_normal.txt -> /symlinks/normal.txt -// /symlinks/recursive -> /symlinks -func createTestDirs(ctx context.Context, t *testing.T, m *fs.MountNamespace) error { - r := m.Root() - defer r.DecRef() - - if err := r.CreateDirectory(ctx, r, "a", fs.FilePermsFromMode(0777)); err != nil { - return err - } - - a, err := r.Walk(ctx, r, "a") - if err != nil { - return err - } - defer a.DecRef() - - a1, err := a.Create(ctx, r, "a1.txt", fs.FileFlags{Read: true, Write: true}, fs.FilePermsFromMode(0666)) - if err != nil { - return err - } - a1.DecRef() - - a2, err := a.Create(ctx, r, "a2.txt", fs.FileFlags{Read: true, Write: true}, fs.FilePermsFromMode(0666)) - if err != nil { - return err - } - a2.DecRef() - - if err := r.CreateDirectory(ctx, r, "b", fs.FilePermsFromMode(0777)); err != nil { - return err - } - - b, err := r.Walk(ctx, r, "b") - if err != nil { - return err - } - defer b.DecRef() - - b1, err := b.Create(ctx, r, "b1.txt", fs.FileFlags{Read: true, Write: true}, fs.FilePermsFromMode(0666)) - if err != nil { - return err - } - b1.DecRef() - - if err := b.CreateDirectory(ctx, r, "c", fs.FilePermsFromMode(0777)); err != nil { - return err - } - - c, err := b.Walk(ctx, r, "c") - if err != nil { - return err - } - defer c.DecRef() - - c1, err := c.Create(ctx, r, "c1.txt", fs.FileFlags{Read: true, Write: true}, fs.FilePermsFromMode(0666)) - if err != nil { - return err - } - c1.DecRef() - - if err := r.CreateDirectory(ctx, r, "symlinks", fs.FilePermsFromMode(0777)); err != nil { - return err - } - - symlinks, err := r.Walk(ctx, r, "symlinks") - if err != nil { - return err - } - defer symlinks.DecRef() - - normal, err := symlinks.Create(ctx, r, "normal.txt", fs.FileFlags{Read: true, Write: true}, fs.FilePermsFromMode(0666)) - if err != nil { - return err - } - normal.DecRef() - - if err := symlinks.CreateLink(ctx, r, "/symlinks/normal.txt", "to_normal.txt"); err != nil { - return err - } - - return symlinks.CreateLink(ctx, r, "/symlinks", "recursive") -} - -// allPaths returns a slice of all paths of entries visible in the rootfs. -func allPaths(ctx context.Context, t *testing.T, m *fs.MountNamespace, base string) ([]string, error) { - var paths []string - root := m.Root() - defer root.DecRef() - - maxTraversals := uint(1) - d, err := m.FindLink(ctx, root, nil, base, &maxTraversals) - if err != nil { - t.Logf("FindLink failed for %q", base) - return paths, err - } - defer d.DecRef() - - if fs.IsDir(d.Inode.StableAttr) { - dir, err := d.Inode.GetFile(ctx, d, fs.FileFlags{Read: true}) - if err != nil { - return nil, fmt.Errorf("failed to open directory %q: %v", base, err) - } - iter, ok := dir.FileOperations.(fs.DirIterator) - if !ok { - return nil, fmt.Errorf("cannot directly iterate on host directory %q", base) - } - dirCtx := &fs.DirCtx{ - Serializer: noopDentrySerializer{}, - } - if _, err := fs.DirentReaddir(ctx, d, iter, root, dirCtx, 0); err != nil { - return nil, err - } - for name := range dirCtx.DentAttrs() { - if name == "." || name == ".." { - continue - } - - fullName := path.Join(base, name) - paths = append(paths, fullName) - - // Recurse. - subpaths, err := allPaths(ctx, t, m, fullName) - if err != nil { - return paths, err - } - paths = append(paths, subpaths...) - } - } - - return paths, nil -} - -type noopDentrySerializer struct{} - -func (noopDentrySerializer) CopyOut(string, fs.DentAttr) error { - return nil -} -func (noopDentrySerializer) Written() int { - return 4096 -} - -// pathsEqual returns true if the two string slices contain the same entries. -func pathsEqual(got, want []string) bool { - sort.Strings(got) - sort.Strings(want) - - if len(got) != len(want) { - return false - } - - for i := range got { - if got[i] != want[i] { - return false - } - } - - return true -} - -func TestWhitelist(t *testing.T) { - for _, test := range []struct { - // description of the test. - desc string - // paths are the paths to whitelist - paths []string - // want are all of the directory entries that should be - // visible (nothing beyond this set should be visible). - want []string - }{ - { - desc: "root", - paths: []string{"/"}, - want: []string{"/a", "/a/a1.txt", "/a/a2.txt", "/b", "/b/b1.txt", "/b/c", "/b/c/c1.txt", "/symlinks", "/symlinks/normal.txt", "/symlinks/to_normal.txt", "/symlinks/recursive"}, - }, - { - desc: "top-level directories", - paths: []string{"/a", "/b"}, - want: []string{"/a", "/a/a1.txt", "/a/a2.txt", "/b", "/b/b1.txt", "/b/c", "/b/c/c1.txt"}, - }, - { - desc: "nested directories (1/2)", - paths: []string{"/b", "/b/c"}, - want: []string{"/b", "/b/b1.txt", "/b/c", "/b/c/c1.txt"}, - }, - { - desc: "nested directories (2/2)", - paths: []string{"/b/c", "/b"}, - want: []string{"/b", "/b/b1.txt", "/b/c", "/b/c/c1.txt"}, - }, - { - desc: "single file", - paths: []string{"/b/c/c1.txt"}, - want: []string{"/b", "/b/c", "/b/c/c1.txt"}, - }, - { - desc: "single file and directory", - paths: []string{"/a/a1.txt", "/b/c"}, - want: []string{"/a", "/a/a1.txt", "/b", "/b/c", "/b/c/c1.txt"}, - }, - { - desc: "symlink", - paths: []string{"/symlinks/to_normal.txt"}, - want: []string{"/symlinks", "/symlinks/normal.txt", "/symlinks/to_normal.txt"}, - }, - { - desc: "recursive symlink", - paths: []string{"/symlinks/recursive/normal.txt"}, - want: []string{"/symlinks", "/symlinks/normal.txt", "/symlinks/recursive"}, - }, - } { - t.Run(test.desc, func(t *testing.T) { - m, p, err := newTestMountNamespace(t) - if err != nil { - t.Errorf("Failed to create MountNamespace: %v", err) - } - defer os.RemoveAll(p) - - ctx := withRoot(contexttest.RootContext(t), m.Root()) - if err := createTestDirs(ctx, t, m); err != nil { - t.Errorf("Failed to create test dirs: %v", err) - } - - if err := installWhitelist(ctx, m, test.paths); err != nil { - t.Errorf("installWhitelist(%v) err got %v want nil", test.paths, err) - } - - got, err := allPaths(ctx, t, m, "/") - if err != nil { - t.Fatalf("Failed to lookup paths (whitelisted: %v): %v", test.paths, err) - } - - if !pathsEqual(got, test.want) { - t.Errorf("For paths %v got %v want %v", test.paths, got, test.want) - } - }) - } -} - -func TestRootPath(t *testing.T) { - // Create a temp dir, which will be the root of our mounted fs. - rootPath, err := ioutil.TempDir(os.TempDir(), "root") - if err != nil { - t.Fatalf("TempDir failed: %v", err) - } - defer os.RemoveAll(rootPath) - - // Create two files inside the new root, one which will be whitelisted - // and one not. - whitelisted, err := ioutil.TempFile(rootPath, "white") - if err != nil { - t.Fatalf("TempFile failed: %v", err) - } - if _, err := ioutil.TempFile(rootPath, "black"); err != nil { - t.Fatalf("TempFile failed: %v", err) - } - - // Create a mount with a root path and single whitelisted file. - hostFS := &Filesystem{} - ctx := contexttest.Context(t) - data := fmt.Sprintf("%s=%s,%s=%s", rootPathKey, rootPath, whitelistKey, whitelisted.Name()) - inode, err := hostFS.Mount(ctx, "", fs.MountSourceFlags{}, data, nil) - if err != nil { - t.Fatalf("Mount failed: %v", err) - } - mm, err := fs.NewMountNamespace(ctx, inode) - if err != nil { - t.Fatalf("NewMountNamespace failed: %v", err) - } - if err := hostFS.InstallWhitelist(ctx, mm); err != nil { - t.Fatalf("InstallWhitelist failed: %v", err) - } - - // Get the contents of the root directory. - rootDir := mm.Root() - rctx := withRoot(ctx, rootDir) - f, err := rootDir.Inode.GetFile(rctx, rootDir, fs.FileFlags{}) - if err != nil { - t.Fatalf("GetFile failed: %v", err) - } - c := &fs.CollectEntriesSerializer{} - if err := f.Readdir(rctx, c); err != nil { - t.Fatalf("Readdir failed: %v", err) - } - - // We should have only our whitelisted file, plus the dots. - want := []string{path.Base(whitelisted.Name()), ".", ".."} - got := c.Order - sort.Strings(want) - sort.Strings(got) - if !reflect.DeepEqual(got, want) { - t.Errorf("Readdir got %v, wanted %v", got, want) - } -} - -type rootContext struct { - context.Context - root *fs.Dirent -} - -// withRoot returns a copy of ctx with the given root. -func withRoot(ctx context.Context, root *fs.Dirent) context.Context { - return &rootContext{ - Context: ctx, - root: root, - } -} - -// Value implements Context.Value. -func (rc rootContext) Value(key interface{}) interface{} { - switch key { - case fs.CtxRoot: - rc.root.IncRef() - return rc.root - default: - return rc.Context.Value(key) - } -} diff --git a/pkg/sentry/fs/host/host.go b/pkg/sentry/fs/host/host.go new file mode 100644 index 000000000..081ba1dd8 --- /dev/null +++ b/pkg/sentry/fs/host/host.go @@ -0,0 +1,59 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// Package host supports file descriptors imported directly. +package host + +import ( + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/syserror" +) + +// filesystem is a host filesystem. +// +// +stateify savable +type filesystem struct{} + +func init() { + fs.RegisterFilesystem(&filesystem{}) +} + +// FilesystemName is the name under which the filesystem is registered. +const FilesystemName = "host" + +// Name is the name of the filesystem. +func (*filesystem) Name() string { + return FilesystemName +} + +// Mount returns an error. Mounting hostfs is not allowed. +func (*filesystem) Mount(ctx context.Context, device string, flags fs.MountSourceFlags, data string, dataObj interface{}) (*fs.Inode, error) { + return nil, syserror.EPERM +} + +// AllowUserMount prohibits users from using mount(2) with this file system. +func (*filesystem) AllowUserMount() bool { + return false +} + +// AllowUserList prohibits this filesystem to be listed in /proc/filesystems. +func (*filesystem) AllowUserList() bool { + return false +} + +// Flags returns that there is nothing special about this file system. +func (*filesystem) Flags() fs.FilesystemFlags { + return 0 +} diff --git a/pkg/sentry/fs/host/inode.go b/pkg/sentry/fs/host/inode.go index 6fa39caab..62f1246aa 100644 --- a/pkg/sentry/fs/host/inode.go +++ b/pkg/sentry/fs/host/inode.go @@ -17,12 +17,10 @@ package host import ( "syscall" - "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/fd" "gvisor.dev/gvisor/pkg/safemem" "gvisor.dev/gvisor/pkg/secio" - "gvisor.dev/gvisor/pkg/sentry/device" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fs/fsutil" "gvisor.dev/gvisor/pkg/sentry/memmap" @@ -69,9 +67,6 @@ type inodeOperations struct { // // +stateify savable type inodeFileState struct { - // Common file system state. - mops *superOperations `state:"wait"` - // descriptor is the backing host FD. descriptor *descriptor `state:"wait"` @@ -160,7 +155,7 @@ func (i *inodeFileState) unstableAttr(ctx context.Context) (fs.UnstableAttr, err if err := syscall.Fstat(i.FD(), &s); err != nil { return fs.UnstableAttr{}, err } - return unstableAttr(i.mops, &s), nil + return unstableAttr(&s), nil } // Allocate implements fsutil.CachedFileObject.Allocate. @@ -172,7 +167,7 @@ func (i *inodeFileState) Allocate(_ context.Context, offset, length int64) error var _ fs.InodeOperations = (*inodeOperations)(nil) // newInode returns a new fs.Inode backed by the host FD. -func newInode(ctx context.Context, msrc *fs.MountSource, fd int, saveable bool, donated bool) (*fs.Inode, error) { +func newInode(ctx context.Context, msrc *fs.MountSource, fd int, saveable bool) (*fs.Inode, error) { // Retrieve metadata. var s syscall.Stat_t err := syscall.Fstat(fd, &s) @@ -181,24 +176,17 @@ func newInode(ctx context.Context, msrc *fs.MountSource, fd int, saveable bool, } fileState := &inodeFileState{ - mops: msrc.MountSourceOperations.(*superOperations), sattr: stableAttr(&s), } // Initialize the wrapped host file descriptor. - fileState.descriptor, err = newDescriptor( - fd, - donated, - saveable, - wouldBlock(&s), - &fileState.queue, - ) + fileState.descriptor, err = newDescriptor(fd, saveable, wouldBlock(&s), &fileState.queue) if err != nil { return nil, err } // Build the fs.InodeOperations. - uattr := unstableAttr(msrc.MountSourceOperations.(*superOperations), &s) + uattr := unstableAttr(&s) iops := &inodeOperations{ fileState: fileState, cachingInodeOps: fsutil.NewCachingInodeOperations(ctx, fileState, uattr, fsutil.CachingInodeOperationsOptions{ @@ -232,54 +220,23 @@ func (i *inodeOperations) Release(context.Context) { // Lookup implements fs.InodeOperations.Lookup. func (i *inodeOperations) Lookup(ctx context.Context, dir *fs.Inode, name string) (*fs.Dirent, error) { - // Get a new FD relative to i at name. - fd, err := open(i, name) - if err != nil { - if err == syserror.ENOENT { - return nil, syserror.ENOENT - } - return nil, err - } - - inode, err := newInode(ctx, dir.MountSource, fd, false /* saveable */, false /* donated */) - if err != nil { - return nil, err - } - - // Return the fs.Dirent. - return fs.NewDirent(ctx, inode, name), nil + return nil, syserror.ENOENT } // Create implements fs.InodeOperations.Create. func (i *inodeOperations) Create(ctx context.Context, dir *fs.Inode, name string, flags fs.FileFlags, perm fs.FilePermissions) (*fs.File, error) { - // Create a file relative to i at name. - // - // N.B. We always open this file O_RDWR regardless of flags because a - // future GetFile might want more access. Open allows this regardless - // of perm. - fd, err := openAt(i, name, syscall.O_RDWR|syscall.O_CREAT|syscall.O_EXCL, perm.LinuxMode()) - if err != nil { - return nil, err - } - - inode, err := newInode(ctx, dir.MountSource, fd, false /* saveable */, false /* donated */) - if err != nil { - return nil, err - } + return nil, syserror.EPERM - d := fs.NewDirent(ctx, inode, name) - defer d.DecRef() - return inode.GetFile(ctx, d, flags) } // CreateDirectory implements fs.InodeOperations.CreateDirectory. func (i *inodeOperations) CreateDirectory(ctx context.Context, dir *fs.Inode, name string, perm fs.FilePermissions) error { - return syscall.Mkdirat(i.fileState.FD(), name, uint32(perm.LinuxMode())) + return syserror.EPERM } // CreateLink implements fs.InodeOperations.CreateLink. func (i *inodeOperations) CreateLink(ctx context.Context, dir *fs.Inode, oldname string, newname string) error { - return createLink(i.fileState.FD(), oldname, newname) + return syserror.EPERM } // CreateHardLink implements fs.InodeOperations.CreateHardLink. @@ -294,25 +251,17 @@ func (*inodeOperations) CreateFifo(context.Context, *fs.Inode, string, fs.FilePe // Remove implements fs.InodeOperations.Remove. func (i *inodeOperations) Remove(ctx context.Context, dir *fs.Inode, name string) error { - return unlinkAt(i.fileState.FD(), name, false /* dir */) + return syserror.EPERM } // RemoveDirectory implements fs.InodeOperations.RemoveDirectory. func (i *inodeOperations) RemoveDirectory(ctx context.Context, dir *fs.Inode, name string) error { - return unlinkAt(i.fileState.FD(), name, true /* dir */) + return syserror.EPERM } // Rename implements fs.InodeOperations.Rename. func (i *inodeOperations) Rename(ctx context.Context, inode *fs.Inode, oldParent *fs.Inode, oldName string, newParent *fs.Inode, newName string, replacement bool) error { - op, ok := oldParent.InodeOperations.(*inodeOperations) - if !ok { - return syscall.EXDEV - } - np, ok := newParent.InodeOperations.(*inodeOperations) - if !ok { - return syscall.EXDEV - } - return syscall.Renameat(op.fileState.FD(), oldName, np.fileState.FD(), newName) + return syserror.EPERM } // Bind implements fs.InodeOperations.Bind. @@ -448,82 +397,17 @@ func (i *inodeOperations) StatFS(context.Context) (fs.Info, error) { } // AddLink implements fs.InodeOperations.AddLink. -// FIXME(b/63117438): Remove this from InodeOperations altogether. func (i *inodeOperations) AddLink() {} // DropLink implements fs.InodeOperations.DropLink. -// FIXME(b/63117438): Remove this from InodeOperations altogether. func (i *inodeOperations) DropLink() {} // NotifyStatusChange implements fs.InodeOperations.NotifyStatusChange. -// FIXME(b/63117438): Remove this from InodeOperations altogether. func (i *inodeOperations) NotifyStatusChange(ctx context.Context) {} // readdirAll returns all of the directory entries in i. func (i *inodeOperations) readdirAll(d *dirInfo) (map[string]fs.DentAttr, error) { - i.readdirMu.Lock() - defer i.readdirMu.Unlock() - - fd := i.fileState.FD() - - // syscall.ReadDirent will use getdents, which will seek the file past - // the last directory entry. To read the directory entries a second - // time, we need to seek back to the beginning. - if _, err := syscall.Seek(fd, 0, 0); err != nil { - if err == syscall.ESPIPE { - // All directories should be seekable. If this file - // isn't seekable, it is not a directory and we should - // return that more sane error. - err = syscall.ENOTDIR - } - return nil, err - } - - names := make([]string, 0, 100) - for { - // Refill the buffer if necessary - if d.bufp >= d.nbuf { - d.bufp = 0 - // ReadDirent will just do a sys_getdents64 to the kernel. - n, err := syscall.ReadDirent(fd, d.buf) - if err != nil { - return nil, err - } - if n == 0 { - break // EOF - } - d.nbuf = n - } - - var nb int - // Parse the dirent buffer we just get and return the directory names along - // with the number of bytes consumed in the buffer. - nb, _, names = syscall.ParseDirent(d.buf[d.bufp:d.nbuf], -1, names) - d.bufp += nb - } - - entries := make(map[string]fs.DentAttr) - for _, filename := range names { - // Lookup the type and host device and inode. - stat, lerr := fstatat(fd, filename, linux.AT_SYMLINK_NOFOLLOW) - if lerr == syscall.ENOENT { - // File disappeared between readdir and lstat. - // Just treat it as if it didn't exist. - continue - } - - // There was a serious problem, we should probably report it. - if lerr != nil { - return nil, lerr - } - - entries[filename] = fs.DentAttr{ - Type: nodeType(&stat), - InodeID: hostFileDevice.Map(device.MultiDeviceKey{ - Device: stat.Dev, - Inode: stat.Ino, - }), - } - } - return entries, nil + // We only support non-directory file descriptors that have been + // imported, so just claim that this isn't a directory, even if it is. + return nil, syscall.ENOTDIR } diff --git a/pkg/sentry/fs/host/inode_state.go b/pkg/sentry/fs/host/inode_state.go index 299e0e0b0..1adbd4562 100644 --- a/pkg/sentry/fs/host/inode_state.go +++ b/pkg/sentry/fs/host/inode_state.go @@ -18,29 +18,14 @@ import ( "fmt" "syscall" - "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/device" "gvisor.dev/gvisor/pkg/sentry/fs" ) -// beforeSave is invoked by stateify. -func (i *inodeFileState) beforeSave() { - if !i.queue.IsEmpty() { - panic("event queue must be empty") - } - if !i.descriptor.donated && i.sattr.Type == fs.RegularFile { - uattr, err := i.unstableAttr(context.Background()) - if err != nil { - panic(fs.ErrSaveRejection{fmt.Errorf("failed to get unstable atttribute of %s: %v", i.mops.inodeMappings[i.sattr.InodeID], err)}) - } - i.savedUAttr = &uattr - } -} - // afterLoad is invoked by stateify. func (i *inodeFileState) afterLoad() { // Initialize the descriptor value. - if err := i.descriptor.initAfterLoad(i.mops, i.sattr.InodeID, &i.queue); err != nil { + if err := i.descriptor.initAfterLoad(i.sattr.InodeID, &i.queue); err != nil { panic(fmt.Sprintf("failed to load value of descriptor: %v", err)) } @@ -61,19 +46,4 @@ func (i *inodeFileState) afterLoad() { // change across save and restore, error out. panic(fs.ErrCorruption{fmt.Errorf("host %s conflict in host device mappings: %s", key, hostFileDevice)}) } - - if !i.descriptor.donated && i.sattr.Type == fs.RegularFile { - env, ok := fs.CurrentRestoreEnvironment() - if !ok { - panic("missing restore environment") - } - uattr := unstableAttr(i.mops, &s) - if env.ValidateFileSize && uattr.Size != i.savedUAttr.Size { - panic(fs.ErrCorruption{fmt.Errorf("file size has changed for %s: previously %d, now %d", i.mops.inodeMappings[i.sattr.InodeID], i.savedUAttr.Size, uattr.Size)}) - } - if env.ValidateFileTimestamp && uattr.ModificationTime != i.savedUAttr.ModificationTime { - panic(fs.ErrCorruption{fmt.Errorf("file modification time has changed for %s: previously %v, now %v", i.mops.inodeMappings[i.sattr.InodeID], i.savedUAttr.ModificationTime, uattr.ModificationTime)}) - } - i.savedUAttr = nil - } } diff --git a/pkg/sentry/fs/host/inode_test.go b/pkg/sentry/fs/host/inode_test.go index 7221bc825..c507f57eb 100644 --- a/pkg/sentry/fs/host/inode_test.go +++ b/pkg/sentry/fs/host/inode_test.go @@ -15,79 +15,12 @@ package host import ( - "io/ioutil" - "os" - "path" "syscall" "testing" "gvisor.dev/gvisor/pkg/sentry/contexttest" - "gvisor.dev/gvisor/pkg/sentry/fs" ) -// TestMultipleReaddir verifies that multiple Readdir calls return the same -// thing if they use different dir contexts. -func TestMultipleReaddir(t *testing.T) { - p, err := ioutil.TempDir("", "readdir") - if err != nil { - t.Fatalf("Failed to create test dir: %v", err) - } - defer os.RemoveAll(p) - - f, err := os.Create(path.Join(p, "a.txt")) - if err != nil { - t.Fatalf("Failed to create a.txt: %v", err) - } - f.Close() - - f, err = os.Create(path.Join(p, "b.txt")) - if err != nil { - t.Fatalf("Failed to create b.txt: %v", err) - } - f.Close() - - fd, err := open(nil, p) - if err != nil { - t.Fatalf("Failed to open %q: %v", p, err) - } - ctx := contexttest.Context(t) - n, err := newInode(ctx, newMountSource(ctx, p, fs.RootOwner, &Filesystem{}, fs.MountSourceFlags{}, false), fd, false, false) - if err != nil { - t.Fatalf("Failed to create inode: %v", err) - } - - dirent := fs.NewDirent(ctx, n, "readdir") - openFile, err := n.GetFile(ctx, dirent, fs.FileFlags{Read: true}) - if err != nil { - t.Fatalf("Failed to get file: %v", err) - } - defer openFile.DecRef() - - c1 := &fs.DirCtx{DirCursor: new(string)} - if _, err := openFile.FileOperations.(*fileOperations).IterateDir(ctx, dirent, c1, 0); err != nil { - t.Fatalf("First Readdir failed: %v", err) - } - - c2 := &fs.DirCtx{DirCursor: new(string)} - if _, err := openFile.FileOperations.(*fileOperations).IterateDir(ctx, dirent, c2, 0); err != nil { - t.Errorf("Second Readdir failed: %v", err) - } - - if _, ok := c1.DentAttrs()["a.txt"]; !ok { - t.Errorf("want a.txt in first Readdir, got %v", c1.DentAttrs()) - } - if _, ok := c1.DentAttrs()["b.txt"]; !ok { - t.Errorf("want b.txt in first Readdir, got %v", c1.DentAttrs()) - } - - if _, ok := c2.DentAttrs()["a.txt"]; !ok { - t.Errorf("want a.txt in second Readdir, got %v", c2.DentAttrs()) - } - if _, ok := c2.DentAttrs()["b.txt"]; !ok { - t.Errorf("want b.txt in second Readdir, got %v", c2.DentAttrs()) - } -} - // TestCloseFD verifies fds will be closed. func TestCloseFD(t *testing.T) { var p [2]int @@ -99,7 +32,7 @@ func TestCloseFD(t *testing.T) { // Use the write-end because we will detect if it's closed on the read end. ctx := contexttest.Context(t) - file, err := NewFile(ctx, p[1], fs.RootOwner) + file, err := NewFile(ctx, p[1]) if err != nil { t.Fatalf("Failed to create File: %v", err) } diff --git a/pkg/sentry/fs/host/ioctl_unsafe.go b/pkg/sentry/fs/host/ioctl_unsafe.go index 271582e54..150ac8e19 100644 --- a/pkg/sentry/fs/host/ioctl_unsafe.go +++ b/pkg/sentry/fs/host/ioctl_unsafe.go @@ -21,6 +21,8 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" ) +// LINT.IfChange + func ioctlGetTermios(fd int) (*linux.Termios, error) { var t linux.Termios _, _, errno := syscall.Syscall(syscall.SYS_IOCTL, uintptr(fd), linux.TCGETS, uintptr(unsafe.Pointer(&t))) @@ -54,3 +56,5 @@ func ioctlSetWinsize(fd int, w *linux.Winsize) error { } return nil } + +// LINT.ThenChange(../../fsimpl/host/ioctl_unsafe.go) diff --git a/pkg/sentry/fs/host/socket_test.go b/pkg/sentry/fs/host/socket_test.go index eb4afe520..affdbcacb 100644 --- a/pkg/sentry/fs/host/socket_test.go +++ b/pkg/sentry/fs/host/socket_test.go @@ -199,14 +199,14 @@ func TestListen(t *testing.T) { } func TestPasscred(t *testing.T) { - e := ConnectedEndpoint{} + e := &ConnectedEndpoint{} if got, want := e.Passcred(), false; got != want { t.Errorf("Got %#v.Passcred() = %t, want = %t", e, got, want) } } func TestGetLocalAddress(t *testing.T) { - e := ConnectedEndpoint{path: "foo"} + e := &ConnectedEndpoint{path: "foo"} want := tcpip.FullAddress{Addr: tcpip.Address("foo")} if got, err := e.GetLocalAddress(); err != nil || got != want { t.Errorf("Got %#v.GetLocalAddress() = %#v, %v, want = %#v, %v", e, got, err, want, nil) @@ -214,7 +214,7 @@ func TestGetLocalAddress(t *testing.T) { } func TestQueuedSize(t *testing.T) { - e := ConnectedEndpoint{} + e := &ConnectedEndpoint{} tests := []struct { name string f func() int64 diff --git a/pkg/sentry/fs/host/tty.go b/pkg/sentry/fs/host/tty.go index 3f218b4a7..cb91355ab 100644 --- a/pkg/sentry/fs/host/tty.go +++ b/pkg/sentry/fs/host/tty.go @@ -26,6 +26,8 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) +// LINT.IfChange + // TTYFileOperations implements fs.FileOperations for a host file descriptor // that wraps a TTY FD. // @@ -43,6 +45,7 @@ type TTYFileOperations struct { // connected to this TTY. fgProcessGroup *kernel.ProcessGroup + // termios contains the terminal attributes for this TTY. termios linux.KernelTermios } @@ -357,3 +360,5 @@ func (t *TTYFileOperations) checkChange(ctx context.Context, sig linux.Signal) e _ = pg.SendSignal(kernel.SignalInfoPriv(sig)) return kernel.ERESTARTSYS } + +// LINT.ThenChange(../../fsimpl/host/tty.go) diff --git a/pkg/sentry/fs/host/util.go b/pkg/sentry/fs/host/util.go index e37e687c6..1b0356930 100644 --- a/pkg/sentry/fs/host/util.go +++ b/pkg/sentry/fs/host/util.go @@ -16,7 +16,6 @@ package host import ( "os" - "path" "syscall" "gvisor.dev/gvisor/pkg/abi/linux" @@ -28,45 +27,6 @@ import ( "gvisor.dev/gvisor/pkg/syserror" ) -func open(parent *inodeOperations, name string) (int, error) { - if parent == nil && !path.IsAbs(name) { - return -1, syserror.EINVAL - } - name = path.Clean(name) - - // Don't follow through symlinks. - flags := syscall.O_NOFOLLOW - - if fd, err := openAt(parent, name, flags|syscall.O_RDWR, 0); err == nil { - return fd, nil - } - // Retry as read-only. - if fd, err := openAt(parent, name, flags|syscall.O_RDONLY, 0); err == nil { - return fd, nil - } - - // Retry as write-only. - if fd, err := openAt(parent, name, flags|syscall.O_WRONLY, 0); err == nil { - return fd, nil - } - - // Retry as a symlink, by including O_PATH as an option. - fd, err := openAt(parent, name, linux.O_PATH|flags, 0) - if err == nil { - return fd, nil - } - - // Everything failed. - return -1, err -} - -func openAt(parent *inodeOperations, name string, flags int, perm linux.FileMode) (int, error) { - if parent == nil { - return syscall.Open(name, flags, uint32(perm)) - } - return syscall.Openat(parent.fileState.FD(), name, flags, uint32(perm)) -} - func nodeType(s *syscall.Stat_t) fs.InodeType { switch x := (s.Mode & syscall.S_IFMT); x { case syscall.S_IFLNK: @@ -107,51 +67,19 @@ func stableAttr(s *syscall.Stat_t) fs.StableAttr { } } -func owner(mo *superOperations, s *syscall.Stat_t) fs.FileOwner { - // User requested no translation, just return actual owner. - if mo.dontTranslateOwnership { - return fs.FileOwner{auth.KUID(s.Uid), auth.KGID(s.Gid)} +func owner(s *syscall.Stat_t) fs.FileOwner { + return fs.FileOwner{ + UID: auth.KUID(s.Uid), + GID: auth.KGID(s.Gid), } - - // Show only IDs relevant to the sandboxed task. I.e. if we not own the - // file, no sandboxed task can own the file. In that case, we - // use OverflowID for UID, implying that the IDs are not mapped in the - // "root" user namespace. - // - // E.g. - // sandbox's host EUID/EGID is 1/1. - // some_dir's host UID/GID is 2/1. - // Task that mounted this fs has virtualized EUID/EGID 5/5. - // - // If you executed `ls -n` in the sandboxed task, it would show: - // drwxwrxwrx [...] 65534 5 [...] some_dir - - // Files are owned by OverflowID by default. - owner := fs.FileOwner{auth.KUID(auth.OverflowUID), auth.KGID(auth.OverflowGID)} - - // If we own file on host, let mounting task's initial EUID own - // the file. - if s.Uid == hostUID { - owner.UID = mo.mounter.UID - } - - // If our group matches file's group, make file's group match - // the mounting task's initial EGID. - for _, gid := range hostGIDs { - if s.Gid == gid { - owner.GID = mo.mounter.GID - break - } - } - return owner } -func unstableAttr(mo *superOperations, s *syscall.Stat_t) fs.UnstableAttr { +func unstableAttr(s *syscall.Stat_t) fs.UnstableAttr { return fs.UnstableAttr{ Size: s.Size, Usage: s.Blocks * 512, Perms: fs.FilePermsFromMode(linux.FileMode(s.Mode)), - Owner: owner(mo, s), + Owner: owner(s), AccessTime: ktime.FromUnix(s.Atim.Sec, s.Atim.Nsec), ModificationTime: ktime.FromUnix(s.Mtim.Sec, s.Mtim.Nsec), StatusChangeTime: ktime.FromUnix(s.Ctim.Sec, s.Ctim.Nsec), @@ -165,6 +93,8 @@ type dirInfo struct { bufp int // location of next record in buf. } +// LINT.IfChange + // isBlockError unwraps os errors and checks if they are caused by EAGAIN or // EWOULDBLOCK. This is so they can be transformed into syserror.ErrWouldBlock. func isBlockError(err error) bool { @@ -177,6 +107,8 @@ func isBlockError(err error) bool { return false } +// LINT.ThenChange(../../fsimpl/host/util.go) + func hostEffectiveKIDs() (uint32, []uint32, error) { gids, err := os.Getgroups() if err != nil { diff --git a/pkg/sentry/fs/host/util_unsafe.go b/pkg/sentry/fs/host/util_unsafe.go index 3ab36b088..23bd35d64 100644 --- a/pkg/sentry/fs/host/util_unsafe.go +++ b/pkg/sentry/fs/host/util_unsafe.go @@ -26,26 +26,6 @@ import ( // NulByte is a single NUL byte. It is passed to readlinkat as an empty string. var NulByte byte = '\x00' -func createLink(fd int, name string, linkName string) error { - namePtr, err := syscall.BytePtrFromString(name) - if err != nil { - return err - } - linkNamePtr, err := syscall.BytePtrFromString(linkName) - if err != nil { - return err - } - _, _, errno := syscall.Syscall( - syscall.SYS_SYMLINKAT, - uintptr(unsafe.Pointer(namePtr)), - uintptr(fd), - uintptr(unsafe.Pointer(linkNamePtr))) - if errno != 0 { - return errno - } - return nil -} - func readLink(fd int) (string, error) { // Buffer sizing copied from os.Readlink. for l := 128; ; l *= 2 { @@ -66,27 +46,6 @@ func readLink(fd int) (string, error) { } } -func unlinkAt(fd int, name string, dir bool) error { - namePtr, err := syscall.BytePtrFromString(name) - if err != nil { - return err - } - var flags uintptr - if dir { - flags = linux.AT_REMOVEDIR - } - _, _, errno := syscall.Syscall( - syscall.SYS_UNLINKAT, - uintptr(fd), - uintptr(unsafe.Pointer(namePtr)), - flags, - ) - if errno != 0 { - return errno - } - return nil -} - func timespecFromTimestamp(t ktime.Time, omit, setSysTime bool) syscall.Timespec { if omit { return syscall.Timespec{0, linux.UTIME_OMIT} diff --git a/pkg/sentry/fs/host/wait_test.go b/pkg/sentry/fs/host/wait_test.go index d49c3a635..ce397a5e3 100644 --- a/pkg/sentry/fs/host/wait_test.go +++ b/pkg/sentry/fs/host/wait_test.go @@ -20,7 +20,6 @@ import ( "time" "gvisor.dev/gvisor/pkg/sentry/contexttest" - "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/waiter" ) @@ -34,7 +33,7 @@ func TestWait(t *testing.T) { defer syscall.Close(fds[1]) ctx := contexttest.Context(t) - file, err := NewFile(ctx, fds[0], fs.RootOwner) + file, err := NewFile(ctx, fds[0]) if err != nil { syscall.Close(fds[0]) t.Fatalf("NewFile failed: %v", err) diff --git a/pkg/sentry/fs/inode.go b/pkg/sentry/fs/inode.go index 55fb71c16..a34fbc946 100644 --- a/pkg/sentry/fs/inode.go +++ b/pkg/sentry/fs/inode.go @@ -102,7 +102,6 @@ func (i *Inode) DecRef() { // destroy releases the Inode and releases the msrc reference taken. func (i *Inode) destroy() { - // FIXME(b/38173783): Context is not plumbed here. ctx := context.Background() if err := i.WriteOut(ctx); err != nil { // FIXME(b/65209558): Mark as warning again once noatime is @@ -397,8 +396,6 @@ func (i *Inode) Getlink(ctx context.Context) (*Dirent, error) { // AddLink calls i.InodeOperations.AddLink. func (i *Inode) AddLink() { if i.overlay != nil { - // FIXME(b/63117438): Remove this from InodeOperations altogether. - // // This interface is only used by ramfs to update metadata of // children. These filesystems should _never_ have overlay // Inodes cached as children. So explicitly disallow this diff --git a/pkg/sentry/fs/inode_overlay.go b/pkg/sentry/fs/inode_overlay.go index 5ada33a32..537c8d257 100644 --- a/pkg/sentry/fs/inode_overlay.go +++ b/pkg/sentry/fs/inode_overlay.go @@ -231,7 +231,8 @@ func overlayCreate(ctx context.Context, o *overlayEntry, parent *Dirent, name st upperFile.Dirent.Inode.IncRef() entry, err := newOverlayEntry(ctx, upperFile.Dirent.Inode, nil, false) if err != nil { - cleanupUpper(ctx, o.upper, name) + werr := fmt.Errorf("newOverlayEntry failed: %v", err) + cleanupUpper(ctx, o.upper, name, werr) return nil, err } diff --git a/pkg/sentry/fs/inotify.go b/pkg/sentry/fs/inotify.go index 928c90aa0..e3a715c1f 100644 --- a/pkg/sentry/fs/inotify.go +++ b/pkg/sentry/fs/inotify.go @@ -143,7 +143,10 @@ func (i *Inotify) Read(ctx context.Context, _ *File, dst usermem.IOSequence, _ i } var writeLen int64 - for event := i.events.Front(); event != nil; event = event.Next() { + for it := i.events.Front(); it != nil; { + event := it + it = it.Next() + // Does the buffer have enough remaining space to hold the event we're // about to write out? if dst.NumBytes() < int64(event.sizeOf()) { diff --git a/pkg/sentry/fs/mount_test.go b/pkg/sentry/fs/mount_test.go index e672a438c..a3d10770b 100644 --- a/pkg/sentry/fs/mount_test.go +++ b/pkg/sentry/fs/mount_test.go @@ -36,11 +36,12 @@ func mountPathsAre(root *Dirent, got []*Mount, want ...string) error { gotPaths := make(map[string]struct{}, len(got)) gotStr := make([]string, len(got)) for i, g := range got { - groot := g.Root() - name, _ := groot.FullName(root) - groot.DecRef() - gotStr[i] = name - gotPaths[name] = struct{}{} + if groot := g.Root(); groot != nil { + name, _ := groot.FullName(root) + groot.DecRef() + gotStr[i] = name + gotPaths[name] = struct{}{} + } } if len(got) != len(want) { return fmt.Errorf("mount paths are different, got: %q, want: %q", gotStr, want) diff --git a/pkg/sentry/fs/mounts.go b/pkg/sentry/fs/mounts.go index 574a2cc91..b414ddaee 100644 --- a/pkg/sentry/fs/mounts.go +++ b/pkg/sentry/fs/mounts.go @@ -100,10 +100,14 @@ func newUndoMount(d *Dirent) *Mount { } } -// Root returns the root dirent of this mount. Callers must call DecRef on the -// returned dirent. +// Root returns the root dirent of this mount. +// +// This may return nil if the mount has already been free. Callers must handle this +// case appropriately. If non-nil, callers must call DecRef on the returned *Dirent. func (m *Mount) Root() *Dirent { - m.root.IncRef() + if !m.root.TryIncRef() { + return nil + } return m.root } @@ -269,19 +273,6 @@ func (mns *MountNamespace) DecRef() { mns.DecRefWithDestructor(mns.destroy) } -// Freeze freezes the entire mount tree. -func (mns *MountNamespace) Freeze() { - mns.mu.Lock() - defer mns.mu.Unlock() - - // We only want to freeze Dirents with active references, not Dirents referenced - // by a mount's MountSource. - mns.flushMountSourceRefsLocked() - - // Freeze the entire shebang. - mns.root.Freeze() -} - // withMountLocked prevents further walks to `node`, because `node` is about to // be a mount point. func (mns *MountNamespace) withMountLocked(node *Dirent, fn func() error) error { diff --git a/pkg/sentry/fs/proc/BUILD b/pkg/sentry/fs/proc/BUILD index 280093c5e..77c2c5c0e 100644 --- a/pkg/sentry/fs/proc/BUILD +++ b/pkg/sentry/fs/proc/BUILD @@ -36,6 +36,7 @@ go_library( "//pkg/sentry/fs/proc/device", "//pkg/sentry/fs/proc/seqfile", "//pkg/sentry/fs/ramfs", + "//pkg/sentry/fsbridge", "//pkg/sentry/inet", "//pkg/sentry/kernel", "//pkg/sentry/kernel/auth", diff --git a/pkg/sentry/fs/proc/mounts.go b/pkg/sentry/fs/proc/mounts.go index c10888100..1fc9c703c 100644 --- a/pkg/sentry/fs/proc/mounts.go +++ b/pkg/sentry/fs/proc/mounts.go @@ -60,13 +60,15 @@ func forEachMount(t *kernel.Task, fn func(string, *fs.Mount)) { }) for _, m := range ms { mroot := m.Root() + if mroot == nil { + continue // No longer valid. + } mountPath, desc := mroot.FullName(rootDir) mroot.DecRef() if !desc { // MountSources that are not descendants of the chroot jail are ignored. continue } - fn(mountPath, m) } } @@ -91,6 +93,12 @@ func (mif *mountInfoFile) ReadSeqFileData(ctx context.Context, handle seqfile.Se var buf bytes.Buffer forEachMount(mif.t, func(mountPath string, m *fs.Mount) { + mroot := m.Root() + if mroot == nil { + return // No longer valid. + } + defer mroot.DecRef() + // Format: // 36 35 98:0 /mnt1 /mnt2 rw,noatime master:1 - ext3 /dev/root rw,errors=continue // (1)(2)(3) (4) (5) (6) (7) (8) (9) (10) (11) @@ -107,9 +115,6 @@ func (mif *mountInfoFile) ReadSeqFileData(ctx context.Context, handle seqfile.Se // (3) Major:Minor device ID. We don't have a superblock, so we // just use the root inode device number. - mroot := m.Root() - defer mroot.DecRef() - sa := mroot.Inode.StableAttr fmt.Fprintf(&buf, "%d:%d ", sa.DeviceFileMajor, sa.DeviceFileMinor) @@ -165,7 +170,8 @@ func superBlockOpts(mountPath string, msrc *fs.MountSource) string { // NOTE(b/147673608): If the mount is a cgroup, we also need to include // the cgroup name in the options. For now we just read that from the // path. - // TODO(gvisor.dev/issues/190): Once gVisor has full cgroup support, we + // + // TODO(gvisor.dev/issue/190): Once gVisor has full cgroup support, we // should get this value from the cgroup itself, and not rely on the // path. if msrc.FilesystemType == "cgroup" { @@ -207,6 +213,9 @@ func (mf *mountsFile) ReadSeqFileData(ctx context.Context, handle seqfile.SeqHan // // The "needs dump"and fsck flags are always 0, which is allowed. root := m.Root() + if root == nil { + return // No longer valid. + } defer root.DecRef() flags := root.Inode.MountSource.Flags diff --git a/pkg/sentry/fs/proc/net.go b/pkg/sentry/fs/proc/net.go index 6f2775344..bd18177d4 100644 --- a/pkg/sentry/fs/proc/net.go +++ b/pkg/sentry/fs/proc/net.go @@ -40,44 +40,48 @@ import ( // LINT.IfChange -// newNet creates a new proc net entry. -func (p *proc) newNetDir(ctx context.Context, k *kernel.Kernel, msrc *fs.MountSource) *fs.Inode { +// newNetDir creates a new proc net entry. +func newNetDir(t *kernel.Task, msrc *fs.MountSource) *fs.Inode { + k := t.Kernel() + var contents map[string]*fs.Inode - if s := p.k.NetworkStack(); s != nil { + if s := t.NetworkNamespace().Stack(); s != nil { + // TODO(gvisor.dev/issue/1833): Make sure file contents reflect the task + // network namespace. contents = map[string]*fs.Inode{ - "dev": seqfile.NewSeqFileInode(ctx, &netDev{s: s}, msrc), - "snmp": seqfile.NewSeqFileInode(ctx, &netSnmp{s: s}, msrc), + "dev": seqfile.NewSeqFileInode(t, &netDev{s: s}, msrc), + "snmp": seqfile.NewSeqFileInode(t, &netSnmp{s: s}, msrc), // The following files are simple stubs until they are // implemented in netstack, if the file contains a // header the stub is just the header otherwise it is // an empty file. - "arp": newStaticProcInode(ctx, msrc, []byte("IP address HW type Flags HW address Mask Device\n")), + "arp": newStaticProcInode(t, msrc, []byte("IP address HW type Flags HW address Mask Device\n")), - "netlink": newStaticProcInode(ctx, msrc, []byte("sk Eth Pid Groups Rmem Wmem Dump Locks Drops Inode\n")), - "netstat": newStaticProcInode(ctx, msrc, []byte("TcpExt: SyncookiesSent SyncookiesRecv SyncookiesFailed EmbryonicRsts PruneCalled RcvPruned OfoPruned OutOfWindowIcmps LockDroppedIcmps ArpFilter TW TWRecycled TWKilled PAWSPassive PAWSActive PAWSEstab DelayedACKs DelayedACKLocked DelayedACKLost ListenOverflows ListenDrops TCPPrequeued TCPDirectCopyFromBacklog TCPDirectCopyFromPrequeue TCPPrequeueDropped TCPHPHits TCPHPHitsToUser TCPPureAcks TCPHPAcks TCPRenoRecovery TCPSackRecovery TCPSACKReneging TCPFACKReorder TCPSACKReorder TCPRenoReorder TCPTSReorder TCPFullUndo TCPPartialUndo TCPDSACKUndo TCPLossUndo TCPLostRetransmit TCPRenoFailures TCPSackFailures TCPLossFailures TCPFastRetrans TCPForwardRetrans TCPSlowStartRetrans TCPTimeouts TCPLossProbes TCPLossProbeRecovery TCPRenoRecoveryFail TCPSackRecoveryFail TCPSchedulerFailed TCPRcvCollapsed TCPDSACKOldSent TCPDSACKOfoSent TCPDSACKRecv TCPDSACKOfoRecv TCPAbortOnData TCPAbortOnClose TCPAbortOnMemory TCPAbortOnTimeout TCPAbortOnLinger TCPAbortFailed TCPMemoryPressures TCPSACKDiscard TCPDSACKIgnoredOld TCPDSACKIgnoredNoUndo TCPSpuriousRTOs TCPMD5NotFound TCPMD5Unexpected TCPMD5Failure TCPSackShifted TCPSackMerged TCPSackShiftFallback TCPBacklogDrop TCPMinTTLDrop TCPDeferAcceptDrop IPReversePathFilter TCPTimeWaitOverflow TCPReqQFullDoCookies TCPReqQFullDrop TCPRetransFail TCPRcvCoalesce TCPOFOQueue TCPOFODrop TCPOFOMerge TCPChallengeACK TCPSYNChallenge TCPFastOpenActive TCPFastOpenActiveFail TCPFastOpenPassive TCPFastOpenPassiveFail TCPFastOpenListenOverflow TCPFastOpenCookieReqd TCPSpuriousRtxHostQueues BusyPollRxPackets TCPAutoCorking TCPFromZeroWindowAdv TCPToZeroWindowAdv TCPWantZeroWindowAdv TCPSynRetrans TCPOrigDataSent TCPHystartTrainDetect TCPHystartTrainCwnd TCPHystartDelayDetect TCPHystartDelayCwnd TCPACKSkippedSynRecv TCPACKSkippedPAWS TCPACKSkippedSeq TCPACKSkippedFinWait2 TCPACKSkippedTimeWait TCPACKSkippedChallenge TCPWinProbe TCPKeepAlive TCPMTUPFail TCPMTUPSuccess\n")), - "packet": newStaticProcInode(ctx, msrc, []byte("sk RefCnt Type Proto Iface R Rmem User Inode\n")), - "protocols": newStaticProcInode(ctx, msrc, []byte("protocol size sockets memory press maxhdr slab module cl co di ac io in de sh ss gs se re sp bi br ha uh gp em\n")), + "netlink": newStaticProcInode(t, msrc, []byte("sk Eth Pid Groups Rmem Wmem Dump Locks Drops Inode\n")), + "netstat": newStaticProcInode(t, msrc, []byte("TcpExt: SyncookiesSent SyncookiesRecv SyncookiesFailed EmbryonicRsts PruneCalled RcvPruned OfoPruned OutOfWindowIcmps LockDroppedIcmps ArpFilter TW TWRecycled TWKilled PAWSPassive PAWSActive PAWSEstab DelayedACKs DelayedACKLocked DelayedACKLost ListenOverflows ListenDrops TCPPrequeued TCPDirectCopyFromBacklog TCPDirectCopyFromPrequeue TCPPrequeueDropped TCPHPHits TCPHPHitsToUser TCPPureAcks TCPHPAcks TCPRenoRecovery TCPSackRecovery TCPSACKReneging TCPFACKReorder TCPSACKReorder TCPRenoReorder TCPTSReorder TCPFullUndo TCPPartialUndo TCPDSACKUndo TCPLossUndo TCPLostRetransmit TCPRenoFailures TCPSackFailures TCPLossFailures TCPFastRetrans TCPForwardRetrans TCPSlowStartRetrans TCPTimeouts TCPLossProbes TCPLossProbeRecovery TCPRenoRecoveryFail TCPSackRecoveryFail TCPSchedulerFailed TCPRcvCollapsed TCPDSACKOldSent TCPDSACKOfoSent TCPDSACKRecv TCPDSACKOfoRecv TCPAbortOnData TCPAbortOnClose TCPAbortOnMemory TCPAbortOnTimeout TCPAbortOnLinger TCPAbortFailed TCPMemoryPressures TCPSACKDiscard TCPDSACKIgnoredOld TCPDSACKIgnoredNoUndo TCPSpuriousRTOs TCPMD5NotFound TCPMD5Unexpected TCPMD5Failure TCPSackShifted TCPSackMerged TCPSackShiftFallback TCPBacklogDrop TCPMinTTLDrop TCPDeferAcceptDrop IPReversePathFilter TCPTimeWaitOverflow TCPReqQFullDoCookies TCPReqQFullDrop TCPRetransFail TCPRcvCoalesce TCPOFOQueue TCPOFODrop TCPOFOMerge TCPChallengeACK TCPSYNChallenge TCPFastOpenActive TCPFastOpenActiveFail TCPFastOpenPassive TCPFastOpenPassiveFail TCPFastOpenListenOverflow TCPFastOpenCookieReqd TCPSpuriousRtxHostQueues BusyPollRxPackets TCPAutoCorking TCPFromZeroWindowAdv TCPToZeroWindowAdv TCPWantZeroWindowAdv TCPSynRetrans TCPOrigDataSent TCPHystartTrainDetect TCPHystartTrainCwnd TCPHystartDelayDetect TCPHystartDelayCwnd TCPACKSkippedSynRecv TCPACKSkippedPAWS TCPACKSkippedSeq TCPACKSkippedFinWait2 TCPACKSkippedTimeWait TCPACKSkippedChallenge TCPWinProbe TCPKeepAlive TCPMTUPFail TCPMTUPSuccess\n")), + "packet": newStaticProcInode(t, msrc, []byte("sk RefCnt Type Proto Iface R Rmem User Inode\n")), + "protocols": newStaticProcInode(t, msrc, []byte("protocol size sockets memory press maxhdr slab module cl co di ac io in de sh ss gs se re sp bi br ha uh gp em\n")), // Linux sets psched values to: nsec per usec, psched // tick in ns, 1000000, high res timer ticks per sec // (ClockGetres returns 1ns resolution). - "psched": newStaticProcInode(ctx, msrc, []byte(fmt.Sprintf("%08x %08x %08x %08x\n", uint64(time.Microsecond/time.Nanosecond), 64, 1000000, uint64(time.Second/time.Nanosecond)))), - "ptype": newStaticProcInode(ctx, msrc, []byte("Type Device Function\n")), - "route": seqfile.NewSeqFileInode(ctx, &netRoute{s: s}, msrc), - "tcp": seqfile.NewSeqFileInode(ctx, &netTCP{k: k}, msrc), - "udp": seqfile.NewSeqFileInode(ctx, &netUDP{k: k}, msrc), - "unix": seqfile.NewSeqFileInode(ctx, &netUnix{k: k}, msrc), + "psched": newStaticProcInode(t, msrc, []byte(fmt.Sprintf("%08x %08x %08x %08x\n", uint64(time.Microsecond/time.Nanosecond), 64, 1000000, uint64(time.Second/time.Nanosecond)))), + "ptype": newStaticProcInode(t, msrc, []byte("Type Device Function\n")), + "route": seqfile.NewSeqFileInode(t, &netRoute{s: s}, msrc), + "tcp": seqfile.NewSeqFileInode(t, &netTCP{k: k}, msrc), + "udp": seqfile.NewSeqFileInode(t, &netUDP{k: k}, msrc), + "unix": seqfile.NewSeqFileInode(t, &netUnix{k: k}, msrc), } if s.SupportsIPv6() { - contents["if_inet6"] = seqfile.NewSeqFileInode(ctx, &ifinet6{s: s}, msrc) - contents["ipv6_route"] = newStaticProcInode(ctx, msrc, []byte("")) - contents["tcp6"] = seqfile.NewSeqFileInode(ctx, &netTCP6{k: k}, msrc) - contents["udp6"] = newStaticProcInode(ctx, msrc, []byte(" sl local_address remote_address st tx_queue rx_queue tr tm->when retrnsmt uid timeout inode\n")) + contents["if_inet6"] = seqfile.NewSeqFileInode(t, &ifinet6{s: s}, msrc) + contents["ipv6_route"] = newStaticProcInode(t, msrc, []byte("")) + contents["tcp6"] = seqfile.NewSeqFileInode(t, &netTCP6{k: k}, msrc) + contents["udp6"] = newStaticProcInode(t, msrc, []byte(" sl local_address remote_address st tx_queue rx_queue tr tm->when retrnsmt uid timeout inode\n")) } } - d := ramfs.NewDir(ctx, contents, fs.RootOwner, fs.FilePermsFromMode(0555)) - return newProcInode(ctx, d, msrc, fs.SpecialDirectory, nil) + d := ramfs.NewDir(t, contents, fs.RootOwner, fs.FilePermsFromMode(0555)) + return newProcInode(t, d, msrc, fs.SpecialDirectory, t) } // ifinet6 implements seqfile.SeqSource for /proc/net/if_inet6. @@ -834,4 +838,4 @@ func (n *netUDP) ReadSeqFileData(ctx context.Context, h seqfile.SeqHandle) ([]se return data, 0 } -// LINT.ThenChange(../../fsimpl/proc/tasks_net.go) +// LINT.ThenChange(../../fsimpl/proc/task_net.go) diff --git a/pkg/sentry/fs/proc/proc.go b/pkg/sentry/fs/proc/proc.go index c8abb5052..c659224a7 100644 --- a/pkg/sentry/fs/proc/proc.go +++ b/pkg/sentry/fs/proc/proc.go @@ -70,6 +70,7 @@ func New(ctx context.Context, msrc *fs.MountSource, cgroupControllers map[string "loadavg": seqfile.NewSeqFileInode(ctx, &loadavgData{}, msrc), "meminfo": seqfile.NewSeqFileInode(ctx, &meminfoData{k}, msrc), "mounts": newProcInode(ctx, ramfs.NewSymlink(ctx, fs.RootOwner, "self/mounts"), msrc, fs.Symlink, nil), + "net": newProcInode(ctx, ramfs.NewSymlink(ctx, fs.RootOwner, "self/net"), msrc, fs.Symlink, nil), "self": newSelf(ctx, pidns, msrc), "stat": seqfile.NewSeqFileInode(ctx, &statData{k}, msrc), "thread-self": newThreadSelf(ctx, pidns, msrc), @@ -86,7 +87,6 @@ func New(ctx context.Context, msrc *fs.MountSource, cgroupControllers map[string } // Add more contents that need proc to be initialized. - p.AddChild(ctx, "net", p.newNetDir(ctx, k, msrc)) p.AddChild(ctx, "sys", p.newSysDir(ctx, msrc)) return newProcInode(ctx, p, msrc, fs.SpecialDirectory, nil), nil diff --git a/pkg/sentry/fs/proc/sys_net.go b/pkg/sentry/fs/proc/sys_net.go index 0772d4ae4..702fdd392 100644 --- a/pkg/sentry/fs/proc/sys_net.go +++ b/pkg/sentry/fs/proc/sys_net.go @@ -80,7 +80,7 @@ func newTCPMemInode(ctx context.Context, msrc *fs.MountSource, s inet.Stack, dir } // Truncate implements fs.InodeOperations.Truncate. -func (tcpMemInode) Truncate(context.Context, *fs.Inode, int64) error { +func (*tcpMemInode) Truncate(context.Context, *fs.Inode, int64) error { return nil } @@ -196,7 +196,7 @@ func newTCPSackInode(ctx context.Context, msrc *fs.MountSource, s inet.Stack) *f } // Truncate implements fs.InodeOperations.Truncate. -func (tcpSack) Truncate(context.Context, *fs.Inode, int64) error { +func (*tcpSack) Truncate(context.Context, *fs.Inode, int64) error { return nil } @@ -357,7 +357,9 @@ func (p *proc) newSysNetIPv4Dir(ctx context.Context, msrc *fs.MountSource, s ine func (p *proc) newSysNetDir(ctx context.Context, msrc *fs.MountSource) *fs.Inode { var contents map[string]*fs.Inode - if s := p.k.NetworkStack(); s != nil { + // TODO(gvisor.dev/issue/1833): Support for using the network stack in the + // network namespace of the calling process. + if s := p.k.RootNetworkNamespace().Stack(); s != nil { contents = map[string]*fs.Inode{ "ipv4": p.newSysNetIPv4Dir(ctx, msrc, s), "core": p.newSysNetCore(ctx, msrc, s), diff --git a/pkg/sentry/fs/proc/task.go b/pkg/sentry/fs/proc/task.go index ca020e11e..4d42eac83 100644 --- a/pkg/sentry/fs/proc/task.go +++ b/pkg/sentry/fs/proc/task.go @@ -28,6 +28,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/fs/proc/device" "gvisor.dev/gvisor/pkg/sentry/fs/proc/seqfile" "gvisor.dev/gvisor/pkg/sentry/fs/ramfs" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/limits" "gvisor.dev/gvisor/pkg/sentry/mm" @@ -56,6 +57,16 @@ func getTaskMM(t *kernel.Task) (*mm.MemoryManager, error) { return m, nil } +func checkTaskState(t *kernel.Task) error { + switch t.ExitState() { + case kernel.TaskExitZombie: + return syserror.EACCES + case kernel.TaskExitDead: + return syserror.ESRCH + } + return nil +} + // taskDir represents a task-level directory. // // +stateify savable @@ -71,24 +82,27 @@ var _ fs.InodeOperations = (*taskDir)(nil) // newTaskDir creates a new proc task entry. func (p *proc) newTaskDir(t *kernel.Task, msrc *fs.MountSource, isThreadGroup bool) *fs.Inode { contents := map[string]*fs.Inode{ - "auxv": newAuxvec(t, msrc), - "cmdline": newExecArgInode(t, msrc, cmdlineExecArg), - "comm": newComm(t, msrc), - "environ": newExecArgInode(t, msrc, environExecArg), - "exe": newExe(t, msrc), - "fd": newFdDir(t, msrc), - "fdinfo": newFdInfoDir(t, msrc), - "gid_map": newGIDMap(t, msrc), - "io": newIO(t, msrc, isThreadGroup), - "maps": newMaps(t, msrc), - "mountinfo": seqfile.NewSeqFileInode(t, &mountInfoFile{t: t}, msrc), - "mounts": seqfile.NewSeqFileInode(t, &mountsFile{t: t}, msrc), - "ns": newNamespaceDir(t, msrc), - "smaps": newSmaps(t, msrc), - "stat": newTaskStat(t, msrc, isThreadGroup, p.pidns), - "statm": newStatm(t, msrc), - "status": newStatus(t, msrc, p.pidns), - "uid_map": newUIDMap(t, msrc), + "auxv": newAuxvec(t, msrc), + "cmdline": newExecArgInode(t, msrc, cmdlineExecArg), + "comm": newComm(t, msrc), + "environ": newExecArgInode(t, msrc, environExecArg), + "exe": newExe(t, msrc), + "fd": newFdDir(t, msrc), + "fdinfo": newFdInfoDir(t, msrc), + "gid_map": newGIDMap(t, msrc), + "io": newIO(t, msrc, isThreadGroup), + "maps": newMaps(t, msrc), + "mountinfo": seqfile.NewSeqFileInode(t, &mountInfoFile{t: t}, msrc), + "mounts": seqfile.NewSeqFileInode(t, &mountsFile{t: t}, msrc), + "net": newNetDir(t, msrc), + "ns": newNamespaceDir(t, msrc), + "oom_score": newOOMScore(t, msrc), + "oom_score_adj": newOOMScoreAdj(t, msrc), + "smaps": newSmaps(t, msrc), + "stat": newTaskStat(t, msrc, isThreadGroup, p.pidns), + "statm": newStatm(t, msrc), + "status": newStatus(t, msrc, p.pidns), + "uid_map": newUIDMap(t, msrc), } if isThreadGroup { contents["task"] = p.newSubtasks(t, msrc) @@ -249,12 +263,13 @@ func newExe(t *kernel.Task, msrc *fs.MountSource) *fs.Inode { return newProcInode(t, exeSymlink, msrc, fs.Symlink, t) } -func (e *exe) executable() (d *fs.Dirent, err error) { +func (e *exe) executable() (file fsbridge.File, err error) { + if err := checkTaskState(e.t); err != nil { + return nil, err + } e.t.WithMuLocked(func(t *kernel.Task) { mm := t.MemoryManager() if mm == nil { - // TODO(b/34851096): Check shouldn't allow Readlink once the - // Task is zombied. err = syserror.EACCES return } @@ -262,9 +277,9 @@ func (e *exe) executable() (d *fs.Dirent, err error) { // The MemoryManager may be destroyed, in which case // MemoryManager.destroy will simply set the executable to nil // (with locks held). - d = mm.Executable() - if d == nil { - err = syserror.ENOENT + file = mm.Executable() + if file == nil { + err = syserror.ESRCH } }) return @@ -283,15 +298,7 @@ func (e *exe) Readlink(ctx context.Context, inode *fs.Inode) (string, error) { } defer exec.DecRef() - root := fs.RootFromContext(ctx) - if root == nil { - // This doesn't correspond to anything in Linux because the vfs is - // global there. - return "", syserror.EINVAL - } - defer root.DecRef() - n, _ := exec.FullName(root) - return n, nil + return exec.PathnameWithDeleted(ctx), nil } // namespaceSymlink represents a symlink in the namespacefs, such as the files @@ -317,11 +324,22 @@ func newNamespaceSymlink(t *kernel.Task, msrc *fs.MountSource, name string) *fs. return newProcInode(t, n, msrc, fs.Symlink, t) } +// Readlink reads the symlink value. +func (n *namespaceSymlink) Readlink(ctx context.Context, inode *fs.Inode) (string, error) { + if err := checkTaskState(n.t); err != nil { + return "", err + } + return n.Symlink.Readlink(ctx, inode) +} + // Getlink implements fs.InodeOperations.Getlink. func (n *namespaceSymlink) Getlink(ctx context.Context, inode *fs.Inode) (*fs.Dirent, error) { if !kernel.ContextCanTrace(ctx, n.t, false) { return nil, syserror.EACCES } + if err := checkTaskState(n.t); err != nil { + return nil, err + } // Create a new regular file to fake the namespace file. iops := fsutil.NewNoReadWriteFileInode(ctx, fs.RootOwner, fs.FilePermsFromMode(0777), linux.PROC_SUPER_MAGIC) @@ -803,4 +821,95 @@ func (f *auxvecFile) Read(ctx context.Context, _ *fs.File, dst usermem.IOSequenc return int64(n), err } +// newOOMScore returns a oom_score file. It is a stub that always returns 0. +// TODO(gvisor.dev/issue/1967) +func newOOMScore(t *kernel.Task, msrc *fs.MountSource) *fs.Inode { + return newStaticProcInode(t, msrc, []byte("0\n")) +} + +// oomScoreAdj is a file containing the oom_score adjustment for a task. +// +// +stateify savable +type oomScoreAdj struct { + fsutil.SimpleFileInode + + t *kernel.Task +} + +// +stateify savable +type oomScoreAdjFile struct { + fsutil.FileGenericSeek `state:"nosave"` + fsutil.FileNoIoctl `state:"nosave"` + fsutil.FileNoMMap `state:"nosave"` + fsutil.FileNoSplice `state:"nosave"` + fsutil.FileNoopFlush `state:"nosave"` + fsutil.FileNoopFsync `state:"nosave"` + fsutil.FileNoopRelease `state:"nosave"` + fsutil.FileNotDirReaddir `state:"nosave"` + fsutil.FileUseInodeUnstableAttr `state:"nosave"` + waiter.AlwaysReady `state:"nosave"` + + t *kernel.Task +} + +// newOOMScoreAdj returns a oom_score_adj file. +func newOOMScoreAdj(t *kernel.Task, msrc *fs.MountSource) *fs.Inode { + i := &oomScoreAdj{ + SimpleFileInode: *fsutil.NewSimpleFileInode(t, fs.RootOwner, fs.FilePermsFromMode(0644), linux.PROC_SUPER_MAGIC), + t: t, + } + return newProcInode(t, i, msrc, fs.SpecialFile, t) +} + +// Truncate implements fs.InodeOperations.Truncate. Truncate is called when +// O_TRUNC is specified for any kind of existing Dirent but is not called via +// (f)truncate for proc files. +func (*oomScoreAdj) Truncate(context.Context, *fs.Inode, int64) error { + return nil +} + +// GetFile implements fs.InodeOperations.GetFile. +func (o *oomScoreAdj) GetFile(ctx context.Context, dirent *fs.Dirent, flags fs.FileFlags) (*fs.File, error) { + return fs.NewFile(ctx, dirent, flags, &oomScoreAdjFile{t: o.t}), nil +} + +// Read implements fs.FileOperations.Read. +func (f *oomScoreAdjFile) Read(ctx context.Context, _ *fs.File, dst usermem.IOSequence, offset int64) (int64, error) { + if f.t.ExitState() == kernel.TaskExitDead { + return 0, syserror.ESRCH + } + var buf bytes.Buffer + fmt.Fprintf(&buf, "%d\n", f.t.OOMScoreAdj()) + if offset >= int64(buf.Len()) { + return 0, io.EOF + } + n, err := dst.CopyOut(ctx, buf.Bytes()[offset:]) + return int64(n), err +} + +// Write implements fs.FileOperations.Write. +func (f *oomScoreAdjFile) Write(ctx context.Context, _ *fs.File, src usermem.IOSequence, offset int64) (int64, error) { + if src.NumBytes() == 0 { + return 0, nil + } + + // Limit input size so as not to impact performance if input size is large. + src = src.TakeFirst(usermem.PageSize - 1) + + var v int32 + n, err := usermem.CopyInt32StringInVec(ctx, src.IO, src.Addrs, &v, src.Opts) + if err != nil { + return 0, err + } + + if f.t.ExitState() == kernel.TaskExitDead { + return 0, syserror.ESRCH + } + if err := f.t.SetOOMScoreAdj(v); err != nil { + return 0, err + } + + return n, nil +} + // LINT.ThenChange(../../fsimpl/proc/task.go|../../fsimpl/proc/task_files.go) diff --git a/pkg/sentry/fs/tmpfs/fs.go b/pkg/sentry/fs/tmpfs/fs.go index d5be56c3f..bc117ca6a 100644 --- a/pkg/sentry/fs/tmpfs/fs.go +++ b/pkg/sentry/fs/tmpfs/fs.go @@ -44,9 +44,6 @@ const ( // lookup. cacheRevalidate = "revalidate" - // TODO(edahlgren/mpratt): support a tmpfs size limit. - // size = "size" - // Permissions that exceed modeMask will be rejected. modeMask = 01777 diff --git a/pkg/sentry/fs/tmpfs/inode_file.go b/pkg/sentry/fs/tmpfs/inode_file.go index 25abbc151..1dc75291d 100644 --- a/pkg/sentry/fs/tmpfs/inode_file.go +++ b/pkg/sentry/fs/tmpfs/inode_file.go @@ -39,7 +39,7 @@ var ( opensRO = metric.MustCreateNewUint64Metric("/in_memory_file/opens_ro", false /* sync */, "Number of times an in-memory file was opened in read-only mode.") opensW = metric.MustCreateNewUint64Metric("/in_memory_file/opens_w", false /* sync */, "Number of times an in-memory file was opened in write mode.") reads = metric.MustCreateNewUint64Metric("/in_memory_file/reads", false /* sync */, "Number of in-memory file reads.") - readWait = metric.MustCreateNewUint64Metric("/in_memory_file/read_wait", false /* sync */, "Time waiting on in-memory file reads, in nanoseconds.") + readWait = metric.MustCreateNewUint64NanosecondsMetric("/in_memory_file/read_wait", false /* sync */, "Time waiting on in-memory file reads, in nanoseconds.") ) // fileInodeOperations implements fs.InodeOperations for a regular tmpfs file. diff --git a/pkg/sentry/fs/tty/slave.go b/pkg/sentry/fs/tty/slave.go index db55cdc48..6a2dbc576 100644 --- a/pkg/sentry/fs/tty/slave.go +++ b/pkg/sentry/fs/tty/slave.go @@ -73,7 +73,7 @@ func (si *slaveInodeOperations) Release(ctx context.Context) { } // Truncate implements fs.InodeOperations.Truncate. -func (slaveInodeOperations) Truncate(context.Context, *fs.Inode, int64) error { +func (*slaveInodeOperations) Truncate(context.Context, *fs.Inode, int64) error { return nil } diff --git a/pkg/sentry/fsbridge/BUILD b/pkg/sentry/fsbridge/BUILD new file mode 100644 index 000000000..6c798f0bd --- /dev/null +++ b/pkg/sentry/fsbridge/BUILD @@ -0,0 +1,24 @@ +load("//tools:defs.bzl", "go_library") + +licenses(["notice"]) + +go_library( + name = "fsbridge", + srcs = [ + "bridge.go", + "fs.go", + "vfs.go", + ], + visibility = ["//pkg/sentry:internal"], + deps = [ + "//pkg/abi/linux", + "//pkg/context", + "//pkg/fspath", + "//pkg/sentry/fs", + "//pkg/sentry/kernel/auth", + "//pkg/sentry/memmap", + "//pkg/sentry/vfs", + "//pkg/syserror", + "//pkg/usermem", + ], +) diff --git a/pkg/sentry/fsbridge/bridge.go b/pkg/sentry/fsbridge/bridge.go new file mode 100644 index 000000000..8e7590721 --- /dev/null +++ b/pkg/sentry/fsbridge/bridge.go @@ -0,0 +1,54 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// Package fsbridge provides common interfaces to bridge between VFS1 and VFS2 +// files. +package fsbridge + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/memmap" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/usermem" +) + +// File provides a common interface to bridge between VFS1 and VFS2 files. +type File interface { + // PathnameWithDeleted returns an absolute pathname to vd, consistent with + // Linux's d_path(). In particular, if vd.Dentry() has been disowned, + // PathnameWithDeleted appends " (deleted)" to the returned pathname. + PathnameWithDeleted(ctx context.Context) string + + // ReadFull read all contents from the file. + ReadFull(ctx context.Context, dst usermem.IOSequence, offset int64) (int64, error) + + // ConfigureMMap mutates opts to implement mmap(2) for the file. + ConfigureMMap(context.Context, *memmap.MMapOpts) error + + // Type returns the file type, e.g. linux.S_IFREG. + Type(context.Context) (linux.FileMode, error) + + // IncRef increments reference. + IncRef() + + // DecRef decrements reference. + DecRef() +} + +// Lookup provides a common interface to open files. +type Lookup interface { + // OpenPath opens a file. + OpenPath(ctx context.Context, path string, opts vfs.OpenOptions, remainingTraversals *uint, resolveFinal bool) (File, error) +} diff --git a/pkg/sentry/fsbridge/fs.go b/pkg/sentry/fsbridge/fs.go new file mode 100644 index 000000000..093ce1fb3 --- /dev/null +++ b/pkg/sentry/fsbridge/fs.go @@ -0,0 +1,181 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package fsbridge + +import ( + "io" + "strings" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/memmap" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +// fsFile implements File interface over fs.File. +// +// +stateify savable +type fsFile struct { + file *fs.File +} + +var _ File = (*fsFile)(nil) + +// NewFSFile creates a new File over fs.File. +func NewFSFile(file *fs.File) File { + return &fsFile{file: file} +} + +// PathnameWithDeleted implements File. +func (f *fsFile) PathnameWithDeleted(ctx context.Context) string { + root := fs.RootFromContext(ctx) + if root == nil { + // This doesn't correspond to anything in Linux because the vfs is + // global there. + return "" + } + defer root.DecRef() + + name, _ := f.file.Dirent.FullName(root) + return name +} + +// ReadFull implements File. +func (f *fsFile) ReadFull(ctx context.Context, dst usermem.IOSequence, offset int64) (int64, error) { + var total int64 + for dst.NumBytes() > 0 { + n, err := f.file.Preadv(ctx, dst, offset+total) + total += n + if err == io.EOF && total != 0 { + return total, io.ErrUnexpectedEOF + } else if err != nil { + return total, err + } + dst = dst.DropFirst64(n) + } + return total, nil +} + +// ConfigureMMap implements File. +func (f *fsFile) ConfigureMMap(ctx context.Context, opts *memmap.MMapOpts) error { + return f.file.ConfigureMMap(ctx, opts) +} + +// Type implements File. +func (f *fsFile) Type(context.Context) (linux.FileMode, error) { + return linux.FileMode(f.file.Dirent.Inode.StableAttr.Type.LinuxType()), nil +} + +// IncRef implements File. +func (f *fsFile) IncRef() { + f.file.IncRef() +} + +// DecRef implements File. +func (f *fsFile) DecRef() { + f.file.DecRef() +} + +// fsLookup implements Lookup interface using fs.File. +// +// +stateify savable +type fsLookup struct { + mntns *fs.MountNamespace + + root *fs.Dirent + workingDir *fs.Dirent +} + +var _ Lookup = (*fsLookup)(nil) + +// NewFSLookup creates a new Lookup using VFS1. +func NewFSLookup(mntns *fs.MountNamespace, root, workingDir *fs.Dirent) Lookup { + return &fsLookup{ + mntns: mntns, + root: root, + workingDir: workingDir, + } +} + +// OpenPath implements Lookup. +func (l *fsLookup) OpenPath(ctx context.Context, path string, opts vfs.OpenOptions, remainingTraversals *uint, resolveFinal bool) (File, error) { + var d *fs.Dirent + var err error + if resolveFinal { + d, err = l.mntns.FindInode(ctx, l.root, l.workingDir, path, remainingTraversals) + } else { + d, err = l.mntns.FindLink(ctx, l.root, l.workingDir, path, remainingTraversals) + } + if err != nil { + return nil, err + } + defer d.DecRef() + + if !resolveFinal && fs.IsSymlink(d.Inode.StableAttr) { + return nil, syserror.ELOOP + } + + fsPerm := openOptionsToPermMask(&opts) + if err := d.Inode.CheckPermission(ctx, fsPerm); err != nil { + return nil, err + } + + // If they claim it's a directory, then make sure. + if strings.HasSuffix(path, "/") { + if d.Inode.StableAttr.Type != fs.Directory { + return nil, syserror.ENOTDIR + } + } + + if opts.FileExec && d.Inode.StableAttr.Type != fs.RegularFile { + ctx.Infof("%q is not a regular file: %v", path, d.Inode.StableAttr.Type) + return nil, syserror.EACCES + } + + f, err := d.Inode.GetFile(ctx, d, flagsToFileFlags(opts.Flags)) + if err != nil { + return nil, err + } + + return &fsFile{file: f}, nil +} + +func openOptionsToPermMask(opts *vfs.OpenOptions) fs.PermMask { + mode := opts.Flags & linux.O_ACCMODE + return fs.PermMask{ + Read: mode == linux.O_RDONLY || mode == linux.O_RDWR, + Write: mode == linux.O_WRONLY || mode == linux.O_RDWR, + Execute: opts.FileExec, + } +} + +func flagsToFileFlags(flags uint32) fs.FileFlags { + return fs.FileFlags{ + Direct: flags&linux.O_DIRECT != 0, + DSync: flags&(linux.O_DSYNC|linux.O_SYNC) != 0, + Sync: flags&linux.O_SYNC != 0, + NonBlocking: flags&linux.O_NONBLOCK != 0, + Read: (flags & linux.O_ACCMODE) != linux.O_WRONLY, + Write: (flags & linux.O_ACCMODE) != linux.O_RDONLY, + Append: flags&linux.O_APPEND != 0, + Directory: flags&linux.O_DIRECTORY != 0, + Async: flags&linux.O_ASYNC != 0, + LargeFile: flags&linux.O_LARGEFILE != 0, + Truncate: flags&linux.O_TRUNC != 0, + } +} diff --git a/pkg/sentry/fsbridge/vfs.go b/pkg/sentry/fsbridge/vfs.go new file mode 100644 index 000000000..89168220a --- /dev/null +++ b/pkg/sentry/fsbridge/vfs.go @@ -0,0 +1,142 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package fsbridge + +import ( + "io" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/memmap" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/usermem" +) + +// VFSFile implements File interface over vfs.FileDescription. +// +// +stateify savable +type VFSFile struct { + file *vfs.FileDescription +} + +var _ File = (*VFSFile)(nil) + +// NewVFSFile creates a new File over fs.File. +func NewVFSFile(file *vfs.FileDescription) File { + return &VFSFile{file: file} +} + +// PathnameWithDeleted implements File. +func (f *VFSFile) PathnameWithDeleted(ctx context.Context) string { + root := vfs.RootFromContext(ctx) + defer root.DecRef() + + vfsObj := f.file.VirtualDentry().Mount().Filesystem().VirtualFilesystem() + name, _ := vfsObj.PathnameWithDeleted(ctx, root, f.file.VirtualDentry()) + return name +} + +// ReadFull implements File. +func (f *VFSFile) ReadFull(ctx context.Context, dst usermem.IOSequence, offset int64) (int64, error) { + var total int64 + for dst.NumBytes() > 0 { + n, err := f.file.PRead(ctx, dst, offset+total, vfs.ReadOptions{}) + total += n + if err == io.EOF && total != 0 { + return total, io.ErrUnexpectedEOF + } else if err != nil { + return total, err + } + dst = dst.DropFirst64(n) + } + return total, nil +} + +// ConfigureMMap implements File. +func (f *VFSFile) ConfigureMMap(ctx context.Context, opts *memmap.MMapOpts) error { + return f.file.ConfigureMMap(ctx, opts) +} + +// Type implements File. +func (f *VFSFile) Type(ctx context.Context) (linux.FileMode, error) { + stat, err := f.file.Stat(ctx, vfs.StatOptions{}) + if err != nil { + return 0, err + } + return linux.FileMode(stat.Mode).FileType(), nil +} + +// IncRef implements File. +func (f *VFSFile) IncRef() { + f.file.IncRef() +} + +// DecRef implements File. +func (f *VFSFile) DecRef() { + f.file.DecRef() +} + +// FileDescription returns the FileDescription represented by f. It does not +// take an additional reference on the returned FileDescription. +func (f *VFSFile) FileDescription() *vfs.FileDescription { + return f.file +} + +// fsLookup implements Lookup interface using fs.File. +// +// +stateify savable +type vfsLookup struct { + mntns *vfs.MountNamespace + + root vfs.VirtualDentry + workingDir vfs.VirtualDentry +} + +var _ Lookup = (*vfsLookup)(nil) + +// NewVFSLookup creates a new Lookup using VFS2. +func NewVFSLookup(mntns *vfs.MountNamespace, root, workingDir vfs.VirtualDentry) Lookup { + return &vfsLookup{ + mntns: mntns, + root: root, + workingDir: workingDir, + } +} + +// OpenPath implements Lookup. +// +// remainingTraversals is not configurable in VFS2, all callers are using the +// default anyways. +func (l *vfsLookup) OpenPath(ctx context.Context, pathname string, opts vfs.OpenOptions, _ *uint, resolveFinal bool) (File, error) { + vfsObj := l.mntns.Root().Mount().Filesystem().VirtualFilesystem() + creds := auth.CredentialsFromContext(ctx) + path := fspath.Parse(pathname) + pop := &vfs.PathOperation{ + Root: l.root, + Start: l.workingDir, + Path: path, + FollowFinalSymlink: resolveFinal, + } + if path.Absolute { + pop.Start = l.root + } + fd, err := vfsObj.OpenAt(ctx, creds, pop, &opts) + if err != nil { + return nil, err + } + return &VFSFile{file: fd}, nil +} diff --git a/pkg/sentry/fsimpl/devtmpfs/devtmpfs.go b/pkg/sentry/fsimpl/devtmpfs/devtmpfs.go index e03a0c665..64f1b142c 100644 --- a/pkg/sentry/fsimpl/devtmpfs/devtmpfs.go +++ b/pkg/sentry/fsimpl/devtmpfs/devtmpfs.go @@ -28,6 +28,9 @@ import ( "gvisor.dev/gvisor/pkg/sync" ) +// Name is the default filesystem name. +const Name = "devtmpfs" + // FilesystemType implements vfs.FilesystemType. type FilesystemType struct { initOnce sync.Once @@ -39,6 +42,11 @@ type FilesystemType struct { root *vfs.Dentry } +// Name implements vfs.FilesystemType.Name. +func (*FilesystemType) Name() string { + return Name +} + // GetFilesystem implements vfs.FilesystemType.GetFilesystem. func (fst *FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { fst.initOnce.Do(func() { @@ -107,6 +115,7 @@ func (a *Accessor) wrapContext(ctx context.Context) *accessorContext { func (ac *accessorContext) Value(key interface{}) interface{} { switch key { case vfs.CtxMountNamespace: + ac.a.mntns.IncRef() return ac.a.mntns case vfs.CtxRoot: ac.a.root.IncRef() diff --git a/pkg/sentry/fsimpl/devtmpfs/devtmpfs_test.go b/pkg/sentry/fsimpl/devtmpfs/devtmpfs_test.go index 73308a2b5..b6d52c015 100644 --- a/pkg/sentry/fsimpl/devtmpfs/devtmpfs_test.go +++ b/pkg/sentry/fsimpl/devtmpfs/devtmpfs_test.go @@ -29,7 +29,10 @@ func TestDevtmpfs(t *testing.T) { ctx := contexttest.Context(t) creds := auth.CredentialsFromContext(ctx) - vfsObj := vfs.New() + vfsObj := &vfs.VirtualFilesystem{} + if err := vfsObj.Init(); err != nil { + t.Fatalf("VFS init: %v", err) + } // Register tmpfs just so that we can have a root filesystem that isn't // devtmpfs. vfsObj.MustRegisterFilesystemType("tmpfs", tmpfs.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ diff --git a/pkg/sentry/fsimpl/ext/BUILD b/pkg/sentry/fsimpl/ext/BUILD index 6f78f478f..a4947c480 100644 --- a/pkg/sentry/fsimpl/ext/BUILD +++ b/pkg/sentry/fsimpl/ext/BUILD @@ -15,6 +15,17 @@ go_template_instance( }, ) +go_template_instance( + name = "fstree", + out = "fstree.go", + package = "ext", + prefix = "generic", + template = "//pkg/sentry/vfs/genericfstree:generic_fstree", + types = { + "Dentry": "dentry", + }, +) + go_library( name = "ext", srcs = [ @@ -26,6 +37,7 @@ go_library( "extent_file.go", "file_description.go", "filesystem.go", + "fstree.go", "inode.go", "regular_file.go", "symlink.go", @@ -45,6 +57,7 @@ go_library( "//pkg/sentry/fsimpl/ext/disklayout", "//pkg/sentry/kernel/auth", "//pkg/sentry/memmap", + "//pkg/sentry/socket/unix/transport", "//pkg/sentry/syscalls/linux", "//pkg/sentry/vfs", "//pkg/sync", diff --git a/pkg/sentry/fsimpl/ext/benchmark/benchmark_test.go b/pkg/sentry/fsimpl/ext/benchmark/benchmark_test.go index 2015a8871..89caee3df 100644 --- a/pkg/sentry/fsimpl/ext/benchmark/benchmark_test.go +++ b/pkg/sentry/fsimpl/ext/benchmark/benchmark_test.go @@ -52,7 +52,10 @@ func setUp(b *testing.B, imagePath string) (context.Context, *vfs.VirtualFilesys creds := auth.CredentialsFromContext(ctx) // Create VFS. - vfsObj := vfs.New() + vfsObj := &vfs.VirtualFilesystem{} + if err := vfsObj.Init(); err != nil { + return nil, nil, nil, nil, err + } vfsObj.MustRegisterFilesystemType("extfs", ext.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ AllowUserMount: true, }) diff --git a/pkg/sentry/fsimpl/ext/dentry.go b/pkg/sentry/fsimpl/ext/dentry.go index a080cb189..bfbd7c3d4 100644 --- a/pkg/sentry/fsimpl/ext/dentry.go +++ b/pkg/sentry/fsimpl/ext/dentry.go @@ -22,6 +22,10 @@ import ( type dentry struct { vfsd vfs.Dentry + // Protected by filesystem.mu. + parent *dentry + name string + // inode is the inode represented by this dentry. Multiple Dentries may // share a single non-directory Inode (with hard links). inode is // immutable. diff --git a/pkg/sentry/fsimpl/ext/directory.go b/pkg/sentry/fsimpl/ext/directory.go index ebb72b75e..12b875c8f 100644 --- a/pkg/sentry/fsimpl/ext/directory.go +++ b/pkg/sentry/fsimpl/ext/directory.go @@ -21,7 +21,6 @@ import ( "gvisor.dev/gvisor/pkg/log" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fsimpl/ext/disklayout" - "gvisor.dev/gvisor/pkg/sentry/memmap" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/syserror" @@ -31,6 +30,10 @@ import ( type directory struct { inode inode + // childCache maps filenames to dentries for children for which dentries + // have been instantiated. childCache is protected by filesystem.mu. + childCache map[string]*dentry + // mu serializes the changes to childList. // Lock Order (outermost locks must be taken first): // directory.mu @@ -50,9 +53,13 @@ type directory struct { childMap map[string]*dirent } -// newDirectroy is the directory constructor. -func newDirectroy(inode inode, newDirent bool) (*directory, error) { - file := &directory{inode: inode, childMap: make(map[string]*dirent)} +// newDirectory is the directory constructor. +func newDirectory(inode inode, newDirent bool) (*directory, error) { + file := &directory{ + inode: inode, + childCache: make(map[string]*dentry), + childMap: make(map[string]*dirent), + } file.inode.impl = file // Initialize childList by reading dirents from the underlying file. @@ -188,14 +195,14 @@ func (fd *directoryFD) IterDirents(ctx context.Context, cb vfs.IterDirentsCallba childType = fs.ToInodeType(childInode.diskInode.Mode().FileType()) } - if !cb.Handle(vfs.Dirent{ + if err := cb.Handle(vfs.Dirent{ Name: child.diskDirent.FileName(), Type: fs.ToDirentType(childType), Ino: uint64(child.diskDirent.Inode()), NextOff: fd.off + 1, - }) { + }); err != nil { dir.childList.InsertBefore(child, fd.iter) - return nil + return err } fd.off++ } @@ -299,9 +306,3 @@ func (fd *directoryFD) Seek(ctx context.Context, offset int64, whence int32) (in fd.off = offset return offset, nil } - -// ConfigureMMap implements vfs.FileDescriptionImpl.ConfigureMMap. -func (fd *directoryFD) ConfigureMMap(ctx context.Context, opts *memmap.MMapOpts) error { - // mmap(2) specifies that EACCESS should be returned for non-regular file fds. - return syserror.EACCES -} diff --git a/pkg/sentry/fsimpl/ext/ext.go b/pkg/sentry/fsimpl/ext/ext.go index 373d23b74..7176af6d1 100644 --- a/pkg/sentry/fsimpl/ext/ext.go +++ b/pkg/sentry/fsimpl/ext/ext.go @@ -30,6 +30,9 @@ import ( "gvisor.dev/gvisor/pkg/syserror" ) +// Name is the name of this filesystem. +const Name = "ext" + // FilesystemType implements vfs.FilesystemType. type FilesystemType struct{} @@ -91,8 +94,13 @@ func isCompatible(sb disklayout.SuperBlock) bool { return true } +// Name implements vfs.FilesystemType.Name. +func (FilesystemType) Name() string { + return Name +} + // GetFilesystem implements vfs.FilesystemType.GetFilesystem. -func (FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { +func (fsType FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { // TODO(b/134676337): Ensure that the user is mounting readonly. If not, // EACCESS should be returned according to mount(2). Filesystem independent // flags (like readonly) are currently not available in pkg/sentry/vfs. @@ -103,7 +111,7 @@ func (FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFile } fs := filesystem{dev: dev, inodeCache: make(map[uint32]*inode)} - fs.vfsfs.Init(vfsObj, &fs) + fs.vfsfs.Init(vfsObj, &fsType, &fs) fs.sb, err = readSuperBlock(dev) if err != nil { return nil, nil, err diff --git a/pkg/sentry/fsimpl/ext/ext_test.go b/pkg/sentry/fsimpl/ext/ext_test.go index 05f992826..29bb73765 100644 --- a/pkg/sentry/fsimpl/ext/ext_test.go +++ b/pkg/sentry/fsimpl/ext/ext_test.go @@ -65,7 +65,10 @@ func setUp(t *testing.T, imagePath string) (context.Context, *vfs.VirtualFilesys creds := auth.CredentialsFromContext(ctx) // Create VFS. - vfsObj := vfs.New() + vfsObj := &vfs.VirtualFilesystem{} + if err := vfsObj.Init(); err != nil { + t.Fatalf("VFS init: %v", err) + } vfsObj.MustRegisterFilesystemType("extfs", FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ AllowUserMount: true, }) @@ -496,9 +499,9 @@ func newIterDirentCb() *iterDirentsCb { } // Handle implements vfs.IterDirentsCallback.Handle. -func (cb *iterDirentsCb) Handle(dirent vfs.Dirent) bool { +func (cb *iterDirentsCb) Handle(dirent vfs.Dirent) error { cb.dirents = append(cb.dirents, dirent) - return true + return nil } // TestIterDirents tests the FileDescriptionImpl.IterDirents functionality. diff --git a/pkg/sentry/fsimpl/ext/filesystem.go b/pkg/sentry/fsimpl/ext/filesystem.go index 07bf58953..2c22a04af 100644 --- a/pkg/sentry/fsimpl/ext/filesystem.go +++ b/pkg/sentry/fsimpl/ext/filesystem.go @@ -22,6 +22,8 @@ import ( "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/fspath" "gvisor.dev/gvisor/pkg/sentry/fsimpl/ext/disklayout" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/syserror" @@ -87,14 +89,33 @@ func stepLocked(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, inode *inode, write boo } for { - nextVFSD, err := rp.ResolveComponent(vfsd) - if err != nil { - return nil, nil, err + name := rp.Component() + if name == "." { + rp.Advance() + return vfsd, inode, nil } - if nextVFSD == nil { - // Since the Dentry tree is not the sole source of truth for extfs, if it's - // not in the Dentry tree, it might need to be pulled from disk. - childDirent, ok := inode.impl.(*directory).childMap[rp.Component()] + d := vfsd.Impl().(*dentry) + if name == ".." { + isRoot, err := rp.CheckRoot(vfsd) + if err != nil { + return nil, nil, err + } + if isRoot || d.parent == nil { + rp.Advance() + return vfsd, inode, nil + } + if err := rp.CheckMount(&d.parent.vfsd); err != nil { + return nil, nil, err + } + rp.Advance() + return &d.parent.vfsd, d.parent.inode, nil + } + + dir := inode.impl.(*directory) + child, ok := dir.childCache[name] + if !ok { + // We may need to instantiate a new dentry for this child. + childDirent, ok := dir.childMap[name] if !ok { // The underlying inode does not exist on disk. return nil, nil, syserror.ENOENT @@ -113,21 +134,22 @@ func stepLocked(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, inode *inode, write boo } // incRef because this is being added to the dentry tree. childInode.incRef() - child := newDentry(childInode) - vfsd.InsertChild(&child.vfsd, rp.Component()) - - // Continue as usual now that nextVFSD is not nil. - nextVFSD = &child.vfsd + child = newDentry(childInode) + child.parent = d + child.name = name + dir.childCache[name] = child + } + if err := rp.CheckMount(&child.vfsd); err != nil { + return nil, nil, err } - nextInode := nextVFSD.Impl().(*dentry).inode - if nextInode.isSymlink() && rp.ShouldFollowSymlink() { - if err := rp.HandleSymlink(inode.impl.(*symlink).target); err != nil { + if child.inode.isSymlink() && rp.ShouldFollowSymlink() { + if err := rp.HandleSymlink(child.inode.impl.(*symlink).target); err != nil { return nil, nil, err } continue } rp.Advance() - return nextVFSD, nextInode, nil + return &child.vfsd, child.inode, nil } } @@ -255,6 +277,15 @@ func (fs *filesystem) statTo(stat *linux.Statfs) { // TODO(b/134676337): Set Statfs.Flags and Statfs.FSID. } +// AccessAt implements vfs.Filesystem.Impl.AccessAt. +func (fs *filesystem) AccessAt(ctx context.Context, rp *vfs.ResolvingPath, creds *auth.Credentials, ats vfs.AccessTypes) error { + _, inode, err := fs.walk(rp, false) + if err != nil { + return err + } + return inode.checkPermissions(rp.Credentials(), ats) +} + // GetDentryAt implements vfs.FilesystemImpl.GetDentryAt. func (fs *filesystem) GetDentryAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.GetDentryOptions) (*vfs.Dentry, error) { vfsd, inode, err := fs.walk(rp, false) @@ -296,7 +327,7 @@ func (fs *filesystem) OpenAt(ctx context.Context, rp *vfs.ResolvingPath, opts vf if vfs.MayWriteFileWithOpenFlags(opts.Flags) || opts.Flags&(linux.O_CREAT|linux.O_EXCL|linux.O_TMPFILE) != 0 { return nil, syserror.EROFS } - return inode.open(rp, vfsd, opts.Flags) + return inode.open(rp, vfsd, &opts) } // ReadlinkAt implements vfs.FilesystemImpl.ReadlinkAt. @@ -453,8 +484,19 @@ func (fs *filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error return syserror.EROFS } +// BoundEndpointAt implements FilesystemImpl.BoundEndpointAt. +func (fs *filesystem) BoundEndpointAt(ctx context.Context, rp *vfs.ResolvingPath) (transport.BoundEndpoint, error) { + _, _, err := fs.walk(rp, false) + if err != nil { + return nil, err + } + + // TODO(b/134676337): Support sockets. + return nil, syserror.ECONNREFUSED +} + // ListxattrAt implements vfs.FilesystemImpl.ListxattrAt. -func (fs *filesystem) ListxattrAt(ctx context.Context, rp *vfs.ResolvingPath) ([]string, error) { +func (fs *filesystem) ListxattrAt(ctx context.Context, rp *vfs.ResolvingPath, size uint64) ([]string, error) { _, _, err := fs.walk(rp, false) if err != nil { return nil, err @@ -463,7 +505,7 @@ func (fs *filesystem) ListxattrAt(ctx context.Context, rp *vfs.ResolvingPath) ([ } // GetxattrAt implements vfs.FilesystemImpl.GetxattrAt. -func (fs *filesystem) GetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) (string, error) { +func (fs *filesystem) GetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.GetxattrOptions) (string, error) { _, _, err := fs.walk(rp, false) if err != nil { return "", err @@ -493,5 +535,5 @@ func (fs *filesystem) RemovexattrAt(ctx context.Context, rp *vfs.ResolvingPath, func (fs *filesystem) PrependPath(ctx context.Context, vfsroot, vd vfs.VirtualDentry, b *fspath.Builder) error { fs.mu.RLock() defer fs.mu.RUnlock() - return vfs.GenericPrependPath(vfsroot, vd, b) + return genericPrependPath(vfsroot, vd.Mount(), vd.Dentry().Impl().(*dentry), b) } diff --git a/pkg/sentry/fsimpl/ext/inode.go b/pkg/sentry/fsimpl/ext/inode.go index 191b39970..a98512350 100644 --- a/pkg/sentry/fsimpl/ext/inode.go +++ b/pkg/sentry/fsimpl/ext/inode.go @@ -136,7 +136,7 @@ func newInode(fs *filesystem, inodeNum uint32) (*inode, error) { } return &f.inode, nil case linux.ModeDirectory: - f, err := newDirectroy(inode, fs.sb.IncompatibleFeatures().DirentFileType) + f, err := newDirectory(inode, fs.sb.IncompatibleFeatures().DirentFileType) if err != nil { return nil, err } @@ -148,8 +148,8 @@ func newInode(fs *filesystem, inodeNum uint32) (*inode, error) { } // open creates and returns a file description for the dentry passed in. -func (in *inode) open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, flags uint32) (*vfs.FileDescription, error) { - ats := vfs.AccessTypesForOpenFlags(flags) +func (in *inode) open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts *vfs.OpenOptions) (*vfs.FileDescription, error) { + ats := vfs.AccessTypesForOpenFlags(opts) if err := in.checkPermissions(rp.Credentials(), ats); err != nil { return nil, err } @@ -157,7 +157,7 @@ func (in *inode) open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, flags uint32) (*v switch in.impl.(type) { case *regularFile: var fd regularFileFD - if err := fd.vfsfd.Init(&fd, flags, mnt, vfsd, &vfs.FileDescriptionOptions{}); err != nil { + if err := fd.vfsfd.Init(&fd, opts.Flags, mnt, vfsd, &vfs.FileDescriptionOptions{}); err != nil { return nil, err } return &fd.vfsfd, nil @@ -168,17 +168,17 @@ func (in *inode) open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, flags uint32) (*v return nil, syserror.EISDIR } var fd directoryFD - if err := fd.vfsfd.Init(&fd, flags, mnt, vfsd, &vfs.FileDescriptionOptions{}); err != nil { + if err := fd.vfsfd.Init(&fd, opts.Flags, mnt, vfsd, &vfs.FileDescriptionOptions{}); err != nil { return nil, err } return &fd.vfsfd, nil case *symlink: - if flags&linux.O_PATH == 0 { + if opts.Flags&linux.O_PATH == 0 { // Can't open symlinks without O_PATH. return nil, syserror.ELOOP } var fd symlinkFD - fd.vfsfd.Init(&fd, flags, mnt, vfsd, &vfs.FileDescriptionOptions{}) + fd.vfsfd.Init(&fd, opts.Flags, mnt, vfsd, &vfs.FileDescriptionOptions{}) return &fd.vfsfd, nil default: panic(fmt.Sprintf("unknown inode type: %T", in.impl)) @@ -186,7 +186,7 @@ func (in *inode) open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, flags uint32) (*v } func (in *inode) checkPermissions(creds *auth.Credentials, ats vfs.AccessTypes) error { - return vfs.GenericCheckPermissions(creds, ats, in.isDir(), uint16(in.diskInode.Mode()), in.diskInode.UID(), in.diskInode.GID()) + return vfs.GenericCheckPermissions(creds, ats, in.diskInode.Mode(), in.diskInode.UID(), in.diskInode.GID()) } // statTo writes the statx fields to the output parameter. diff --git a/pkg/sentry/fsimpl/gofer/BUILD b/pkg/sentry/fsimpl/gofer/BUILD index 4ba76a1e8..acd061905 100644 --- a/pkg/sentry/fsimpl/gofer/BUILD +++ b/pkg/sentry/fsimpl/gofer/BUILD @@ -1,4 +1,4 @@ -load("//tools:defs.bzl", "go_library") +load("//tools:defs.bzl", "go_library", "go_test") load("//tools/go_generics:defs.bzl", "go_template_instance") licenses(["notice"]) @@ -15,12 +15,24 @@ go_template_instance( }, ) +go_template_instance( + name = "fstree", + out = "fstree.go", + package = "gofer", + prefix = "generic", + template = "//pkg/sentry/vfs/genericfstree:generic_fstree", + types = { + "Dentry": "dentry", + }, +) + go_library( name = "gofer", srcs = [ "dentry_list.go", "directory.go", "filesystem.go", + "fstree.go", "gofer.go", "handle.go", "handle_unsafe.go", @@ -46,6 +58,7 @@ go_library( "//pkg/sentry/memmap", "//pkg/sentry/pgalloc", "//pkg/sentry/platform", + "//pkg/sentry/socket/unix/transport", "//pkg/sentry/usage", "//pkg/sentry/vfs", "//pkg/syserror", @@ -53,3 +66,13 @@ go_library( "//pkg/usermem", ], ) + +go_test( + name = "gofer_test", + srcs = ["gofer_test.go"], + library = ":gofer", + deps = [ + "//pkg/p9", + "//pkg/sentry/contexttest", + ], +) diff --git a/pkg/sentry/fsimpl/gofer/directory.go b/pkg/sentry/fsimpl/gofer/directory.go index 6d4ebc2bf..d02691232 100644 --- a/pkg/sentry/fsimpl/gofer/directory.go +++ b/pkg/sentry/fsimpl/gofer/directory.go @@ -29,13 +29,25 @@ func (d *dentry) isDir() bool { return d.fileType() == linux.S_IFDIR } +// Preconditions: filesystem.renameMu must be locked. d.dirMu must be locked. +// d.isDir(). child must be a newly-created dentry that has never had a parent. +func (d *dentry) cacheNewChildLocked(child *dentry, name string) { + d.IncRef() // reference held by child on its parent + child.parent = d + child.name = name + if d.children == nil { + d.children = make(map[string]*dentry) + } + d.children[name] = child +} + // Preconditions: d.dirMu must be locked. d.isDir(). fs.opts.interop != // InteropModeShared. func (d *dentry) cacheNegativeChildLocked(name string) { - if d.negativeChildren == nil { - d.negativeChildren = make(map[string]struct{}) + if d.children == nil { + d.children = make(map[string]*dentry) } - d.negativeChildren[name] = struct{}{} + d.children[name] = nil } type directoryFD struct { @@ -56,17 +68,22 @@ func (fd *directoryFD) IterDirents(ctx context.Context, cb vfs.IterDirentsCallba fd.mu.Lock() defer fd.mu.Unlock() + d := fd.dentry() if fd.dirents == nil { - ds, err := fd.dentry().getDirents(ctx) + ds, err := d.getDirents(ctx) if err != nil { return err } fd.dirents = ds } + if d.fs.opts.interop != InteropModeShared { + d.touchAtime(fd.vfsfd.Mount()) + } + for fd.off < int64(len(fd.dirents)) { - if !cb.Handle(fd.dirents[fd.off]) { - return nil + if err := cb.Handle(fd.dirents[fd.off]); err != nil { + return err } fd.off++ } @@ -75,34 +92,32 @@ func (fd *directoryFD) IterDirents(ctx context.Context, cb vfs.IterDirentsCallba // Preconditions: d.isDir(). There exists at least one directoryFD representing d. func (d *dentry) getDirents(ctx context.Context) ([]vfs.Dirent, error) { - // 9P2000.L's readdir does not specify behavior in the presence of - // concurrent mutation of an iterated directory, so implementations may - // duplicate or omit entries in this case, which violates POSIX semantics. - // Thus we read all directory entries while holding d.dirMu to exclude - // directory mutations. (Note that it is impossible for the client to - // exclude concurrent mutation from other remote filesystem users. Since - // there is no way to detect if the server has incorrectly omitted - // directory entries, we simply assume that the server is well-behaved - // under InteropModeShared.) This is inconsistent with Linux (which appears - // to assume that directory fids have the correct semantics, and translates - // struct file_operations::readdir calls directly to readdir RPCs), but is - // consistent with VFS1. - // - // NOTE(b/135560623): In particular, some gofer implementations may not - // retain state between calls to Readdir, so may not provide a coherent - // directory stream across in the presence of mutation. - + // NOTE(b/135560623): 9P2000.L's readdir does not specify behavior in the + // presence of concurrent mutation of an iterated directory, so + // implementations may duplicate or omit entries in this case, which + // violates POSIX semantics. Thus we read all directory entries while + // holding d.dirMu to exclude directory mutations. (Note that it is + // impossible for the client to exclude concurrent mutation from other + // remote filesystem users. Since there is no way to detect if the server + // has incorrectly omitted directory entries, we simply assume that the + // server is well-behaved under InteropModeShared.) This is inconsistent + // with Linux (which appears to assume that directory fids have the correct + // semantics, and translates struct file_operations::readdir calls directly + // to readdir RPCs), but is consistent with VFS1. + + // filesystem.renameMu is needed for d.parent, and must be locked before + // dentry.dirMu. d.fs.renameMu.RLock() - defer d.fs.renameMu.RUnlock() d.dirMu.Lock() defer d.dirMu.Unlock() if d.dirents != nil { + d.fs.renameMu.RUnlock() return d.dirents, nil } // It's not clear if 9P2000.L's readdir is expected to return "." and "..", // so we generate them here. - parent := d.vfsd.ParentOrSelf().Impl().(*dentry) + parent := genericParentOrSelf(d) dirents := []vfs.Dirent{ { Name: ".", @@ -117,6 +132,7 @@ func (d *dentry) getDirents(ctx context.Context) ([]vfs.Dirent, error) { NextOff: 2, }, } + d.fs.renameMu.RUnlock() off := uint64(0) const count = 64 * 1024 // for consistency with the vfs1 client d.handleMu.RLock() diff --git a/pkg/sentry/fsimpl/gofer/filesystem.go b/pkg/sentry/fsimpl/gofer/filesystem.go index 8eb61debf..43e863c61 100644 --- a/pkg/sentry/fsimpl/gofer/filesystem.go +++ b/pkg/sentry/fsimpl/gofer/filesystem.go @@ -21,6 +21,8 @@ import ( "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/fspath" "gvisor.dev/gvisor/pkg/p9" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserror" ) @@ -114,11 +116,13 @@ func putDentrySlice(ds *[]*dentry) { // Preconditions: fs.renameMu must be locked. d.dirMu must be locked. // !rp.Done(). If fs.opts.interop == InteropModeShared, then d's cached // metadata must be up to date. +// +// Postconditions: The returned dentry's cached metadata is up to date. func (fs *filesystem) stepLocked(ctx context.Context, rp *vfs.ResolvingPath, d *dentry, ds **[]*dentry) (*dentry, error) { if !d.isDir() { return nil, syserror.ENOTDIR } - if err := d.checkPermissions(rp.Credentials(), vfs.MayExec, true); err != nil { + if err := d.checkPermissions(rp.Credentials(), vfs.MayExec); err != nil { return nil, err } afterSymlink: @@ -128,39 +132,42 @@ afterSymlink: return d, nil } if name == ".." { - parentVFSD, err := rp.ResolveParent(&d.vfsd) - if err != nil { + if isRoot, err := rp.CheckRoot(&d.vfsd); err != nil { + return nil, err + } else if isRoot || d.parent == nil { + rp.Advance() + return d, nil + } + // We must assume that d.parent is correct, because if d has been moved + // elsewhere in the remote filesystem so that its parent has changed, + // we have no way of determining its new parent's location in the + // filesystem. + // + // Call rp.CheckMount() before updating d.parent's metadata, since if + // we traverse to another mount then d.parent's metadata is irrelevant. + if err := rp.CheckMount(&d.parent.vfsd); err != nil { return nil, err } - parent := parentVFSD.Impl().(*dentry) - if fs.opts.interop == InteropModeShared { - // We must assume that parentVFSD is correct, because if d has been - // moved elsewhere in the remote filesystem so that its parent has - // changed, we have no way of determining its new parent's location - // in the filesystem. Get updated metadata for parentVFSD. - _, attrMask, attr, err := parent.file.getAttr(ctx, dentryAttrMask()) + if fs.opts.interop == InteropModeShared && d != d.parent { + _, attrMask, attr, err := d.parent.file.getAttr(ctx, dentryAttrMask()) if err != nil { return nil, err } - parent.updateFromP9Attrs(attrMask, &attr) + d.parent.updateFromP9Attrs(attrMask, &attr) } rp.Advance() - return parent, nil - } - childVFSD, err := rp.ResolveChild(&d.vfsd, name) - if err != nil { - return nil, err + return d.parent, nil } - // FIXME(jamieliu): Linux performs revalidation before mount lookup - // (fs/namei.c:lookup_fast() => __d_lookup_rcu(), d_revalidate(), - // __follow_mount_rcu()). - child, err := fs.revalidateChildLocked(ctx, rp.VirtualFilesystem(), d, name, childVFSD, ds) + child, err := fs.getChildLocked(ctx, rp.VirtualFilesystem(), d, name, ds) if err != nil { return nil, err } if child == nil { return nil, syserror.ENOENT } + if err := rp.CheckMount(&child.vfsd); err != nil { + return nil, err + } if child.isSymlink() && rp.ShouldFollowSymlink() { target, err := child.readlink(ctx, rp.Mount()) if err != nil { @@ -175,38 +182,37 @@ afterSymlink: return child, nil } -// revalidateChildLocked must be called after a call to parent.vfsd.Child(name) -// or vfs.ResolvingPath.ResolveChild(name) returns childVFSD (which may be -// nil) to verify that the returned child (or lack thereof) is correct. If no file -// exists at name, revalidateChildLocked returns (nil, nil). +// getChildLocked returns a dentry representing the child of parent with the +// given name. If no such child exists, getChildLocked returns (nil, nil). // // Preconditions: fs.renameMu must be locked. parent.dirMu must be locked. // parent.isDir(). name is not "." or "..". // -// Postconditions: If revalidateChildLocked returns a non-nil dentry, its -// cached metadata is up to date. -func (fs *filesystem) revalidateChildLocked(ctx context.Context, vfsObj *vfs.VirtualFilesystem, parent *dentry, name string, childVFSD *vfs.Dentry, ds **[]*dentry) (*dentry, error) { - if childVFSD != nil && fs.opts.interop != InteropModeShared { - // We have a cached dentry that is assumed to be correct. - return childVFSD.Impl().(*dentry), nil - } - // We either don't have a cached dentry or need to verify that it's still - // correct, either of which requires a remote lookup. Check if this name is - // valid before performing the lookup. +// Postconditions: If getChildLocked returns a non-nil dentry, its cached +// metadata is up to date. +func (fs *filesystem) getChildLocked(ctx context.Context, vfsObj *vfs.VirtualFilesystem, parent *dentry, name string, ds **[]*dentry) (*dentry, error) { if len(name) > maxFilenameLen { return nil, syserror.ENAMETOOLONG } - // Check if we've already cached this lookup with a negative result. - if _, ok := parent.negativeChildren[name]; ok { - return nil, nil + child, ok := parent.children[name] + if ok && fs.opts.interop != InteropModeShared { + // Whether child is nil or not, it is cached information that is + // assumed to be correct. + return child, nil } - // Perform the remote lookup. + // We either don't have cached information or need to verify that it's + // still correct, either of which requires a remote lookup. Check if this + // name is valid before performing the lookup. + return fs.revalidateChildLocked(ctx, vfsObj, parent, name, child, ds) +} + +// Preconditions: As for getChildLocked. +func (fs *filesystem) revalidateChildLocked(ctx context.Context, vfsObj *vfs.VirtualFilesystem, parent *dentry, name string, child *dentry, ds **[]*dentry) (*dentry, error) { qid, file, attrMask, attr, err := parent.file.walkGetAttrOne(ctx, name) if err != nil && err != syserror.ENOENT { return nil, err } - if childVFSD != nil { - child := childVFSD.Impl().(*dentry) + if child != nil { if !file.isNil() && qid.Path == child.ino { // The file at this path hasn't changed. Just update cached // metadata. @@ -217,9 +223,8 @@ func (fs *filesystem) revalidateChildLocked(ctx context.Context, vfsObj *vfs.Vir // The file at this path has changed or no longer exists. Remove // the stale dentry from the tree, and re-evaluate its caching // status (i.e. if it has 0 references, drop it). - vfsObj.ForceDeleteDentry(childVFSD) + vfsObj.InvalidateDentry(&child.vfsd) *ds = appendDentry(*ds, child) - childVFSD = nil } if file.isNil() { // No file exists at this path now. Cache the negative lookup if @@ -230,13 +235,12 @@ func (fs *filesystem) revalidateChildLocked(ctx context.Context, vfsObj *vfs.Vir return nil, nil } // Create a new dentry representing the file. - child, err := fs.newDentry(ctx, file, qid, attrMask, &attr) + child, err = fs.newDentry(ctx, file, qid, attrMask, &attr) if err != nil { file.close(ctx) return nil, err } - parent.IncRef() // reference held by child on its parent - parent.vfsd.InsertChild(&child.vfsd, name) + parent.cacheNewChildLocked(child, name) // For now, child has 0 references, so our caller should call // child.checkCachingLocked(). *ds = appendDentry(*ds, child) @@ -313,12 +317,9 @@ func (fs *filesystem) doCreateAt(ctx context.Context, rp *vfs.ResolvingPath, dir if err != nil { return err } - if err := parent.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true); err != nil { + if err := parent.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec); err != nil { return err } - if parent.isDeleted() { - return syserror.ENOENT - } name := rp.Component() if name == "." || name == ".." { return syserror.EEXIST @@ -329,6 +330,9 @@ func (fs *filesystem) doCreateAt(ctx context.Context, rp *vfs.ResolvingPath, dir if !dir && rp.MustBeDir() { return syserror.ENOENT } + if parent.isDeleted() { + return syserror.ENOENT + } mnt := rp.Mount() if err := mnt.CheckBeginWrite(); err != nil { return err @@ -346,7 +350,7 @@ func (fs *filesystem) doCreateAt(ctx context.Context, rp *vfs.ResolvingPath, dir // it's used. return create(parent, name) } - if parent.vfsd.Child(name) != nil { + if child := parent.children[name]; child != nil { return syserror.EEXIST } // No cached dentry exists; however, there might still be an existing file @@ -354,8 +358,11 @@ func (fs *filesystem) doCreateAt(ctx context.Context, rp *vfs.ResolvingPath, dir if err := create(parent, name); err != nil { return err } - parent.touchCMtime(ctx) - delete(parent.negativeChildren, name) + parent.touchCMtime() + // Either parent.children[name] doesn't exist (in which case this is a + // no-op) or is nil (in which case this erases the now-stale information + // that the file doesn't exist). + delete(parent.children, name) parent.dirents = nil return nil } @@ -377,7 +384,7 @@ func (fs *filesystem) unlinkAt(ctx context.Context, rp *vfs.ResolvingPath, dir b if err != nil { return err } - if err := parent.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true); err != nil { + if err := parent.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec); err != nil { return err } if err := rp.Mount().CheckBeginWrite(); err != nil { @@ -400,64 +407,74 @@ func (fs *filesystem) unlinkAt(ctx context.Context, rp *vfs.ResolvingPath, dir b } vfsObj := rp.VirtualFilesystem() mntns := vfs.MountNamespaceFromContext(ctx) + defer mntns.DecRef() parent.dirMu.Lock() defer parent.dirMu.Unlock() - childVFSD := parent.vfsd.Child(name) - var child *dentry + child, ok := parent.children[name] + if ok && child == nil { + return syserror.ENOENT + } // We only need a dentry representing the file at name if it can be a mount - // point. If childVFSD is nil, then it can't be a mount point. If childVFSD - // is non-nil but stale, the actual file can't be a mount point either; we + // point. If child is nil, then it can't be a mount point. If child is + // non-nil but stale, the actual file can't be a mount point either; we // detect this case by just speculatively calling PrepareDeleteDentry and // only revalidating the dentry if that fails (indicating that the existing // dentry is a mount point). - if childVFSD != nil { - child = childVFSD.Impl().(*dentry) - if err := vfsObj.PrepareDeleteDentry(mntns, childVFSD); err != nil { - child, err = fs.revalidateChildLocked(ctx, vfsObj, parent, name, childVFSD, &ds) + if child != nil { + if err := vfsObj.PrepareDeleteDentry(mntns, &child.vfsd); err != nil { + if fs.opts.interop != InteropModeShared { + return err + } + child, err = fs.revalidateChildLocked(ctx, vfsObj, parent, name, child, &ds) if err != nil { return err } if child != nil { - childVFSD = &child.vfsd - if err := vfsObj.PrepareDeleteDentry(mntns, childVFSD); err != nil { + if err := vfsObj.PrepareDeleteDentry(mntns, &child.vfsd); err != nil { return err } - } else { - childVFSD = nil } } - } else if _, ok := parent.negativeChildren[name]; ok { - return syserror.ENOENT } flags := uint32(0) if dir { if child != nil && !child.isDir() { + vfsObj.AbortDeleteDentry(&child.vfsd) return syserror.ENOTDIR } flags = linux.AT_REMOVEDIR } else { if child != nil && child.isDir() { + vfsObj.AbortDeleteDentry(&child.vfsd) return syserror.EISDIR } if rp.MustBeDir() { + if child != nil { + vfsObj.AbortDeleteDentry(&child.vfsd) + } return syserror.ENOTDIR } } err = parent.file.unlinkAt(ctx, name, flags) if err != nil { - if childVFSD != nil { - vfsObj.AbortDeleteDentry(childVFSD) + if child != nil { + vfsObj.AbortDeleteDentry(&child.vfsd) } return err } if fs.opts.interop != InteropModeShared { - parent.touchCMtime(ctx) + parent.touchCMtime() + if dir { + parent.decLinks() + } parent.cacheNegativeChildLocked(name) parent.dirents = nil + } else { + delete(parent.children, name) } if child != nil { child.setDeleted() - vfsObj.CommitDeleteDentry(childVFSD) + vfsObj.CommitDeleteDentry(&child.vfsd) ds = appendDentry(ds, child) } return nil @@ -498,6 +515,18 @@ func (fs *filesystem) renameMuUnlockAndCheckCaching(ds **[]*dentry) { putDentrySlice(*ds) } +// AccessAt implements vfs.Filesystem.Impl.AccessAt. +func (fs *filesystem) AccessAt(ctx context.Context, rp *vfs.ResolvingPath, creds *auth.Credentials, ats vfs.AccessTypes) error { + var ds *[]*dentry + fs.renameMu.RLock() + defer fs.renameMuRUnlockAndCheckCaching(&ds) + d, err := fs.resolveLocked(ctx, rp, &ds) + if err != nil { + return err + } + return d.checkPermissions(creds, ats) +} + // GetDentryAt implements vfs.FilesystemImpl.GetDentryAt. func (fs *filesystem) GetDentryAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.GetDentryOptions) (*vfs.Dentry, error) { var ds *[]*dentry @@ -511,7 +540,7 @@ func (fs *filesystem) GetDentryAt(ctx context.Context, rp *vfs.ResolvingPath, op if !d.isDir() { return nil, syserror.ENOTDIR } - if err := d.checkPermissions(rp.Credentials(), vfs.MayExec, true); err != nil { + if err := d.checkPermissions(rp.Credentials(), vfs.MayExec); err != nil { return nil, err } } @@ -555,8 +584,13 @@ func (fs *filesystem) LinkAt(ctx context.Context, rp *vfs.ResolvingPath, vd vfs. func (fs *filesystem) MkdirAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.MkdirOptions) error { return fs.doCreateAt(ctx, rp, true /* dir */, func(parent *dentry, name string) error { creds := rp.Credentials() - _, err := parent.file.mkdir(ctx, name, (p9.FileMode)(opts.Mode), (p9.UID)(creds.EffectiveKUID), (p9.GID)(creds.EffectiveKGID)) - return err + if _, err := parent.file.mkdir(ctx, name, (p9.FileMode)(opts.Mode), (p9.UID)(creds.EffectiveKUID), (p9.GID)(creds.EffectiveKGID)); err != nil { + return err + } + if fs.opts.interop != InteropModeShared { + parent.incLinks() + } + return nil }) } @@ -593,7 +627,7 @@ func (fs *filesystem) OpenAt(ctx context.Context, rp *vfs.ResolvingPath, opts vf } } if rp.Done() { - return start.openLocked(ctx, rp, opts.Flags) + return start.openLocked(ctx, rp, &opts) } afterTrailingSymlink: @@ -602,7 +636,7 @@ afterTrailingSymlink: return nil, err } // Check for search permission in the parent directory. - if err := parent.checkPermissions(rp.Credentials(), vfs.MayExec, true); err != nil { + if err := parent.checkPermissions(rp.Credentials(), vfs.MayExec); err != nil { return nil, err } // Determine whether or not we need to create a file. @@ -633,24 +667,24 @@ afterTrailingSymlink: start = parent goto afterTrailingSymlink } - return child.openLocked(ctx, rp, opts.Flags) + return child.openLocked(ctx, rp, &opts) } // Preconditions: fs.renameMu must be locked. -func (d *dentry) openLocked(ctx context.Context, rp *vfs.ResolvingPath, flags uint32) (*vfs.FileDescription, error) { - ats := vfs.AccessTypesForOpenFlags(flags) - if err := d.checkPermissions(rp.Credentials(), ats, d.isDir()); err != nil { +func (d *dentry) openLocked(ctx context.Context, rp *vfs.ResolvingPath, opts *vfs.OpenOptions) (*vfs.FileDescription, error) { + ats := vfs.AccessTypesForOpenFlags(opts) + if err := d.checkPermissions(rp.Credentials(), ats); err != nil { return nil, err } mnt := rp.Mount() filetype := d.fileType() switch { case filetype == linux.S_IFREG && !d.fs.opts.regularFilesUseSpecialFileFD: - if err := d.ensureSharedHandle(ctx, ats&vfs.MayRead != 0, ats&vfs.MayWrite != 0, flags&linux.O_TRUNC != 0); err != nil { + if err := d.ensureSharedHandle(ctx, ats&vfs.MayRead != 0, ats&vfs.MayWrite != 0, opts.Flags&linux.O_TRUNC != 0); err != nil { return nil, err } fd := ®ularFileFD{} - if err := fd.vfsfd.Init(fd, flags, mnt, &d.vfsd, &vfs.FileDescriptionOptions{ + if err := fd.vfsfd.Init(fd, opts.Flags, mnt, &d.vfsd, &vfs.FileDescriptionOptions{ AllowDirectIO: true, }); err != nil { return nil, err @@ -658,21 +692,21 @@ func (d *dentry) openLocked(ctx context.Context, rp *vfs.ResolvingPath, flags ui return &fd.vfsfd, nil case filetype == linux.S_IFDIR: // Can't open directories with O_CREAT. - if flags&linux.O_CREAT != 0 { + if opts.Flags&linux.O_CREAT != 0 { return nil, syserror.EISDIR } // Can't open directories writably. if ats&vfs.MayWrite != 0 { return nil, syserror.EISDIR } - if flags&linux.O_DIRECT != 0 { + if opts.Flags&linux.O_DIRECT != 0 { return nil, syserror.EINVAL } if err := d.ensureSharedHandle(ctx, ats&vfs.MayRead != 0, false /* write */, false /* trunc */); err != nil { return nil, err } fd := &directoryFD{} - if err := fd.vfsfd.Init(fd, flags, mnt, &d.vfsd, &vfs.FileDescriptionOptions{}); err != nil { + if err := fd.vfsfd.Init(fd, opts.Flags, mnt, &d.vfsd, &vfs.FileDescriptionOptions{}); err != nil { return nil, err } return &fd.vfsfd, nil @@ -680,17 +714,17 @@ func (d *dentry) openLocked(ctx context.Context, rp *vfs.ResolvingPath, flags ui // Can't open symlinks without O_PATH (which is unimplemented). return nil, syserror.ELOOP default: - if flags&linux.O_DIRECT != 0 { + if opts.Flags&linux.O_DIRECT != 0 { return nil, syserror.EINVAL } - h, err := openHandle(ctx, d.file, ats&vfs.MayRead != 0, ats&vfs.MayWrite != 0, flags&linux.O_TRUNC != 0) + h, err := openHandle(ctx, d.file, ats&vfs.MayRead != 0, ats&vfs.MayWrite != 0, opts.Flags&linux.O_TRUNC != 0) if err != nil { return nil, err } fd := &specialFileFD{ handle: h, } - if err := fd.vfsfd.Init(fd, flags, mnt, &d.vfsd, &vfs.FileDescriptionOptions{}); err != nil { + if err := fd.vfsfd.Init(fd, opts.Flags, mnt, &d.vfsd, &vfs.FileDescriptionOptions{}); err != nil { h.close(ctx) return nil, err } @@ -700,7 +734,7 @@ func (d *dentry) openLocked(ctx context.Context, rp *vfs.ResolvingPath, flags ui // Preconditions: d.fs.renameMu must be locked. d.dirMu must be locked. func (d *dentry) createAndOpenChildLocked(ctx context.Context, rp *vfs.ResolvingPath, opts *vfs.OpenOptions) (*vfs.FileDescription, error) { - if err := d.checkPermissions(rp.Credentials(), vfs.MayWrite, true); err != nil { + if err := d.checkPermissions(rp.Credentials(), vfs.MayWrite); err != nil { return nil, err } if d.isDeleted() { @@ -776,17 +810,14 @@ func (d *dentry) createAndOpenChildLocked(ctx context.Context, rp *vfs.Resolving // eligible for caching yet, so we don't need to append to a dentry slice.) child.refs = 1 // Insert the dentry into the tree. - d.IncRef() // reference held by child on its parent d - d.vfsd.InsertChild(&child.vfsd, name) + d.cacheNewChildLocked(child, name) if d.fs.opts.interop != InteropModeShared { - d.touchCMtime(ctx) - delete(d.negativeChildren, name) + d.touchCMtime() d.dirents = nil } // Finally, construct a file description representing the created file. var childVFSFD *vfs.FileDescription - mnt.IncRef() if useRegularFileFD { fd := ®ularFileFD{} if err := fd.vfsfd.Init(fd, opts.Flags, mnt, &child.vfsd, &vfs.FileDescriptionOptions{ @@ -862,7 +893,7 @@ func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa return err } } - if err := oldParent.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true); err != nil { + if err := oldParent.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec); err != nil { return err } vfsObj := rp.VirtualFilesystem() @@ -870,7 +901,7 @@ func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa // directory, we need to check for write permission on it. oldParent.dirMu.Lock() defer oldParent.dirMu.Unlock() - renamed, err := fs.revalidateChildLocked(ctx, vfsObj, oldParent, oldName, oldParent.vfsd.Child(oldName), &ds) + renamed, err := fs.getChildLocked(ctx, vfsObj, oldParent, oldName, &ds) if err != nil { return err } @@ -878,11 +909,11 @@ func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa return syserror.ENOENT } if renamed.isDir() { - if renamed == newParent || renamed.vfsd.IsAncestorOf(&newParent.vfsd) { + if renamed == newParent || genericIsAncestorDentry(renamed, newParent) { return syserror.EINVAL } if oldParent != newParent { - if err := renamed.checkPermissions(rp.Credentials(), vfs.MayWrite, true); err != nil { + if err := renamed.checkPermissions(rp.Credentials(), vfs.MayWrite); err != nil { return err } } @@ -893,7 +924,7 @@ func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa } if oldParent != newParent { - if err := newParent.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true); err != nil { + if err := newParent.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec); err != nil { return err } newParent.dirMu.Lock() @@ -902,16 +933,17 @@ func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa if newParent.isDeleted() { return syserror.ENOENT } - replacedVFSD := newParent.vfsd.Child(newName) - var replaced *dentry + replaced := newParent.children[newName] // This is similar to unlinkAt, except: // - // - We revalidate the replaced dentry unconditionally for simplicity. + // - If a dentry exists for the file to be replaced, we revalidate it + // unconditionally (instead of only if PrepareRenameDentry fails) for + // simplicity. // // - If rp.MustBeDir(), then we need a dentry representing the replaced // file regardless to confirm that it's a directory. - if replacedVFSD != nil || rp.MustBeDir() { - replaced, err = fs.revalidateChildLocked(ctx, vfsObj, newParent, newName, replacedVFSD, &ds) + if replaced != nil || rp.MustBeDir() { + replaced, err = fs.getChildLocked(ctx, rp.VirtualFilesystem(), newParent, newName, &ds) if err != nil { return err } @@ -925,16 +957,19 @@ func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa return syserror.ENOTDIR } } - replacedVFSD = &replaced.vfsd - } else { - replacedVFSD = nil } } + var replacedVFSD *vfs.Dentry + if replaced != nil { + replacedVFSD = &replaced.vfsd + } if oldParent == newParent && oldName == newName { return nil } - if err := vfsObj.PrepareRenameDentry(vfs.MountNamespaceFromContext(ctx), &renamed.vfsd, replacedVFSD); err != nil { + mntns := vfs.MountNamespaceFromContext(ctx) + defer mntns.DecRef() + if err := vfsObj.PrepareRenameDentry(mntns, &renamed.vfsd, replacedVFSD); err != nil { return err } if err := renamed.file.rename(ctx, newParent.file, newName); err != nil { @@ -944,10 +979,32 @@ func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa if fs.opts.interop != InteropModeShared { oldParent.cacheNegativeChildLocked(oldName) oldParent.dirents = nil - delete(newParent.negativeChildren, newName) newParent.dirents = nil + if renamed.isDir() { + oldParent.decLinks() + newParent.incLinks() + } + oldParent.touchCMtime() + newParent.touchCMtime() + renamed.touchCtime() + } else { + delete(oldParent.children, oldName) } - vfsObj.CommitRenameReplaceDentry(&renamed.vfsd, &newParent.vfsd, newName, replacedVFSD) + if oldParent != newParent { + appendDentry(ds, oldParent) + newParent.IncRef() + } + renamed.parent = newParent + renamed.name = newName + if newParent.children == nil { + newParent.children = make(map[string]*dentry) + } + newParent.children[newName] = renamed + if replaced != nil { + replaced.setDeleted() + appendDentry(ds, replaced) + } + vfsObj.CommitRenameReplaceDentry(&renamed.vfsd, replacedVFSD) return nil } @@ -1031,8 +1088,15 @@ func (fs *filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error return fs.unlinkAt(ctx, rp, false /* dir */) } +// BoundEndpointAt implements FilesystemImpl.BoundEndpointAt. +// +// TODO(gvisor.dev/issue/1476): Implement BoundEndpointAt. +func (fs *filesystem) BoundEndpointAt(ctx context.Context, rp *vfs.ResolvingPath) (transport.BoundEndpoint, error) { + return nil, syserror.ECONNREFUSED +} + // ListxattrAt implements vfs.FilesystemImpl.ListxattrAt. -func (fs *filesystem) ListxattrAt(ctx context.Context, rp *vfs.ResolvingPath) ([]string, error) { +func (fs *filesystem) ListxattrAt(ctx context.Context, rp *vfs.ResolvingPath, size uint64) ([]string, error) { var ds *[]*dentry fs.renameMu.RLock() defer fs.renameMuRUnlockAndCheckCaching(&ds) @@ -1040,11 +1104,11 @@ func (fs *filesystem) ListxattrAt(ctx context.Context, rp *vfs.ResolvingPath) ([ if err != nil { return nil, err } - return d.listxattr(ctx) + return d.listxattr(ctx, rp.Credentials(), size) } // GetxattrAt implements vfs.FilesystemImpl.GetxattrAt. -func (fs *filesystem) GetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) (string, error) { +func (fs *filesystem) GetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.GetxattrOptions) (string, error) { var ds *[]*dentry fs.renameMu.RLock() defer fs.renameMuRUnlockAndCheckCaching(&ds) @@ -1052,7 +1116,7 @@ func (fs *filesystem) GetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, nam if err != nil { return "", err } - return d.getxattr(ctx, name) + return d.getxattr(ctx, rp.Credentials(), &opts) } // SetxattrAt implements vfs.FilesystemImpl.SetxattrAt. @@ -1064,7 +1128,7 @@ func (fs *filesystem) SetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opt if err != nil { return err } - return d.setxattr(ctx, &opts) + return d.setxattr(ctx, rp.Credentials(), &opts) } // RemovexattrAt implements vfs.FilesystemImpl.RemovexattrAt. @@ -1076,12 +1140,12 @@ func (fs *filesystem) RemovexattrAt(ctx context.Context, rp *vfs.ResolvingPath, if err != nil { return err } - return d.removexattr(ctx, name) + return d.removexattr(ctx, rp.Credentials(), name) } // PrependPath implements vfs.FilesystemImpl.PrependPath. func (fs *filesystem) PrependPath(ctx context.Context, vfsroot, vd vfs.VirtualDentry, b *fspath.Builder) error { fs.renameMu.RLock() defer fs.renameMu.RUnlock() - return vfs.GenericPrependPath(vfsroot, vd, b) + return genericPrependPath(vfsroot, vd.Mount(), vd.Dentry().Impl().(*dentry), b) } diff --git a/pkg/sentry/fsimpl/gofer/gofer.go b/pkg/sentry/fsimpl/gofer/gofer.go index d0552bd99..293df2545 100644 --- a/pkg/sentry/fsimpl/gofer/gofer.go +++ b/pkg/sentry/fsimpl/gofer/gofer.go @@ -34,6 +34,7 @@ package gofer import ( "fmt" "strconv" + "strings" "sync" "sync/atomic" "syscall" @@ -44,6 +45,7 @@ import ( "gvisor.dev/gvisor/pkg/p9" "gvisor.dev/gvisor/pkg/sentry/fs/fsutil" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" "gvisor.dev/gvisor/pkg/sentry/memmap" "gvisor.dev/gvisor/pkg/sentry/pgalloc" "gvisor.dev/gvisor/pkg/sentry/vfs" @@ -52,6 +54,9 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) +// Name is the default filesystem name. +const Name = "9p" + // FilesystemType implements vfs.FilesystemType. type FilesystemType struct{} @@ -69,6 +74,9 @@ type filesystem struct { // client is the client used by this filesystem. client is immutable. client *p9.Client + // clock is a realtime clock used to set timestamps in file operations. + clock ktime.Clock + // uid and gid are the effective KUID and KGID of the filesystem's creator, // and are used as the owner and group for files that don't specify one. // uid and gid are immutable. @@ -196,6 +204,11 @@ const ( InteropModeShared ) +// Name implements vfs.FilesystemType.Name. +func (FilesystemType) Name() string { + return Name +} + // GetFilesystem implements vfs.FilesystemType.GetFilesystem. func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { mfp := pgalloc.MemoryFileProviderFromContext(ctx) @@ -368,10 +381,11 @@ func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virt uid: creds.EffectiveKUID, gid: creds.EffectiveKGID, client: client, + clock: ktime.RealtimeClockFromContext(ctx), dentries: make(map[*dentry]struct{}), specialFileFDs: make(map[*specialFileFD]struct{}), } - fs.vfsfs.Init(vfsObj, fs) + fs.vfsfs.Init(vfsObj, &fstype, fs) // Construct the root dentry. root, err := fs.newDentry(ctx, attachFile, qid, attrMask, &attr) @@ -431,12 +445,23 @@ type dentry struct { // refs is the reference count. Each dentry holds a reference on its // parent, even if disowned. refs is accessed using atomic memory - // operations. + // operations. When refs reaches 0, the dentry may be added to the cache or + // destroyed. If refs==-1 the dentry has already been destroyed. refs int64 // fs is the owning filesystem. fs is immutable. fs *filesystem + // parent is this dentry's parent directory. Each dentry holds a reference + // on its parent. If this dentry is a filesystem root, parent is nil. + // parent is protected by filesystem.renameMu. + parent *dentry + + // name is the name of this dentry in its parent. If this dentry is a + // filesystem root, name is the empty string. name is protected by + // filesystem.renameMu. + name string + // We don't support hard links, so each dentry maps 1:1 to an inode. // file is the unopened p9.File that backs this dentry. file is immutable. @@ -454,10 +479,15 @@ type dentry struct { dirMu sync.Mutex - // If this dentry represents a directory, and InteropModeShared is not in - // effect, negativeChildren is a set of child names in this directory that - // are known not to exist. negativeChildren is protected by dirMu. - negativeChildren map[string]struct{} + // If this dentry represents a directory, children contains: + // + // - Mappings of child filenames to dentries representing those children. + // + // - Mappings of child filenames that are known not to exist to nil + // dentries (only if InteropModeShared is not in effect). + // + // children is protected by dirMu. + children map[string]*dentry // If this dentry represents a directory, InteropModeShared is not in // effect, and dirents is not nil, it is a cache of all entries in the @@ -482,6 +512,11 @@ type dentry struct { // locked to mutate it). size uint64 + // nlink counts the number of hard links to this dentry. It's updated and + // accessed using atomic operations. It's not protected by metadataMu like the + // other metadata fields. + nlink uint32 + mapsMu sync.Mutex // If this dentry represents a regular file, mappings tracks mappings of @@ -601,6 +636,9 @@ func (fs *filesystem) newDentry(ctx context.Context, file p9file, qid p9.QID, ma if mask.BTime { d.btime = dentryTimestampFromP9(attr.BTimeSeconds, attr.BTimeNanoSeconds) } + if mask.NLink { + d.nlink = uint32(attr.NLink) + } d.vfsd.Init(d) fs.syncMu.Lock() @@ -642,6 +680,9 @@ func (d *dentry) updateFromP9Attrs(mask p9.AttrMask, attr *p9.Attr) { if mask.BTime { atomic.StoreInt64(&d.btime, dentryTimestampFromP9(attr.BTimeSeconds, attr.BTimeNanoSeconds)) } + if mask.NLink { + atomic.StoreUint32(&d.nlink, uint32(attr.NLink)) + } if mask.Size { d.dataMu.Lock() atomic.StoreUint64(&d.size, attr.Size) @@ -684,10 +725,7 @@ func (d *dentry) fileType() uint32 { func (d *dentry) statTo(stat *linux.Statx) { stat.Mask = linux.STATX_TYPE | linux.STATX_MODE | linux.STATX_NLINK | linux.STATX_UID | linux.STATX_GID | linux.STATX_ATIME | linux.STATX_MTIME | linux.STATX_CTIME | linux.STATX_INO | linux.STATX_SIZE | linux.STATX_BLOCKS | linux.STATX_BTIME stat.Blksize = atomic.LoadUint32(&d.blockSize) - stat.Nlink = 1 - if d.isDir() { - stat.Nlink = 2 - } + stat.Nlink = atomic.LoadUint32(&d.nlink) stat.UID = atomic.LoadUint32(&d.uid) stat.GID = atomic.LoadUint32(&d.gid) stat.Mode = uint16(atomic.LoadUint32(&d.mode)) @@ -700,7 +738,7 @@ func (d *dentry) statTo(stat *linux.Statx) { stat.Btime = statxTimestampFromDentry(atomic.LoadInt64(&d.btime)) stat.Ctime = statxTimestampFromDentry(atomic.LoadInt64(&d.ctime)) stat.Mtime = statxTimestampFromDentry(atomic.LoadInt64(&d.mtime)) - // TODO(jamieliu): device number + // TODO(gvisor.dev/issue/1198): device number } func (d *dentry) setStat(ctx context.Context, creds *auth.Credentials, stat *linux.Statx, mnt *vfs.Mount) error { @@ -710,7 +748,8 @@ func (d *dentry) setStat(ctx context.Context, creds *auth.Credentials, stat *lin if stat.Mask&^(linux.STATX_MODE|linux.STATX_UID|linux.STATX_GID|linux.STATX_ATIME|linux.STATX_MTIME|linux.STATX_SIZE) != 0 { return syserror.EPERM } - if err := vfs.CheckSetStat(creds, stat, uint16(atomic.LoadUint32(&d.mode))&^linux.S_IFMT, auth.KUID(atomic.LoadUint32(&d.uid)), auth.KGID(atomic.LoadUint32(&d.gid))); err != nil { + mode := linux.FileMode(atomic.LoadUint32(&d.mode)) + if err := vfs.CheckSetStat(ctx, creds, stat, mode, auth.KUID(atomic.LoadUint32(&d.uid)), auth.KGID(atomic.LoadUint32(&d.gid))); err != nil { return err } if err := mnt.CheckBeginWrite(); err != nil { @@ -762,10 +801,7 @@ func (d *dentry) setStat(ctx context.Context, creds *auth.Credentials, stat *lin // data, so there's no cache to truncate either.) return nil } - now, haveNow := nowFromContext(ctx) - if !haveNow { - ctx.Warningf("gofer.dentry.setStat: current time not available") - } + now := d.fs.clock.Now().Nanoseconds() if stat.Mask&linux.STATX_MODE != 0 { atomic.StoreUint32(&d.mode, d.fileType()|uint32(stat.Mode)) } @@ -777,25 +813,19 @@ func (d *dentry) setStat(ctx context.Context, creds *auth.Credentials, stat *lin } if setLocalAtime { if stat.Atime.Nsec == linux.UTIME_NOW { - if haveNow { - atomic.StoreInt64(&d.atime, now) - } + atomic.StoreInt64(&d.atime, now) } else { atomic.StoreInt64(&d.atime, dentryTimestampFromStatx(stat.Atime)) } } if setLocalMtime { if stat.Mtime.Nsec == linux.UTIME_NOW { - if haveNow { - atomic.StoreInt64(&d.mtime, now) - } + atomic.StoreInt64(&d.mtime, now) } else { atomic.StoreInt64(&d.mtime, dentryTimestampFromStatx(stat.Mtime)) } } - if haveNow { - atomic.StoreInt64(&d.ctime, now) - } + atomic.StoreInt64(&d.ctime, now) if stat.Mask&linux.STATX_SIZE != 0 { d.dataMu.Lock() oldSize := d.size @@ -832,8 +862,8 @@ func (d *dentry) setStat(ctx context.Context, creds *auth.Credentials, stat *lin return nil } -func (d *dentry) checkPermissions(creds *auth.Credentials, ats vfs.AccessTypes, isDir bool) error { - return vfs.GenericCheckPermissions(creds, ats, isDir, uint16(atomic.LoadUint32(&d.mode))&0777, auth.KUID(atomic.LoadUint32(&d.uid)), auth.KGID(atomic.LoadUint32(&d.gid))) +func (d *dentry) checkPermissions(creds *auth.Credentials, ats vfs.AccessTypes) error { + return vfs.GenericCheckPermissions(creds, ats, linux.FileMode(atomic.LoadUint32(&d.mode)), auth.KUID(atomic.LoadUint32(&d.uid)), auth.KGID(atomic.LoadUint32(&d.gid))) } // IncRef implements vfs.DentryImpl.IncRef. @@ -847,7 +877,7 @@ func (d *dentry) IncRef() { func (d *dentry) TryIncRef() bool { for { refs := atomic.LoadInt64(&d.refs) - if refs == 0 { + if refs <= 0 { return false } if atomic.CompareAndSwapInt64(&d.refs, refs, refs+1) { @@ -870,13 +900,20 @@ func (d *dentry) DecRef() { // checkCachingLocked should be called after d's reference count becomes 0 or it // becomes disowned. // +// It may be called on a destroyed dentry. For example, +// renameMu[R]UnlockAndCheckCaching may call checkCachingLocked multiple times +// for the same dentry when the dentry is visited more than once in the same +// operation. One of the calls may destroy the dentry, so subsequent calls will +// do nothing. +// // Preconditions: d.fs.renameMu must be locked for writing. func (d *dentry) checkCachingLocked() { // Dentries with a non-zero reference count must be retained. (The only way // to obtain a reference on a dentry with zero references is via path // resolution, which requires renameMu, so if d.refs is zero then it will // remain zero while we hold renameMu for writing.) - if atomic.LoadInt64(&d.refs) != 0 { + refs := atomic.LoadInt64(&d.refs) + if refs > 0 { if d.cached { d.fs.cachedDentries.Remove(d) d.fs.cachedDentriesLen-- @@ -884,9 +921,13 @@ func (d *dentry) checkCachingLocked() { } return } - // Non-child dentries with zero references are no longer reachable by path - // resolution and should be dropped immediately. - if d.vfsd.Parent() == nil || d.vfsd.IsDisowned() { + if refs == -1 { + // Dentry has already been destroyed. + return + } + // Deleted and invalidated dentries with zero references are no longer + // reachable by path resolution and should be dropped immediately. + if d.vfsd.IsDead() { if d.cached { d.fs.cachedDentries.Remove(d) d.fs.cachedDentriesLen-- @@ -911,34 +952,45 @@ func (d *dentry) checkCachingLocked() { d.fs.cachedDentries.Remove(victim) d.fs.cachedDentriesLen-- victim.cached = false - // victim.refs may have become non-zero from an earlier path - // resolution since it was inserted into fs.cachedDentries; see - // dentry.incRefLocked(). Either way, we brought - // fs.cachedDentriesLen back down to fs.opts.maxCachedDentries, so - // we don't loop. + // victim.refs may have become non-zero from an earlier path resolution + // since it was inserted into fs.cachedDentries. if atomic.LoadInt64(&victim.refs) == 0 { - if victimParentVFSD := victim.vfsd.Parent(); victimParentVFSD != nil { - victimParent := victimParentVFSD.Impl().(*dentry) - victimParent.dirMu.Lock() - if !victim.vfsd.IsDisowned() { - // victim can't be a mount point (in any mount - // namespace), since VFS holds references on mount - // points. - d.fs.vfsfs.VirtualFilesystem().ForceDeleteDentry(&victim.vfsd) + if victim.parent != nil { + victim.parent.dirMu.Lock() + if !victim.vfsd.IsDead() { + // Note that victim can't be a mount point (in any mount + // namespace), since VFS holds references on mount points. + d.fs.vfsfs.VirtualFilesystem().InvalidateDentry(&victim.vfsd) + delete(victim.parent.children, victim.name) // We're only deleting the dentry, not the file it // represents, so we don't need to update // victimParent.dirents etc. } - victimParent.dirMu.Unlock() + victim.parent.dirMu.Unlock() } victim.destroyLocked() } + // Whether or not victim was destroyed, we brought fs.cachedDentriesLen + // back down to fs.opts.maxCachedDentries, so we don't loop. } } +// destroyLocked destroys the dentry. It may flushes dirty pages from cache, +// close p9 file and remove reference on parent dentry. +// // Preconditions: d.fs.renameMu must be locked for writing. d.refs == 0. d is // not a child dentry. func (d *dentry) destroyLocked() { + switch atomic.LoadInt64(&d.refs) { + case 0: + // Mark the dentry destroyed. + atomic.StoreInt64(&d.refs, -1) + case -1: + panic("dentry.destroyLocked() called on already destroyed dentry") + default: + panic("dentry.destroyLocked() called with references on the dentry") + } + ctx := context.Background() d.handleMu.Lock() if !d.handle.file.isNil() { @@ -958,17 +1010,19 @@ func (d *dentry) destroyLocked() { d.handle.close(ctx) } d.handleMu.Unlock() - d.file.close(ctx) + if !d.file.isNil() { + d.file.close(ctx) + d.file = p9file{} + } // Remove d from the set of all dentries. d.fs.syncMu.Lock() delete(d.fs.dentries, d) d.fs.syncMu.Unlock() - // Drop the reference held by d on its parent. - if parentVFSD := d.vfsd.Parent(); parentVFSD != nil { - parent := parentVFSD.Impl().(*dentry) - // This is parent.DecRef() without recursive locking of d.fs.renameMu. - if refs := atomic.AddInt64(&parent.refs, -1); refs == 0 { - parent.checkCachingLocked() + // Drop the reference held by d on its parent without recursively locking + // d.fs.renameMu. + if d.parent != nil { + if refs := atomic.AddInt64(&d.parent.refs, -1); refs == 0 { + d.parent.checkCachingLocked() } else if refs < 0 { panic("gofer.dentry.DecRef() called without holding a reference") } @@ -983,21 +1037,50 @@ func (d *dentry) setDeleted() { atomic.StoreUint32(&d.deleted, 1) } -func (d *dentry) listxattr(ctx context.Context) ([]string, error) { - return nil, syserror.ENOTSUP +// We only support xattrs prefixed with "user." (see b/148380782). Currently, +// there is no need to expose any other xattrs through a gofer. +func (d *dentry) listxattr(ctx context.Context, creds *auth.Credentials, size uint64) ([]string, error) { + xattrMap, err := d.file.listXattr(ctx, size) + if err != nil { + return nil, err + } + xattrs := make([]string, 0, len(xattrMap)) + for x := range xattrMap { + if strings.HasPrefix(x, linux.XATTR_USER_PREFIX) { + xattrs = append(xattrs, x) + } + } + return xattrs, nil } -func (d *dentry) getxattr(ctx context.Context, name string) (string, error) { - // TODO(jamieliu): add vfs.GetxattrOptions.Size - return d.file.getXattr(ctx, name, linux.XATTR_SIZE_MAX) +func (d *dentry) getxattr(ctx context.Context, creds *auth.Credentials, opts *vfs.GetxattrOptions) (string, error) { + if err := d.checkPermissions(creds, vfs.MayRead); err != nil { + return "", err + } + if !strings.HasPrefix(opts.Name, linux.XATTR_USER_PREFIX) { + return "", syserror.EOPNOTSUPP + } + return d.file.getXattr(ctx, opts.Name, opts.Size) } -func (d *dentry) setxattr(ctx context.Context, opts *vfs.SetxattrOptions) error { +func (d *dentry) setxattr(ctx context.Context, creds *auth.Credentials, opts *vfs.SetxattrOptions) error { + if err := d.checkPermissions(creds, vfs.MayWrite); err != nil { + return err + } + if !strings.HasPrefix(opts.Name, linux.XATTR_USER_PREFIX) { + return syserror.EOPNOTSUPP + } return d.file.setXattr(ctx, opts.Name, opts.Value, opts.Flags) } -func (d *dentry) removexattr(ctx context.Context, name string) error { - return syserror.ENOTSUP +func (d *dentry) removexattr(ctx context.Context, creds *auth.Credentials, name string) error { + if err := d.checkPermissions(creds, vfs.MayWrite); err != nil { + return err + } + if !strings.HasPrefix(name, linux.XATTR_USER_PREFIX) { + return syserror.EOPNOTSUPP + } + return d.file.removeXattr(ctx, name) } // Preconditions: d.isRegularFile() || d.isDirectory(). @@ -1042,13 +1125,13 @@ func (d *dentry) ensureSharedHandle(ctx context.Context, read, write, trunc bool // using the old file descriptor, preventing us from safely // closing it. We could handle this by invalidating existing // memmap.Translations, but this is expensive. Instead, use - // dup2() to make the old file descriptor refer to the new file + // dup3 to make the old file descriptor refer to the new file // description, then close the new file descriptor (which is no // longer needed). Racing callers may use the old or new file // description, but this doesn't matter since they refer to the // same file (unless d.fs.opts.overlayfsStaleRead is true, // which we handle separately). - if err := syscall.Dup2(int(h.fd), int(d.handle.fd)); err != nil { + if err := syscall.Dup3(int(h.fd), int(d.handle.fd), syscall.O_CLOEXEC); err != nil { d.handleMu.Unlock() ctx.Warningf("gofer.dentry.ensureSharedHandle: failed to dup fd %d to fd %d: %v", h.fd, d.handle.fd, err) h.close(ctx) @@ -1091,6 +1174,26 @@ func (d *dentry) ensureSharedHandle(ctx context.Context, read, write, trunc bool return nil } +// incLinks increments link count. +// +// Preconditions: d.nlink != 0 && d.nlink < math.MaxUint32. +func (d *dentry) incLinks() { + v := atomic.AddUint32(&d.nlink, 1) + if v < 2 { + panic(fmt.Sprintf("dentry.nlink is invalid (was 0 or overflowed): %d", v)) + } +} + +// decLinks decrements link count. +// +// Preconditions: d.nlink > 1. +func (d *dentry) decLinks() { + v := atomic.AddUint32(&d.nlink, ^uint32(0)) + if v == 0 { + panic(fmt.Sprintf("dentry.nlink must be greater than 0: %d", v)) + } +} + // fileDescription is embedded by gofer implementations of // vfs.FileDescriptionImpl. type fileDescription struct { @@ -1109,7 +1212,8 @@ func (fd *fileDescription) dentry() *dentry { // Stat implements vfs.FileDescriptionImpl.Stat. func (fd *fileDescription) Stat(ctx context.Context, opts vfs.StatOptions) (linux.Statx, error) { d := fd.dentry() - if d.fs.opts.interop == InteropModeShared && opts.Mask&(linux.STATX_MODE|linux.STATX_UID|linux.STATX_GID|linux.STATX_ATIME|linux.STATX_MTIME|linux.STATX_CTIME|linux.STATX_SIZE|linux.STATX_BLOCKS|linux.STATX_BTIME) != 0 && opts.Sync != linux.AT_STATX_DONT_SYNC { + const validMask = uint32(linux.STATX_MODE | linux.STATX_UID | linux.STATX_GID | linux.STATX_ATIME | linux.STATX_MTIME | linux.STATX_CTIME | linux.STATX_SIZE | linux.STATX_BLOCKS | linux.STATX_BTIME) + if d.fs.opts.interop == InteropModeShared && opts.Mask&(validMask) != 0 && opts.Sync != linux.AT_STATX_DONT_SYNC { // TODO(jamieliu): Use specialFileFD.handle.file for the getattr if // available? if err := d.updateFromGetattr(ctx); err != nil { @@ -1127,21 +1231,21 @@ func (fd *fileDescription) SetStat(ctx context.Context, opts vfs.SetStatOptions) } // Listxattr implements vfs.FileDescriptionImpl.Listxattr. -func (fd *fileDescription) Listxattr(ctx context.Context) ([]string, error) { - return fd.dentry().listxattr(ctx) +func (fd *fileDescription) Listxattr(ctx context.Context, size uint64) ([]string, error) { + return fd.dentry().listxattr(ctx, auth.CredentialsFromContext(ctx), size) } // Getxattr implements vfs.FileDescriptionImpl.Getxattr. -func (fd *fileDescription) Getxattr(ctx context.Context, name string) (string, error) { - return fd.dentry().getxattr(ctx, name) +func (fd *fileDescription) Getxattr(ctx context.Context, opts vfs.GetxattrOptions) (string, error) { + return fd.dentry().getxattr(ctx, auth.CredentialsFromContext(ctx), &opts) } // Setxattr implements vfs.FileDescriptionImpl.Setxattr. func (fd *fileDescription) Setxattr(ctx context.Context, opts vfs.SetxattrOptions) error { - return fd.dentry().setxattr(ctx, &opts) + return fd.dentry().setxattr(ctx, auth.CredentialsFromContext(ctx), &opts) } // Removexattr implements vfs.FileDescriptionImpl.Removexattr. func (fd *fileDescription) Removexattr(ctx context.Context, name string) error { - return fd.dentry().removexattr(ctx, name) + return fd.dentry().removexattr(ctx, auth.CredentialsFromContext(ctx), name) } diff --git a/pkg/sentry/fsimpl/gofer/gofer_test.go b/pkg/sentry/fsimpl/gofer/gofer_test.go new file mode 100644 index 000000000..4041fb252 --- /dev/null +++ b/pkg/sentry/fsimpl/gofer/gofer_test.go @@ -0,0 +1,63 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package gofer + +import ( + "sync/atomic" + "testing" + + "gvisor.dev/gvisor/pkg/p9" + "gvisor.dev/gvisor/pkg/sentry/contexttest" +) + +func TestDestroyIdempotent(t *testing.T) { + fs := filesystem{ + dentries: make(map[*dentry]struct{}), + opts: filesystemOptions{ + // Test relies on no dentry being held in the cache. + maxCachedDentries: 0, + }, + } + + ctx := contexttest.Context(t) + attr := &p9.Attr{ + Mode: p9.ModeRegular, + } + mask := p9.AttrMask{ + Mode: true, + Size: true, + } + parent, err := fs.newDentry(ctx, p9file{}, p9.QID{}, mask, attr) + if err != nil { + t.Fatalf("fs.newDentry(): %v", err) + } + + child, err := fs.newDentry(ctx, p9file{}, p9.QID{}, mask, attr) + if err != nil { + t.Fatalf("fs.newDentry(): %v", err) + } + parent.cacheNewChildLocked(child, "child") + + child.checkCachingLocked() + if got := atomic.LoadInt64(&child.refs); got != -1 { + t.Fatalf("child.refs=%d, want: -1", got) + } + // Parent will also be destroyed when child reference is removed. + if got := atomic.LoadInt64(&parent.refs); got != -1 { + t.Fatalf("parent.refs=%d, want: -1", got) + } + child.checkCachingLocked() + child.checkCachingLocked() +} diff --git a/pkg/sentry/fsimpl/gofer/p9file.go b/pkg/sentry/fsimpl/gofer/p9file.go index 755ac2985..87f0b877f 100644 --- a/pkg/sentry/fsimpl/gofer/p9file.go +++ b/pkg/sentry/fsimpl/gofer/p9file.go @@ -85,6 +85,13 @@ func (f p9file) setAttr(ctx context.Context, valid p9.SetAttrMask, attr p9.SetAt return err } +func (f p9file) listXattr(ctx context.Context, size uint64) (map[string]struct{}, error) { + ctx.UninterruptibleSleepStart(false) + xattrs, err := f.file.ListXattr(size) + ctx.UninterruptibleSleepFinish(false) + return xattrs, err +} + func (f p9file) getXattr(ctx context.Context, name string, size uint64) (string, error) { ctx.UninterruptibleSleepStart(false) val, err := f.file.GetXattr(name, size) @@ -99,6 +106,13 @@ func (f p9file) setXattr(ctx context.Context, name, value string, flags uint32) return err } +func (f p9file) removeXattr(ctx context.Context, name string) error { + ctx.UninterruptibleSleepStart(false) + err := f.file.RemoveXattr(name) + ctx.UninterruptibleSleepFinish(false) + return err +} + func (f p9file) allocate(ctx context.Context, mode p9.AllocateMode, offset, length uint64) error { ctx.UninterruptibleSleepStart(false) err := f.file.Allocate(mode, offset, length) diff --git a/pkg/sentry/fsimpl/gofer/regular_file.go b/pkg/sentry/fsimpl/gofer/regular_file.go index 8e11e06b3..857f7c74e 100644 --- a/pkg/sentry/fsimpl/gofer/regular_file.go +++ b/pkg/sentry/fsimpl/gofer/regular_file.go @@ -104,7 +104,7 @@ func (fd *regularFileFD) PRead(ctx context.Context, dst usermem.IOSequence, offs putDentryReadWriter(rw) if d.fs.opts.interop != InteropModeShared { // Compare Linux's mm/filemap.c:do_generic_file_read() => file_accessed(). - d.touchAtime(ctx, fd.vfsfd.Mount()) + d.touchAtime(fd.vfsfd.Mount()) } return n, err } @@ -126,6 +126,11 @@ func (fd *regularFileFD) PWrite(ctx context.Context, src usermem.IOSequence, off if opts.Flags != 0 { return 0, syserror.EOPNOTSUPP } + limit, err := vfs.CheckLimit(ctx, offset, src.NumBytes()) + if err != nil { + return 0, err + } + src = src.TakeFirst64(limit) d := fd.dentry() d.metadataMu.Lock() @@ -134,10 +139,7 @@ func (fd *regularFileFD) PWrite(ctx context.Context, src usermem.IOSequence, off // Compare Linux's mm/filemap.c:__generic_file_write_iter() => // file_update_time(). This is d.touchCMtime(), but without locking // d.metadataMu (recursively). - if now, ok := nowFromContext(ctx); ok { - atomic.StoreInt64(&d.mtime, now) - atomic.StoreInt64(&d.ctime, now) - } + d.touchCMtimeLocked() } if fd.vfsfd.StatusFlags()&linux.O_DIRECT != 0 { // Write dirty cached pages that will be touched by the write back to @@ -361,8 +363,15 @@ func (rw *dentryReadWriter) WriteFromBlocks(srcs safemem.BlockSeq) (uint64, erro rw.d.handleMu.RLock() if (rw.d.handle.fd >= 0 && !rw.d.fs.opts.forcePageCache) || rw.d.fs.opts.interop == InteropModeShared || rw.direct { n, err := rw.d.handle.writeFromBlocksAt(rw.ctx, srcs, rw.off) - rw.d.handleMu.RUnlock() rw.off += n + rw.d.dataMu.Lock() + if rw.off > rw.d.size { + atomic.StoreUint64(&rw.d.size, rw.off) + // The remote file's size will implicitly be extended to the correct + // value when we write back to it. + } + rw.d.dataMu.Unlock() + rw.d.handleMu.RUnlock() return n, err } @@ -571,6 +580,8 @@ func (fd *regularFileFD) ConfigureMMap(ctx context.Context, opts *memmap.MMapOpt default: panic(fmt.Sprintf("unknown InteropMode %v", d.fs.opts.interop)) } + // After this point, d may be used as a memmap.Mappable. + d.pf.hostFileMapperInitOnce.Do(d.pf.hostFileMapper.Init) return vfs.GenericConfigureMMap(&fd.vfsfd, d, opts) } @@ -799,6 +810,9 @@ type dentryPlatformFile struct { // If this dentry represents a regular file, and handle.fd >= 0, // hostFileMapper caches mappings of handle.fd. hostFileMapper fsutil.HostFileMapper + + // hostFileMapperInitOnce is used to lazily initialize hostFileMapper. + hostFileMapperInitOnce sync.Once } // IncRef implements platform.File.IncRef. diff --git a/pkg/sentry/fsimpl/gofer/special_file.go b/pkg/sentry/fsimpl/gofer/special_file.go index 08c691c47..507e0e276 100644 --- a/pkg/sentry/fsimpl/gofer/special_file.go +++ b/pkg/sentry/fsimpl/gofer/special_file.go @@ -76,7 +76,7 @@ func (fd *specialFileFD) PRead(ctx context.Context, dst usermem.IOSequence, offs // hold here since specialFileFD doesn't client-cache data. Just buffer the // read instead. if d := fd.dentry(); d.fs.opts.interop != InteropModeShared { - d.touchAtime(ctx, fd.vfsfd.Mount()) + d.touchAtime(fd.vfsfd.Mount()) } buf := make([]byte, dst.NumBytes()) n, err := fd.handle.readToBlocksAt(ctx, safemem.BlockSeqOf(safemem.BlockFromSafeSlice(buf)), uint64(offset)) @@ -107,9 +107,17 @@ func (fd *specialFileFD) PWrite(ctx context.Context, src usermem.IOSequence, off return 0, syserror.EOPNOTSUPP } + if fd.dentry().fileType() == linux.S_IFREG { + limit, err := vfs.CheckLimit(ctx, offset, src.NumBytes()) + if err != nil { + return 0, err + } + src = src.TakeFirst64(limit) + } + // Do a buffered write. See rationale in PRead. if d := fd.dentry(); d.fs.opts.interop != InteropModeShared { - d.touchCMtime(ctx) + d.touchCMtime() } buf := make([]byte, src.NumBytes()) // Don't do partial writes if we get a partial read from src. diff --git a/pkg/sentry/fsimpl/gofer/symlink.go b/pkg/sentry/fsimpl/gofer/symlink.go index adf43be60..2ec819f86 100644 --- a/pkg/sentry/fsimpl/gofer/symlink.go +++ b/pkg/sentry/fsimpl/gofer/symlink.go @@ -27,7 +27,7 @@ func (d *dentry) isSymlink() bool { // Precondition: d.isSymlink(). func (d *dentry) readlink(ctx context.Context, mnt *vfs.Mount) (string, error) { if d.fs.opts.interop != InteropModeShared { - d.touchAtime(ctx, mnt) + d.touchAtime(mnt) d.dataMu.Lock() if d.haveTarget { target := d.target diff --git a/pkg/sentry/fsimpl/gofer/time.go b/pkg/sentry/fsimpl/gofer/time.go index 7598ec6a8..2608e7e1d 100644 --- a/pkg/sentry/fsimpl/gofer/time.go +++ b/pkg/sentry/fsimpl/gofer/time.go @@ -18,8 +18,6 @@ import ( "sync/atomic" "gvisor.dev/gvisor/pkg/abi/linux" - "gvisor.dev/gvisor/pkg/context" - ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" "gvisor.dev/gvisor/pkg/sentry/vfs" ) @@ -38,23 +36,12 @@ func statxTimestampFromDentry(ns int64) linux.StatxTimestamp { } } -func nowFromContext(ctx context.Context) (int64, bool) { - if clock := ktime.RealtimeClockFromContext(ctx); clock != nil { - return clock.Now().Nanoseconds(), true - } - return 0, false -} - // Preconditions: fs.interop != InteropModeShared. -func (d *dentry) touchAtime(ctx context.Context, mnt *vfs.Mount) { +func (d *dentry) touchAtime(mnt *vfs.Mount) { if err := mnt.CheckBeginWrite(); err != nil { return } - now, ok := nowFromContext(ctx) - if !ok { - mnt.EndWrite() - return - } + now := d.fs.clock.Now().Nanoseconds() d.metadataMu.Lock() atomic.StoreInt64(&d.atime, now) d.metadataMu.Unlock() @@ -63,13 +50,25 @@ func (d *dentry) touchAtime(ctx context.Context, mnt *vfs.Mount) { // Preconditions: fs.interop != InteropModeShared. The caller has successfully // called vfs.Mount.CheckBeginWrite(). -func (d *dentry) touchCMtime(ctx context.Context) { - now, ok := nowFromContext(ctx) - if !ok { - return - } +func (d *dentry) touchCtime() { + now := d.fs.clock.Now().Nanoseconds() + d.metadataMu.Lock() + atomic.StoreInt64(&d.ctime, now) + d.metadataMu.Unlock() +} + +// Preconditions: fs.interop != InteropModeShared. The caller has successfully +// called vfs.Mount.CheckBeginWrite(). +func (d *dentry) touchCMtime() { + now := d.fs.clock.Now().Nanoseconds() d.metadataMu.Lock() atomic.StoreInt64(&d.mtime, now) atomic.StoreInt64(&d.ctime, now) d.metadataMu.Unlock() } + +func (d *dentry) touchCMtimeLocked() { + now := d.fs.clock.Now().Nanoseconds() + atomic.StoreInt64(&d.mtime, now) + atomic.StoreInt64(&d.ctime, now) +} diff --git a/pkg/sentry/fsimpl/host/BUILD b/pkg/sentry/fsimpl/host/BUILD new file mode 100644 index 000000000..82e1fb74b --- /dev/null +++ b/pkg/sentry/fsimpl/host/BUILD @@ -0,0 +1,34 @@ +load("//tools:defs.bzl", "go_library") + +licenses(["notice"]) + +go_library( + name = "host", + srcs = [ + "host.go", + "ioctl_unsafe.go", + "tty.go", + "util.go", + "util_unsafe.go", + ], + visibility = ["//pkg/sentry:internal"], + deps = [ + "//pkg/abi/linux", + "//pkg/context", + "//pkg/fd", + "//pkg/log", + "//pkg/refs", + "//pkg/safemem", + "//pkg/sentry/arch", + "//pkg/sentry/fsimpl/kernfs", + "//pkg/sentry/kernel", + "//pkg/sentry/kernel/auth", + "//pkg/sentry/memmap", + "//pkg/sentry/unimpl", + "//pkg/sentry/vfs", + "//pkg/sync", + "//pkg/syserror", + "//pkg/usermem", + "@org_golang_x_sys//unix:go_default_library", + ], +) diff --git a/pkg/sentry/fsimpl/host/host.go b/pkg/sentry/fsimpl/host/host.go new file mode 100644 index 000000000..fe14476f1 --- /dev/null +++ b/pkg/sentry/fsimpl/host/host.go @@ -0,0 +1,639 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// Package host provides a filesystem implementation for host files imported as +// file descriptors. +package host + +import ( + "errors" + "fmt" + "math" + "syscall" + + "golang.org/x/sys/unix" + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/fd" + "gvisor.dev/gvisor/pkg/log" + "gvisor.dev/gvisor/pkg/refs" + "gvisor.dev/gvisor/pkg/safemem" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/memmap" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +// filesystemType implements vfs.FilesystemType. +type filesystemType struct{} + +// GetFilesystem implements FilesystemType.GetFilesystem. +func (filesystemType) GetFilesystem(context.Context, *vfs.VirtualFilesystem, *auth.Credentials, string, vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { + panic("cannot instaniate a host filesystem") +} + +// Name implements FilesystemType.Name. +func (filesystemType) Name() string { + return "none" +} + +// filesystem implements vfs.FilesystemImpl. +type filesystem struct { + kernfs.Filesystem +} + +// NewMount returns a new disconnected mount in vfsObj that may be passed to ImportFD. +func NewMount(vfsObj *vfs.VirtualFilesystem) (*vfs.Mount, error) { + fs := &filesystem{} + fs.Init(vfsObj, &filesystemType{}) + vfsfs := fs.VFSFilesystem() + // NewDisconnectedMount will take an additional reference on vfsfs. + defer vfsfs.DecRef() + return vfsObj.NewDisconnectedMount(vfsfs, nil, &vfs.MountOptions{}) +} + +// ImportFD sets up and returns a vfs.FileDescription from a donated fd. +func ImportFD(mnt *vfs.Mount, hostFD int, isTTY bool) (*vfs.FileDescription, error) { + fs, ok := mnt.Filesystem().Impl().(*kernfs.Filesystem) + if !ok { + return nil, fmt.Errorf("can't import host FDs into filesystems of type %T", mnt.Filesystem().Impl()) + } + + // Retrieve metadata. + var s unix.Stat_t + if err := unix.Fstat(hostFD, &s); err != nil { + return nil, err + } + + fileMode := linux.FileMode(s.Mode) + fileType := fileMode.FileType() + + // Determine if hostFD is seekable. If not, this syscall will return ESPIPE + // (see fs/read_write.c:llseek), e.g. for pipes, sockets, and some character + // devices. + _, err := unix.Seek(hostFD, 0, linux.SEEK_CUR) + seekable := err != syserror.ESPIPE + + i := &inode{ + hostFD: hostFD, + seekable: seekable, + isTTY: isTTY, + canMap: canMap(uint32(fileType)), + ino: fs.NextIno(), + // For simplicity, set offset to 0. Technically, we should use the existing + // offset on the host if the file is seekable. + offset: 0, + } + + // Non-seekable files can't be memory mapped, assert this. + if !i.seekable && i.canMap { + panic("files that can return EWOULDBLOCK (sockets, pipes, etc.) cannot be memory mapped") + } + + d := &kernfs.Dentry{} + d.Init(i) + // i.open will take a reference on d. + defer d.DecRef() + + return i.open(d.VFSDentry(), mnt) +} + +// inode implements kernfs.Inode. +type inode struct { + kernfs.InodeNotDirectory + kernfs.InodeNotSymlink + + // When the reference count reaches zero, the host fd is closed. + refs.AtomicRefCount + + // hostFD contains the host fd that this file was originally created from, + // which must be available at time of restore. + // + // This field is initialized at creation time and is immutable. + hostFD int + + // seekable is false if the host fd points to a file representing a stream, + // e.g. a socket or a pipe. Such files are not seekable and can return + // EWOULDBLOCK for I/O operations. + // + // This field is initialized at creation time and is immutable. + seekable bool + + // isTTY is true if this file represents a TTY. + // + // This field is initialized at creation time and is immutable. + isTTY bool + + // canMap specifies whether we allow the file to be memory mapped. + // + // This field is initialized at creation time and is immutable. + canMap bool + + // ino is an inode number unique within this filesystem. + // + // This field is initialized at creation time and is immutable. + ino uint64 + + // offsetMu protects offset. + offsetMu sync.Mutex + + // offset specifies the current file offset. + offset int64 +} + +// Note that these flags may become out of date, since they can be modified +// on the host, e.g. with fcntl. +func fileFlagsFromHostFD(fd int) (int, error) { + flags, err := unix.FcntlInt(uintptr(fd), syscall.F_GETFL, 0) + if err != nil { + log.Warningf("Failed to get file flags for donated FD %d: %v", fd, err) + return 0, err + } + // TODO(gvisor.dev/issue/1672): implement behavior corresponding to these allowed flags. + flags &= syscall.O_ACCMODE | syscall.O_DIRECT | syscall.O_NONBLOCK | syscall.O_DSYNC | syscall.O_SYNC | syscall.O_APPEND + return flags, nil +} + +// CheckPermissions implements kernfs.Inode. +func (i *inode) CheckPermissions(ctx context.Context, creds *auth.Credentials, ats vfs.AccessTypes) error { + mode, uid, gid, err := i.getPermissions() + if err != nil { + return err + } + return vfs.GenericCheckPermissions(creds, ats, mode, uid, gid) +} + +// Mode implements kernfs.Inode. +func (i *inode) Mode() linux.FileMode { + mode, _, _, err := i.getPermissions() + // Retrieving the mode from the host fd using fstat(2) should not fail. + // If the syscall does not succeed, something is fundamentally wrong. + if err != nil { + panic(fmt.Sprintf("failed to retrieve mode from host fd %d: %v", i.hostFD, err)) + } + return linux.FileMode(mode) +} + +func (i *inode) getPermissions() (linux.FileMode, auth.KUID, auth.KGID, error) { + // Retrieve metadata. + var s syscall.Stat_t + if err := syscall.Fstat(i.hostFD, &s); err != nil { + return 0, 0, 0, err + } + return linux.FileMode(s.Mode), auth.KUID(s.Uid), auth.KGID(s.Gid), nil +} + +// Stat implements kernfs.Inode. +func (i *inode) Stat(_ *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) { + if opts.Mask&linux.STATX__RESERVED != 0 { + return linux.Statx{}, syserror.EINVAL + } + if opts.Sync&linux.AT_STATX_SYNC_TYPE == linux.AT_STATX_SYNC_TYPE { + return linux.Statx{}, syserror.EINVAL + } + + // Limit our host call only to known flags. + mask := opts.Mask & linux.STATX_ALL + var s unix.Statx_t + err := unix.Statx(i.hostFD, "", int(unix.AT_EMPTY_PATH|opts.Sync), int(mask), &s) + // Fallback to fstat(2), if statx(2) is not supported on the host. + // + // TODO(b/151263641): Remove fallback. + if err == syserror.ENOSYS { + return i.fstat(opts) + } else if err != nil { + return linux.Statx{}, err + } + + ls := linux.Statx{Mask: mask} + // Unconditionally fill blksize, attributes, and device numbers, as indicated + // by /include/uapi/linux/stat.h. + // + // RdevMajor/RdevMinor are left as zero, so as not to expose host device + // numbers. + // + // TODO(gvisor.dev/issue/1672): Use kernfs-specific, internally defined + // device numbers. If we use the device number from the host, it may collide + // with another sentry-internal device number. We handle device/inode + // numbers without relying on the host to prevent collisions. + ls.Blksize = s.Blksize + ls.Attributes = s.Attributes + ls.AttributesMask = s.Attributes_mask + + if mask&linux.STATX_TYPE != 0 { + ls.Mode |= s.Mode & linux.S_IFMT + } + if mask&linux.STATX_MODE != 0 { + ls.Mode |= s.Mode &^ linux.S_IFMT + } + if mask&linux.STATX_NLINK != 0 { + ls.Nlink = s.Nlink + } + if mask&linux.STATX_UID != 0 { + ls.UID = s.Uid + } + if mask&linux.STATX_GID != 0 { + ls.GID = s.Gid + } + if mask&linux.STATX_ATIME != 0 { + ls.Atime = unixToLinuxStatxTimestamp(s.Atime) + } + if mask&linux.STATX_BTIME != 0 { + ls.Btime = unixToLinuxStatxTimestamp(s.Btime) + } + if mask&linux.STATX_CTIME != 0 { + ls.Ctime = unixToLinuxStatxTimestamp(s.Ctime) + } + if mask&linux.STATX_MTIME != 0 { + ls.Mtime = unixToLinuxStatxTimestamp(s.Mtime) + } + if mask&linux.STATX_SIZE != 0 { + ls.Size = s.Size + } + if mask&linux.STATX_BLOCKS != 0 { + ls.Blocks = s.Blocks + } + + // Use our own internal inode number. + if mask&linux.STATX_INO != 0 { + ls.Ino = i.ino + } + + return ls, nil +} + +// fstat is a best-effort fallback for inode.Stat() if the host does not +// support statx(2). +// +// We ignore the mask and sync flags in opts and simply supply +// STATX_BASIC_STATS, as fstat(2) itself does not allow the specification +// of a mask or sync flags. fstat(2) does not provide any metadata +// equivalent to Statx.Attributes, Statx.AttributesMask, or Statx.Btime, so +// those fields remain empty. +func (i *inode) fstat(opts vfs.StatOptions) (linux.Statx, error) { + var s unix.Stat_t + if err := unix.Fstat(i.hostFD, &s); err != nil { + return linux.Statx{}, err + } + + // Note that rdev numbers are left as 0; do not expose host device numbers. + ls := linux.Statx{ + Mask: linux.STATX_BASIC_STATS, + Blksize: uint32(s.Blksize), + Nlink: uint32(s.Nlink), + UID: s.Uid, + GID: s.Gid, + Mode: uint16(s.Mode), + Size: uint64(s.Size), + Blocks: uint64(s.Blocks), + Atime: timespecToStatxTimestamp(s.Atim), + Ctime: timespecToStatxTimestamp(s.Ctim), + Mtime: timespecToStatxTimestamp(s.Mtim), + } + + // Use our own internal inode number. + // + // TODO(gvisor.dev/issue/1672): Use a kernfs-specific device number as well. + // If we use the device number from the host, it may collide with another + // sentry-internal device number. We handle device/inode numbers without + // relying on the host to prevent collisions. + ls.Ino = i.ino + + return ls, nil +} + +// SetStat implements kernfs.Inode. +func (i *inode) SetStat(ctx context.Context, fs *vfs.Filesystem, creds *auth.Credentials, opts vfs.SetStatOptions) error { + s := opts.Stat + + m := s.Mask + if m == 0 { + return nil + } + if m&^(linux.STATX_MODE|linux.STATX_SIZE|linux.STATX_ATIME|linux.STATX_MTIME) != 0 { + return syserror.EPERM + } + mode, uid, gid, err := i.getPermissions() + if err != nil { + return err + } + if err := vfs.CheckSetStat(ctx, creds, &s, mode.Permissions(), uid, gid); err != nil { + return err + } + + if m&linux.STATX_MODE != 0 { + if err := syscall.Fchmod(i.hostFD, uint32(s.Mode)); err != nil { + return err + } + } + if m&linux.STATX_SIZE != 0 { + if err := syscall.Ftruncate(i.hostFD, int64(s.Size)); err != nil { + return err + } + } + if m&(linux.STATX_ATIME|linux.STATX_MTIME) != 0 { + ts := [2]syscall.Timespec{ + toTimespec(s.Atime, m&linux.STATX_ATIME == 0), + toTimespec(s.Mtime, m&linux.STATX_MTIME == 0), + } + if err := setTimestamps(i.hostFD, &ts); err != nil { + return err + } + } + return nil +} + +// DecRef implements kernfs.Inode. +func (i *inode) DecRef() { + i.AtomicRefCount.DecRefWithDestructor(i.Destroy) +} + +// Destroy implements kernfs.Inode. +func (i *inode) Destroy() { + if err := unix.Close(i.hostFD); err != nil { + log.Warningf("failed to close host fd %d: %v", i.hostFD, err) + } +} + +// Open implements kernfs.Inode. +func (i *inode) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + return i.open(vfsd, rp.Mount()) +} + +func (i *inode) open(d *vfs.Dentry, mnt *vfs.Mount) (*vfs.FileDescription, error) { + mode, _, _, err := i.getPermissions() + if err != nil { + return nil, err + } + fileType := mode.FileType() + if fileType == syscall.S_IFSOCK { + if i.isTTY { + return nil, errors.New("cannot use host socket as TTY") + } + // TODO(gvisor.dev/issue/1672): support importing sockets. + return nil, errors.New("importing host sockets not supported") + } + + // TODO(gvisor.dev/issue/1672): Whitelist specific file types here, so that + // we don't allow importing arbitrary file types without proper support. + var ( + vfsfd *vfs.FileDescription + fdImpl vfs.FileDescriptionImpl + ) + if i.isTTY { + fd := &ttyFD{ + fileDescription: fileDescription{inode: i}, + termios: linux.DefaultSlaveTermios, + } + vfsfd = &fd.vfsfd + fdImpl = fd + } else { + // For simplicity, set offset to 0. Technically, we should + // only set to 0 on files that are not seekable (sockets, pipes, etc.), + // and use the offset from the host fd otherwise. + fd := &fileDescription{inode: i} + vfsfd = &fd.vfsfd + fdImpl = fd + } + + flags, err := fileFlagsFromHostFD(i.hostFD) + if err != nil { + return nil, err + } + + if err := vfsfd.Init(fdImpl, uint32(flags), mnt, d, &vfs.FileDescriptionOptions{}); err != nil { + return nil, err + } + return vfsfd, nil +} + +// fileDescription is embedded by host fd implementations of FileDescriptionImpl. +// +// TODO(gvisor.dev/issue/1672): Implement Waitable interface. +type fileDescription struct { + vfsfd vfs.FileDescription + vfs.FileDescriptionDefaultImpl + + // inode is vfsfd.Dentry().Impl().(*kernfs.Dentry).Inode().(*inode), but + // cached to reduce indirections and casting. fileDescription does not hold + // a reference on the inode through the inode field (since one is already + // held via the Dentry). + // + // inode is immutable after fileDescription creation. + inode *inode +} + +// SetStat implements vfs.FileDescriptionImpl. +func (f *fileDescription) SetStat(ctx context.Context, opts vfs.SetStatOptions) error { + creds := auth.CredentialsFromContext(ctx) + return f.inode.SetStat(ctx, nil, creds, opts) +} + +// Stat implements vfs.FileDescriptionImpl. +func (f *fileDescription) Stat(_ context.Context, opts vfs.StatOptions) (linux.Statx, error) { + return f.inode.Stat(nil, opts) +} + +// Release implements vfs.FileDescriptionImpl. +func (f *fileDescription) Release() { + // noop +} + +// PRead implements FileDescriptionImpl. +func (f *fileDescription) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { + i := f.inode + if !i.seekable { + return 0, syserror.ESPIPE + } + + return readFromHostFD(ctx, i.hostFD, dst, offset, opts.Flags) +} + +// Read implements FileDescriptionImpl. +func (f *fileDescription) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { + i := f.inode + if !i.seekable { + n, err := readFromHostFD(ctx, i.hostFD, dst, -1, opts.Flags) + if isBlockError(err) { + // If we got any data at all, return it as a "completed" partial read + // rather than retrying until complete. + if n != 0 { + err = nil + } else { + err = syserror.ErrWouldBlock + } + } + return n, err + } + // TODO(gvisor.dev/issue/1672): Cache pages, when forced to do so. + i.offsetMu.Lock() + n, err := readFromHostFD(ctx, i.hostFD, dst, i.offset, opts.Flags) + i.offset += n + i.offsetMu.Unlock() + return n, err +} + +func readFromHostFD(ctx context.Context, hostFD int, dst usermem.IOSequence, offset int64, flags uint32) (int64, error) { + // TODO(gvisor.dev/issue/1672): Support select preadv2 flags. + if flags != 0 { + return 0, syserror.EOPNOTSUPP + } + + var reader safemem.Reader + if offset == -1 { + reader = safemem.FromIOReader{fd.NewReadWriter(hostFD)} + } else { + reader = safemem.FromVecReaderFunc{ + func(srcs [][]byte) (int64, error) { + n, err := unix.Preadv(hostFD, srcs, offset) + return int64(n), err + }, + } + } + n, err := dst.CopyOutFrom(ctx, reader) + return int64(n), err +} + +// PWrite implements FileDescriptionImpl. +func (f *fileDescription) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) { + i := f.inode + if !i.seekable { + return 0, syserror.ESPIPE + } + + return writeToHostFD(ctx, i.hostFD, src, offset, opts.Flags) +} + +// Write implements FileDescriptionImpl. +func (f *fileDescription) Write(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) { + i := f.inode + if !i.seekable { + n, err := writeToHostFD(ctx, i.hostFD, src, -1, opts.Flags) + if isBlockError(err) { + err = syserror.ErrWouldBlock + } + return n, err + } + // TODO(gvisor.dev/issue/1672): Cache pages, when forced to do so. + // TODO(gvisor.dev/issue/1672): Write to end of file and update offset if O_APPEND is set on this file. + i.offsetMu.Lock() + n, err := writeToHostFD(ctx, i.hostFD, src, i.offset, opts.Flags) + i.offset += n + i.offsetMu.Unlock() + return n, err +} + +func writeToHostFD(ctx context.Context, hostFD int, src usermem.IOSequence, offset int64, flags uint32) (int64, error) { + // TODO(gvisor.dev/issue/1672): Support select pwritev2 flags. + if flags != 0 { + return 0, syserror.EOPNOTSUPP + } + + var writer safemem.Writer + if offset == -1 { + writer = safemem.FromIOWriter{fd.NewReadWriter(hostFD)} + } else { + writer = safemem.FromVecWriterFunc{ + func(srcs [][]byte) (int64, error) { + n, err := unix.Pwritev(hostFD, srcs, offset) + return int64(n), err + }, + } + } + n, err := src.CopyInTo(ctx, writer) + return int64(n), err +} + +// Seek implements FileDescriptionImpl. +// +// Note that we do not support seeking on directories, since we do not even +// allow directory fds to be imported at all. +func (f *fileDescription) Seek(_ context.Context, offset int64, whence int32) (int64, error) { + i := f.inode + if !i.seekable { + return 0, syserror.ESPIPE + } + + i.offsetMu.Lock() + defer i.offsetMu.Unlock() + + switch whence { + case linux.SEEK_SET: + if offset < 0 { + return i.offset, syserror.EINVAL + } + i.offset = offset + + case linux.SEEK_CUR: + // Check for overflow. Note that underflow cannot occur, since i.offset >= 0. + if offset > math.MaxInt64-i.offset { + return i.offset, syserror.EOVERFLOW + } + if i.offset+offset < 0 { + return i.offset, syserror.EINVAL + } + i.offset += offset + + case linux.SEEK_END: + var s syscall.Stat_t + if err := syscall.Fstat(i.hostFD, &s); err != nil { + return i.offset, err + } + size := s.Size + + // Check for overflow. Note that underflow cannot occur, since size >= 0. + if offset > math.MaxInt64-size { + return i.offset, syserror.EOVERFLOW + } + if size+offset < 0 { + return i.offset, syserror.EINVAL + } + i.offset = size + offset + + case linux.SEEK_DATA, linux.SEEK_HOLE: + // Modifying the offset in the host file table should not matter, since + // this is the only place where we use it. + // + // For reading and writing, we always rely on our internal offset. + n, err := unix.Seek(i.hostFD, offset, int(whence)) + if err != nil { + return i.offset, err + } + i.offset = n + + default: + // Invalid whence. + return i.offset, syserror.EINVAL + } + + return i.offset, nil +} + +// Sync implements FileDescriptionImpl. +func (f *fileDescription) Sync(context.Context) error { + // TODO(gvisor.dev/issue/1672): Currently we do not support the SyncData optimization, so we always sync everything. + return unix.Fsync(f.inode.hostFD) +} + +// ConfigureMMap implements FileDescriptionImpl. +func (f *fileDescription) ConfigureMMap(_ context.Context, opts *memmap.MMapOpts) error { + if !f.inode.canMap { + return syserror.ENODEV + } + // TODO(gvisor.dev/issue/1672): Implement ConfigureMMap and Mappable interface. + return syserror.ENODEV +} diff --git a/pkg/sentry/fsimpl/host/ioctl_unsafe.go b/pkg/sentry/fsimpl/host/ioctl_unsafe.go new file mode 100644 index 000000000..0983bf7d8 --- /dev/null +++ b/pkg/sentry/fsimpl/host/ioctl_unsafe.go @@ -0,0 +1,56 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package host + +import ( + "syscall" + "unsafe" + + "gvisor.dev/gvisor/pkg/abi/linux" +) + +func ioctlGetTermios(fd int) (*linux.Termios, error) { + var t linux.Termios + _, _, errno := syscall.Syscall(syscall.SYS_IOCTL, uintptr(fd), linux.TCGETS, uintptr(unsafe.Pointer(&t))) + if errno != 0 { + return nil, errno + } + return &t, nil +} + +func ioctlSetTermios(fd int, req uint64, t *linux.Termios) error { + _, _, errno := syscall.Syscall(syscall.SYS_IOCTL, uintptr(fd), uintptr(req), uintptr(unsafe.Pointer(t))) + if errno != 0 { + return errno + } + return nil +} + +func ioctlGetWinsize(fd int) (*linux.Winsize, error) { + var w linux.Winsize + _, _, errno := syscall.Syscall(syscall.SYS_IOCTL, uintptr(fd), linux.TIOCGWINSZ, uintptr(unsafe.Pointer(&w))) + if errno != 0 { + return nil, errno + } + return &w, nil +} + +func ioctlSetWinsize(fd int, w *linux.Winsize) error { + _, _, errno := syscall.Syscall(syscall.SYS_IOCTL, uintptr(fd), linux.TIOCSWINSZ, uintptr(unsafe.Pointer(w))) + if errno != 0 { + return errno + } + return nil +} diff --git a/pkg/sentry/fsimpl/host/tty.go b/pkg/sentry/fsimpl/host/tty.go new file mode 100644 index 000000000..8936afb06 --- /dev/null +++ b/pkg/sentry/fsimpl/host/tty.go @@ -0,0 +1,379 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package host + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/unimpl" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +// ttyFD implements vfs.FileDescriptionImpl for a host file descriptor +// that wraps a TTY FD. +type ttyFD struct { + fileDescription + + // mu protects the fields below. + mu sync.Mutex `state:"nosave"` + + // session is the session attached to this ttyFD. + session *kernel.Session + + // fgProcessGroup is the foreground process group that is currently + // connected to this TTY. + fgProcessGroup *kernel.ProcessGroup + + // termios contains the terminal attributes for this TTY. + termios linux.KernelTermios +} + +// InitForegroundProcessGroup sets the foreground process group and session for +// the TTY. This should only be called once, after the foreground process group +// has been created, but before it has started running. +func (t *ttyFD) InitForegroundProcessGroup(pg *kernel.ProcessGroup) { + t.mu.Lock() + defer t.mu.Unlock() + if t.fgProcessGroup != nil { + panic("foreground process group is already set") + } + t.fgProcessGroup = pg + t.session = pg.Session() +} + +// ForegroundProcessGroup returns the foreground process for the TTY. +func (t *ttyFD) ForegroundProcessGroup() *kernel.ProcessGroup { + t.mu.Lock() + defer t.mu.Unlock() + return t.fgProcessGroup +} + +// Release implements fs.FileOperations.Release. +func (t *ttyFD) Release() { + t.mu.Lock() + t.fgProcessGroup = nil + t.mu.Unlock() + + t.fileDescription.Release() +} + +// PRead implements vfs.FileDescriptionImpl. +// +// Reading from a TTY is only allowed for foreground process groups. Background +// process groups will either get EIO or a SIGTTIN. +func (t *ttyFD) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { + t.mu.Lock() + defer t.mu.Unlock() + + // Are we allowed to do the read? + // drivers/tty/n_tty.c:n_tty_read()=>job_control()=>tty_check_change(). + if err := t.checkChange(ctx, linux.SIGTTIN); err != nil { + return 0, err + } + + // Do the read. + return t.fileDescription.PRead(ctx, dst, offset, opts) +} + +// Read implements vfs.FileDescriptionImpl. +// +// Reading from a TTY is only allowed for foreground process groups. Background +// process groups will either get EIO or a SIGTTIN. +func (t *ttyFD) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { + t.mu.Lock() + defer t.mu.Unlock() + + // Are we allowed to do the read? + // drivers/tty/n_tty.c:n_tty_read()=>job_control()=>tty_check_change(). + if err := t.checkChange(ctx, linux.SIGTTIN); err != nil { + return 0, err + } + + // Do the read. + return t.fileDescription.Read(ctx, dst, opts) +} + +// PWrite implements vfs.FileDescriptionImpl. +func (t *ttyFD) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) { + t.mu.Lock() + defer t.mu.Unlock() + + // Check whether TOSTOP is enabled. This corresponds to the check in + // drivers/tty/n_tty.c:n_tty_write(). + if t.termios.LEnabled(linux.TOSTOP) { + if err := t.checkChange(ctx, linux.SIGTTOU); err != nil { + return 0, err + } + } + return t.fileDescription.PWrite(ctx, src, offset, opts) +} + +// Write implements vfs.FileDescriptionImpl. +func (t *ttyFD) Write(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) { + t.mu.Lock() + defer t.mu.Unlock() + + // Check whether TOSTOP is enabled. This corresponds to the check in + // drivers/tty/n_tty.c:n_tty_write(). + if t.termios.LEnabled(linux.TOSTOP) { + if err := t.checkChange(ctx, linux.SIGTTOU); err != nil { + return 0, err + } + } + return t.fileDescription.Write(ctx, src, opts) +} + +// Ioctl implements vfs.FileDescriptionImpl. +func (t *ttyFD) Ioctl(ctx context.Context, io usermem.IO, args arch.SyscallArguments) (uintptr, error) { + // Ignore arg[0]. This is the real FD: + fd := t.inode.hostFD + ioctl := args[1].Uint64() + switch ioctl { + case linux.TCGETS: + termios, err := ioctlGetTermios(fd) + if err != nil { + return 0, err + } + _, err = usermem.CopyObjectOut(ctx, io, args[2].Pointer(), termios, usermem.IOOpts{ + AddressSpaceActive: true, + }) + return 0, err + + case linux.TCSETS, linux.TCSETSW, linux.TCSETSF: + t.mu.Lock() + defer t.mu.Unlock() + + if err := t.checkChange(ctx, linux.SIGTTOU); err != nil { + return 0, err + } + + var termios linux.Termios + if _, err := usermem.CopyObjectIn(ctx, io, args[2].Pointer(), &termios, usermem.IOOpts{ + AddressSpaceActive: true, + }); err != nil { + return 0, err + } + err := ioctlSetTermios(fd, ioctl, &termios) + if err == nil { + t.termios.FromTermios(termios) + } + return 0, err + + case linux.TIOCGPGRP: + // Args: pid_t *argp + // When successful, equivalent to *argp = tcgetpgrp(fd). + // Get the process group ID of the foreground process group on this + // terminal. + + pidns := kernel.PIDNamespaceFromContext(ctx) + if pidns == nil { + return 0, syserror.ENOTTY + } + + t.mu.Lock() + defer t.mu.Unlock() + + // Map the ProcessGroup into a ProcessGroupID in the task's PID namespace. + pgID := pidns.IDOfProcessGroup(t.fgProcessGroup) + _, err := usermem.CopyObjectOut(ctx, io, args[2].Pointer(), &pgID, usermem.IOOpts{ + AddressSpaceActive: true, + }) + return 0, err + + case linux.TIOCSPGRP: + // Args: const pid_t *argp + // Equivalent to tcsetpgrp(fd, *argp). + // Set the foreground process group ID of this terminal. + + task := kernel.TaskFromContext(ctx) + if task == nil { + return 0, syserror.ENOTTY + } + + t.mu.Lock() + defer t.mu.Unlock() + + // Check that we are allowed to set the process group. + if err := t.checkChange(ctx, linux.SIGTTOU); err != nil { + // drivers/tty/tty_io.c:tiocspgrp() converts -EIO from tty_check_change() + // to -ENOTTY. + if err == syserror.EIO { + return 0, syserror.ENOTTY + } + return 0, err + } + + // Check that calling task's process group is in the TTY session. + if task.ThreadGroup().Session() != t.session { + return 0, syserror.ENOTTY + } + + var pgID kernel.ProcessGroupID + if _, err := usermem.CopyObjectIn(ctx, io, args[2].Pointer(), &pgID, usermem.IOOpts{ + AddressSpaceActive: true, + }); err != nil { + return 0, err + } + + // pgID must be non-negative. + if pgID < 0 { + return 0, syserror.EINVAL + } + + // Process group with pgID must exist in this PID namespace. + pidns := task.PIDNamespace() + pg := pidns.ProcessGroupWithID(pgID) + if pg == nil { + return 0, syserror.ESRCH + } + + // Check that new process group is in the TTY session. + if pg.Session() != t.session { + return 0, syserror.EPERM + } + + t.fgProcessGroup = pg + return 0, nil + + case linux.TIOCGWINSZ: + // Args: struct winsize *argp + // Get window size. + winsize, err := ioctlGetWinsize(fd) + if err != nil { + return 0, err + } + _, err = usermem.CopyObjectOut(ctx, io, args[2].Pointer(), winsize, usermem.IOOpts{ + AddressSpaceActive: true, + }) + return 0, err + + case linux.TIOCSWINSZ: + // Args: const struct winsize *argp + // Set window size. + + // Unlike setting the termios, any process group (even background ones) can + // set the winsize. + + var winsize linux.Winsize + if _, err := usermem.CopyObjectIn(ctx, io, args[2].Pointer(), &winsize, usermem.IOOpts{ + AddressSpaceActive: true, + }); err != nil { + return 0, err + } + err := ioctlSetWinsize(fd, &winsize) + return 0, err + + // Unimplemented commands. + case linux.TIOCSETD, + linux.TIOCSBRK, + linux.TIOCCBRK, + linux.TCSBRK, + linux.TCSBRKP, + linux.TIOCSTI, + linux.TIOCCONS, + linux.FIONBIO, + linux.TIOCEXCL, + linux.TIOCNXCL, + linux.TIOCGEXCL, + linux.TIOCNOTTY, + linux.TIOCSCTTY, + linux.TIOCGSID, + linux.TIOCGETD, + linux.TIOCVHANGUP, + linux.TIOCGDEV, + linux.TIOCMGET, + linux.TIOCMSET, + linux.TIOCMBIC, + linux.TIOCMBIS, + linux.TIOCGICOUNT, + linux.TCFLSH, + linux.TIOCSSERIAL, + linux.TIOCGPTPEER: + + unimpl.EmitUnimplementedEvent(ctx) + fallthrough + default: + return 0, syserror.ENOTTY + } +} + +// checkChange checks that the process group is allowed to read, write, or +// change the state of the TTY. +// +// This corresponds to Linux drivers/tty/tty_io.c:tty_check_change(). The logic +// is a bit convoluted, but documented inline. +// +// Preconditions: t.mu must be held. +func (t *ttyFD) checkChange(ctx context.Context, sig linux.Signal) error { + task := kernel.TaskFromContext(ctx) + if task == nil { + // No task? Linux does not have an analog for this case, but + // tty_check_change is more of a blacklist of cases than a + // whitelist, and is surprisingly permissive. Allowing the + // change seems most appropriate. + return nil + } + + tg := task.ThreadGroup() + pg := tg.ProcessGroup() + + // If the session for the task is different than the session for the + // controlling TTY, then the change is allowed. Seems like a bad idea, + // but that's exactly what linux does. + if tg.Session() != t.fgProcessGroup.Session() { + return nil + } + + // If we are the foreground process group, then the change is allowed. + if pg == t.fgProcessGroup { + return nil + } + + // We are not the foreground process group. + + // Is the provided signal blocked or ignored? + if (task.SignalMask()&linux.SignalSetOf(sig) != 0) || tg.SignalHandlers().IsIgnored(sig) { + // If the signal is SIGTTIN, then we are attempting to read + // from the TTY. Don't send the signal and return EIO. + if sig == linux.SIGTTIN { + return syserror.EIO + } + + // Otherwise, we are writing or changing terminal state. This is allowed. + return nil + } + + // If the process group is an orphan, return EIO. + if pg.IsOrphan() { + return syserror.EIO + } + + // Otherwise, send the signal to the process group and return ERESTARTSYS. + // + // Note that Linux also unconditionally sets TIF_SIGPENDING on current, + // but this isn't necessary in gVisor because the rationale given in + // 040b6362d58f "tty: fix leakage of -ERESTARTSYS to userland" doesn't + // apply: the sentry will handle -ERESTARTSYS in + // kernel.runApp.execute() even if the kernel.Task isn't interrupted. + // + // Linux ignores the result of kill_pgrp(). + _ = pg.SendSignal(kernel.SignalInfoPriv(sig)) + return kernel.ERESTARTSYS +} diff --git a/pkg/sentry/fsimpl/host/util.go b/pkg/sentry/fsimpl/host/util.go new file mode 100644 index 000000000..2bc757b1a --- /dev/null +++ b/pkg/sentry/fsimpl/host/util.go @@ -0,0 +1,66 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package host + +import ( + "syscall" + + "golang.org/x/sys/unix" + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/syserror" +) + +func toTimespec(ts linux.StatxTimestamp, omit bool) syscall.Timespec { + if omit { + return syscall.Timespec{ + Sec: 0, + Nsec: unix.UTIME_OMIT, + } + } + return syscall.Timespec{ + Sec: ts.Sec, + Nsec: int64(ts.Nsec), + } +} + +func unixToLinuxStatxTimestamp(ts unix.StatxTimestamp) linux.StatxTimestamp { + return linux.StatxTimestamp{Sec: ts.Sec, Nsec: ts.Nsec} +} + +func timespecToStatxTimestamp(ts unix.Timespec) linux.StatxTimestamp { + return linux.StatxTimestamp{Sec: int64(ts.Sec), Nsec: uint32(ts.Nsec)} +} + +// wouldBlock returns true for file types that can return EWOULDBLOCK +// for blocking operations, e.g. pipes, character devices, and sockets. +func wouldBlock(fileType uint32) bool { + return fileType == syscall.S_IFIFO || fileType == syscall.S_IFCHR || fileType == syscall.S_IFSOCK +} + +// canMap returns true if a file with fileType is allowed to be memory mapped. +// This is ported over from VFS1, but it's probably not the best way for us +// to check if a file can be memory mapped. +func canMap(fileType uint32) bool { + // TODO(gvisor.dev/issue/1672): Also allow "special files" to be mapped (see fs/host:canMap()). + // + // TODO(b/38213152): Some obscure character devices can be mapped. + return fileType == syscall.S_IFREG +} + +// isBlockError checks if an error is EAGAIN or EWOULDBLOCK. +// If so, they can be transformed into syserror.ErrWouldBlock. +func isBlockError(err error) bool { + return err == syserror.EAGAIN || err == syserror.EWOULDBLOCK +} diff --git a/pkg/sentry/kernel/pipe/buffer_test.go b/pkg/sentry/fsimpl/host/util_unsafe.go index 4d54b8b8f..5136ac844 100644 --- a/pkg/sentry/kernel/pipe/buffer_test.go +++ b/pkg/sentry/fsimpl/host/util_unsafe.go @@ -1,4 +1,4 @@ -// Copyright 2019 The gVisor Authors. +// Copyright 2020 The gVisor Authors. // // Licensed under the Apache License, Version 2.0 (the "License"); // you may not use this file except in compliance with the License. @@ -12,21 +12,23 @@ // See the License for the specific language governing permissions and // limitations under the License. -package pipe +package host import ( - "testing" + "syscall" "unsafe" - - "gvisor.dev/gvisor/pkg/usermem" ) -func TestBufferSize(t *testing.T) { - bufferSize := unsafe.Sizeof(buffer{}) - if bufferSize < usermem.PageSize { - t.Errorf("buffer is less than a page") - } - if bufferSize > (2 * usermem.PageSize) { - t.Errorf("buffer is greater than two pages") +func setTimestamps(fd int, ts *[2]syscall.Timespec) error { + _, _, errno := syscall.Syscall6( + syscall.SYS_UTIMENSAT, + uintptr(fd), + 0, /* path */ + uintptr(unsafe.Pointer(ts)), + 0, /* flags */ + 0, 0) + if errno != 0 { + return errno } + return nil } diff --git a/pkg/sentry/fsimpl/kernfs/BUILD b/pkg/sentry/fsimpl/kernfs/BUILD index e73f1f857..ef34cb28a 100644 --- a/pkg/sentry/fsimpl/kernfs/BUILD +++ b/pkg/sentry/fsimpl/kernfs/BUILD @@ -4,6 +4,17 @@ load("//tools/go_generics:defs.bzl", "go_template_instance") licenses(["notice"]) go_template_instance( + name = "fstree", + out = "fstree.go", + package = "kernfs", + prefix = "generic", + template = "//pkg/sentry/vfs/genericfstree:generic_fstree", + types = { + "Dentry": "Dentry", + }, +) + +go_template_instance( name = "slot_list", out = "slot_list.go", package = "kernfs", @@ -21,6 +32,7 @@ go_library( "dynamic_bytes_file.go", "fd_impl_util.go", "filesystem.go", + "fstree.go", "inode_impl_util.go", "kernfs.go", "slot_list.go", @@ -35,6 +47,7 @@ go_library( "//pkg/refs", "//pkg/sentry/kernel/auth", "//pkg/sentry/memmap", + "//pkg/sentry/socket/unix/transport", "//pkg/sentry/vfs", "//pkg/sync", "//pkg/syserror", diff --git a/pkg/sentry/fsimpl/kernfs/dynamic_bytes_file.go b/pkg/sentry/fsimpl/kernfs/dynamic_bytes_file.go index 733792c78..d8bddbafa 100644 --- a/pkg/sentry/fsimpl/kernfs/dynamic_bytes_file.go +++ b/pkg/sentry/fsimpl/kernfs/dynamic_bytes_file.go @@ -53,17 +53,18 @@ func (f *DynamicBytesFile) Init(creds *auth.Credentials, ino uint64, data vfs.Dy } // Open implements Inode.Open. -func (f *DynamicBytesFile) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, flags uint32) (*vfs.FileDescription, error) { +func (f *DynamicBytesFile) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { fd := &DynamicBytesFD{} - if err := fd.Init(rp.Mount(), vfsd, f.data, flags); err != nil { + if err := fd.Init(rp.Mount(), vfsd, f.data, opts.Flags); err != nil { return nil, err } return &fd.vfsfd, nil } -// SetStat implements Inode.SetStat. -func (f *DynamicBytesFile) SetStat(*vfs.Filesystem, vfs.SetStatOptions) error { - // DynamicBytesFiles are immutable. +// SetStat implements Inode.SetStat. By default DynamicBytesFile doesn't allow +// inode attributes to be changed. Override SetStat() making it call +// f.InodeAttrs to allow it. +func (*DynamicBytesFile) SetStat(context.Context, *vfs.Filesystem, *auth.Credentials, vfs.SetStatOptions) error { return syserror.EPERM } @@ -122,7 +123,7 @@ func (fd *DynamicBytesFD) Release() {} // Stat implements vfs.FileDescriptionImpl.Stat. func (fd *DynamicBytesFD) Stat(ctx context.Context, opts vfs.StatOptions) (linux.Statx, error) { fs := fd.vfsfd.VirtualDentry().Mount().Filesystem() - return fd.inode.Stat(fs), nil + return fd.inode.Stat(fs, opts) } // SetStat implements vfs.FileDescriptionImpl.SetStat. diff --git a/pkg/sentry/fsimpl/kernfs/fd_impl_util.go b/pkg/sentry/fsimpl/kernfs/fd_impl_util.go index 6104751c8..e8a4670b8 100644 --- a/pkg/sentry/fsimpl/kernfs/fd_impl_util.go +++ b/pkg/sentry/fsimpl/kernfs/fd_impl_util.go @@ -17,6 +17,7 @@ package kernfs import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/memmap" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserror" @@ -43,12 +44,12 @@ type GenericDirectoryFD struct { } // Init initializes a GenericDirectoryFD. -func (fd *GenericDirectoryFD) Init(m *vfs.Mount, d *vfs.Dentry, children *OrderedChildren, flags uint32) error { - if vfs.AccessTypesForOpenFlags(flags)&vfs.MayWrite != 0 { +func (fd *GenericDirectoryFD) Init(m *vfs.Mount, d *vfs.Dentry, children *OrderedChildren, opts *vfs.OpenOptions) error { + if vfs.AccessTypesForOpenFlags(opts)&vfs.MayWrite != 0 { // Can't open directories for writing. return syserror.EISDIR } - if err := fd.vfsfd.Init(fd, flags, m, d, &vfs.FileDescriptionOptions{}); err != nil { + if err := fd.vfsfd.Init(fd, opts.Flags, m, d, &vfs.FileDescriptionOptions{}); err != nil { return err } fd.children = children @@ -107,33 +108,40 @@ func (fd *GenericDirectoryFD) IterDirents(ctx context.Context, cb vfs.IterDirent fs.mu.Lock() defer fs.mu.Unlock() + opts := vfs.StatOptions{Mask: linux.STATX_INO} // Handle ".". if fd.off == 0 { - stat := fd.inode().Stat(vfsFS) + stat, err := fd.inode().Stat(vfsFS, opts) + if err != nil { + return err + } dirent := vfs.Dirent{ Name: ".", Type: linux.DT_DIR, Ino: stat.Ino, NextOff: 1, } - if !cb.Handle(dirent) { - return nil + if err := cb.Handle(dirent); err != nil { + return err } fd.off++ } // Handle "..". if fd.off == 1 { - parentInode := vfsd.ParentOrSelf().Impl().(*Dentry).inode - stat := parentInode.Stat(vfsFS) + parentInode := genericParentOrSelf(vfsd.Impl().(*Dentry)).inode + stat, err := parentInode.Stat(vfsFS, opts) + if err != nil { + return err + } dirent := vfs.Dirent{ Name: "..", Type: linux.FileMode(stat.Mode).DirentType(), Ino: stat.Ino, NextOff: 2, } - if !cb.Handle(dirent) { - return nil + if err := cb.Handle(dirent); err != nil { + return err } fd.off++ } @@ -146,15 +154,18 @@ func (fd *GenericDirectoryFD) IterDirents(ctx context.Context, cb vfs.IterDirent childIdx := fd.off - 2 for it := fd.children.nthLocked(childIdx); it != nil; it = it.Next() { inode := it.Dentry.Impl().(*Dentry).inode - stat := inode.Stat(vfsFS) + stat, err := inode.Stat(vfsFS, opts) + if err != nil { + return err + } dirent := vfs.Dirent{ Name: it.Name, Type: linux.FileMode(stat.Mode).DirentType(), Ino: stat.Ino, NextOff: fd.off + 1, } - if !cb.Handle(dirent) { - return nil + if err := cb.Handle(dirent); err != nil { + return err } fd.off++ } @@ -190,12 +201,12 @@ func (fd *GenericDirectoryFD) Seek(ctx context.Context, offset int64, whence int func (fd *GenericDirectoryFD) Stat(ctx context.Context, opts vfs.StatOptions) (linux.Statx, error) { fs := fd.filesystem() inode := fd.inode() - return inode.Stat(fs), nil + return inode.Stat(fs, opts) } // SetStat implements vfs.FileDescriptionImpl.SetStat. func (fd *GenericDirectoryFD) SetStat(ctx context.Context, opts vfs.SetStatOptions) error { - fs := fd.filesystem() + creds := auth.CredentialsFromContext(ctx) inode := fd.vfsfd.VirtualDentry().Dentry().Impl().(*Dentry).inode - return inode.SetStat(fs, opts) + return inode.SetStat(ctx, fd.filesystem(), creds, opts) } diff --git a/pkg/sentry/fsimpl/kernfs/filesystem.go b/pkg/sentry/fsimpl/kernfs/filesystem.go index e49303c26..01c23d192 100644 --- a/pkg/sentry/fsimpl/kernfs/filesystem.go +++ b/pkg/sentry/fsimpl/kernfs/filesystem.go @@ -12,16 +12,18 @@ // See the License for the specific language governing permissions and // limitations under the License. -// This file implements vfs.FilesystemImpl for kernfs. - package kernfs +// This file implements vfs.FilesystemImpl for kernfs. + import ( "fmt" "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserror" ) @@ -54,36 +56,48 @@ afterSymlink: return vfsd, nil } if name == ".." { - nextVFSD, err := rp.ResolveParent(vfsd) - if err != nil { + if isRoot, err := rp.CheckRoot(vfsd); err != nil { + return nil, err + } else if isRoot || d.parent == nil { + rp.Advance() + return vfsd, nil + } + if err := rp.CheckMount(&d.parent.vfsd); err != nil { return nil, err } rp.Advance() - return nextVFSD, nil + return &d.parent.vfsd, nil + } + if len(name) > linux.NAME_MAX { + return nil, syserror.ENAMETOOLONG } d.dirMu.Lock() - nextVFSD, err := rp.ResolveChild(vfsd, name) + next, err := fs.revalidateChildLocked(ctx, rp.VirtualFilesystem(), d, name, d.children[name]) + d.dirMu.Unlock() if err != nil { - d.dirMu.Unlock() return nil, err } - next, err := fs.revalidateChildLocked(ctx, rp.VirtualFilesystem(), d, name, nextVFSD) - d.dirMu.Unlock() - if err != nil { + if err := rp.CheckMount(&next.vfsd); err != nil { return nil, err } // Resolve any symlink at current path component. if rp.ShouldFollowSymlink() && next.isSymlink() { - // TODO: VFS2 needs something extra for /proc/[pid]/fd/ "magic symlinks". - target, err := next.inode.Readlink(ctx) + targetVD, targetPathname, err := next.inode.Getlink(ctx) if err != nil { return nil, err } - if err := rp.HandleSymlink(target); err != nil { - return nil, err + if targetVD.Ok() { + err := rp.HandleJump(targetVD) + targetVD.DecRef() + if err != nil { + return nil, err + } + } else { + if err := rp.HandleSymlink(targetPathname); err != nil { + return nil, err + } } goto afterSymlink - } rp.Advance() return &next.vfsd, nil @@ -97,17 +111,17 @@ afterSymlink: // parent.dirMu must be locked. parent.isDir(). name is not "." or "..". // // Postconditions: Caller must call fs.processDeferredDecRefs*. -func (fs *Filesystem) revalidateChildLocked(ctx context.Context, vfsObj *vfs.VirtualFilesystem, parent *Dentry, name string, childVFSD *vfs.Dentry) (*Dentry, error) { - if childVFSD != nil { +func (fs *Filesystem) revalidateChildLocked(ctx context.Context, vfsObj *vfs.VirtualFilesystem, parent *Dentry, name string, child *Dentry) (*Dentry, error) { + if child != nil { // Cached dentry exists, revalidate. - child := childVFSD.Impl().(*Dentry) if !child.inode.Valid(ctx) { - vfsObj.ForceDeleteDentry(childVFSD) - fs.deferDecRef(childVFSD) // Reference from Lookup. - childVFSD = nil + delete(parent.children, name) + vfsObj.InvalidateDentry(&child.vfsd) + fs.deferDecRef(&child.vfsd) // Reference from Lookup. + child = nil } } - if childVFSD == nil { + if child == nil { // Dentry isn't cached; it either doesn't exist or failed // revalidation. Attempt to resolve it via Lookup. // @@ -115,15 +129,15 @@ func (fs *Filesystem) revalidateChildLocked(ctx context.Context, vfsObj *vfs.Vir // *(kernfs.)Dentry, not *vfs.Dentry, since (kernfs.)Filesystem assumes // that all dentries in the filesystem are (kernfs.)Dentry and performs // vfs.DentryImpl casts accordingly. - var err error - childVFSD, err = parent.inode.Lookup(ctx, name) + childVFSD, err := parent.inode.Lookup(ctx, name) if err != nil { return nil, err } // Reference on childVFSD dropped by a corresponding Valid. - parent.insertChildLocked(name, childVFSD) + child = childVFSD.Impl().(*Dentry) + parent.insertChildLocked(name, child) } - return childVFSD.Impl().(*Dentry), nil + return child, nil } // walkExistingLocked resolves rp to an existing file. @@ -189,14 +203,14 @@ func checkCreateLocked(ctx context.Context, rp *vfs.ResolvingPath, parentVFSD *v if pc == "." || pc == ".." { return "", syserror.EEXIST } - childVFSD, err := rp.ResolveChild(parentVFSD, pc) - if err != nil { - return "", err + if len(pc) > linux.NAME_MAX { + return "", syserror.ENAMETOOLONG } - if childVFSD != nil { + // FIXME(gvisor.dev/issue/1193): Data race due to not holding dirMu. + if _, ok := parentVFSD.Impl().(*Dentry).children[pc]; ok { return "", syserror.EEXIST } - if parentVFSD.IsDisowned() { + if parentVFSD.IsDead() { return "", syserror.ENOENT } return pc, nil @@ -206,14 +220,14 @@ func checkCreateLocked(ctx context.Context, rp *vfs.ResolvingPath, parentVFSD *v // // Preconditions: Filesystem.mu must be locked for at least reading. func checkDeleteLocked(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry) error { - parentVFSD := vfsd.Parent() - if parentVFSD == nil { + parent := vfsd.Impl().(*Dentry).parent + if parent == nil { return syserror.EBUSY } - if parentVFSD.IsDisowned() { + if parent.vfsd.IsDead() { return syserror.ENOENT } - if err := parentVFSD.Impl().(*Dentry).inode.CheckPermissions(ctx, rp.Credentials(), vfs.MayWrite|vfs.MayExec); err != nil { + if err := parent.inode.CheckPermissions(ctx, rp.Credentials(), vfs.MayWrite|vfs.MayExec); err != nil { return err } return nil @@ -229,6 +243,19 @@ func (fs *Filesystem) Sync(ctx context.Context) error { return nil } +// AccessAt implements vfs.Filesystem.Impl.AccessAt. +func (fs *Filesystem) AccessAt(ctx context.Context, rp *vfs.ResolvingPath, creds *auth.Credentials, ats vfs.AccessTypes) error { + fs.mu.RLock() + defer fs.mu.RUnlock() + defer fs.processDeferredDecRefs() + + _, inode, err := fs.walkExistingLocked(ctx, rp) + if err != nil { + return err + } + return inode.CheckPermissions(ctx, creds, ats) +} + // GetDentryAt implements vfs.FilesystemImpl.GetDentryAt. func (fs *Filesystem) GetDentryAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.GetDentryOptions) (*vfs.Dentry, error) { fs.mu.RLock() @@ -294,11 +321,11 @@ func (fs *Filesystem) LinkAt(ctx context.Context, rp *vfs.ResolvingPath, vd vfs. return syserror.EPERM } - child, err := parentInode.NewLink(ctx, pc, d.inode) + childVFSD, err := parentInode.NewLink(ctx, pc, d.inode) if err != nil { return err } - parentVFSD.Impl().(*Dentry).InsertChild(pc, child) + parentVFSD.Impl().(*Dentry).InsertChild(pc, childVFSD.Impl().(*Dentry)) return nil } @@ -322,11 +349,11 @@ func (fs *Filesystem) MkdirAt(ctx context.Context, rp *vfs.ResolvingPath, opts v return err } defer rp.Mount().EndWrite() - child, err := parentInode.NewDir(ctx, pc, opts) + childVFSD, err := parentInode.NewDir(ctx, pc, opts) if err != nil { return err } - parentVFSD.Impl().(*Dentry).InsertChild(pc, child) + parentVFSD.Impl().(*Dentry).InsertChild(pc, childVFSD.Impl().(*Dentry)) return nil } @@ -350,11 +377,11 @@ func (fs *Filesystem) MknodAt(ctx context.Context, rp *vfs.ResolvingPath, opts v return err } defer rp.Mount().EndWrite() - new, err := parentInode.NewNode(ctx, pc, opts) + newVFSD, err := parentInode.NewNode(ctx, pc, opts) if err != nil { return err } - parentVFSD.Impl().(*Dentry).InsertChild(pc, new) + parentVFSD.Impl().(*Dentry).InsertChild(pc, newVFSD.Impl().(*Dentry)) return nil } @@ -365,7 +392,7 @@ func (fs *Filesystem) OpenAt(ctx context.Context, rp *vfs.ResolvingPath, opts vf // appropriate bits in rp), but are returned by // FileDescriptionImpl.StatusFlags(). opts.Flags &= linux.O_ACCMODE | linux.O_CREAT | linux.O_EXCL | linux.O_TRUNC | linux.O_DIRECTORY | linux.O_NOFOLLOW - ats := vfs.AccessTypesForOpenFlags(opts.Flags) + ats := vfs.AccessTypesForOpenFlags(&opts) // Do not create new file. if opts.Flags&linux.O_CREAT == 0 { @@ -379,7 +406,7 @@ func (fs *Filesystem) OpenAt(ctx context.Context, rp *vfs.ResolvingPath, opts vf if err := inode.CheckPermissions(ctx, rp.Credentials(), ats); err != nil { return nil, err } - return inode.Open(rp, vfsd, opts.Flags) + return inode.Open(rp, vfsd, opts) } // May create new file. @@ -398,7 +425,7 @@ func (fs *Filesystem) OpenAt(ctx context.Context, rp *vfs.ResolvingPath, opts vf if err := inode.CheckPermissions(ctx, rp.Credentials(), ats); err != nil { return nil, err } - return inode.Open(rp, vfsd, opts.Flags) + return inode.Open(rp, vfsd, opts) } afterTrailingSymlink: parentVFSD, parentInode, err := fs.walkParentDirLocked(ctx, rp) @@ -418,12 +445,12 @@ afterTrailingSymlink: if pc == "." || pc == ".." { return nil, syserror.EISDIR } - // Determine whether or not we need to create a file. - childVFSD, err := rp.ResolveChild(parentVFSD, pc) - if err != nil { - return nil, err + if len(pc) > linux.NAME_MAX { + return nil, syserror.ENAMETOOLONG } - if childVFSD == nil { + // Determine whether or not we need to create a file. + childVFSD, err := fs.stepExistingLocked(ctx, rp, parentVFSD) + if err == syserror.ENOENT { // Already checked for searchability above; now check for writability. if err := parentInode.CheckPermissions(ctx, rp.Credentials(), vfs.MayWrite); err != nil { return nil, err @@ -433,37 +460,46 @@ afterTrailingSymlink: } defer rp.Mount().EndWrite() // Create and open the child. - child, err := parentInode.NewFile(ctx, pc, opts) + childVFSD, err = parentInode.NewFile(ctx, pc, opts) if err != nil { return nil, err } + child := childVFSD.Impl().(*Dentry) parentVFSD.Impl().(*Dentry).InsertChild(pc, child) - return child.Impl().(*Dentry).inode.Open(rp, child, opts.Flags) + return child.inode.Open(rp, childVFSD, opts) + } + if err != nil { + return nil, err } // Open existing file or follow symlink. if mustCreate { return nil, syserror.EEXIST } - childDentry := childVFSD.Impl().(*Dentry) - childInode := childDentry.inode - if rp.ShouldFollowSymlink() { - if childDentry.isSymlink() { - target, err := childInode.Readlink(ctx) + child := childVFSD.Impl().(*Dentry) + if rp.ShouldFollowSymlink() && child.isSymlink() { + targetVD, targetPathname, err := child.inode.Getlink(ctx) + if err != nil { + return nil, err + } + if targetVD.Ok() { + err := rp.HandleJump(targetVD) + targetVD.DecRef() if err != nil { return nil, err } - if err := rp.HandleSymlink(target); err != nil { + } else { + if err := rp.HandleSymlink(targetPathname); err != nil { return nil, err } - // rp.Final() may no longer be true since we now need to resolve the - // symlink target. - goto afterTrailingSymlink } + // rp.Final() may no longer be true since we now need to resolve the + // symlink target. + goto afterTrailingSymlink } - if err := childInode.CheckPermissions(ctx, rp.Credentials(), ats); err != nil { + if err := child.inode.CheckPermissions(ctx, rp.Credentials(), ats); err != nil { return nil, err } - return childInode.Open(rp, childVFSD, opts.Flags) + return child.inode.Open(rp, &child.vfsd, opts) } // ReadlinkAt implements vfs.FilesystemImpl.ReadlinkAt. @@ -490,15 +526,16 @@ func (fs *Filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa noReplace := opts.Flags&linux.RENAME_NOREPLACE != 0 fs.mu.Lock() - defer fs.mu.Lock() + defer fs.processDeferredDecRefsLocked() + defer fs.mu.Unlock() // Resolve the destination directory first to verify that it's on this // Mount. dstDirVFSD, dstDirInode, err := fs.walkParentDirLocked(ctx, rp) - fs.processDeferredDecRefsLocked() if err != nil { return err } + dstDir := dstDirVFSD.Impl().(*Dentry) mnt := rp.Mount() if mnt != oldParentVD.Mount() { return syserror.EXDEV @@ -511,9 +548,8 @@ func (fs *Filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa srcDirVFSD := oldParentVD.Dentry() srcDir := srcDirVFSD.Impl().(*Dentry) srcDir.dirMu.Lock() - src, err := fs.revalidateChildLocked(ctx, rp.VirtualFilesystem(), srcDir, oldName, srcDirVFSD.Child(oldName)) + src, err := fs.revalidateChildLocked(ctx, rp.VirtualFilesystem(), srcDir, oldName, srcDir.children[oldName]) srcDir.dirMu.Unlock() - fs.processDeferredDecRefsLocked() if err != nil { return err } @@ -525,7 +561,7 @@ func (fs *Filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa } // Can we create the dst dentry? - var dstVFSD *vfs.Dentry + var dst *Dentry pc, err := checkCreateLocked(ctx, rp, dstDirVFSD, dstDirInode) switch err { case nil: @@ -535,37 +571,51 @@ func (fs *Filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa // Won't overwrite existing node since RENAME_NOREPLACE was requested. return syserror.EEXIST } - dstVFSD, err = rp.ResolveChild(dstDirVFSD, pc) - if err != nil { + dst = dstDir.children[pc] + if dst == nil { panic(fmt.Sprintf("Child %q for parent Dentry %+v disappeared inside atomic section?", pc, dstDirVFSD)) } default: return err } + var dstVFSD *vfs.Dentry + if dst != nil { + dstVFSD = &dst.vfsd + } mntns := vfs.MountNamespaceFromContext(ctx) + defer mntns.DecRef() virtfs := rp.VirtualFilesystem() - srcDirDentry := srcDirVFSD.Impl().(*Dentry) - dstDirDentry := dstDirVFSD.Impl().(*Dentry) - // We can't deadlock here due to lock ordering because we're protected from // concurrent renames by fs.mu held for writing. - srcDirDentry.dirMu.Lock() - defer srcDirDentry.dirMu.Unlock() - dstDirDentry.dirMu.Lock() - defer dstDirDentry.dirMu.Unlock() + srcDir.dirMu.Lock() + defer srcDir.dirMu.Unlock() + if srcDir != dstDir { + dstDir.dirMu.Lock() + defer dstDir.dirMu.Unlock() + } if err := virtfs.PrepareRenameDentry(mntns, srcVFSD, dstVFSD); err != nil { return err } - srcDirInode := srcDirDentry.inode - replaced, err := srcDirInode.Rename(ctx, srcVFSD.Name(), pc, srcVFSD, dstDirVFSD) + replaced, err := srcDir.inode.Rename(ctx, src.name, pc, srcVFSD, dstDirVFSD) if err != nil { virtfs.AbortRenameDentry(srcVFSD, dstVFSD) return err } - virtfs.CommitRenameReplaceDentry(srcVFSD, dstDirVFSD, pc, replaced) + delete(srcDir.children, src.name) + if srcDir != dstDir { + fs.deferDecRef(srcDirVFSD) + dstDir.IncRef() + } + src.parent = dstDir + src.name = pc + if dstDir.children == nil { + dstDir.children = make(map[string]*Dentry) + } + dstDir.children[pc] = src + virtfs.CommitRenameReplaceDentry(srcVFSD, replaced) return nil } @@ -585,17 +635,21 @@ func (fs *Filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error if err := checkDeleteLocked(ctx, rp, vfsd); err != nil { return err } - if !vfsd.Impl().(*Dentry).isDir() { + d := vfsd.Impl().(*Dentry) + if !d.isDir() { return syserror.ENOTDIR } if inode.HasChildren() { return syserror.ENOTEMPTY } virtfs := rp.VirtualFilesystem() - parentDentry := vfsd.Parent().Impl().(*Dentry) + parentDentry := d.parent parentDentry.dirMu.Lock() defer parentDentry.dirMu.Unlock() - if err := virtfs.PrepareDeleteDentry(vfs.MountNamespaceFromContext(ctx), vfsd); err != nil { + + mntns := vfs.MountNamespaceFromContext(ctx) + defer mntns.DecRef() + if err := virtfs.PrepareDeleteDentry(mntns, vfsd); err != nil { return err } if err := parentDentry.inode.RmDir(ctx, rp.Component(), vfsd); err != nil { @@ -618,7 +672,7 @@ func (fs *Filesystem) SetStatAt(ctx context.Context, rp *vfs.ResolvingPath, opts if opts.Stat.Mask == 0 { return nil } - return inode.SetStat(fs.VFSFilesystem(), opts) + return inode.SetStat(ctx, fs.VFSFilesystem(), rp.Credentials(), opts) } // StatAt implements vfs.FilesystemImpl.StatAt. @@ -630,7 +684,7 @@ func (fs *Filesystem) StatAt(ctx context.Context, rp *vfs.ResolvingPath, opts vf if err != nil { return linux.Statx{}, err } - return inode.Stat(fs.VFSFilesystem()), nil + return inode.Stat(fs.VFSFilesystem(), opts) } // StatFSAt implements vfs.FilesystemImpl.StatFSAt. @@ -642,7 +696,7 @@ func (fs *Filesystem) StatFSAt(ctx context.Context, rp *vfs.ResolvingPath) (linu if err != nil { return linux.Statfs{}, err } - // TODO: actually implement statfs + // TODO(gvisor.dev/issue/1193): actually implement statfs. return linux.Statfs{}, syserror.ENOSYS } @@ -666,11 +720,11 @@ func (fs *Filesystem) SymlinkAt(ctx context.Context, rp *vfs.ResolvingPath, targ return err } defer rp.Mount().EndWrite() - child, err := parentInode.NewSymlink(ctx, pc, target) + childVFSD, err := parentInode.NewSymlink(ctx, pc, target) if err != nil { return err } - parentVFSD.Impl().(*Dentry).InsertChild(pc, child) + parentVFSD.Impl().(*Dentry).InsertChild(pc, childVFSD.Impl().(*Dentry)) return nil } @@ -690,14 +744,17 @@ func (fs *Filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error if err := checkDeleteLocked(ctx, rp, vfsd); err != nil { return err } - if vfsd.Impl().(*Dentry).isDir() { + d := vfsd.Impl().(*Dentry) + if d.isDir() { return syserror.EISDIR } virtfs := rp.VirtualFilesystem() - parentDentry := vfsd.Parent().Impl().(*Dentry) + parentDentry := d.parent parentDentry.dirMu.Lock() defer parentDentry.dirMu.Unlock() - if err := virtfs.PrepareDeleteDentry(vfs.MountNamespaceFromContext(ctx), vfsd); err != nil { + mntns := vfs.MountNamespaceFromContext(ctx) + defer mntns.DecRef() + if err := virtfs.PrepareDeleteDentry(mntns, vfsd); err != nil { return err } if err := parentDentry.inode.Unlink(ctx, rp.Component(), vfsd); err != nil { @@ -708,8 +765,20 @@ func (fs *Filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error return nil } +// BoundEndpointAt implements FilesystemImpl.BoundEndpointAt. +func (fs *Filesystem) BoundEndpointAt(ctx context.Context, rp *vfs.ResolvingPath) (transport.BoundEndpoint, error) { + fs.mu.RLock() + _, _, err := fs.walkExistingLocked(ctx, rp) + fs.mu.RUnlock() + fs.processDeferredDecRefs() + if err != nil { + return nil, err + } + return nil, syserror.ECONNREFUSED +} + // ListxattrAt implements vfs.FilesystemImpl.ListxattrAt. -func (fs *Filesystem) ListxattrAt(ctx context.Context, rp *vfs.ResolvingPath) ([]string, error) { +func (fs *Filesystem) ListxattrAt(ctx context.Context, rp *vfs.ResolvingPath, size uint64) ([]string, error) { fs.mu.RLock() _, _, err := fs.walkExistingLocked(ctx, rp) fs.mu.RUnlock() @@ -722,7 +791,7 @@ func (fs *Filesystem) ListxattrAt(ctx context.Context, rp *vfs.ResolvingPath) ([ } // GetxattrAt implements vfs.FilesystemImpl.GetxattrAt. -func (fs *Filesystem) GetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) (string, error) { +func (fs *Filesystem) GetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.GetxattrOptions) (string, error) { fs.mu.RLock() _, _, err := fs.walkExistingLocked(ctx, rp) fs.mu.RUnlock() @@ -764,5 +833,5 @@ func (fs *Filesystem) RemovexattrAt(ctx context.Context, rp *vfs.ResolvingPath, func (fs *Filesystem) PrependPath(ctx context.Context, vfsroot, vd vfs.VirtualDentry, b *fspath.Builder) error { fs.mu.RLock() defer fs.mu.RUnlock() - return vfs.GenericPrependPath(vfsroot, vd, b) + return genericPrependPath(vfsroot, vd.Mount(), vd.Dentry().Impl().(*Dentry), b) } diff --git a/pkg/sentry/fsimpl/kernfs/inode_impl_util.go b/pkg/sentry/fsimpl/kernfs/inode_impl_util.go index adca2313f..9f526359e 100644 --- a/pkg/sentry/fsimpl/kernfs/inode_impl_util.go +++ b/pkg/sentry/fsimpl/kernfs/inode_impl_util.go @@ -36,20 +36,20 @@ type InodeNoopRefCount struct { } // IncRef implements Inode.IncRef. -func (n *InodeNoopRefCount) IncRef() { +func (InodeNoopRefCount) IncRef() { } // DecRef implements Inode.DecRef. -func (n *InodeNoopRefCount) DecRef() { +func (InodeNoopRefCount) DecRef() { } // TryIncRef implements Inode.TryIncRef. -func (n *InodeNoopRefCount) TryIncRef() bool { +func (InodeNoopRefCount) TryIncRef() bool { return true } // Destroy implements Inode.Destroy. -func (n *InodeNoopRefCount) Destroy() { +func (InodeNoopRefCount) Destroy() { } // InodeDirectoryNoNewChildren partially implements the Inode interface. @@ -58,27 +58,27 @@ func (n *InodeNoopRefCount) Destroy() { type InodeDirectoryNoNewChildren struct{} // NewFile implements Inode.NewFile. -func (*InodeDirectoryNoNewChildren) NewFile(context.Context, string, vfs.OpenOptions) (*vfs.Dentry, error) { +func (InodeDirectoryNoNewChildren) NewFile(context.Context, string, vfs.OpenOptions) (*vfs.Dentry, error) { return nil, syserror.EPERM } // NewDir implements Inode.NewDir. -func (*InodeDirectoryNoNewChildren) NewDir(context.Context, string, vfs.MkdirOptions) (*vfs.Dentry, error) { +func (InodeDirectoryNoNewChildren) NewDir(context.Context, string, vfs.MkdirOptions) (*vfs.Dentry, error) { return nil, syserror.EPERM } // NewLink implements Inode.NewLink. -func (*InodeDirectoryNoNewChildren) NewLink(context.Context, string, Inode) (*vfs.Dentry, error) { +func (InodeDirectoryNoNewChildren) NewLink(context.Context, string, Inode) (*vfs.Dentry, error) { return nil, syserror.EPERM } // NewSymlink implements Inode.NewSymlink. -func (*InodeDirectoryNoNewChildren) NewSymlink(context.Context, string, string) (*vfs.Dentry, error) { +func (InodeDirectoryNoNewChildren) NewSymlink(context.Context, string, string) (*vfs.Dentry, error) { return nil, syserror.EPERM } // NewNode implements Inode.NewNode. -func (*InodeDirectoryNoNewChildren) NewNode(context.Context, string, vfs.MknodOptions) (*vfs.Dentry, error) { +func (InodeDirectoryNoNewChildren) NewNode(context.Context, string, vfs.MknodOptions) (*vfs.Dentry, error) { return nil, syserror.EPERM } @@ -90,62 +90,62 @@ type InodeNotDirectory struct { } // HasChildren implements Inode.HasChildren. -func (*InodeNotDirectory) HasChildren() bool { +func (InodeNotDirectory) HasChildren() bool { return false } // NewFile implements Inode.NewFile. -func (*InodeNotDirectory) NewFile(context.Context, string, vfs.OpenOptions) (*vfs.Dentry, error) { +func (InodeNotDirectory) NewFile(context.Context, string, vfs.OpenOptions) (*vfs.Dentry, error) { panic("NewFile called on non-directory inode") } // NewDir implements Inode.NewDir. -func (*InodeNotDirectory) NewDir(context.Context, string, vfs.MkdirOptions) (*vfs.Dentry, error) { +func (InodeNotDirectory) NewDir(context.Context, string, vfs.MkdirOptions) (*vfs.Dentry, error) { panic("NewDir called on non-directory inode") } // NewLink implements Inode.NewLinkink. -func (*InodeNotDirectory) NewLink(context.Context, string, Inode) (*vfs.Dentry, error) { +func (InodeNotDirectory) NewLink(context.Context, string, Inode) (*vfs.Dentry, error) { panic("NewLink called on non-directory inode") } // NewSymlink implements Inode.NewSymlink. -func (*InodeNotDirectory) NewSymlink(context.Context, string, string) (*vfs.Dentry, error) { +func (InodeNotDirectory) NewSymlink(context.Context, string, string) (*vfs.Dentry, error) { panic("NewSymlink called on non-directory inode") } // NewNode implements Inode.NewNode. -func (*InodeNotDirectory) NewNode(context.Context, string, vfs.MknodOptions) (*vfs.Dentry, error) { +func (InodeNotDirectory) NewNode(context.Context, string, vfs.MknodOptions) (*vfs.Dentry, error) { panic("NewNode called on non-directory inode") } // Unlink implements Inode.Unlink. -func (*InodeNotDirectory) Unlink(context.Context, string, *vfs.Dentry) error { +func (InodeNotDirectory) Unlink(context.Context, string, *vfs.Dentry) error { panic("Unlink called on non-directory inode") } // RmDir implements Inode.RmDir. -func (*InodeNotDirectory) RmDir(context.Context, string, *vfs.Dentry) error { +func (InodeNotDirectory) RmDir(context.Context, string, *vfs.Dentry) error { panic("RmDir called on non-directory inode") } // Rename implements Inode.Rename. -func (*InodeNotDirectory) Rename(context.Context, string, string, *vfs.Dentry, *vfs.Dentry) (*vfs.Dentry, error) { +func (InodeNotDirectory) Rename(context.Context, string, string, *vfs.Dentry, *vfs.Dentry) (*vfs.Dentry, error) { panic("Rename called on non-directory inode") } // Lookup implements Inode.Lookup. -func (*InodeNotDirectory) Lookup(ctx context.Context, name string) (*vfs.Dentry, error) { +func (InodeNotDirectory) Lookup(ctx context.Context, name string) (*vfs.Dentry, error) { panic("Lookup called on non-directory inode") } // IterDirents implements Inode.IterDirents. -func (*InodeNotDirectory) IterDirents(ctx context.Context, callback vfs.IterDirentsCallback, offset, relOffset int64) (newOffset int64, err error) { +func (InodeNotDirectory) IterDirents(ctx context.Context, callback vfs.IterDirentsCallback, offset, relOffset int64) (newOffset int64, err error) { panic("IterDirents called on non-directory inode") } // Valid implements Inode.Valid. -func (*InodeNotDirectory) Valid(context.Context) bool { +func (InodeNotDirectory) Valid(context.Context) bool { return true } @@ -157,17 +157,17 @@ func (*InodeNotDirectory) Valid(context.Context) bool { type InodeNoDynamicLookup struct{} // Lookup implements Inode.Lookup. -func (*InodeNoDynamicLookup) Lookup(ctx context.Context, name string) (*vfs.Dentry, error) { +func (InodeNoDynamicLookup) Lookup(ctx context.Context, name string) (*vfs.Dentry, error) { return nil, syserror.ENOENT } // IterDirents implements Inode.IterDirents. -func (*InodeNoDynamicLookup) IterDirents(ctx context.Context, callback vfs.IterDirentsCallback, offset, relOffset int64) (int64, error) { +func (InodeNoDynamicLookup) IterDirents(ctx context.Context, callback vfs.IterDirentsCallback, offset, relOffset int64) (int64, error) { return offset, nil } // Valid implements Inode.Valid. -func (*InodeNoDynamicLookup) Valid(ctx context.Context) bool { +func (InodeNoDynamicLookup) Valid(ctx context.Context) bool { return true } @@ -177,10 +177,15 @@ func (*InodeNoDynamicLookup) Valid(ctx context.Context) bool { type InodeNotSymlink struct{} // Readlink implements Inode.Readlink. -func (*InodeNotSymlink) Readlink(context.Context) (string, error) { +func (InodeNotSymlink) Readlink(context.Context) (string, error) { return "", syserror.EINVAL } +// Getlink implements Inode.Getlink. +func (InodeNotSymlink) Getlink(context.Context) (vfs.VirtualDentry, string, error) { + return vfs.VirtualDentry{}, "", syserror.EINVAL +} + // InodeAttrs partially implements the Inode interface, specifically the // inodeMetadata sub interface. InodeAttrs provides functionality related to // inode attributes. @@ -219,7 +224,7 @@ func (a *InodeAttrs) Mode() linux.FileMode { // Stat partially implements Inode.Stat. Note that this function doesn't provide // all the stat fields, and the embedder should consider extending the result // with filesystem-specific fields. -func (a *InodeAttrs) Stat(*vfs.Filesystem) linux.Statx { +func (a *InodeAttrs) Stat(*vfs.Filesystem, vfs.StatOptions) (linux.Statx, error) { var stat linux.Statx stat.Mask = linux.STATX_TYPE | linux.STATX_MODE | linux.STATX_UID | linux.STATX_GID | linux.STATX_INO | linux.STATX_NLINK stat.Ino = atomic.LoadUint64(&a.ino) @@ -228,13 +233,23 @@ func (a *InodeAttrs) Stat(*vfs.Filesystem) linux.Statx { stat.GID = atomic.LoadUint32(&a.gid) stat.Nlink = atomic.LoadUint32(&a.nlink) - // TODO: Implement other stat fields like timestamps. + // TODO(gvisor.dev/issue/1193): Implement other stat fields like timestamps. - return stat + return stat, nil } // SetStat implements Inode.SetStat. -func (a *InodeAttrs) SetStat(_ *vfs.Filesystem, opts vfs.SetStatOptions) error { +func (a *InodeAttrs) SetStat(ctx context.Context, fs *vfs.Filesystem, creds *auth.Credentials, opts vfs.SetStatOptions) error { + if opts.Stat.Mask == 0 { + return nil + } + if opts.Stat.Mask&^(linux.STATX_MODE|linux.STATX_UID|linux.STATX_GID) != 0 { + return syserror.EPERM + } + if err := vfs.CheckSetStat(ctx, creds, &opts.Stat, a.Mode(), auth.KUID(atomic.LoadUint32(&a.uid)), auth.KGID(atomic.LoadUint32(&a.gid))); err != nil { + return err + } + stat := opts.Stat if stat.Mask&linux.STATX_MODE != 0 { for { @@ -256,19 +271,17 @@ func (a *InodeAttrs) SetStat(_ *vfs.Filesystem, opts vfs.SetStatOptions) error { // Note that not all fields are modifiable. For example, the file type and // inode numbers are immutable after node creation. - // TODO: Implement other stat fields like timestamps. + // TODO(gvisor.dev/issue/1193): Implement other stat fields like timestamps. return nil } // CheckPermissions implements Inode.CheckPermissions. func (a *InodeAttrs) CheckPermissions(_ context.Context, creds *auth.Credentials, ats vfs.AccessTypes) error { - mode := a.Mode() return vfs.GenericCheckPermissions( creds, ats, - mode.FileType() == linux.ModeDirectory, - uint16(mode), + a.Mode(), auth.KUID(atomic.LoadUint32(&a.uid)), auth.KGID(atomic.LoadUint32(&a.gid)), ) @@ -357,7 +370,7 @@ func (o *OrderedChildren) Populate(d *Dentry, children map[string]*Dentry) uint3 if err := o.Insert(name, child.VFSDentry()); err != nil { panic(fmt.Sprintf("Collision when attempting to insert child %q (%+v) into %+v", name, child, d)) } - d.InsertChild(name, child.VFSDentry()) + d.InsertChild(name, child) } return links } @@ -507,7 +520,7 @@ type InodeSymlink struct { } // Open implements Inode.Open. -func (InodeSymlink) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, flags uint32) (*vfs.FileDescription, error) { +func (InodeSymlink) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { return nil, syserror.ELOOP } @@ -549,8 +562,21 @@ func (s *StaticDirectory) Init(creds *auth.Credentials, ino uint64, perm linux.F } // Open implements kernfs.Inode. -func (s *StaticDirectory) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, flags uint32) (*vfs.FileDescription, error) { +func (s *StaticDirectory) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { fd := &GenericDirectoryFD{} - fd.Init(rp.Mount(), vfsd, &s.OrderedChildren, flags) + fd.Init(rp.Mount(), vfsd, &s.OrderedChildren, &opts) return fd.VFSFileDescription(), nil } + +// SetStat implements Inode.SetStat not allowing inode attributes to be changed. +func (*StaticDirectory) SetStat(context.Context, *vfs.Filesystem, *auth.Credentials, vfs.SetStatOptions) error { + return syserror.EPERM +} + +// AlwaysValid partially implements kernfs.inodeDynamicLookup. +type AlwaysValid struct{} + +// Valid implements kernfs.inodeDynamicLookup. +func (*AlwaysValid) Valid(context.Context) bool { + return true +} diff --git a/pkg/sentry/fsimpl/kernfs/kernfs.go b/pkg/sentry/fsimpl/kernfs/kernfs.go index 79ebea8a5..f5041824f 100644 --- a/pkg/sentry/fsimpl/kernfs/kernfs.go +++ b/pkg/sentry/fsimpl/kernfs/kernfs.go @@ -63,9 +63,6 @@ import ( "gvisor.dev/gvisor/pkg/sync" ) -// FilesystemType implements vfs.FilesystemType. -type FilesystemType struct{} - // Filesystem mostly implements vfs.FilesystemImpl for a generic in-memory // filesystem. Concrete implementations are expected to embed this in their own // Filesystem type. @@ -138,8 +135,8 @@ func (fs *Filesystem) processDeferredDecRefsLocked() { // Init initializes a kernfs filesystem. This should be called from during // vfs.FilesystemType.NewFilesystem for the concrete filesystem embedding // kernfs. -func (fs *Filesystem) Init(vfsObj *vfs.VirtualFilesystem) { - fs.vfsfs.Init(vfsObj, fs) +func (fs *Filesystem) Init(vfsObj *vfs.VirtualFilesystem, fsType vfs.FilesystemType) { + fs.vfsfs.Init(vfsObj, fsType, fs) } // VFSFilesystem returns the generic vfs filesystem object. @@ -171,19 +168,22 @@ const ( // // Must be initialized by Init prior to first use. type Dentry struct { - refs.AtomicRefCount - - vfsd vfs.Dentry - inode Inode + vfsd vfs.Dentry - refs uint64 + refs.AtomicRefCount // flags caches useful information about the dentry from the inode. See the // dflags* consts above. Must be accessed by atomic ops. flags uint32 - // dirMu protects vfsd.children for directory dentries. - dirMu sync.Mutex + parent *Dentry + name string + + // dirMu protects children and the names of child Dentries. + dirMu sync.Mutex + children map[string]*Dentry + + inode Inode } // Init initializes this dentry. @@ -227,8 +227,8 @@ func (d *Dentry) DecRef() { func (d *Dentry) destroy() { d.inode.DecRef() // IncRef from Init. d.inode = nil - if parent := d.vfsd.Parent(); parent != nil { - parent.DecRef() // IncRef from Dentry.InsertChild. + if d.parent != nil { + d.parent.DecRef() // IncRef from Dentry.InsertChild. } } @@ -238,7 +238,7 @@ func (d *Dentry) destroy() { // updates the link count on d if required. // // Precondition: d must represent a directory inode. -func (d *Dentry) InsertChild(name string, child *vfs.Dentry) { +func (d *Dentry) InsertChild(name string, child *Dentry) { d.dirMu.Lock() d.insertChildLocked(name, child) d.dirMu.Unlock() @@ -248,13 +248,17 @@ func (d *Dentry) InsertChild(name string, child *vfs.Dentry) { // preconditions. // // Precondition: d.dirMu must be locked. -func (d *Dentry) insertChildLocked(name string, child *vfs.Dentry) { +func (d *Dentry) insertChildLocked(name string, child *Dentry) { if !d.isDir() { panic(fmt.Sprintf("InsertChild called on non-directory Dentry: %+v.", d)) } - vfsDentry := d.VFSDentry() - vfsDentry.IncRef() // DecRef in child's Dentry.destroy. - vfsDentry.InsertChild(child, name) + d.IncRef() // DecRef in child's Dentry.destroy. + child.parent = d + child.name = name + if d.children == nil { + d.children = make(map[string]*Dentry) + } + d.children[name] = child } // The Inode interface maps filesystem-level operations that operate on paths to @@ -302,8 +306,9 @@ type Inode interface { // this inode. The returned file description should hold a reference on the // inode for its lifetime. // - // Precondition: !rp.Done(). vfsd.Impl() must be a kernfs Dentry. - Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, flags uint32) (*vfs.FileDescription, error) + // Precondition: rp.Done(). vfsd.Impl() must be the kernfs Dentry containing + // the inode on which Open() is being called. + Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) } type inodeRefs interface { @@ -320,7 +325,7 @@ type inodeMetadata interface { // CheckPermissions checks that creds may access this inode for the // requested access type, per the the rules of // fs/namei.c:generic_permission(). - CheckPermissions(ctx context.Context, creds *auth.Credentials, atx vfs.AccessTypes) error + CheckPermissions(ctx context.Context, creds *auth.Credentials, ats vfs.AccessTypes) error // Mode returns the (struct stat)::st_mode value for this inode. This is // separated from Stat for performance. @@ -328,11 +333,13 @@ type inodeMetadata interface { // Stat returns the metadata for this inode. This corresponds to // vfs.FilesystemImpl.StatAt. - Stat(fs *vfs.Filesystem) linux.Statx + Stat(fs *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) // SetStat updates the metadata for this inode. This corresponds to - // vfs.FilesystemImpl.SetStatAt. - SetStat(fs *vfs.Filesystem, opts vfs.SetStatOptions) error + // vfs.FilesystemImpl.SetStatAt. Implementations are responsible for checking + // if the operation can be performed (see vfs.CheckSetStat() for common + // checks). + SetStat(ctx context.Context, fs *vfs.Filesystem, creds *auth.Credentials, opts vfs.SetStatOptions) error } // Precondition: All methods in this interface may only be called on directory @@ -416,7 +423,21 @@ type inodeDynamicLookup interface { } type inodeSymlink interface { - // Readlink resolves the target of a symbolic link. If an inode is not a + // Readlink returns the target of a symbolic link. If an inode is not a // symlink, the implementation should return EINVAL. Readlink(ctx context.Context) (string, error) + + // Getlink returns the target of a symbolic link, as used by path + // resolution: + // + // - If the inode is a "magic link" (a link whose target is most accurately + // represented as a VirtualDentry), Getlink returns (ok VirtualDentry, "", + // nil). A reference is taken on the returned VirtualDentry. + // + // - If the inode is an ordinary symlink, Getlink returns (zero-value + // VirtualDentry, symlink target, nil). + // + // - If the inode is not a symlink, Getlink returns (zero-value + // VirtualDentry, "", EINVAL). + Getlink(ctx context.Context) (vfs.VirtualDentry, string, error) } diff --git a/pkg/sentry/fsimpl/kernfs/kernfs_test.go b/pkg/sentry/fsimpl/kernfs/kernfs_test.go index ee65cf491..465451f35 100644 --- a/pkg/sentry/fsimpl/kernfs/kernfs_test.go +++ b/pkg/sentry/fsimpl/kernfs/kernfs_test.go @@ -45,7 +45,10 @@ type RootDentryFn func(*auth.Credentials, *filesystem) *kernfs.Dentry func newTestSystem(t *testing.T, rootFn RootDentryFn) *testutil.System { ctx := contexttest.Context(t) creds := auth.CredentialsFromContext(ctx) - v := vfs.New() + v := &vfs.VirtualFilesystem{} + if err := v.Init(); err != nil { + t.Fatalf("VFS init: %v", err) + } v.MustRegisterFilesystemType("testfs", &fsType{rootFn: rootFn}, &vfs.RegisterFilesystemTypeOptions{ AllowUserMount: true, }) @@ -88,7 +91,7 @@ type attrs struct { kernfs.InodeAttrs } -func (a *attrs) SetStat(fs *vfs.Filesystem, opt vfs.SetStatOptions) error { +func (*attrs) SetStat(context.Context, *vfs.Filesystem, *auth.Credentials, vfs.SetStatOptions) error { return syserror.EPERM } @@ -113,9 +116,9 @@ func (fs *filesystem) newReadonlyDir(creds *auth.Credentials, mode linux.FileMod return &dir.dentry } -func (d *readonlyDir) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, flags uint32) (*vfs.FileDescription, error) { +func (d *readonlyDir) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { fd := &kernfs.GenericDirectoryFD{} - if err := fd.Init(rp.Mount(), vfsd, &d.OrderedChildren, flags); err != nil { + if err := fd.Init(rp.Mount(), vfsd, &d.OrderedChildren, &opts); err != nil { return nil, err } return fd.VFSFileDescription(), nil @@ -143,9 +146,9 @@ func (fs *filesystem) newDir(creds *auth.Credentials, mode linux.FileMode, conte return &dir.dentry } -func (d *dir) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, flags uint32) (*vfs.FileDescription, error) { +func (d *dir) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { fd := &kernfs.GenericDirectoryFD{} - fd.Init(rp.Mount(), vfsd, &d.OrderedChildren, flags) + fd.Init(rp.Mount(), vfsd, &d.OrderedChildren, &opts) return fd.VFSFileDescription(), nil } @@ -184,9 +187,13 @@ func (*dir) NewNode(context.Context, string, vfs.MknodOptions) (*vfs.Dentry, err return nil, syserror.EPERM } -func (fst *fsType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opt vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { +func (fsType) Name() string { + return "kernfs" +} + +func (fst fsType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opt vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { fs := &filesystem{} - fs.Init(vfsObj) + fs.Init(vfsObj, &fst) root := fst.rootFn(creds, fs) return fs.VFSFilesystem(), root.VFSDentry(), nil } diff --git a/pkg/sentry/fsimpl/kernfs/symlink.go b/pkg/sentry/fsimpl/kernfs/symlink.go index 0ee7eb9b7..018aa503c 100644 --- a/pkg/sentry/fsimpl/kernfs/symlink.go +++ b/pkg/sentry/fsimpl/kernfs/symlink.go @@ -18,6 +18,8 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" ) // StaticSymlink provides an Inode implementation for symlinks that point to @@ -52,3 +54,13 @@ func (s *StaticSymlink) Init(creds *auth.Credentials, ino uint64, target string) func (s *StaticSymlink) Readlink(_ context.Context) (string, error) { return s.target, nil } + +// Getlink implements Inode.Getlink. +func (s *StaticSymlink) Getlink(_ context.Context) (vfs.VirtualDentry, string, error) { + return vfs.VirtualDentry{}, s.target, nil +} + +// SetStat implements Inode.SetStat not allowing inode attributes to be changed. +func (*StaticSymlink) SetStat(context.Context, *vfs.Filesystem, *auth.Credentials, vfs.SetStatOptions) error { + return syserror.EPERM +} diff --git a/pkg/sentry/fsimpl/pipefs/BUILD b/pkg/sentry/fsimpl/pipefs/BUILD new file mode 100644 index 000000000..0d411606f --- /dev/null +++ b/pkg/sentry/fsimpl/pipefs/BUILD @@ -0,0 +1,20 @@ +load("//tools:defs.bzl", "go_library") + +licenses(["notice"]) + +go_library( + name = "pipefs", + srcs = ["pipefs.go"], + visibility = ["//pkg/sentry:internal"], + deps = [ + "//pkg/abi/linux", + "//pkg/context", + "//pkg/sentry/fsimpl/kernfs", + "//pkg/sentry/kernel/auth", + "//pkg/sentry/kernel/pipe", + "//pkg/sentry/kernel/time", + "//pkg/sentry/vfs", + "//pkg/syserror", + "//pkg/usermem", + ], +) diff --git a/pkg/sentry/fsimpl/pipefs/pipefs.go b/pkg/sentry/fsimpl/pipefs/pipefs.go new file mode 100644 index 000000000..faf3179bc --- /dev/null +++ b/pkg/sentry/fsimpl/pipefs/pipefs.go @@ -0,0 +1,148 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// Package pipefs provides the filesystem implementation backing +// Kernel.PipeMount. +package pipefs + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/kernel/pipe" + ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +type filesystemType struct{} + +// Name implements vfs.FilesystemType.Name. +func (filesystemType) Name() string { + return "pipefs" +} + +// GetFilesystem implements vfs.FilesystemType.GetFilesystem. +func (filesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { + panic("pipefs.filesystemType.GetFilesystem should never be called") +} + +// filesystem implements vfs.FilesystemImpl. +type filesystem struct { + kernfs.Filesystem + + // TODO(gvisor.dev/issue/1193): + // + // - kernfs does not provide a way to implement statfs, from which we + // should indicate PIPEFS_MAGIC. + // + // - kernfs does not provide a way to override names for + // vfs.FilesystemImpl.PrependPath(); pipefs inodes should use synthetic + // name fmt.Sprintf("pipe:[%d]", inode.ino). +} + +// NewFilesystem sets up and returns a new vfs.Filesystem implemented by +// pipefs. +func NewFilesystem(vfsObj *vfs.VirtualFilesystem) *vfs.Filesystem { + fs := &filesystem{} + fs.Init(vfsObj, filesystemType{}) + return fs.VFSFilesystem() +} + +// inode implements kernfs.Inode. +type inode struct { + kernfs.InodeNotDirectory + kernfs.InodeNotSymlink + kernfs.InodeNoopRefCount + + pipe *pipe.VFSPipe + + ino uint64 + uid auth.KUID + gid auth.KGID + // We use the creation timestamp for all of atime, mtime, and ctime. + ctime ktime.Time +} + +func newInode(ctx context.Context, fs *kernfs.Filesystem) *inode { + creds := auth.CredentialsFromContext(ctx) + return &inode{ + pipe: pipe.NewVFSPipe(false /* isNamed */, pipe.DefaultPipeSize, usermem.PageSize), + ino: fs.NextIno(), + uid: creds.EffectiveKUID, + gid: creds.EffectiveKGID, + ctime: ktime.NowFromContext(ctx), + } +} + +const pipeMode = 0600 | linux.S_IFIFO + +// CheckPermissions implements kernfs.Inode.CheckPermissions. +func (i *inode) CheckPermissions(ctx context.Context, creds *auth.Credentials, ats vfs.AccessTypes) error { + return vfs.GenericCheckPermissions(creds, ats, pipeMode, i.uid, i.gid) +} + +// Mode implements kernfs.Inode.Mode. +func (i *inode) Mode() linux.FileMode { + return pipeMode +} + +// Stat implements kernfs.Inode.Stat. +func (i *inode) Stat(vfsfs *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) { + ts := linux.NsecToStatxTimestamp(i.ctime.Nanoseconds()) + return linux.Statx{ + Mask: linux.STATX_TYPE | linux.STATX_MODE | linux.STATX_NLINK | linux.STATX_UID | linux.STATX_GID | linux.STATX_ATIME | linux.STATX_MTIME | linux.STATX_CTIME | linux.STATX_INO | linux.STATX_SIZE | linux.STATX_BLOCKS, + Blksize: usermem.PageSize, + Nlink: 1, + UID: uint32(i.uid), + GID: uint32(i.gid), + Mode: pipeMode, + Ino: i.ino, + Size: 0, + Blocks: 0, + Atime: ts, + Ctime: ts, + Mtime: ts, + // TODO(gvisor.dev/issue/1197): Device number. + }, nil +} + +// SetStat implements kernfs.Inode.SetStat. +func (i *inode) SetStat(ctx context.Context, vfsfs *vfs.Filesystem, creds *auth.Credentials, opts vfs.SetStatOptions) error { + if opts.Stat.Mask == 0 { + return nil + } + return syserror.EPERM +} + +// Open implements kernfs.Inode.Open. +func (i *inode) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + // FIXME(b/38173783): kernfs does not plumb Context here. + return i.pipe.Open(context.Background(), rp.Mount(), vfsd, opts.Flags) +} + +// NewConnectedPipeFDs returns a pair of FileDescriptions representing the read +// and write ends of a newly-created pipe, as for pipe(2) and pipe2(2). +// +// Preconditions: mnt.Filesystem() must have been returned by NewFilesystem(). +func NewConnectedPipeFDs(ctx context.Context, mnt *vfs.Mount, flags uint32) (*vfs.FileDescription, *vfs.FileDescription) { + fs := mnt.Filesystem().Impl().(*kernfs.Filesystem) + inode := newInode(ctx, fs) + var d kernfs.Dentry + d.Init(inode) + defer d.DecRef() + return inode.pipe.ReaderWriterPair(mnt, d.VFSDentry(), flags) +} diff --git a/pkg/sentry/fsimpl/proc/BUILD b/pkg/sentry/fsimpl/proc/BUILD index 12aac2e6a..17c1342b5 100644 --- a/pkg/sentry/fsimpl/proc/BUILD +++ b/pkg/sentry/fsimpl/proc/BUILD @@ -8,18 +8,21 @@ go_library( "filesystem.go", "subtasks.go", "task.go", + "task_fds.go", "task_files.go", + "task_net.go", "tasks.go", "tasks_files.go", - "tasks_net.go", "tasks_sys.go", ], + visibility = ["//pkg/sentry:internal"], deps = [ "//pkg/abi/linux", "//pkg/context", "//pkg/log", + "//pkg/refs", "//pkg/safemem", - "//pkg/sentry/fs", + "//pkg/sentry/fsbridge", "//pkg/sentry/fsimpl/kernfs", "//pkg/sentry/inet", "//pkg/sentry/kernel", @@ -52,6 +55,7 @@ go_test( "//pkg/fspath", "//pkg/sentry/contexttest", "//pkg/sentry/fsimpl/testutil", + "//pkg/sentry/fsimpl/tmpfs", "//pkg/sentry/inet", "//pkg/sentry/kernel", "//pkg/sentry/kernel/auth", diff --git a/pkg/sentry/fsimpl/proc/filesystem.go b/pkg/sentry/fsimpl/proc/filesystem.go index 11477b6a9..104fc9030 100644 --- a/pkg/sentry/fsimpl/proc/filesystem.go +++ b/pkg/sentry/fsimpl/proc/filesystem.go @@ -26,15 +26,23 @@ import ( "gvisor.dev/gvisor/pkg/sentry/vfs" ) -// procFSType is the factory class for procfs. +// Name is the default filesystem name. +const Name = "proc" + +// FilesystemType is the factory class for procfs. // // +stateify savable -type procFSType struct{} +type FilesystemType struct{} + +var _ vfs.FilesystemType = (*FilesystemType)(nil) -var _ vfs.FilesystemType = (*procFSType)(nil) +// Name implements vfs.FilesystemType.Name. +func (FilesystemType) Name() string { + return Name +} -// GetFilesystem implements vfs.FilesystemType. -func (ft *procFSType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { +// GetFilesystem implements vfs.FilesystemType.GetFilesystem. +func (ft FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { k := kernel.KernelFromContext(ctx) if k == nil { return nil, nil, fmt.Errorf("procfs requires a kernel") @@ -45,14 +53,15 @@ func (ft *procFSType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFile } procfs := &kernfs.Filesystem{} - procfs.VFSFilesystem().Init(vfsObj, procfs) + procfs.VFSFilesystem().Init(vfsObj, &ft, procfs) - var data *InternalData + var cgroups map[string]string if opts.InternalData != nil { - data = opts.InternalData.(*InternalData) + data := opts.InternalData.(*InternalData) + cgroups = data.Cgroups } - _, dentry := newTasksInode(procfs, k, pidns, data.Cgroups) + _, dentry := newTasksInode(procfs, k, pidns, cgroups) return procfs.VFSFilesystem(), dentry.VFSDentry(), nil } diff --git a/pkg/sentry/fsimpl/proc/subtasks.go b/pkg/sentry/fsimpl/proc/subtasks.go index 353e37195..a21313666 100644 --- a/pkg/sentry/fsimpl/proc/subtasks.go +++ b/pkg/sentry/fsimpl/proc/subtasks.go @@ -22,6 +22,7 @@ import ( "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs" "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserror" ) @@ -34,6 +35,7 @@ type subtasksInode struct { kernfs.InodeDirectoryNoNewChildren kernfs.InodeAttrs kernfs.OrderedChildren + kernfs.AlwaysValid task *kernel.Task pidns *kernel.PIDNamespace @@ -61,11 +63,6 @@ func newSubtasks(task *kernel.Task, pidns *kernel.PIDNamespace, inoGen InoGenera return dentry } -// Valid implements kernfs.inodeDynamicLookup. -func (i *subtasksInode) Valid(ctx context.Context) bool { - return true -} - // Lookup implements kernfs.inodeDynamicLookup. func (i *subtasksInode) Lookup(ctx context.Context, name string) (*vfs.Dentry, error) { tid, err := strconv.ParseUint(name, 10, 32) @@ -105,8 +102,8 @@ func (i *subtasksInode) IterDirents(ctx context.Context, cb vfs.IterDirentsCallb Ino: i.inoGen.NextIno(), NextOff: offset + 1, } - if !cb.Handle(dirent) { - return offset, nil + if err := cb.Handle(dirent); err != nil { + return offset, err } offset++ } @@ -114,15 +111,25 @@ func (i *subtasksInode) IterDirents(ctx context.Context, cb vfs.IterDirentsCallb } // Open implements kernfs.Inode. -func (i *subtasksInode) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, flags uint32) (*vfs.FileDescription, error) { +func (i *subtasksInode) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { fd := &kernfs.GenericDirectoryFD{} - fd.Init(rp.Mount(), vfsd, &i.OrderedChildren, flags) + fd.Init(rp.Mount(), vfsd, &i.OrderedChildren, &opts) return fd.VFSFileDescription(), nil } // Stat implements kernfs.Inode. -func (i *subtasksInode) Stat(vsfs *vfs.Filesystem) linux.Statx { - stat := i.InodeAttrs.Stat(vsfs) - stat.Nlink += uint32(i.task.ThreadGroup().Count()) - return stat +func (i *subtasksInode) Stat(vsfs *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) { + stat, err := i.InodeAttrs.Stat(vsfs, opts) + if err != nil { + return linux.Statx{}, err + } + if opts.Mask&linux.STATX_NLINK != 0 { + stat.Nlink += uint32(i.task.ThreadGroup().Count()) + } + return stat, nil +} + +// SetStat implements Inode.SetStat not allowing inode attributes to be changed. +func (*subtasksInode) SetStat(context.Context, *vfs.Filesystem, *auth.Credentials, vfs.SetStatOptions) error { + return syserror.EPERM } diff --git a/pkg/sentry/fsimpl/proc/task.go b/pkg/sentry/fsimpl/proc/task.go index eb5bc62c0..888afc0fd 100644 --- a/pkg/sentry/fsimpl/proc/task.go +++ b/pkg/sentry/fsimpl/proc/task.go @@ -45,28 +45,31 @@ var _ kernfs.Inode = (*taskInode)(nil) func newTaskInode(inoGen InoGenerator, task *kernel.Task, pidns *kernel.PIDNamespace, isThreadGroup bool, cgroupControllers map[string]string) *kernfs.Dentry { contents := map[string]*kernfs.Dentry{ - "auxv": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &auxvData{task: task}), - "cmdline": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &cmdlineData{task: task, arg: cmdlineDataArg}), - "comm": newComm(task, inoGen.NextIno(), 0444), - "environ": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &cmdlineData{task: task, arg: environDataArg}), - //"exe": newExe(t, msrc), - //"fd": newFdDir(t, msrc), - //"fdinfo": newFdInfoDir(t, msrc), - "gid_map": newTaskOwnedFile(task, inoGen.NextIno(), 0644, &idMapData{task: task, gids: true}), - "io": newTaskOwnedFile(task, inoGen.NextIno(), 0400, newIO(task, isThreadGroup)), - "maps": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &mapsData{task: task}), - //"mountinfo": seqfile.NewSeqFileInode(t, &mountInfoFile{t: t}, msrc), - //"mounts": seqfile.NewSeqFileInode(t, &mountsFile{t: t}, msrc), + "auxv": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &auxvData{task: task}), + "cmdline": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &cmdlineData{task: task, arg: cmdlineDataArg}), + "comm": newComm(task, inoGen.NextIno(), 0444), + "environ": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &cmdlineData{task: task, arg: environDataArg}), + "exe": newExeSymlink(task, inoGen.NextIno()), + "fd": newFDDirInode(task, inoGen), + "fdinfo": newFDInfoDirInode(task, inoGen), + "gid_map": newTaskOwnedFile(task, inoGen.NextIno(), 0644, &idMapData{task: task, gids: true}), + "io": newTaskOwnedFile(task, inoGen.NextIno(), 0400, newIO(task, isThreadGroup)), + "maps": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &mapsData{task: task}), + "mountinfo": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &mountInfoData{task: task}), + "mounts": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &mountsData{task: task}), + "net": newTaskNetDir(task, inoGen), "ns": newTaskOwnedDir(task, inoGen.NextIno(), 0511, map[string]*kernfs.Dentry{ "net": newNamespaceSymlink(task, inoGen.NextIno(), "net"), "pid": newNamespaceSymlink(task, inoGen.NextIno(), "pid"), "user": newNamespaceSymlink(task, inoGen.NextIno(), "user"), }), - "smaps": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &smapsData{task: task}), - "stat": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &taskStatData{task: task, pidns: pidns, tgstats: isThreadGroup}), - "statm": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &statmData{task: task}), - "status": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &statusData{task: task, pidns: pidns}), - "uid_map": newTaskOwnedFile(task, inoGen.NextIno(), 0644, &idMapData{task: task, gids: false}), + "oom_score": newTaskOwnedFile(task, inoGen.NextIno(), 0444, newStaticFile("0\n")), + "oom_score_adj": newTaskOwnedFile(task, inoGen.NextIno(), 0644, &oomScoreAdj{task: task}), + "smaps": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &smapsData{task: task}), + "stat": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &taskStatData{task: task, pidns: pidns, tgstats: isThreadGroup}), + "statm": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &statmData{task: task}), + "status": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &statusData{task: task, pidns: pidns}), + "uid_map": newTaskOwnedFile(task, inoGen.NextIno(), 0644, &idMapData{task: task, gids: false}), } if isThreadGroup { contents["task"] = newSubtasks(task, pidns, inoGen, cgroupControllers) @@ -98,19 +101,15 @@ func (i *taskInode) Valid(ctx context.Context) bool { } // Open implements kernfs.Inode. -func (i *taskInode) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, flags uint32) (*vfs.FileDescription, error) { +func (i *taskInode) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { fd := &kernfs.GenericDirectoryFD{} - fd.Init(rp.Mount(), vfsd, &i.OrderedChildren, flags) + fd.Init(rp.Mount(), vfsd, &i.OrderedChildren, &opts) return fd.VFSFileDescription(), nil } -// SetStat implements kernfs.Inode. -func (i *taskInode) SetStat(_ *vfs.Filesystem, opts vfs.SetStatOptions) error { - stat := opts.Stat - if stat.Mask&linux.STATX_MODE != 0 { - return syserror.EPERM - } - return nil +// SetStat implements Inode.SetStat not allowing inode attributes to be changed. +func (*taskInode) SetStat(context.Context, *vfs.Filesystem, *auth.Credentials, vfs.SetStatOptions) error { + return syserror.EPERM } // taskOwnedInode implements kernfs.Inode and overrides inode owner with task @@ -152,26 +151,28 @@ func newTaskOwnedDir(task *kernel.Task, ino uint64, perm linux.FileMode, childre } // Stat implements kernfs.Inode. -func (i *taskOwnedInode) Stat(fs *vfs.Filesystem) linux.Statx { - stat := i.Inode.Stat(fs) - uid, gid := i.getOwner(linux.FileMode(stat.Mode)) - stat.UID = uint32(uid) - stat.GID = uint32(gid) - return stat +func (i *taskOwnedInode) Stat(fs *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) { + stat, err := i.Inode.Stat(fs, opts) + if err != nil { + return linux.Statx{}, err + } + if opts.Mask&(linux.STATX_UID|linux.STATX_GID) != 0 { + uid, gid := i.getOwner(linux.FileMode(stat.Mode)) + if opts.Mask&linux.STATX_UID != 0 { + stat.UID = uint32(uid) + } + if opts.Mask&linux.STATX_GID != 0 { + stat.GID = uint32(gid) + } + } + return stat, nil } // CheckPermissions implements kernfs.Inode. func (i *taskOwnedInode) CheckPermissions(_ context.Context, creds *auth.Credentials, ats vfs.AccessTypes) error { mode := i.Mode() uid, gid := i.getOwner(mode) - return vfs.GenericCheckPermissions( - creds, - ats, - mode.FileType() == linux.ModeDirectory, - uint16(mode), - uid, - gid, - ) + return vfs.GenericCheckPermissions(creds, ats, mode, uid, gid) } func (i *taskOwnedInode) getOwner(mode linux.FileMode) (auth.KUID, auth.KGID) { @@ -213,28 +214,12 @@ func newIO(t *kernel.Task, isThreadGroup bool) *ioData { return &ioData{ioUsage: t} } -func newNamespaceSymlink(task *kernel.Task, ino uint64, ns string) *kernfs.Dentry { - // Namespace symlinks should contain the namespace name and the inode number - // for the namespace instance, so for example user:[123456]. We currently fake - // the inode number by sticking the symlink inode in its place. - target := fmt.Sprintf("%s:[%d]", ns, ino) - - inode := &kernfs.StaticSymlink{} - // Note: credentials are overridden by taskOwnedInode. - inode.Init(task.Credentials(), ino, target) - - taskInode := &taskOwnedInode{Inode: inode, owner: task} - d := &kernfs.Dentry{} - d.Init(taskInode) - return d -} - // newCgroupData creates inode that shows cgroup information. // From man 7 cgroups: "For each cgroup hierarchy of which the process is a // member, there is one entry containing three colon-separated fields: // hierarchy-ID:controller-list:cgroup-path" func newCgroupData(controllers map[string]string) dynamicInode { - buf := bytes.Buffer{} + var buf bytes.Buffer // The hierarchy ids must be positive integers (for cgroup v1), but the // exact number does not matter, so long as they are unique. We can diff --git a/pkg/sentry/fsimpl/proc/task_fds.go b/pkg/sentry/fsimpl/proc/task_fds.go new file mode 100644 index 000000000..046265eca --- /dev/null +++ b/pkg/sentry/fsimpl/proc/task_fds.go @@ -0,0 +1,302 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package proc + +import ( + "bytes" + "fmt" + "sort" + "strconv" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/refs" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" +) + +func getTaskFD(t *kernel.Task, fd int32) (*vfs.FileDescription, kernel.FDFlags) { + var ( + file *vfs.FileDescription + flags kernel.FDFlags + ) + t.WithMuLocked(func(t *kernel.Task) { + if fdt := t.FDTable(); fdt != nil { + file, flags = fdt.GetVFS2(fd) + } + }) + return file, flags +} + +func taskFDExists(t *kernel.Task, fd int32) bool { + file, _ := getTaskFD(t, fd) + if file == nil { + return false + } + file.DecRef() + return true +} + +type fdDir struct { + inoGen InoGenerator + task *kernel.Task + + // When produceSymlinks is set, dirents produces for the FDs are reported + // as symlink. Otherwise, they are reported as regular files. + produceSymlink bool +} + +// IterDirents implements kernfs.inodeDynamicLookup. +func (i *fdDir) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback, absOffset, relOffset int64) (int64, error) { + var fds []int32 + i.task.WithMuLocked(func(t *kernel.Task) { + if fdTable := t.FDTable(); fdTable != nil { + fds = fdTable.GetFDs() + } + }) + + offset := absOffset + relOffset + typ := uint8(linux.DT_REG) + if i.produceSymlink { + typ = linux.DT_LNK + } + + // Find the appropriate starting point. + idx := sort.Search(len(fds), func(i int) bool { return fds[i] >= int32(relOffset) }) + if idx >= len(fds) { + return offset, nil + } + for _, fd := range fds[idx:] { + dirent := vfs.Dirent{ + Name: strconv.FormatUint(uint64(fd), 10), + Type: typ, + Ino: i.inoGen.NextIno(), + NextOff: offset + 1, + } + if err := cb.Handle(dirent); err != nil { + return offset, err + } + offset++ + } + return offset, nil +} + +// fdDirInode represents the inode for /proc/[pid]/fd directory. +// +// +stateify savable +type fdDirInode struct { + kernfs.InodeNotSymlink + kernfs.InodeDirectoryNoNewChildren + kernfs.InodeAttrs + kernfs.OrderedChildren + kernfs.AlwaysValid + fdDir +} + +var _ kernfs.Inode = (*fdDirInode)(nil) + +func newFDDirInode(task *kernel.Task, inoGen InoGenerator) *kernfs.Dentry { + inode := &fdDirInode{ + fdDir: fdDir{ + inoGen: inoGen, + task: task, + produceSymlink: true, + }, + } + inode.InodeAttrs.Init(task.Credentials(), inoGen.NextIno(), linux.ModeDirectory|0555) + + dentry := &kernfs.Dentry{} + dentry.Init(inode) + inode.OrderedChildren.Init(kernfs.OrderedChildrenOptions{}) + + return dentry +} + +// Lookup implements kernfs.inodeDynamicLookup. +func (i *fdDirInode) Lookup(ctx context.Context, name string) (*vfs.Dentry, error) { + fdInt, err := strconv.ParseInt(name, 10, 32) + if err != nil { + return nil, syserror.ENOENT + } + fd := int32(fdInt) + if !taskFDExists(i.task, fd) { + return nil, syserror.ENOENT + } + taskDentry := newFDSymlink(i.task, fd, i.inoGen.NextIno()) + return taskDentry.VFSDentry(), nil +} + +// Open implements kernfs.Inode. +func (i *fdDirInode) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd := &kernfs.GenericDirectoryFD{} + fd.Init(rp.Mount(), vfsd, &i.OrderedChildren, &opts) + return fd.VFSFileDescription(), nil +} + +// CheckPermissions implements kernfs.Inode. +// +// This is to match Linux, which uses a special permission handler to guarantee +// that a process can still access /proc/self/fd after it has executed +// setuid. See fs/proc/fd.c:proc_fd_permission. +func (i *fdDirInode) CheckPermissions(ctx context.Context, creds *auth.Credentials, ats vfs.AccessTypes) error { + err := i.InodeAttrs.CheckPermissions(ctx, creds, ats) + if err == nil { + // Access granted, no extra check needed. + return nil + } + if t := kernel.TaskFromContext(ctx); t != nil { + // Allow access if the task trying to access it is in the thread group + // corresponding to this directory. + if i.task.ThreadGroup() == t.ThreadGroup() { + // Access granted (overridden). + return nil + } + } + return err +} + +// fdSymlink is an symlink for the /proc/[pid]/fd/[fd] file. +// +// +stateify savable +type fdSymlink struct { + kernfs.InodeAttrs + kernfs.InodeNoopRefCount + kernfs.InodeSymlink + + task *kernel.Task + fd int32 +} + +var _ kernfs.Inode = (*fdSymlink)(nil) + +func newFDSymlink(task *kernel.Task, fd int32, ino uint64) *kernfs.Dentry { + inode := &fdSymlink{ + task: task, + fd: fd, + } + inode.Init(task.Credentials(), ino, linux.ModeSymlink|0777) + + d := &kernfs.Dentry{} + d.Init(inode) + return d +} + +func (s *fdSymlink) Readlink(ctx context.Context) (string, error) { + file, _ := getTaskFD(s.task, s.fd) + if file == nil { + return "", syserror.ENOENT + } + defer file.DecRef() + root := vfs.RootFromContext(ctx) + defer root.DecRef() + return s.task.Kernel().VFS().PathnameWithDeleted(ctx, root, file.VirtualDentry()) +} + +func (s *fdSymlink) Getlink(ctx context.Context) (vfs.VirtualDentry, string, error) { + file, _ := getTaskFD(s.task, s.fd) + if file == nil { + return vfs.VirtualDentry{}, "", syserror.ENOENT + } + defer file.DecRef() + vd := file.VirtualDentry() + vd.IncRef() + return vd, "", nil +} + +// fdInfoDirInode represents the inode for /proc/[pid]/fdinfo directory. +// +// +stateify savable +type fdInfoDirInode struct { + kernfs.InodeNotSymlink + kernfs.InodeDirectoryNoNewChildren + kernfs.InodeAttrs + kernfs.OrderedChildren + kernfs.AlwaysValid + fdDir +} + +var _ kernfs.Inode = (*fdInfoDirInode)(nil) + +func newFDInfoDirInode(task *kernel.Task, inoGen InoGenerator) *kernfs.Dentry { + inode := &fdInfoDirInode{ + fdDir: fdDir{ + inoGen: inoGen, + task: task, + }, + } + inode.InodeAttrs.Init(task.Credentials(), inoGen.NextIno(), linux.ModeDirectory|0555) + + dentry := &kernfs.Dentry{} + dentry.Init(inode) + inode.OrderedChildren.Init(kernfs.OrderedChildrenOptions{}) + + return dentry +} + +// Lookup implements kernfs.inodeDynamicLookup. +func (i *fdInfoDirInode) Lookup(ctx context.Context, name string) (*vfs.Dentry, error) { + fdInt, err := strconv.ParseInt(name, 10, 32) + if err != nil { + return nil, syserror.ENOENT + } + fd := int32(fdInt) + if !taskFDExists(i.task, fd) { + return nil, syserror.ENOENT + } + data := &fdInfoData{ + task: i.task, + fd: fd, + } + dentry := newTaskOwnedFile(i.task, i.inoGen.NextIno(), 0444, data) + return dentry.VFSDentry(), nil +} + +// Open implements kernfs.Inode. +func (i *fdInfoDirInode) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd := &kernfs.GenericDirectoryFD{} + fd.Init(rp.Mount(), vfsd, &i.OrderedChildren, &opts) + return fd.VFSFileDescription(), nil +} + +// fdInfoData implements vfs.DynamicBytesSource for /proc/[pid]/fdinfo/[fd]. +// +// +stateify savable +type fdInfoData struct { + kernfs.DynamicBytesFile + refs.AtomicRefCount + + task *kernel.Task + fd int32 +} + +var _ dynamicInode = (*fdInfoData)(nil) + +// Generate implements vfs.DynamicBytesSource.Generate. +func (d *fdInfoData) Generate(ctx context.Context, buf *bytes.Buffer) error { + file, descriptorFlags := getTaskFD(d.task, d.fd) + if file == nil { + return syserror.ENOENT + } + defer file.DecRef() + // TODO(b/121266871): Include pos, locks, and other data. For now we only + // have flags. + // See https://www.kernel.org/doc/Documentation/filesystems/proc.txt + flags := uint(file.StatusFlags()) | descriptorFlags.ToLinuxFileFlags() + fmt.Fprintf(buf, "flags:\t0%o\n", flags) + return nil +} diff --git a/pkg/sentry/fsimpl/proc/task_files.go b/pkg/sentry/fsimpl/proc/task_files.go index efd3b3453..2c6f8bdfc 100644 --- a/pkg/sentry/fsimpl/proc/task_files.go +++ b/pkg/sentry/fsimpl/proc/task_files.go @@ -22,6 +22,7 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/safemem" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" "gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" @@ -63,6 +64,16 @@ func getMMIncRef(task *kernel.Task) (*mm.MemoryManager, error) { return m, nil } +func checkTaskState(t *kernel.Task) error { + switch t.ExitState() { + case kernel.TaskExitZombie: + return syserror.EACCES + case kernel.TaskExitDead: + return syserror.ESRCH + } + return nil +} + type bufferWriter struct { buf *bytes.Buffer } @@ -496,7 +507,7 @@ func (s *statusData) Generate(ctx context.Context, buf *bytes.Buffer) error { return nil } -// ioUsage is the /proc/<pid>/io and /proc/<pid>/task/<tid>/io data provider. +// ioUsage is the /proc/[pid]/io and /proc/[pid]/task/[tid]/io data provider. type ioUsage interface { // IOUsage returns the io usage data. IOUsage() *usage.IO @@ -525,3 +536,226 @@ func (i *ioData) Generate(ctx context.Context, buf *bytes.Buffer) error { fmt.Fprintf(buf, "cancelled_write_bytes: %d\n", io.BytesWriteCancelled) return nil } + +// oomScoreAdj is a stub of the /proc/<pid>/oom_score_adj file. +// +// +stateify savable +type oomScoreAdj struct { + kernfs.DynamicBytesFile + + task *kernel.Task +} + +var _ vfs.WritableDynamicBytesSource = (*oomScoreAdj)(nil) + +// Generate implements vfs.DynamicBytesSource.Generate. +func (o *oomScoreAdj) Generate(ctx context.Context, buf *bytes.Buffer) error { + if o.task.ExitState() == kernel.TaskExitDead { + return syserror.ESRCH + } + fmt.Fprintf(buf, "%d\n", o.task.OOMScoreAdj()) + return nil +} + +// Write implements vfs.WritableDynamicBytesSource.Write. +func (o *oomScoreAdj) Write(ctx context.Context, src usermem.IOSequence, offset int64) (int64, error) { + if src.NumBytes() == 0 { + return 0, nil + } + + // Limit input size so as not to impact performance if input size is large. + src = src.TakeFirst(usermem.PageSize - 1) + + var v int32 + n, err := usermem.CopyInt32StringInVec(ctx, src.IO, src.Addrs, &v, src.Opts) + if err != nil { + return 0, err + } + + if o.task.ExitState() == kernel.TaskExitDead { + return 0, syserror.ESRCH + } + if err := o.task.SetOOMScoreAdj(v); err != nil { + return 0, err + } + + return n, nil +} + +// exeSymlink is an symlink for the /proc/[pid]/exe file. +// +// +stateify savable +type exeSymlink struct { + kernfs.InodeAttrs + kernfs.InodeNoopRefCount + kernfs.InodeSymlink + + task *kernel.Task +} + +var _ kernfs.Inode = (*exeSymlink)(nil) + +func newExeSymlink(task *kernel.Task, ino uint64) *kernfs.Dentry { + inode := &exeSymlink{task: task} + inode.Init(task.Credentials(), ino, linux.ModeSymlink|0777) + + d := &kernfs.Dentry{} + d.Init(inode) + return d +} + +// Readlink implements kernfs.Inode. +func (s *exeSymlink) Readlink(ctx context.Context) (string, error) { + if !kernel.ContextCanTrace(ctx, s.task, false) { + return "", syserror.EACCES + } + + // Pull out the executable for /proc/[pid]/exe. + exec, err := s.executable() + if err != nil { + return "", err + } + defer exec.DecRef() + + return exec.PathnameWithDeleted(ctx), nil +} + +// Getlink implements kernfs.Inode.Getlink. +func (s *exeSymlink) Getlink(ctx context.Context) (vfs.VirtualDentry, string, error) { + if !kernel.ContextCanTrace(ctx, s.task, false) { + return vfs.VirtualDentry{}, "", syserror.EACCES + } + + exec, err := s.executable() + if err != nil { + return vfs.VirtualDentry{}, "", err + } + defer exec.DecRef() + + vd := exec.(*fsbridge.VFSFile).FileDescription().VirtualDentry() + vd.IncRef() + return vd, "", nil +} + +func (s *exeSymlink) executable() (file fsbridge.File, err error) { + if err := checkTaskState(s.task); err != nil { + return nil, err + } + + s.task.WithMuLocked(func(t *kernel.Task) { + mm := t.MemoryManager() + if mm == nil { + err = syserror.EACCES + return + } + + // The MemoryManager may be destroyed, in which case + // MemoryManager.destroy will simply set the executable to nil + // (with locks held). + file = mm.Executable() + if file == nil { + err = syserror.ESRCH + } + }) + return +} + +// mountInfoData is used to implement /proc/[pid]/mountinfo. +// +// +stateify savable +type mountInfoData struct { + kernfs.DynamicBytesFile + + task *kernel.Task +} + +var _ dynamicInode = (*mountInfoData)(nil) + +// Generate implements vfs.DynamicBytesSource.Generate. +func (i *mountInfoData) Generate(ctx context.Context, buf *bytes.Buffer) error { + var fsctx *kernel.FSContext + i.task.WithMuLocked(func(t *kernel.Task) { + fsctx = t.FSContext() + }) + if fsctx == nil { + // The task has been destroyed. Nothing to show here. + return nil + } + rootDir := fsctx.RootDirectoryVFS2() + if !rootDir.Ok() { + // Root has been destroyed. Don't try to read mounts. + return nil + } + defer rootDir.DecRef() + i.task.Kernel().VFS().GenerateProcMountInfo(ctx, rootDir, buf) + return nil +} + +// mountsData is used to implement /proc/[pid]/mounts. +// +// +stateify savable +type mountsData struct { + kernfs.DynamicBytesFile + + task *kernel.Task +} + +var _ dynamicInode = (*mountsData)(nil) + +// Generate implements vfs.DynamicBytesSource.Generate. +func (i *mountsData) Generate(ctx context.Context, buf *bytes.Buffer) error { + var fsctx *kernel.FSContext + i.task.WithMuLocked(func(t *kernel.Task) { + fsctx = t.FSContext() + }) + if fsctx == nil { + // The task has been destroyed. Nothing to show here. + return nil + } + rootDir := fsctx.RootDirectoryVFS2() + if !rootDir.Ok() { + // Root has been destroyed. Don't try to read mounts. + return nil + } + defer rootDir.DecRef() + i.task.Kernel().VFS().GenerateProcMounts(ctx, rootDir, buf) + return nil +} + +type namespaceSymlink struct { + kernfs.StaticSymlink + + task *kernel.Task +} + +func newNamespaceSymlink(task *kernel.Task, ino uint64, ns string) *kernfs.Dentry { + // Namespace symlinks should contain the namespace name and the inode number + // for the namespace instance, so for example user:[123456]. We currently fake + // the inode number by sticking the symlink inode in its place. + target := fmt.Sprintf("%s:[%d]", ns, ino) + + inode := &namespaceSymlink{task: task} + // Note: credentials are overridden by taskOwnedInode. + inode.Init(task.Credentials(), ino, target) + + taskInode := &taskOwnedInode{Inode: inode, owner: task} + d := &kernfs.Dentry{} + d.Init(taskInode) + return d +} + +// Readlink implements Inode. +func (s *namespaceSymlink) Readlink(ctx context.Context) (string, error) { + if err := checkTaskState(s.task); err != nil { + return "", err + } + return s.StaticSymlink.Readlink(ctx) +} + +// Getlink implements Inode.Getlink. +func (s *namespaceSymlink) Getlink(ctx context.Context) (vfs.VirtualDentry, string, error) { + if err := checkTaskState(s.task); err != nil { + return vfs.VirtualDentry{}, "", err + } + return s.StaticSymlink.Getlink(ctx) +} diff --git a/pkg/sentry/fsimpl/proc/tasks_net.go b/pkg/sentry/fsimpl/proc/task_net.go index 608fec017..6595fcee6 100644 --- a/pkg/sentry/fsimpl/proc/tasks_net.go +++ b/pkg/sentry/fsimpl/proc/task_net.go @@ -24,7 +24,6 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/log" - "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs" "gvisor.dev/gvisor/pkg/sentry/inet" "gvisor.dev/gvisor/pkg/sentry/kernel" @@ -32,14 +31,19 @@ import ( "gvisor.dev/gvisor/pkg/sentry/socket" "gvisor.dev/gvisor/pkg/sentry/socket/unix" "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" + "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/tcpip/header" "gvisor.dev/gvisor/pkg/usermem" ) -func newNetDir(root *auth.Credentials, inoGen InoGenerator, k *kernel.Kernel) *kernfs.Dentry { +func newTaskNetDir(task *kernel.Task, inoGen InoGenerator) *kernfs.Dentry { + k := task.Kernel() + pidns := task.PIDNamespace() + root := auth.NewRootCredentials(pidns.UserNamespace()) + var contents map[string]*kernfs.Dentry - if stack := k.NetworkStack(); stack != nil { + if stack := task.NetworkNamespace().Stack(); stack != nil { const ( arp = "IP address HW type Flags HW address Mask Device\n" netlink = "sk Eth Pid Groups Rmem Wmem Dump Locks Drops Inode\n" @@ -50,6 +54,8 @@ func newNetDir(root *auth.Credentials, inoGen InoGenerator, k *kernel.Kernel) *k ) psched := fmt.Sprintf("%08x %08x %08x %08x\n", uint64(time.Microsecond/time.Nanosecond), 64, 1000000, uint64(time.Second/time.Nanosecond)) + // TODO(gvisor.dev/issue/1833): Make sure file contents reflect the task + // network namespace. contents = map[string]*kernfs.Dentry{ "dev": newDentry(root, inoGen.NextIno(), 0444, &netDevData{stack: stack}), "snmp": newDentry(root, inoGen.NextIno(), 0444, &netSnmpData{stack: stack}), @@ -81,7 +87,7 @@ func newNetDir(root *auth.Credentials, inoGen InoGenerator, k *kernel.Kernel) *k } } - return kernfs.NewStaticDir(root, inoGen.NextIno(), 0555, contents) + return newTaskOwnedDir(task, inoGen.NextIno(), 0555, contents) } // ifinet6 implements vfs.DynamicBytesSource for /proc/net/if_inet6. @@ -200,22 +206,21 @@ var _ dynamicInode = (*netUnixData)(nil) func (n *netUnixData) Generate(ctx context.Context, buf *bytes.Buffer) error { buf.WriteString("Num RefCount Protocol Flags Type St Inode Path\n") for _, se := range n.kernel.ListSockets() { - s := se.Sock.Get() - if s == nil { - log.Debugf("Couldn't resolve weakref %v in socket table, racing with destruction?", se.Sock) + s := se.SockVFS2 + if !s.TryIncRef() { + log.Debugf("Couldn't get reference on %v in socket table, racing with destruction?", s) continue } - sfile := s.(*fs.File) - if family, _, _ := sfile.FileOperations.(socket.Socket).Type(); family != linux.AF_UNIX { + if family, _, _ := s.Impl().(socket.SocketVFS2).Type(); family != linux.AF_UNIX { s.DecRef() // Not a unix socket. continue } - sops := sfile.FileOperations.(*unix.SocketOperations) + sops := s.Impl().(*unix.SocketVFS2) addr, err := sops.Endpoint().GetLocalAddress() if err != nil { - log.Warningf("Failed to retrieve socket name from %+v: %v", sfile, err) + log.Warningf("Failed to retrieve socket name from %+v: %v", s, err) addr.Addr = "<unknown>" } @@ -228,6 +233,15 @@ func (n *netUnixData) Generate(ctx context.Context, buf *bytes.Buffer) error { } } + // Get inode number. + var ino uint64 + stat, statErr := s.Stat(ctx, vfs.StatOptions{Mask: linux.STATX_INO}) + if statErr != nil || stat.Mask&linux.STATX_INO == 0 { + log.Warningf("Failed to retrieve ino for socket file: %v", statErr) + } else { + ino = stat.Ino + } + // In the socket entry below, the value for the 'Num' field requires // some consideration. Linux prints the address to the struct // unix_sock representing a socket in the kernel, but may redact the @@ -246,14 +260,14 @@ func (n *netUnixData) Generate(ctx context.Context, buf *bytes.Buffer) error { // the definition of this struct changes over time. // // For now, we always redact this pointer. - fmt.Fprintf(buf, "%#016p: %08X %08X %08X %04X %02X %5d", + fmt.Fprintf(buf, "%#016p: %08X %08X %08X %04X %02X %8d", (*unix.SocketOperations)(nil), // Num, pointer to kernel socket struct. - sfile.ReadRefs()-1, // RefCount, don't count our own ref. + s.Refs()-1, // RefCount, don't count our own ref. 0, // Protocol, always 0 for UDS. sockFlags, // Flags. sops.Endpoint().Type(), // Type. sops.State(), // State. - sfile.InodeID(), // Inode. + ino, // Inode. ) // Path @@ -335,15 +349,14 @@ func commonGenerateTCP(ctx context.Context, buf *bytes.Buffer, k *kernel.Kernel, t := kernel.TaskFromContext(ctx) for _, se := range k.ListSockets() { - s := se.Sock.Get() - if s == nil { - log.Debugf("Couldn't resolve weakref with ID %v in socket table, racing with destruction?", se.ID) + s := se.SockVFS2 + if !s.TryIncRef() { + log.Debugf("Couldn't get reference on %v in socket table, racing with destruction?", s) continue } - sfile := s.(*fs.File) - sops, ok := sfile.FileOperations.(socket.Socket) + sops, ok := s.Impl().(socket.SocketVFS2) if !ok { - panic(fmt.Sprintf("Found non-socket file in socket table: %+v", sfile)) + panic(fmt.Sprintf("Found non-socket file in socket table: %+v", s)) } if fa, stype, _ := sops.Type(); !(family == fa && stype == linux.SOCK_STREAM) { s.DecRef() @@ -392,14 +405,15 @@ func commonGenerateTCP(ctx context.Context, buf *bytes.Buffer, k *kernel.Kernel, // Unimplemented. fmt.Fprintf(buf, "%08X ", 0) + stat, statErr := s.Stat(ctx, vfs.StatOptions{Mask: linux.STATX_UID | linux.STATX_INO}) + // Field: uid. - uattr, err := sfile.Dirent.Inode.UnstableAttr(ctx) - if err != nil { - log.Warningf("Failed to retrieve unstable attr for socket file: %v", err) + if statErr != nil || stat.Mask&linux.STATX_UID == 0 { + log.Warningf("Failed to retrieve uid for socket file: %v", statErr) fmt.Fprintf(buf, "%5d ", 0) } else { creds := auth.CredentialsFromContext(ctx) - fmt.Fprintf(buf, "%5d ", uint32(uattr.Owner.UID.In(creds.UserNamespace).OrOverflow())) + fmt.Fprintf(buf, "%5d ", uint32(auth.KUID(stat.UID).In(creds.UserNamespace).OrOverflow())) } // Field: timeout; number of unanswered 0-window probes. @@ -407,11 +421,16 @@ func commonGenerateTCP(ctx context.Context, buf *bytes.Buffer, k *kernel.Kernel, fmt.Fprintf(buf, "%8d ", 0) // Field: inode. - fmt.Fprintf(buf, "%8d ", sfile.InodeID()) + if statErr != nil || stat.Mask&linux.STATX_INO == 0 { + log.Warningf("Failed to retrieve inode for socket file: %v", statErr) + fmt.Fprintf(buf, "%8d ", 0) + } else { + fmt.Fprintf(buf, "%8d ", stat.Ino) + } // Field: refcount. Don't count the ref we obtain while deferencing // the weakref to this socket. - fmt.Fprintf(buf, "%d ", sfile.ReadRefs()-1) + fmt.Fprintf(buf, "%d ", s.Refs()-1) // Field: Socket struct address. Redacted due to the same reason as // the 'Num' field in /proc/net/unix, see netUnix.ReadSeqFileData. @@ -493,15 +512,14 @@ func (d *netUDPData) Generate(ctx context.Context, buf *bytes.Buffer) error { t := kernel.TaskFromContext(ctx) for _, se := range d.kernel.ListSockets() { - s := se.Sock.Get() - if s == nil { - log.Debugf("Couldn't resolve weakref with ID %v in socket table, racing with destruction?", se.ID) + s := se.SockVFS2 + if !s.TryIncRef() { + log.Debugf("Couldn't get reference on %v in socket table, racing with destruction?", s) continue } - sfile := s.(*fs.File) - sops, ok := sfile.FileOperations.(socket.Socket) + sops, ok := s.Impl().(socket.SocketVFS2) if !ok { - panic(fmt.Sprintf("Found non-socket file in socket table: %+v", sfile)) + panic(fmt.Sprintf("Found non-socket file in socket table: %+v", s)) } if family, stype, _ := sops.Type(); family != linux.AF_INET || stype != linux.SOCK_DGRAM { s.DecRef() @@ -545,25 +563,31 @@ func (d *netUDPData) Generate(ctx context.Context, buf *bytes.Buffer) error { // Field: retrnsmt. Always 0 for UDP. fmt.Fprintf(buf, "%08X ", 0) + stat, statErr := s.Stat(ctx, vfs.StatOptions{Mask: linux.STATX_UID | linux.STATX_INO}) + // Field: uid. - uattr, err := sfile.Dirent.Inode.UnstableAttr(ctx) - if err != nil { - log.Warningf("Failed to retrieve unstable attr for socket file: %v", err) + if statErr != nil || stat.Mask&linux.STATX_UID == 0 { + log.Warningf("Failed to retrieve uid for socket file: %v", statErr) fmt.Fprintf(buf, "%5d ", 0) } else { creds := auth.CredentialsFromContext(ctx) - fmt.Fprintf(buf, "%5d ", uint32(uattr.Owner.UID.In(creds.UserNamespace).OrOverflow())) + fmt.Fprintf(buf, "%5d ", uint32(auth.KUID(stat.UID).In(creds.UserNamespace).OrOverflow())) } // Field: timeout. Always 0 for UDP. fmt.Fprintf(buf, "%8d ", 0) // Field: inode. - fmt.Fprintf(buf, "%8d ", sfile.InodeID()) + if statErr != nil || stat.Mask&linux.STATX_INO == 0 { + log.Warningf("Failed to retrieve inode for socket file: %v", statErr) + fmt.Fprintf(buf, "%8d ", 0) + } else { + fmt.Fprintf(buf, "%8d ", stat.Ino) + } // Field: ref; reference count on the socket inode. Don't count the ref // we obtain while deferencing the weakref to this socket. - fmt.Fprintf(buf, "%d ", sfile.ReadRefs()-1) + fmt.Fprintf(buf, "%d ", s.Refs()-1) // Field: Socket struct address. Redacted due to the same reason as // the 'Num' field in /proc/net/unix, see netUnix.ReadSeqFileData. @@ -664,9 +688,9 @@ func (d *netSnmpData) Generate(ctx context.Context, buf *bytes.Buffer) error { if line.prefix == "Tcp" { tcp := stat.(*inet.StatSNMPTCP) // "Tcp" needs special processing because MaxConn is signed. RFC 2012. - fmt.Sprintf("%s: %s %d %s\n", line.prefix, sprintSlice(tcp[:3]), int64(tcp[3]), sprintSlice(tcp[4:])) + fmt.Fprintf(buf, "%s: %s %d %s\n", line.prefix, sprintSlice(tcp[:3]), int64(tcp[3]), sprintSlice(tcp[4:])) } else { - fmt.Sprintf("%s: %s\n", line.prefix, sprintSlice(toSlice(stat))) + fmt.Fprintf(buf, "%s: %s\n", line.prefix, sprintSlice(toSlice(stat))) } } return nil diff --git a/pkg/sentry/fsimpl/proc/tasks.go b/pkg/sentry/fsimpl/proc/tasks.go index 14bd334e8..9f2ef8200 100644 --- a/pkg/sentry/fsimpl/proc/tasks.go +++ b/pkg/sentry/fsimpl/proc/tasks.go @@ -46,6 +46,7 @@ type tasksInode struct { kernfs.InodeDirectoryNoNewChildren kernfs.InodeAttrs kernfs.OrderedChildren + kernfs.AlwaysValid inoGen InoGenerator pidns *kernel.PIDNamespace @@ -66,23 +67,23 @@ var _ kernfs.Inode = (*tasksInode)(nil) func newTasksInode(inoGen InoGenerator, k *kernel.Kernel, pidns *kernel.PIDNamespace, cgroupControllers map[string]string) (*tasksInode, *kernfs.Dentry) { root := auth.NewRootCredentials(pidns.UserNamespace()) contents := map[string]*kernfs.Dentry{ - "cpuinfo": newDentry(root, inoGen.NextIno(), 0444, newStaticFile(cpuInfoData(k))), - //"filesystems": newDentry(root, inoGen.NextIno(), 0444, &filesystemsData{}), - "loadavg": newDentry(root, inoGen.NextIno(), 0444, &loadavgData{}), - "sys": newSysDir(root, inoGen, k), - "meminfo": newDentry(root, inoGen.NextIno(), 0444, &meminfoData{}), - "mounts": kernfs.NewStaticSymlink(root, inoGen.NextIno(), "self/mounts"), - "net": newNetDir(root, inoGen, k), - "stat": newDentry(root, inoGen.NextIno(), 0444, &statData{}), - "uptime": newDentry(root, inoGen.NextIno(), 0444, &uptimeData{}), - "version": newDentry(root, inoGen.NextIno(), 0444, &versionData{}), + "cpuinfo": newDentry(root, inoGen.NextIno(), 0444, newStaticFileSetStat(cpuInfoData(k))), + "filesystems": newDentry(root, inoGen.NextIno(), 0444, &filesystemsData{}), + "loadavg": newDentry(root, inoGen.NextIno(), 0444, &loadavgData{}), + "sys": newSysDir(root, inoGen, k), + "meminfo": newDentry(root, inoGen.NextIno(), 0444, &meminfoData{}), + "mounts": kernfs.NewStaticSymlink(root, inoGen.NextIno(), "self/mounts"), + "net": kernfs.NewStaticSymlink(root, inoGen.NextIno(), "self/net"), + "stat": newDentry(root, inoGen.NextIno(), 0444, &statData{}), + "uptime": newDentry(root, inoGen.NextIno(), 0444, &uptimeData{}), + "version": newDentry(root, inoGen.NextIno(), 0444, &versionData{}), } inode := &tasksInode{ pidns: pidns, inoGen: inoGen, - selfSymlink: newSelfSymlink(root, inoGen.NextIno(), 0444, pidns).VFSDentry(), - threadSelfSymlink: newThreadSelfSymlink(root, inoGen.NextIno(), 0444, pidns).VFSDentry(), + selfSymlink: newSelfSymlink(root, inoGen.NextIno(), pidns).VFSDentry(), + threadSelfSymlink: newThreadSelfSymlink(root, inoGen.NextIno(), pidns).VFSDentry(), cgroupControllers: cgroupControllers, } inode.InodeAttrs.Init(root, inoGen.NextIno(), linux.ModeDirectory|0555) @@ -121,11 +122,6 @@ func (i *tasksInode) Lookup(ctx context.Context, name string) (*vfs.Dentry, erro return taskDentry.VFSDentry(), nil } -// Valid implements kernfs.inodeDynamicLookup. -func (i *tasksInode) Valid(ctx context.Context) bool { - return true -} - // IterDirents implements kernfs.inodeDynamicLookup. func (i *tasksInode) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback, offset, _ int64) (int64, error) { // fs/proc/internal.h: #define FIRST_PROCESS_ENTRY 256 @@ -151,8 +147,8 @@ func (i *tasksInode) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback Ino: i.inoGen.NextIno(), NextOff: offset + 1, } - if !cb.Handle(dirent) { - return offset, nil + if err := cb.Handle(dirent); err != nil { + return offset, err } offset++ } @@ -163,8 +159,8 @@ func (i *tasksInode) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback Ino: i.inoGen.NextIno(), NextOff: offset + 1, } - if !cb.Handle(dirent) { - return offset, nil + if err := cb.Handle(dirent); err != nil { + return offset, err } offset++ } @@ -196,8 +192,8 @@ func (i *tasksInode) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback Ino: i.inoGen.NextIno(), NextOff: FIRST_PROCESS_ENTRY + 2 + int64(tid) + 1, } - if !cb.Handle(dirent) { - return offset, nil + if err := cb.Handle(dirent); err != nil { + return offset, err } offset++ } @@ -205,23 +201,42 @@ func (i *tasksInode) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback } // Open implements kernfs.Inode. -func (i *tasksInode) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, flags uint32) (*vfs.FileDescription, error) { +func (i *tasksInode) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { fd := &kernfs.GenericDirectoryFD{} - fd.Init(rp.Mount(), vfsd, &i.OrderedChildren, flags) + fd.Init(rp.Mount(), vfsd, &i.OrderedChildren, &opts) return fd.VFSFileDescription(), nil } -func (i *tasksInode) Stat(vsfs *vfs.Filesystem) linux.Statx { - stat := i.InodeAttrs.Stat(vsfs) +func (i *tasksInode) Stat(vsfs *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) { + stat, err := i.InodeAttrs.Stat(vsfs, opts) + if err != nil { + return linux.Statx{}, err + } - // Add dynamic children to link count. - for _, tg := range i.pidns.ThreadGroups() { - if leader := tg.Leader(); leader != nil { - stat.Nlink++ + if opts.Mask&linux.STATX_NLINK != 0 { + // Add dynamic children to link count. + for _, tg := range i.pidns.ThreadGroups() { + if leader := tg.Leader(); leader != nil { + stat.Nlink++ + } } } - return stat + return stat, nil +} + +// staticFileSetStat implements a special static file that allows inode +// attributes to be set. This is to support /proc files that are readonly, but +// allow attributes to be set. +type staticFileSetStat struct { + dynamicBytesFileSetAttr + vfs.StaticData +} + +var _ dynamicInode = (*staticFileSetStat)(nil) + +func newStaticFileSetStat(data string) *staticFileSetStat { + return &staticFileSetStat{StaticData: vfs.StaticData{Data: data}} } func cpuInfoData(k *kernel.Kernel) string { diff --git a/pkg/sentry/fsimpl/proc/tasks_files.go b/pkg/sentry/fsimpl/proc/tasks_files.go index 434998910..4621e2de0 100644 --- a/pkg/sentry/fsimpl/proc/tasks_files.go +++ b/pkg/sentry/fsimpl/proc/tasks_files.go @@ -26,6 +26,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/kernel/time" "gvisor.dev/gvisor/pkg/sentry/usage" + "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" ) @@ -40,9 +41,9 @@ type selfSymlink struct { var _ kernfs.Inode = (*selfSymlink)(nil) -func newSelfSymlink(creds *auth.Credentials, ino uint64, perm linux.FileMode, pidns *kernel.PIDNamespace) *kernfs.Dentry { +func newSelfSymlink(creds *auth.Credentials, ino uint64, pidns *kernel.PIDNamespace) *kernfs.Dentry { inode := &selfSymlink{pidns: pidns} - inode.Init(creds, ino, linux.ModeSymlink|perm) + inode.Init(creds, ino, linux.ModeSymlink|0777) d := &kernfs.Dentry{} d.Init(inode) @@ -62,6 +63,16 @@ func (s *selfSymlink) Readlink(ctx context.Context) (string, error) { return strconv.FormatUint(uint64(tgid), 10), nil } +func (s *selfSymlink) Getlink(ctx context.Context) (vfs.VirtualDentry, string, error) { + target, err := s.Readlink(ctx) + return vfs.VirtualDentry{}, target, err +} + +// SetStat implements Inode.SetStat not allowing inode attributes to be changed. +func (*selfSymlink) SetStat(context.Context, *vfs.Filesystem, *auth.Credentials, vfs.SetStatOptions) error { + return syserror.EPERM +} + type threadSelfSymlink struct { kernfs.InodeAttrs kernfs.InodeNoopRefCount @@ -72,9 +83,9 @@ type threadSelfSymlink struct { var _ kernfs.Inode = (*threadSelfSymlink)(nil) -func newThreadSelfSymlink(creds *auth.Credentials, ino uint64, perm linux.FileMode, pidns *kernel.PIDNamespace) *kernfs.Dentry { +func newThreadSelfSymlink(creds *auth.Credentials, ino uint64, pidns *kernel.PIDNamespace) *kernfs.Dentry { inode := &threadSelfSymlink{pidns: pidns} - inode.Init(creds, ino, linux.ModeSymlink|perm) + inode.Init(creds, ino, linux.ModeSymlink|0777) d := &kernfs.Dentry{} d.Init(inode) @@ -95,6 +106,28 @@ func (s *threadSelfSymlink) Readlink(ctx context.Context) (string, error) { return fmt.Sprintf("%d/task/%d", tgid, tid), nil } +func (s *threadSelfSymlink) Getlink(ctx context.Context) (vfs.VirtualDentry, string, error) { + target, err := s.Readlink(ctx) + return vfs.VirtualDentry{}, target, err +} + +// SetStat implements Inode.SetStat not allowing inode attributes to be changed. +func (*threadSelfSymlink) SetStat(context.Context, *vfs.Filesystem, *auth.Credentials, vfs.SetStatOptions) error { + return syserror.EPERM +} + +// dynamicBytesFileSetAttr implements a special file that allows inode +// attributes to be set. This is to support /proc files that are readonly, but +// allow attributes to be set. +type dynamicBytesFileSetAttr struct { + kernfs.DynamicBytesFile +} + +// SetStat implements Inode.SetStat. +func (d *dynamicBytesFileSetAttr) SetStat(ctx context.Context, fs *vfs.Filesystem, creds *auth.Credentials, opts vfs.SetStatOptions) error { + return d.DynamicBytesFile.InodeAttrs.SetStat(ctx, fs, creds, opts) +} + // cpuStats contains the breakdown of CPU time for /proc/stat. type cpuStats struct { // user is time spent in userspace tasks with non-positive niceness. @@ -137,22 +170,20 @@ func (c cpuStats) String() string { // // +stateify savable type statData struct { - kernfs.DynamicBytesFile - - // k is the owning Kernel. - k *kernel.Kernel + dynamicBytesFileSetAttr } var _ dynamicInode = (*statData)(nil) // Generate implements vfs.DynamicBytesSource.Generate. -func (s *statData) Generate(ctx context.Context, buf *bytes.Buffer) error { +func (*statData) Generate(ctx context.Context, buf *bytes.Buffer) error { // TODO(b/37226836): We currently export only zero CPU stats. We could // at least provide some aggregate stats. var cpu cpuStats fmt.Fprintf(buf, "cpu %s\n", cpu) - for c, max := uint(0), s.k.ApplicationCores(); c < max; c++ { + k := kernel.KernelFromContext(ctx) + for c, max := uint(0), k.ApplicationCores(); c < max; c++ { fmt.Fprintf(buf, "cpu%d %s\n", c, cpu) } @@ -176,7 +207,7 @@ func (s *statData) Generate(ctx context.Context, buf *bytes.Buffer) error { fmt.Fprintf(buf, "ctxt 0\n") // CLOCK_REALTIME timestamp from boot, in seconds. - fmt.Fprintf(buf, "btime %d\n", s.k.Timekeeper().BootTime().Seconds()) + fmt.Fprintf(buf, "btime %d\n", k.Timekeeper().BootTime().Seconds()) // Total number of clones. // TODO(b/37226836): Count this. @@ -203,13 +234,13 @@ func (s *statData) Generate(ctx context.Context, buf *bytes.Buffer) error { // // +stateify savable type loadavgData struct { - kernfs.DynamicBytesFile + dynamicBytesFileSetAttr } var _ dynamicInode = (*loadavgData)(nil) // Generate implements vfs.DynamicBytesSource.Generate. -func (d *loadavgData) Generate(ctx context.Context, buf *bytes.Buffer) error { +func (*loadavgData) Generate(ctx context.Context, buf *bytes.Buffer) error { // TODO(b/62345059): Include real data in fields. // Column 1-3: CPU and IO utilization of the last 1, 5, and 10 minute periods. // Column 4-5: currently running processes and the total number of processes. @@ -222,17 +253,15 @@ func (d *loadavgData) Generate(ctx context.Context, buf *bytes.Buffer) error { // // +stateify savable type meminfoData struct { - kernfs.DynamicBytesFile - - // k is the owning Kernel. - k *kernel.Kernel + dynamicBytesFileSetAttr } var _ dynamicInode = (*meminfoData)(nil) // Generate implements vfs.DynamicBytesSource.Generate. -func (d *meminfoData) Generate(ctx context.Context, buf *bytes.Buffer) error { - mf := d.k.MemoryFile() +func (*meminfoData) Generate(ctx context.Context, buf *bytes.Buffer) error { + k := kernel.KernelFromContext(ctx) + mf := k.MemoryFile() mf.UpdateUsage() snapshot, totalUsage := usage.MemoryAccounting.Copy() totalSize := usage.TotalMemory(mf.TotalSize(), totalUsage) @@ -275,7 +304,7 @@ func (d *meminfoData) Generate(ctx context.Context, buf *bytes.Buffer) error { // // +stateify savable type uptimeData struct { - kernfs.DynamicBytesFile + dynamicBytesFileSetAttr } var _ dynamicInode = (*uptimeData)(nil) @@ -294,17 +323,15 @@ func (*uptimeData) Generate(ctx context.Context, buf *bytes.Buffer) error { // // +stateify savable type versionData struct { - kernfs.DynamicBytesFile - - // k is the owning Kernel. - k *kernel.Kernel + dynamicBytesFileSetAttr } var _ dynamicInode = (*versionData)(nil) // Generate implements vfs.DynamicBytesSource.Generate. -func (v *versionData) Generate(ctx context.Context, buf *bytes.Buffer) error { - init := v.k.GlobalInit() +func (*versionData) Generate(ctx context.Context, buf *bytes.Buffer) error { + k := kernel.KernelFromContext(ctx) + init := k.GlobalInit() if init == nil { // Attempted to read before the init Task is created. This can // only occur during startup, which should never need to read @@ -335,3 +362,19 @@ func (v *versionData) Generate(ctx context.Context, buf *bytes.Buffer) error { fmt.Fprintf(buf, "%s version %s %s\n", ver.Sysname, ver.Release, ver.Version) return nil } + +// filesystemsData backs /proc/filesystems. +// +// +stateify savable +type filesystemsData struct { + kernfs.DynamicBytesFile +} + +var _ dynamicInode = (*filesystemsData)(nil) + +// Generate implements vfs.DynamicBytesSource.Generate. +func (d *filesystemsData) Generate(ctx context.Context, buf *bytes.Buffer) error { + k := kernel.KernelFromContext(ctx) + k.VFS().GenerateProcFilesystems(buf) + return nil +} diff --git a/pkg/sentry/fsimpl/proc/tasks_sys.go b/pkg/sentry/fsimpl/proc/tasks_sys.go index c7ce74883..3d5dc463c 100644 --- a/pkg/sentry/fsimpl/proc/tasks_sys.go +++ b/pkg/sentry/fsimpl/proc/tasks_sys.go @@ -50,7 +50,9 @@ func newSysDir(root *auth.Credentials, inoGen InoGenerator, k *kernel.Kernel) *k func newSysNetDir(root *auth.Credentials, inoGen InoGenerator, k *kernel.Kernel) *kernfs.Dentry { var contents map[string]*kernfs.Dentry - if stack := k.NetworkStack(); stack != nil { + // TODO(gvisor.dev/issue/1833): Support for using the network stack in the + // network namespace of the calling process. + if stack := k.RootNetworkNamespace().Stack(); stack != nil { contents = map[string]*kernfs.Dentry{ "ipv4": kernfs.NewStaticDir(root, inoGen.NextIno(), 0555, map[string]*kernfs.Dentry{ "tcp_sack": newDentry(root, inoGen.NextIno(), 0644, &tcpSackData{stack: stack}), diff --git a/pkg/sentry/fsimpl/proc/tasks_test.go b/pkg/sentry/fsimpl/proc/tasks_test.go index 6fc3524db..19abb5034 100644 --- a/pkg/sentry/fsimpl/proc/tasks_test.go +++ b/pkg/sentry/fsimpl/proc/tasks_test.go @@ -25,6 +25,7 @@ import ( "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/fspath" "gvisor.dev/gvisor/pkg/sentry/fsimpl/testutil" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/tmpfs" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/vfs" @@ -47,10 +48,11 @@ var ( var ( tasksStaticFiles = map[string]testutil.DirentType{ "cpuinfo": linux.DT_REG, + "filesystems": linux.DT_REG, "loadavg": linux.DT_REG, "meminfo": linux.DT_REG, "mounts": linux.DT_LNK, - "net": linux.DT_DIR, + "net": linux.DT_LNK, "self": linux.DT_LNK, "stat": linux.DT_REG, "sys": linux.DT_DIR, @@ -63,21 +65,29 @@ var ( "thread-self": threadSelfLink.NextOff, } taskStaticFiles = map[string]testutil.DirentType{ - "auxv": linux.DT_REG, - "cgroup": linux.DT_REG, - "cmdline": linux.DT_REG, - "comm": linux.DT_REG, - "environ": linux.DT_REG, - "gid_map": linux.DT_REG, - "io": linux.DT_REG, - "maps": linux.DT_REG, - "ns": linux.DT_DIR, - "smaps": linux.DT_REG, - "stat": linux.DT_REG, - "statm": linux.DT_REG, - "status": linux.DT_REG, - "task": linux.DT_DIR, - "uid_map": linux.DT_REG, + "auxv": linux.DT_REG, + "cgroup": linux.DT_REG, + "cmdline": linux.DT_REG, + "comm": linux.DT_REG, + "environ": linux.DT_REG, + "exe": linux.DT_LNK, + "fd": linux.DT_DIR, + "fdinfo": linux.DT_DIR, + "gid_map": linux.DT_REG, + "io": linux.DT_REG, + "maps": linux.DT_REG, + "mountinfo": linux.DT_REG, + "mounts": linux.DT_REG, + "net": linux.DT_DIR, + "ns": linux.DT_DIR, + "oom_score": linux.DT_REG, + "oom_score_adj": linux.DT_REG, + "smaps": linux.DT_REG, + "stat": linux.DT_REG, + "statm": linux.DT_REG, + "status": linux.DT_REG, + "task": linux.DT_DIR, + "uid_map": linux.DT_REG, } ) @@ -90,30 +100,49 @@ func setup(t *testing.T) *testutil.System { ctx := k.SupervisorContext() creds := auth.CredentialsFromContext(ctx) - vfsObj := vfs.New() - vfsObj.MustRegisterFilesystemType("procfs", &procFSType{}, &vfs.RegisterFilesystemTypeOptions{ + k.VFS().MustRegisterFilesystemType(Name, &FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ AllowUserMount: true, }) - fsOpts := vfs.GetFilesystemOptions{ - InternalData: &InternalData{ - Cgroups: map[string]string{ - "cpuset": "/foo/cpuset", - "memory": "/foo/memory", + + mntns, err := k.VFS().NewMountNamespace(ctx, creds, "", tmpfs.Name, &vfs.GetFilesystemOptions{}) + if err != nil { + t.Fatalf("NewMountNamespace(): %v", err) + } + pop := &vfs.PathOperation{ + Root: mntns.Root(), + Start: mntns.Root(), + Path: fspath.Parse("/proc"), + } + if err := k.VFS().MkdirAt(ctx, creds, pop, &vfs.MkdirOptions{Mode: 0777}); err != nil { + t.Fatalf("MkDir(/proc): %v", err) + } + + pop = &vfs.PathOperation{ + Root: mntns.Root(), + Start: mntns.Root(), + Path: fspath.Parse("/proc"), + } + mntOpts := &vfs.MountOptions{ + GetFilesystemOptions: vfs.GetFilesystemOptions{ + InternalData: &InternalData{ + Cgroups: map[string]string{ + "cpuset": "/foo/cpuset", + "memory": "/foo/memory", + }, }, }, } - mntns, err := vfsObj.NewMountNamespace(ctx, creds, "", "procfs", &fsOpts) - if err != nil { - t.Fatalf("NewMountNamespace(): %v", err) + if err := k.VFS().MountAt(ctx, creds, "", pop, Name, mntOpts); err != nil { + t.Fatalf("MountAt(/proc): %v", err) } - return testutil.NewSystem(ctx, t, vfsObj, mntns) + return testutil.NewSystem(ctx, t, k.VFS(), mntns) } func TestTasksEmpty(t *testing.T) { s := setup(t) defer s.Destroy() - collector := s.ListDirents(s.PathOpAtRoot("/")) + collector := s.ListDirents(s.PathOpAtRoot("/proc")) s.AssertAllDirentTypes(collector, tasksStaticFiles) s.AssertDirentOffsets(collector, tasksStaticFilesNextOffs) } @@ -131,7 +160,7 @@ func TestTasks(t *testing.T) { var tasks []*kernel.Task for i := 0; i < 5; i++ { tc := k.NewThreadGroup(nil, k.RootPIDNamespace(), kernel.NewSignalHandlers(), linux.SIGCHLD, k.GlobalInit().Limits()) - task, err := testutil.CreateTask(s.Ctx, fmt.Sprintf("name-%d", i), tc) + task, err := testutil.CreateTask(s.Ctx, fmt.Sprintf("name-%d", i), tc, s.MntNs, s.Root, s.Root) if err != nil { t.Fatalf("CreateTask(): %v", err) } @@ -139,7 +168,7 @@ func TestTasks(t *testing.T) { expectedDirents[fmt.Sprintf("%d", i+1)] = linux.DT_DIR } - collector := s.ListDirents(s.PathOpAtRoot("/")) + collector := s.ListDirents(s.PathOpAtRoot("/proc")) s.AssertAllDirentTypes(collector, expectedDirents) s.AssertDirentOffsets(collector, tasksStaticFilesNextOffs) @@ -179,7 +208,7 @@ func TestTasks(t *testing.T) { } // Test lookup. - for _, path := range []string{"/1", "/2"} { + for _, path := range []string{"/proc/1", "/proc/2"} { fd, err := s.VFS.OpenAt( s.Ctx, s.Creds, @@ -189,6 +218,7 @@ func TestTasks(t *testing.T) { if err != nil { t.Fatalf("vfsfs.OpenAt(%q) failed: %v", path, err) } + defer fd.DecRef() buf := make([]byte, 1) bufIOSeq := usermem.BytesIOSequence(buf) if _, err := fd.Read(s.Ctx, bufIOSeq, vfs.ReadOptions{}); err != syserror.EISDIR { @@ -199,10 +229,10 @@ func TestTasks(t *testing.T) { if _, err := s.VFS.OpenAt( s.Ctx, s.Creds, - s.PathOpAtRoot("/9999"), + s.PathOpAtRoot("/proc/9999"), &vfs.OpenOptions{}, ); err != syserror.ENOENT { - t.Fatalf("wrong error from vfsfs.OpenAt(/9999): %v", err) + t.Fatalf("wrong error from vfsfs.OpenAt(/proc/9999): %v", err) } } @@ -213,7 +243,7 @@ func TestTasksOffset(t *testing.T) { k := kernel.KernelFromContext(s.Ctx) for i := 0; i < 3; i++ { tc := k.NewThreadGroup(nil, k.RootPIDNamespace(), kernel.NewSignalHandlers(), linux.SIGCHLD, k.GlobalInit().Limits()) - if _, err := testutil.CreateTask(s.Ctx, fmt.Sprintf("name-%d", i), tc); err != nil { + if _, err := testutil.CreateTask(s.Ctx, fmt.Sprintf("name-%d", i), tc, s.MntNs, s.Root, s.Root); err != nil { t.Fatalf("CreateTask(): %v", err) } } @@ -300,12 +330,13 @@ func TestTasksOffset(t *testing.T) { fd, err := s.VFS.OpenAt( s.Ctx, s.Creds, - s.PathOpAtRoot("/"), + s.PathOpAtRoot("/proc"), &vfs.OpenOptions{}, ) if err != nil { t.Fatalf("vfsfs.OpenAt(/) failed: %v", err) } + defer fd.DecRef() if _, err := fd.Seek(s.Ctx, tc.offset, linux.SEEK_SET); err != nil { t.Fatalf("Seek(%d, SEEK_SET): %v", tc.offset, err) } @@ -337,12 +368,12 @@ func TestTask(t *testing.T) { k := kernel.KernelFromContext(s.Ctx) tc := k.NewThreadGroup(nil, k.RootPIDNamespace(), kernel.NewSignalHandlers(), linux.SIGCHLD, k.GlobalInit().Limits()) - _, err := testutil.CreateTask(s.Ctx, "name", tc) + _, err := testutil.CreateTask(s.Ctx, "name", tc, s.MntNs, s.Root, s.Root) if err != nil { t.Fatalf("CreateTask(): %v", err) } - collector := s.ListDirents(s.PathOpAtRoot("/1")) + collector := s.ListDirents(s.PathOpAtRoot("/proc/1")) s.AssertAllDirentTypes(collector, taskStaticFiles) } @@ -352,7 +383,7 @@ func TestProcSelf(t *testing.T) { k := kernel.KernelFromContext(s.Ctx) tc := k.NewThreadGroup(nil, k.RootPIDNamespace(), kernel.NewSignalHandlers(), linux.SIGCHLD, k.GlobalInit().Limits()) - task, err := testutil.CreateTask(s.Ctx, "name", tc) + task, err := testutil.CreateTask(s.Ctx, "name", tc, s.MntNs, s.Root, s.Root) if err != nil { t.Fatalf("CreateTask(): %v", err) } @@ -360,14 +391,14 @@ func TestProcSelf(t *testing.T) { collector := s.WithTemporaryContext(task).ListDirents(&vfs.PathOperation{ Root: s.Root, Start: s.Root, - Path: fspath.Parse("/self/"), + Path: fspath.Parse("/proc/self/"), FollowFinalSymlink: true, }) s.AssertAllDirentTypes(collector, taskStaticFiles) } func iterateDir(ctx context.Context, t *testing.T, s *testutil.System, fd *vfs.FileDescription) { - t.Logf("Iterating: /proc%s", fd.MappedName(ctx)) + t.Logf("Iterating: %s", fd.MappedName(ctx)) var collector testutil.DirentCollector if err := fd.IterDirents(ctx, &collector); err != nil { @@ -384,35 +415,36 @@ func iterateDir(ctx context.Context, t *testing.T, s *testutil.System, fd *vfs.F if d.Name == "." || d.Name == ".." { continue } - childPath := path.Join(fd.MappedName(ctx), d.Name) + absPath := path.Join(fd.MappedName(ctx), d.Name) if d.Type == linux.DT_LNK { link, err := s.VFS.ReadlinkAt( ctx, auth.CredentialsFromContext(ctx), - &vfs.PathOperation{Root: s.Root, Start: s.Root, Path: fspath.Parse(childPath)}, + &vfs.PathOperation{Root: s.Root, Start: s.Root, Path: fspath.Parse(absPath)}, ) if err != nil { - t.Errorf("vfsfs.ReadlinkAt(%v) failed: %v", childPath, err) + t.Errorf("vfsfs.ReadlinkAt(%v) failed: %v", absPath, err) } else { - t.Logf("Skipping symlink: /proc%s => %s", childPath, link) + t.Logf("Skipping symlink: %s => %s", absPath, link) } continue } - t.Logf("Opening: /proc%s", childPath) + t.Logf("Opening: %s", absPath) child, err := s.VFS.OpenAt( ctx, auth.CredentialsFromContext(ctx), - &vfs.PathOperation{Root: s.Root, Start: s.Root, Path: fspath.Parse(childPath)}, + &vfs.PathOperation{Root: s.Root, Start: s.Root, Path: fspath.Parse(absPath)}, &vfs.OpenOptions{}, ) if err != nil { - t.Errorf("vfsfs.OpenAt(%v) failed: %v", childPath, err) + t.Errorf("vfsfs.OpenAt(%v) failed: %v", absPath, err) continue } + defer child.DecRef() stat, err := child.Stat(ctx, vfs.StatOptions{}) if err != nil { - t.Errorf("Stat(%v) failed: %v", childPath, err) + t.Errorf("Stat(%v) failed: %v", absPath, err) } if got := linux.FileMode(stat.Mode).DirentType(); got != d.Type { t.Errorf("wrong file mode, stat: %v, dirent: %v", got, d.Type) @@ -430,13 +462,31 @@ func TestTree(t *testing.T) { defer s.Destroy() k := kernel.KernelFromContext(s.Ctx) + + pop := &vfs.PathOperation{ + Root: s.Root, + Start: s.Root, + Path: fspath.Parse("test-file"), + } + opts := &vfs.OpenOptions{ + Flags: linux.O_RDONLY | linux.O_CREAT, + Mode: 0777, + } + file, err := s.VFS.OpenAt(s.Ctx, s.Creds, pop, opts) + if err != nil { + t.Fatalf("failed to create test file: %v", err) + } + defer file.DecRef() + var tasks []*kernel.Task for i := 0; i < 5; i++ { tc := k.NewThreadGroup(nil, k.RootPIDNamespace(), kernel.NewSignalHandlers(), linux.SIGCHLD, k.GlobalInit().Limits()) - task, err := testutil.CreateTask(s.Ctx, fmt.Sprintf("name-%d", i), tc) + task, err := testutil.CreateTask(s.Ctx, fmt.Sprintf("name-%d", i), tc, s.MntNs, s.Root, s.Root) if err != nil { t.Fatalf("CreateTask(): %v", err) } + // Add file to populate /proc/[pid]/fd and fdinfo directories. + task.FDTable().NewFDVFS2(task, 0, file, kernel.FDFlags{}) tasks = append(tasks, task) } @@ -444,11 +494,12 @@ func TestTree(t *testing.T) { fd, err := s.VFS.OpenAt( ctx, auth.CredentialsFromContext(s.Ctx), - &vfs.PathOperation{Root: s.Root, Start: s.Root, Path: fspath.Parse("/")}, + &vfs.PathOperation{Root: s.Root, Start: s.Root, Path: fspath.Parse("/proc")}, &vfs.OpenOptions{}, ) if err != nil { - t.Fatalf("vfsfs.OpenAt(/) failed: %v", err) + t.Fatalf("vfsfs.OpenAt(/proc) failed: %v", err) } iterateDir(ctx, t, s, fd) + fd.DecRef() } diff --git a/pkg/sentry/fsimpl/sockfs/BUILD b/pkg/sentry/fsimpl/sockfs/BUILD new file mode 100644 index 000000000..52084ddb5 --- /dev/null +++ b/pkg/sentry/fsimpl/sockfs/BUILD @@ -0,0 +1,17 @@ +load("//tools:defs.bzl", "go_library") + +licenses(["notice"]) + +go_library( + name = "sockfs", + srcs = ["sockfs.go"], + visibility = ["//pkg/sentry:internal"], + deps = [ + "//pkg/abi/linux", + "//pkg/context", + "//pkg/sentry/fsimpl/kernfs", + "//pkg/sentry/kernel/auth", + "//pkg/sentry/vfs", + "//pkg/syserror", + ], +) diff --git a/pkg/sentry/fsimpl/sockfs/sockfs.go b/pkg/sentry/fsimpl/sockfs/sockfs.go new file mode 100644 index 000000000..3f7ad1d65 --- /dev/null +++ b/pkg/sentry/fsimpl/sockfs/sockfs.go @@ -0,0 +1,102 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// Package sockfs provides a filesystem implementation for anonymous sockets. +package sockfs + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" +) + +// NewFilesystem creates a new sockfs filesystem. +// +// Note that there should only ever be one instance of sockfs.Filesystem, +// backing a global socket mount. +func NewFilesystem(vfsObj *vfs.VirtualFilesystem) *vfs.Filesystem { + fs, _, err := filesystemType{}.GetFilesystem(nil, vfsObj, nil, "", vfs.GetFilesystemOptions{}) + if err != nil { + panic("failed to create sockfs filesystem") + } + return fs +} + +// filesystemType implements vfs.FilesystemType. +type filesystemType struct{} + +// GetFilesystem implements FilesystemType.GetFilesystem. +func (fsType filesystemType) GetFilesystem(_ context.Context, vfsObj *vfs.VirtualFilesystem, _ *auth.Credentials, _ string, _ vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { + fs := &filesystem{} + fs.Init(vfsObj, fsType) + return fs.VFSFilesystem(), nil, nil +} + +// Name implements FilesystemType.Name. +// +// Note that registering sockfs is unnecessary, except for the fact that it +// will not show up under /proc/filesystems as a result. This is a very minor +// discrepancy from Linux. +func (filesystemType) Name() string { + return "sockfs" +} + +// filesystem implements vfs.FilesystemImpl. +type filesystem struct { + kernfs.Filesystem +} + +// inode implements kernfs.Inode. +// +// TODO(gvisor.dev/issue/1476): Add device numbers to this inode (which are +// not included in InodeAttrs) to store the numbers of the appropriate +// socket device. Override InodeAttrs.Stat() accordingly. +type inode struct { + kernfs.InodeNotDirectory + kernfs.InodeNotSymlink + kernfs.InodeAttrs + kernfs.InodeNoopRefCount +} + +// Open implements kernfs.Inode.Open. +func (i *inode) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + return nil, syserror.ENXIO +} + +// InitSocket initializes a socket FileDescription, with a corresponding +// Dentry in mnt. +// +// fd should be the FileDescription associated with socketImpl, i.e. its first +// field. mnt should be the global socket mount, Kernel.socketMount. +func InitSocket(socketImpl vfs.FileDescriptionImpl, fd *vfs.FileDescription, mnt *vfs.Mount, creds *auth.Credentials) error { + fsimpl := mnt.Filesystem().Impl() + fs := fsimpl.(*kernfs.Filesystem) + + // File mode matches net/socket.c:sock_alloc. + filemode := linux.FileMode(linux.S_IFSOCK | 0600) + i := &inode{} + i.Init(creds, fs.NextIno(), filemode) + + d := &kernfs.Dentry{} + d.Init(i) + + opts := &vfs.FileDescriptionOptions{UseDentryMetadata: true} + if err := fd.Init(socketImpl, linux.O_RDWR, mnt, d.VFSDentry(), opts); err != nil { + return err + } + return nil +} diff --git a/pkg/sentry/fsimpl/sys/BUILD b/pkg/sentry/fsimpl/sys/BUILD index 66c0d8bc8..a741e2bb6 100644 --- a/pkg/sentry/fsimpl/sys/BUILD +++ b/pkg/sentry/fsimpl/sys/BUILD @@ -7,6 +7,7 @@ go_library( srcs = [ "sys.go", ], + visibility = ["//pkg/sentry:internal"], deps = [ "//pkg/abi/linux", "//pkg/context", diff --git a/pkg/sentry/fsimpl/sys/sys.go b/pkg/sentry/fsimpl/sys/sys.go index e35d52d17..5c617270e 100644 --- a/pkg/sentry/fsimpl/sys/sys.go +++ b/pkg/sentry/fsimpl/sys/sys.go @@ -28,6 +28,9 @@ import ( "gvisor.dev/gvisor/pkg/syserror" ) +// Name is the default filesystem name. +const Name = "sysfs" + // FilesystemType implements vfs.FilesystemType. type FilesystemType struct{} @@ -36,10 +39,15 @@ type filesystem struct { kernfs.Filesystem } +// Name implements vfs.FilesystemType.Name. +func (FilesystemType) Name() string { + return Name +} + // GetFilesystem implements vfs.FilesystemType.GetFilesystem. -func (FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { +func (fsType FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { fs := &filesystem{} - fs.Filesystem.Init(vfsObj) + fs.Filesystem.Init(vfsObj, &fsType) k := kernel.KernelFromContext(ctx) maxCPUCores := k.ApplicationCores() defaultSysDirMode := linux.FileMode(0755) @@ -91,15 +99,17 @@ func (fs *filesystem) newDir(creds *auth.Credentials, mode linux.FileMode, conte return &d.dentry } -// SetStat implements kernfs.Inode.SetStat. -func (d *dir) SetStat(fs *vfs.Filesystem, opts vfs.SetStatOptions) error { +// SetStat implements Inode.SetStat not allowing inode attributes to be changed. +func (*dir) SetStat(context.Context, *vfs.Filesystem, *auth.Credentials, vfs.SetStatOptions) error { return syserror.EPERM } // Open implements kernfs.Inode.Open. -func (d *dir) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, flags uint32) (*vfs.FileDescription, error) { +func (d *dir) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { fd := &kernfs.GenericDirectoryFD{} - fd.Init(rp.Mount(), vfsd, &d.OrderedChildren, flags) + if err := fd.Init(rp.Mount(), vfsd, &d.OrderedChildren, &opts); err != nil { + return nil, err + } return fd.VFSFileDescription(), nil } diff --git a/pkg/sentry/fsimpl/sys/sys_test.go b/pkg/sentry/fsimpl/sys/sys_test.go index 8b1cf0bd0..4b3602d47 100644 --- a/pkg/sentry/fsimpl/sys/sys_test.go +++ b/pkg/sentry/fsimpl/sys/sys_test.go @@ -34,16 +34,15 @@ func newTestSystem(t *testing.T) *testutil.System { } ctx := k.SupervisorContext() creds := auth.CredentialsFromContext(ctx) - v := vfs.New() - v.MustRegisterFilesystemType("sysfs", sys.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ + k.VFS().MustRegisterFilesystemType(sys.Name, sys.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ AllowUserMount: true, }) - mns, err := v.NewMountNamespace(ctx, creds, "", "sysfs", &vfs.GetFilesystemOptions{}) + mns, err := k.VFS().NewMountNamespace(ctx, creds, "", sys.Name, &vfs.GetFilesystemOptions{}) if err != nil { t.Fatalf("Failed to create new mount namespace: %v", err) } - return testutil.NewSystem(ctx, t, v, mns) + return testutil.NewSystem(ctx, t, k.VFS(), mns) } func TestReadCPUFile(t *testing.T) { diff --git a/pkg/sentry/fsimpl/testutil/BUILD b/pkg/sentry/fsimpl/testutil/BUILD index efd5974c4..0e4053a46 100644 --- a/pkg/sentry/fsimpl/testutil/BUILD +++ b/pkg/sentry/fsimpl/testutil/BUILD @@ -16,12 +16,14 @@ go_library( "//pkg/cpuid", "//pkg/fspath", "//pkg/memutil", - "//pkg/sentry/fs", + "//pkg/sentry/fsbridge", + "//pkg/sentry/fsimpl/tmpfs", "//pkg/sentry/kernel", "//pkg/sentry/kernel/auth", "//pkg/sentry/kernel/sched", "//pkg/sentry/limits", "//pkg/sentry/loader", + "//pkg/sentry/mm", "//pkg/sentry/pgalloc", "//pkg/sentry/platform", "//pkg/sentry/platform/kvm", diff --git a/pkg/sentry/fsimpl/testutil/kernel.go b/pkg/sentry/fsimpl/testutil/kernel.go index 89f8c4915..c16a36cdb 100644 --- a/pkg/sentry/fsimpl/testutil/kernel.go +++ b/pkg/sentry/fsimpl/testutil/kernel.go @@ -23,16 +23,20 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/cpuid" + "gvisor.dev/gvisor/pkg/fspath" "gvisor.dev/gvisor/pkg/memutil" - "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/tmpfs" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/kernel/sched" "gvisor.dev/gvisor/pkg/sentry/limits" "gvisor.dev/gvisor/pkg/sentry/loader" + "gvisor.dev/gvisor/pkg/sentry/mm" "gvisor.dev/gvisor/pkg/sentry/pgalloc" "gvisor.dev/gvisor/pkg/sentry/platform" "gvisor.dev/gvisor/pkg/sentry/time" + "gvisor.dev/gvisor/pkg/sentry/vfs" // Platforms are plugable. _ "gvisor.dev/gvisor/pkg/sentry/platform/kvm" @@ -99,40 +103,67 @@ func Boot() (*kernel.Kernel, error) { return nil, fmt.Errorf("initializing kernel: %v", err) } - ctx := k.SupervisorContext() + kernel.VFS2Enabled = true - // Create mount namespace without root as it's the minimum required to create - // the global thread group. - mntns, err := fs.NewMountNamespace(ctx, nil) - if err != nil { - return nil, err + if err := k.VFS().Init(); err != nil { + return nil, fmt.Errorf("VFS init: %v", err) } + k.VFS().MustRegisterFilesystemType(tmpfs.Name, &tmpfs.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ + AllowUserMount: true, + AllowUserList: true, + }) + ls, err := limits.NewLinuxLimitSet() if err != nil { return nil, err } - tg := k.NewThreadGroup(mntns, k.RootPIDNamespace(), kernel.NewSignalHandlers(), linux.SIGCHLD, ls) + tg := k.NewThreadGroup(nil, k.RootPIDNamespace(), kernel.NewSignalHandlers(), linux.SIGCHLD, ls) k.TestOnly_SetGlobalInit(tg) return k, nil } // CreateTask creates a new bare bones task for tests. -func CreateTask(ctx context.Context, name string, tc *kernel.ThreadGroup) (*kernel.Task, error) { +func CreateTask(ctx context.Context, name string, tc *kernel.ThreadGroup, mntns *vfs.MountNamespace, root, cwd vfs.VirtualDentry) (*kernel.Task, error) { k := kernel.KernelFromContext(ctx) + exe, err := newFakeExecutable(ctx, k.VFS(), auth.CredentialsFromContext(ctx), root) + if err != nil { + return nil, err + } + m := mm.NewMemoryManager(k, k, k.SleepForAddressSpaceActivation) + m.SetExecutable(fsbridge.NewVFSFile(exe)) + config := &kernel.TaskConfig{ Kernel: k, ThreadGroup: tc, - TaskContext: &kernel.TaskContext{Name: name}, + TaskContext: &kernel.TaskContext{Name: name, MemoryManager: m}, Credentials: auth.CredentialsFromContext(ctx), + NetworkNamespace: k.RootNetworkNamespace(), AllowedCPUMask: sched.NewFullCPUSet(k.ApplicationCores()), UTSNamespace: kernel.UTSNamespaceFromContext(ctx), IPCNamespace: kernel.IPCNamespaceFromContext(ctx), AbstractSocketNamespace: kernel.NewAbstractSocketNamespace(), + MountNamespaceVFS2: mntns, + FSContext: kernel.NewFSContextVFS2(root, cwd, 0022), + FDTable: k.NewFDTable(), } return k.TaskSet().NewTask(config) } +func newFakeExecutable(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, root vfs.VirtualDentry) (*vfs.FileDescription, error) { + const name = "executable" + pop := &vfs.PathOperation{ + Root: root, + Start: root, + Path: fspath.Parse(name), + } + opts := &vfs.OpenOptions{ + Flags: linux.O_RDONLY | linux.O_CREAT, + Mode: 0777, + } + return vfsObj.OpenAt(ctx, creds, pop, opts) +} + func createMemoryFile() (*pgalloc.MemoryFile, error) { const memfileName = "test-memory" memfd, err := memutil.CreateMemFD(memfileName, 0) diff --git a/pkg/sentry/fsimpl/testutil/testutil.go b/pkg/sentry/fsimpl/testutil/testutil.go index 69fd84ddd..0556af877 100644 --- a/pkg/sentry/fsimpl/testutil/testutil.go +++ b/pkg/sentry/fsimpl/testutil/testutil.go @@ -41,12 +41,12 @@ type System struct { Creds *auth.Credentials VFS *vfs.VirtualFilesystem Root vfs.VirtualDentry - mns *vfs.MountNamespace + MntNs *vfs.MountNamespace } // NewSystem constructs a System. // -// Precondition: Caller must hold a reference on mns, whose ownership +// Precondition: Caller must hold a reference on MntNs, whose ownership // is transferred to the new System. func NewSystem(ctx context.Context, t *testing.T, v *vfs.VirtualFilesystem, mns *vfs.MountNamespace) *System { s := &System{ @@ -54,7 +54,7 @@ func NewSystem(ctx context.Context, t *testing.T, v *vfs.VirtualFilesystem, mns Ctx: ctx, Creds: auth.CredentialsFromContext(ctx), VFS: v, - mns: mns, + MntNs: mns, Root: mns.Root(), } return s @@ -75,7 +75,7 @@ func (s *System) WithSubtest(t *testing.T) *System { Ctx: s.Ctx, Creds: s.Creds, VFS: s.VFS, - mns: s.mns, + MntNs: s.MntNs, Root: s.Root, } } @@ -90,7 +90,7 @@ func (s *System) WithTemporaryContext(ctx context.Context) *System { Ctx: ctx, Creds: s.Creds, VFS: s.VFS, - mns: s.mns, + MntNs: s.MntNs, Root: s.Root, } } @@ -98,7 +98,7 @@ func (s *System) WithTemporaryContext(ctx context.Context) *System { // Destroy release resources associated with a test system. func (s *System) Destroy() { s.Root.DecRef() - s.mns.DecRef() // Reference on mns passed to NewSystem. + s.MntNs.DecRef() // Reference on MntNs passed to NewSystem. } // ReadToEnd reads the contents of fd until EOF to a string. @@ -162,6 +162,9 @@ func (s *System) ListDirents(pop *vfs.PathOperation) *DirentCollector { // exactly the specified set of expected entries. AssertAllDirentTypes respects // collector.skipDots, and implicitly checks for "." and ".." accordingly. func (s *System) AssertAllDirentTypes(collector *DirentCollector, expected map[string]DirentType) { + if expected == nil { + expected = make(map[string]DirentType) + } // Also implicitly check for "." and "..", if enabled. if !collector.skipDots { expected["."] = linux.DT_DIR @@ -226,7 +229,7 @@ func (d *DirentCollector) SkipDotsChecks(value bool) { } // Handle implements vfs.IterDirentsCallback.Handle. -func (d *DirentCollector) Handle(dirent vfs.Dirent) bool { +func (d *DirentCollector) Handle(dirent vfs.Dirent) error { d.mu.Lock() if d.dirents == nil { d.dirents = make(map[string]*vfs.Dirent) @@ -234,7 +237,7 @@ func (d *DirentCollector) Handle(dirent vfs.Dirent) bool { d.order = append(d.order, &dirent) d.dirents[dirent.Name] = &dirent d.mu.Unlock() - return true + return nil } // Count returns the number of dirents currently in the collector. diff --git a/pkg/sentry/fsimpl/tmpfs/BUILD b/pkg/sentry/fsimpl/tmpfs/BUILD index 57abd5583..a2d9649e7 100644 --- a/pkg/sentry/fsimpl/tmpfs/BUILD +++ b/pkg/sentry/fsimpl/tmpfs/BUILD @@ -15,6 +15,17 @@ go_template_instance( }, ) +go_template_instance( + name = "fstree", + out = "fstree.go", + package = "tmpfs", + prefix = "generic", + template = "//pkg/sentry/vfs/genericfstree:generic_fstree", + types = { + "Dentry": "dentry", + }, +) + go_library( name = "tmpfs", srcs = [ @@ -22,8 +33,10 @@ go_library( "device_file.go", "directory.go", "filesystem.go", + "fstree.go", "named_pipe.go", "regular_file.go", + "socket_file.go", "symlink.go", "tmpfs.go", ], @@ -46,9 +59,11 @@ go_library( "//pkg/sentry/memmap", "//pkg/sentry/pgalloc", "//pkg/sentry/platform", + "//pkg/sentry/socket/unix/transport", "//pkg/sentry/usage", "//pkg/sentry/vfs", "//pkg/sentry/vfs/lock", + "//pkg/sentry/vfs/memxattr", "//pkg/sync", "//pkg/syserror", "//pkg/usermem", diff --git a/pkg/sentry/fsimpl/tmpfs/benchmark_test.go b/pkg/sentry/fsimpl/tmpfs/benchmark_test.go index 9fce5e4b4..2fb5c4d84 100644 --- a/pkg/sentry/fsimpl/tmpfs/benchmark_test.go +++ b/pkg/sentry/fsimpl/tmpfs/benchmark_test.go @@ -168,14 +168,17 @@ func BenchmarkVFS1TmpfsStat(b *testing.B) { } } -func BenchmarkVFS2MemfsStat(b *testing.B) { +func BenchmarkVFS2TmpfsStat(b *testing.B) { for _, depth := range depths { b.Run(fmt.Sprintf("%d", depth), func(b *testing.B) { ctx := contexttest.Context(b) creds := auth.CredentialsFromContext(ctx) // Create VFS. - vfsObj := vfs.New() + vfsObj := vfs.VirtualFilesystem{} + if err := vfsObj.Init(); err != nil { + b.Fatalf("VFS init: %v", err) + } vfsObj.MustRegisterFilesystemType("tmpfs", tmpfs.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ AllowUserMount: true, }) @@ -359,14 +362,17 @@ func BenchmarkVFS1TmpfsMountStat(b *testing.B) { } } -func BenchmarkVFS2MemfsMountStat(b *testing.B) { +func BenchmarkVFS2TmpfsMountStat(b *testing.B) { for _, depth := range depths { b.Run(fmt.Sprintf("%d", depth), func(b *testing.B) { ctx := contexttest.Context(b) creds := auth.CredentialsFromContext(ctx) // Create VFS. - vfsObj := vfs.New() + vfsObj := vfs.VirtualFilesystem{} + if err := vfsObj.Init(); err != nil { + b.Fatalf("VFS init: %v", err) + } vfsObj.MustRegisterFilesystemType("tmpfs", tmpfs.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ AllowUserMount: true, }) @@ -432,13 +438,6 @@ func BenchmarkVFS2MemfsMountStat(b *testing.B) { filePathBuilder.WriteByte('/') } - // Verify that we didn't create any directories under the mount - // point (i.e. they were all created on the submount). - firstDirName := fmt.Sprintf("%d", depth) - if child := mountPoint.Dentry().Child(firstDirName); child != nil { - b.Fatalf("created directory %q under root mount, not submount", firstDirName) - } - // Create the file that will be stat'd. fd, err := vfsObj.OpenAt(ctx, creds, &vfs.PathOperation{ Root: root, diff --git a/pkg/sentry/fsimpl/tmpfs/device_file.go b/pkg/sentry/fsimpl/tmpfs/device_file.go index 84b181b90..83bf885ee 100644 --- a/pkg/sentry/fsimpl/tmpfs/device_file.go +++ b/pkg/sentry/fsimpl/tmpfs/device_file.go @@ -15,6 +15,8 @@ package tmpfs import ( + "fmt" + "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/vfs" @@ -33,6 +35,14 @@ func (fs *filesystem) newDeviceFile(creds *auth.Credentials, mode linux.FileMode major: major, minor: minor, } + switch kind { + case vfs.BlockDevice: + mode |= linux.S_IFBLK + case vfs.CharDevice: + mode |= linux.S_IFCHR + default: + panic(fmt.Sprintf("invalid DeviceKind: %v", kind)) + } file.inode.init(file, fs, creds, mode) file.inode.nlink = 1 // from parent directory return &file.inode diff --git a/pkg/sentry/fsimpl/tmpfs/directory.go b/pkg/sentry/fsimpl/tmpfs/directory.go index dc0d27cf9..f2399981b 100644 --- a/pkg/sentry/fsimpl/tmpfs/directory.go +++ b/pkg/sentry/fsimpl/tmpfs/directory.go @@ -15,40 +15,77 @@ package tmpfs import ( + "sync/atomic" + "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/syserror" ) type directory struct { - inode inode + // Since directories can't be hard-linked, each directory can only be + // associated with a single dentry, which we can store in the directory + // struct. + dentry dentry + inode inode + + // childMap maps the names of the directory's children to their dentries. + // childMap is protected by filesystem.mu. + childMap map[string]*dentry - // childList is a list containing (1) child Dentries and (2) fake Dentries + // numChildren is len(childMap), but accessed using atomic memory + // operations to avoid locking in inode.statTo(). + numChildren int64 + + // childList is a list containing (1) child dentries and (2) fake dentries // (with inode == nil) that represent the iteration position of // directoryFDs. childList is used to support directoryFD.IterDirents() - // efficiently. childList is protected by filesystem.mu. + // efficiently. childList is protected by iterMu. + iterMu sync.Mutex childList dentryList } -func (fs *filesystem) newDirectory(creds *auth.Credentials, mode linux.FileMode) *inode { +func (fs *filesystem) newDirectory(creds *auth.Credentials, mode linux.FileMode) *directory { dir := &directory{} - dir.inode.init(dir, fs, creds, mode) + dir.inode.init(dir, fs, creds, linux.S_IFDIR|mode) dir.inode.nlink = 2 // from "." and parent directory or ".." for root - return &dir.inode + dir.dentry.inode = &dir.inode + dir.dentry.vfsd.Init(&dir.dentry) + return dir } -func (i *inode) isDir() bool { - _, ok := i.impl.(*directory) - return ok +// Preconditions: filesystem.mu must be locked for writing. dir must not +// already contain a child with the given name. +func (dir *directory) insertChildLocked(child *dentry, name string) { + child.parent = &dir.dentry + child.name = name + if dir.childMap == nil { + dir.childMap = make(map[string]*dentry) + } + dir.childMap[name] = child + atomic.AddInt64(&dir.numChildren, 1) + dir.iterMu.Lock() + dir.childList.PushBack(child) + dir.iterMu.Unlock() +} + +// Preconditions: filesystem.mu must be locked for writing. +func (dir *directory) removeChildLocked(child *dentry) { + delete(dir.childMap, child.name) + atomic.AddInt64(&dir.numChildren, -1) + dir.iterMu.Lock() + dir.childList.Remove(child) + dir.iterMu.Unlock() } type directoryFD struct { fileDescription vfs.DirectoryFileDescriptionDefaultImpl - // Protected by filesystem.mu. + // Protected by directory.iterMu. iter *dentry off int64 } @@ -56,11 +93,10 @@ type directoryFD struct { // Release implements vfs.FileDescriptionImpl.Release. func (fd *directoryFD) Release() { if fd.iter != nil { - fs := fd.filesystem() dir := fd.inode().impl.(*directory) - fs.mu.Lock() + dir.iterMu.Lock() dir.childList.Remove(fd.iter) - fs.mu.Unlock() + dir.iterMu.Unlock() fd.iter = nil } } @@ -68,36 +104,41 @@ func (fd *directoryFD) Release() { // IterDirents implements vfs.FileDescriptionImpl.IterDirents. func (fd *directoryFD) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback) error { fs := fd.filesystem() - vfsd := fd.vfsfd.VirtualDentry().Dentry() + dir := fd.inode().impl.(*directory) - fs.mu.Lock() - defer fs.mu.Unlock() + // fs.mu is required to read d.parent and dentry.name. + fs.mu.RLock() + defer fs.mu.RUnlock() + dir.iterMu.Lock() + defer dir.iterMu.Unlock() + + fd.inode().touchAtime(fd.vfsfd.Mount()) if fd.off == 0 { - if !cb.Handle(vfs.Dirent{ + if err := cb.Handle(vfs.Dirent{ Name: ".", Type: linux.DT_DIR, - Ino: vfsd.Impl().(*dentry).inode.ino, + Ino: dir.inode.ino, NextOff: 1, - }) { - return nil + }); err != nil { + return err } fd.off++ } + if fd.off == 1 { - parentInode := vfsd.ParentOrSelf().Impl().(*dentry).inode - if !cb.Handle(vfs.Dirent{ + parentInode := genericParentOrSelf(&dir.dentry).inode + if err := cb.Handle(vfs.Dirent{ Name: "..", Type: parentInode.direntType(), Ino: parentInode.ino, NextOff: 2, - }) { - return nil + }); err != nil { + return err } fd.off++ } - dir := vfsd.Impl().(*dentry).inode.impl.(*directory) var child *dentry if fd.iter == nil { // Start iteration at the beginning of dir. @@ -111,14 +152,14 @@ func (fd *directoryFD) IterDirents(ctx context.Context, cb vfs.IterDirentsCallba for child != nil { // Skip other directoryFD iterators. if child.inode != nil { - if !cb.Handle(vfs.Dirent{ - Name: child.vfsd.Name(), + if err := cb.Handle(vfs.Dirent{ + Name: child.name, Type: child.inode.direntType(), Ino: child.inode.ino, NextOff: fd.off + 1, - }) { + }); err != nil { dir.childList.InsertBefore(child, fd.iter) - return nil + return err } fd.off++ } @@ -130,9 +171,9 @@ func (fd *directoryFD) IterDirents(ctx context.Context, cb vfs.IterDirentsCallba // Seek implements vfs.FileDescriptionImpl.Seek. func (fd *directoryFD) Seek(ctx context.Context, offset int64, whence int32) (int64, error) { - fs := fd.filesystem() - fs.mu.Lock() - defer fs.mu.Unlock() + dir := fd.inode().impl.(*directory) + dir.iterMu.Lock() + defer dir.iterMu.Unlock() switch whence { case linux.SEEK_SET: @@ -160,8 +201,6 @@ func (fd *directoryFD) Seek(ctx context.Context, offset int64, whence int32) (in remChildren = offset - 2 } - dir := fd.inode().impl.(*directory) - // Ensure that fd.iter exists and is not linked into dir.childList. if fd.iter == nil { fd.iter = &dentry{} diff --git a/pkg/sentry/fsimpl/tmpfs/filesystem.go b/pkg/sentry/fsimpl/tmpfs/filesystem.go index 72bc15264..5b62f9ebb 100644 --- a/pkg/sentry/fsimpl/tmpfs/filesystem.go +++ b/pkg/sentry/fsimpl/tmpfs/filesystem.go @@ -16,11 +16,12 @@ package tmpfs import ( "fmt" - "sync/atomic" "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserror" ) @@ -38,25 +39,44 @@ func (fs *filesystem) Sync(ctx context.Context) error { // // Preconditions: filesystem.mu must be locked. !rp.Done(). func stepLocked(rp *vfs.ResolvingPath, d *dentry) (*dentry, error) { - if !d.inode.isDir() { + dir, ok := d.inode.impl.(*directory) + if !ok { return nil, syserror.ENOTDIR } - if err := d.inode.checkPermissions(rp.Credentials(), vfs.MayExec, true); err != nil { + if err := d.inode.checkPermissions(rp.Credentials(), vfs.MayExec); err != nil { return nil, err } afterSymlink: - nextVFSD, err := rp.ResolveComponent(&d.vfsd) - if err != nil { - return nil, err + name := rp.Component() + if name == "." { + rp.Advance() + return d, nil + } + if name == ".." { + if isRoot, err := rp.CheckRoot(&d.vfsd); err != nil { + return nil, err + } else if isRoot || d.parent == nil { + rp.Advance() + return d, nil + } + if err := rp.CheckMount(&d.parent.vfsd); err != nil { + return nil, err + } + rp.Advance() + return d.parent, nil + } + if len(name) > linux.NAME_MAX { + return nil, syserror.ENAMETOOLONG } - if nextVFSD == nil { - // Since the Dentry tree is the sole source of truth for tmpfs, if it's - // not in the Dentry tree, it doesn't exist. + child, ok := dir.childMap[name] + if !ok { return nil, syserror.ENOENT } - next := nextVFSD.Impl().(*dentry) - if symlink, ok := next.inode.impl.(*symlink); ok && rp.ShouldFollowSymlink() { - // TODO(gvisor.dev/issues/1197): Symlink traversals updates + if err := rp.CheckMount(&child.vfsd); err != nil { + return nil, err + } + if symlink, ok := child.inode.impl.(*symlink); ok && rp.ShouldFollowSymlink() { + // TODO(gvisor.dev/issue/1197): Symlink traversals updates // access time. if err := rp.HandleSymlink(symlink.target); err != nil { return nil, err @@ -64,7 +84,7 @@ afterSymlink: goto afterSymlink // don't check the current directory again } rp.Advance() - return next, nil + return child, nil } // walkParentDirLocked resolves all but the last path component of rp to an @@ -76,7 +96,7 @@ afterSymlink: // fs/namei.c:path_parentat(). // // Preconditions: filesystem.mu must be locked. !rp.Done(). -func walkParentDirLocked(rp *vfs.ResolvingPath, d *dentry) (*dentry, error) { +func walkParentDirLocked(rp *vfs.ResolvingPath, d *dentry) (*directory, error) { for !rp.Final() { next, err := stepLocked(rp, d) if err != nil { @@ -84,10 +104,11 @@ func walkParentDirLocked(rp *vfs.ResolvingPath, d *dentry) (*dentry, error) { } d = next } - if !d.inode.isDir() { + dir, ok := d.inode.impl.(*directory) + if !ok { return nil, syserror.ENOTDIR } - return d, nil + return dir, nil } // resolveLocked resolves rp to an existing file. @@ -118,33 +139,32 @@ func resolveLocked(rp *vfs.ResolvingPath) (*dentry, error) { // // Preconditions: !rp.Done(). For the final path component in rp, // !rp.ShouldFollowSymlink(). -func (fs *filesystem) doCreateAt(rp *vfs.ResolvingPath, dir bool, create func(parent *dentry, name string) error) error { +func (fs *filesystem) doCreateAt(rp *vfs.ResolvingPath, dir bool, create func(parentDir *directory, name string) error) error { fs.mu.Lock() defer fs.mu.Unlock() - parent, err := walkParentDirLocked(rp, rp.Start().Impl().(*dentry)) + parentDir, err := walkParentDirLocked(rp, rp.Start().Impl().(*dentry)) if err != nil { return err } - if err := parent.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true /* isDir */); err != nil { + if err := parentDir.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec); err != nil { return err } name := rp.Component() if name == "." || name == ".." { return syserror.EEXIST } - // Call parent.vfsd.Child() instead of stepLocked() or rp.ResolveChild(), - // because if the child exists we want to return EEXIST immediately instead - // of attempting symlink/mount traversal. - if parent.vfsd.Child(name) != nil { + if len(name) > linux.NAME_MAX { + return syserror.ENAMETOOLONG + } + if _, ok := parentDir.childMap[name]; ok { return syserror.EEXIST } if !dir && rp.MustBeDir() { return syserror.ENOENT } - // In memfs, the only way to cause a dentry to be disowned is by removing - // it from the filesystem, so this check is equivalent to checking if - // parent has been removed. - if parent.vfsd.IsDisowned() { + // tmpfs never calls VFS.InvalidateDentry(), so parentDir.dentry can only + // be dead if it was deleted. + if parentDir.dentry.vfsd.IsDead() { return syserror.ENOENT } mnt := rp.Mount() @@ -152,7 +172,22 @@ func (fs *filesystem) doCreateAt(rp *vfs.ResolvingPath, dir bool, create func(pa return err } defer mnt.EndWrite() - return create(parent, name) + if err := create(parentDir, name); err != nil { + return err + } + parentDir.inode.touchCMtime() + return nil +} + +// AccessAt implements vfs.Filesystem.Impl.AccessAt. +func (fs *filesystem) AccessAt(ctx context.Context, rp *vfs.ResolvingPath, creds *auth.Credentials, ats vfs.AccessTypes) error { + fs.mu.RLock() + defer fs.mu.RUnlock() + d, err := resolveLocked(rp) + if err != nil { + return err + } + return d.inode.checkPermissions(creds, ats) } // GetDentryAt implements vfs.FilesystemImpl.GetDentryAt. @@ -167,7 +202,7 @@ func (fs *filesystem) GetDentryAt(ctx context.Context, rp *vfs.ResolvingPath, op if !d.inode.isDir() { return nil, syserror.ENOTDIR } - if err := d.inode.checkPermissions(rp.Credentials(), vfs.MayExec, true /* isDir */); err != nil { + if err := d.inode.checkPermissions(rp.Credentials(), vfs.MayExec); err != nil { return nil, err } } @@ -179,17 +214,17 @@ func (fs *filesystem) GetDentryAt(ctx context.Context, rp *vfs.ResolvingPath, op func (fs *filesystem) GetParentDentryAt(ctx context.Context, rp *vfs.ResolvingPath) (*vfs.Dentry, error) { fs.mu.RLock() defer fs.mu.RUnlock() - d, err := walkParentDirLocked(rp, rp.Start().Impl().(*dentry)) + dir, err := walkParentDirLocked(rp, rp.Start().Impl().(*dentry)) if err != nil { return nil, err } - d.IncRef() - return &d.vfsd, nil + dir.dentry.IncRef() + return &dir.dentry.vfsd, nil } // LinkAt implements vfs.FilesystemImpl.LinkAt. func (fs *filesystem) LinkAt(ctx context.Context, rp *vfs.ResolvingPath, vd vfs.VirtualDentry) error { - return fs.doCreateAt(rp, false /* dir */, func(parent *dentry, name string) error { + return fs.doCreateAt(rp, false /* dir */, func(parentDir *directory, name string) error { if rp.Mount() != vd.Mount() { return syserror.EXDEV } @@ -204,30 +239,27 @@ func (fs *filesystem) LinkAt(ctx context.Context, rp *vfs.ResolvingPath, vd vfs. return syserror.EMLINK } d.inode.incLinksLocked() - child := fs.newDentry(d.inode) - parent.vfsd.InsertChild(&child.vfsd, name) - parent.inode.impl.(*directory).childList.PushBack(child) + parentDir.insertChildLocked(fs.newDentry(d.inode), name) return nil }) } // MkdirAt implements vfs.FilesystemImpl.MkdirAt. func (fs *filesystem) MkdirAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.MkdirOptions) error { - return fs.doCreateAt(rp, true /* dir */, func(parent *dentry, name string) error { - if parent.inode.nlink == maxLinks { + return fs.doCreateAt(rp, true /* dir */, func(parentDir *directory, name string) error { + if parentDir.inode.nlink == maxLinks { return syserror.EMLINK } - parent.inode.incLinksLocked() // from child's ".." - child := fs.newDentry(fs.newDirectory(rp.Credentials(), opts.Mode)) - parent.vfsd.InsertChild(&child.vfsd, name) - parent.inode.impl.(*directory).childList.PushBack(child) + parentDir.inode.incLinksLocked() // from child's ".." + childDir := fs.newDirectory(rp.Credentials(), opts.Mode) + parentDir.insertChildLocked(&childDir.dentry, name) return nil }) } // MknodAt implements vfs.FilesystemImpl.MknodAt. func (fs *filesystem) MknodAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.MknodOptions) error { - return fs.doCreateAt(rp, false /* dir */, func(parent *dentry, name string) error { + return fs.doCreateAt(rp, false /* dir */, func(parentDir *directory, name string) error { var childInode *inode switch opts.Mode.FileType() { case 0, linux.S_IFREG: @@ -239,14 +271,12 @@ func (fs *filesystem) MknodAt(ctx context.Context, rp *vfs.ResolvingPath, opts v case linux.S_IFCHR: childInode = fs.newDeviceFile(rp.Credentials(), opts.Mode, vfs.CharDevice, opts.DevMajor, opts.DevMinor) case linux.S_IFSOCK: - // Not yet supported. - return syserror.EPERM + childInode = fs.newSocketFile(rp.Credentials(), opts.Mode, opts.Endpoint) default: return syserror.EINVAL } child := fs.newDentry(childInode) - parent.vfsd.InsertChild(&child.vfsd, name) - parent.inode.impl.(*directory).childList.PushBack(child) + parentDir.insertChildLocked(child, name) return nil }) } @@ -285,12 +315,12 @@ func (fs *filesystem) OpenAt(ctx context.Context, rp *vfs.ResolvingPath, opts vf return start.open(ctx, rp, &opts, false /* afterCreate */) } afterTrailingSymlink: - parent, err := walkParentDirLocked(rp, start) + parentDir, err := walkParentDirLocked(rp, start) if err != nil { return nil, err } // Check for search permission in the parent directory. - if err := parent.inode.checkPermissions(rp.Credentials(), vfs.MayExec, true); err != nil { + if err := parentDir.inode.checkPermissions(rp.Credentials(), vfs.MayExec); err != nil { return nil, err } // Reject attempts to open directories with O_CREAT. @@ -301,11 +331,14 @@ afterTrailingSymlink: if name == "." || name == ".." { return nil, syserror.EISDIR } + if len(name) > linux.NAME_MAX { + return nil, syserror.ENAMETOOLONG + } // Determine whether or not we need to create a file. - child, err := stepLocked(rp, parent) - if err == syserror.ENOENT { + child, ok := parentDir.childMap[name] + if !ok { // Already checked for searchability above; now check for writability. - if err := parent.inode.checkPermissions(rp.Credentials(), vfs.MayWrite, true); err != nil { + if err := parentDir.inode.checkPermissions(rp.Credentials(), vfs.MayWrite); err != nil { return nil, err } if err := rp.Mount().CheckBeginWrite(); err != nil { @@ -314,16 +347,26 @@ afterTrailingSymlink: defer rp.Mount().EndWrite() // Create and open the child. child := fs.newDentry(fs.newRegularFile(rp.Credentials(), opts.Mode)) - parent.vfsd.InsertChild(&child.vfsd, name) - parent.inode.impl.(*directory).childList.PushBack(child) - return child.open(ctx, rp, &opts, true) + parentDir.insertChildLocked(child, name) + fd, err := child.open(ctx, rp, &opts, true) + if err != nil { + return nil, err + } + parentDir.inode.touchCMtime() + return fd, nil } - if err != nil { + // Is the file mounted over? + if err := rp.CheckMount(&child.vfsd); err != nil { return nil, err } // Do we need to resolve a trailing symlink? - if !rp.Done() { - start = parent + if symlink, ok := child.inode.impl.(*symlink); ok && rp.ShouldFollowSymlink() { + // TODO(gvisor.dev/issue/1197): Symlink traversals updates + // access time. + if err := rp.HandleSymlink(symlink.target); err != nil { + return nil, err + } + start = &parentDir.dentry goto afterTrailingSymlink } // Open existing file. @@ -334,9 +377,9 @@ afterTrailingSymlink: } func (d *dentry) open(ctx context.Context, rp *vfs.ResolvingPath, opts *vfs.OpenOptions, afterCreate bool) (*vfs.FileDescription, error) { - ats := vfs.AccessTypesForOpenFlags(opts.Flags) + ats := vfs.AccessTypesForOpenFlags(opts) if !afterCreate { - if err := d.inode.checkPermissions(rp.Credentials(), ats, d.inode.isDir()); err != nil { + if err := d.inode.checkPermissions(rp.Credentials(), ats); err != nil { return nil, err } } @@ -347,10 +390,9 @@ func (d *dentry) open(ctx context.Context, rp *vfs.ResolvingPath, opts *vfs.Open return nil, err } if opts.Flags&linux.O_TRUNC != 0 { - impl.mu.Lock() - impl.data.Truncate(0, impl.memFile) - atomic.StoreUint64(&impl.size, 0) - impl.mu.Unlock() + if _, err := impl.truncate(0); err != nil { + return nil, err + } } return &fd.vfsfd, nil case *directory: @@ -367,9 +409,11 @@ func (d *dentry) open(ctx context.Context, rp *vfs.ResolvingPath, opts *vfs.Open // Can't open symlinks without O_PATH (which is unimplemented). return nil, syserror.ELOOP case *namedPipe: - return newNamedPipeFD(ctx, impl, rp, &d.vfsd, opts.Flags) + return impl.pipe.Open(ctx, rp.Mount(), &d.vfsd, opts.Flags) case *deviceFile: return rp.VirtualFilesystem().OpenDeviceSpecialFile(ctx, rp.Mount(), &d.vfsd, impl.kind, impl.major, impl.minor, opts) + case *socketFile: + return nil, syserror.ENXIO default: panic(fmt.Sprintf("unknown inode type: %T", d.inode.impl)) } @@ -387,6 +431,7 @@ func (fs *filesystem) ReadlinkAt(ctx context.Context, rp *vfs.ResolvingPath) (st if !ok { return "", syserror.EINVAL } + symlink.inode.touchAtime(rp.Mount()) return symlink.target, nil } @@ -400,7 +445,7 @@ func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa // Resolve newParent first to verify that it's on this Mount. fs.mu.Lock() defer fs.mu.Unlock() - newParent, err := walkParentDirLocked(rp, rp.Start().Impl().(*dentry)) + newParentDir, err := walkParentDirLocked(rp, rp.Start().Impl().(*dentry)) if err != nil { return err } @@ -417,25 +462,24 @@ func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa } defer mnt.EndWrite() - oldParent := oldParentVD.Dentry().Impl().(*dentry) - if err := oldParent.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true /* isDir */); err != nil { + oldParentDir := oldParentVD.Dentry().Impl().(*dentry).inode.impl.(*directory) + if err := oldParentDir.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec); err != nil { return err } - // Call vfs.Dentry.Child() instead of stepLocked() or rp.ResolveChild(), - // because if the existing child is a symlink or mount point then we want - // to rename over it rather than follow it. - renamedVFSD := oldParent.vfsd.Child(oldName) - if renamedVFSD == nil { + renamed, ok := oldParentDir.childMap[oldName] + if !ok { return syserror.ENOENT } - renamed := renamedVFSD.Impl().(*dentry) + // Note that we don't need to call rp.CheckMount(), since if renamed is a + // mount point then we want to rename the mount point, not anything in the + // mounted filesystem. if renamed.inode.isDir() { - if renamed == newParent || renamedVFSD.IsAncestorOf(&newParent.vfsd) { + if renamed == &newParentDir.dentry || genericIsAncestorDentry(renamed, &newParentDir.dentry) { return syserror.EINVAL } - if oldParent != newParent { + if oldParentDir != newParentDir { // Writability is needed to change renamed's "..". - if err := renamed.inode.checkPermissions(rp.Credentials(), vfs.MayWrite, true /* isDir */); err != nil { + if err := renamed.inode.checkPermissions(rp.Credentials(), vfs.MayWrite); err != nil { return err } } @@ -445,18 +489,17 @@ func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa } } - if err := newParent.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true /* isDir */); err != nil { + if err := newParentDir.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec); err != nil { return err } - replacedVFSD := newParent.vfsd.Child(newName) - var replaced *dentry - if replacedVFSD != nil { - replaced = replacedVFSD.Impl().(*dentry) - if replaced.inode.isDir() { + replaced, ok := newParentDir.childMap[newName] + if ok { + replacedDir, ok := replaced.inode.impl.(*directory) + if ok { if !renamed.inode.isDir() { return syserror.EISDIR } - if replaced.vfsd.HasChildren() { + if len(replacedDir.childMap) != 0 { return syserror.ENOTEMPTY } } else { @@ -468,11 +511,13 @@ func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa } } } else { - if renamed.inode.isDir() && newParent.inode.nlink == maxLinks { + if renamed.inode.isDir() && newParentDir.inode.nlink == maxLinks { return syserror.EMLINK } } - if newParent.vfsd.IsDisowned() { + // tmpfs never calls VFS.InvalidateDentry(), so newParentDir.dentry can + // only be dead if it was deleted. + if newParentDir.dentry.vfsd.IsDead() { return syserror.ENOENT } @@ -480,31 +525,38 @@ func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa // simplicity, under the assumption that applications are not intentionally // doing noop renames expecting them to succeed where non-noop renames // would fail. - if renamedVFSD == replacedVFSD { + if renamed == replaced { return nil } vfsObj := rp.VirtualFilesystem() - oldParentDir := oldParent.inode.impl.(*directory) - newParentDir := newParent.inode.impl.(*directory) - if err := vfsObj.PrepareRenameDentry(vfs.MountNamespaceFromContext(ctx), renamedVFSD, replacedVFSD); err != nil { + mntns := vfs.MountNamespaceFromContext(ctx) + defer mntns.DecRef() + var replacedVFSD *vfs.Dentry + if replaced != nil { + replacedVFSD = &replaced.vfsd + } + if err := vfsObj.PrepareRenameDentry(mntns, &renamed.vfsd, replacedVFSD); err != nil { return err } if replaced != nil { - newParentDir.childList.Remove(replaced) + newParentDir.removeChildLocked(replaced) if replaced.inode.isDir() { - newParent.inode.decLinksLocked() // from replaced's ".." + newParentDir.inode.decLinksLocked() // from replaced's ".." } replaced.inode.decLinksLocked() } - oldParentDir.childList.Remove(renamed) - newParentDir.childList.PushBack(renamed) - if renamed.inode.isDir() { - oldParent.inode.decLinksLocked() - newParent.inode.incLinksLocked() + oldParentDir.removeChildLocked(renamed) + newParentDir.insertChildLocked(renamed, newName) + vfsObj.CommitRenameReplaceDentry(&renamed.vfsd, replacedVFSD) + oldParentDir.inode.touchCMtime() + if oldParentDir != newParentDir { + if renamed.inode.isDir() { + oldParentDir.inode.decLinksLocked() + newParentDir.inode.incLinksLocked() + } + newParentDir.inode.touchCMtime() } - // TODO(gvisor.dev/issues/1197): Update timestamps and parent directory - // sizes. - vfsObj.CommitRenameReplaceDentry(renamedVFSD, &newParent.vfsd, newName, replacedVFSD) + renamed.inode.touchCtime() return nil } @@ -512,11 +564,11 @@ func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa func (fs *filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error { fs.mu.Lock() defer fs.mu.Unlock() - parent, err := walkParentDirLocked(rp, rp.Start().Impl().(*dentry)) + parentDir, err := walkParentDirLocked(rp, rp.Start().Impl().(*dentry)) if err != nil { return err } - if err := parent.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true /* isDir */); err != nil { + if err := parentDir.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec); err != nil { return err } name := rp.Component() @@ -526,15 +578,15 @@ func (fs *filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error if name == ".." { return syserror.ENOTEMPTY } - childVFSD := parent.vfsd.Child(name) - if childVFSD == nil { + child, ok := parentDir.childMap[name] + if !ok { return syserror.ENOENT } - child := childVFSD.Impl().(*dentry) - if !child.inode.isDir() { + childDir, ok := child.inode.impl.(*directory) + if !ok { return syserror.ENOTDIR } - if childVFSD.HasChildren() { + if len(childDir.childMap) != 0 { return syserror.ENOTEMPTY } mnt := rp.Mount() @@ -543,13 +595,16 @@ func (fs *filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error } defer mnt.EndWrite() vfsObj := rp.VirtualFilesystem() - if err := vfsObj.PrepareDeleteDentry(vfs.MountNamespaceFromContext(ctx), childVFSD); err != nil { + mntns := vfs.MountNamespaceFromContext(ctx) + defer mntns.DecRef() + if err := vfsObj.PrepareDeleteDentry(mntns, &child.vfsd); err != nil { return err } - parent.inode.impl.(*directory).childList.Remove(child) - parent.inode.decLinksLocked() // from child's ".." + parentDir.removeChildLocked(child) + parentDir.inode.decLinksLocked() // from child's ".." child.inode.decLinksLocked() - vfsObj.CommitDeleteDentry(childVFSD) + vfsObj.CommitDeleteDentry(&child.vfsd) + parentDir.inode.touchCMtime() return nil } @@ -561,7 +616,7 @@ func (fs *filesystem) SetStatAt(ctx context.Context, rp *vfs.ResolvingPath, opts if err != nil { return err } - return d.inode.setStat(opts.Stat) + return d.inode.setStat(ctx, rp.Credentials(), &opts.Stat) } // StatAt implements vfs.FilesystemImpl.StatAt. @@ -585,16 +640,15 @@ func (fs *filesystem) StatFSAt(ctx context.Context, rp *vfs.ResolvingPath) (linu if err != nil { return linux.Statfs{}, err } - // TODO(gvisor.dev/issues/1197): Actually implement statfs. + // TODO(gvisor.dev/issue/1197): Actually implement statfs. return linux.Statfs{}, syserror.ENOSYS } // SymlinkAt implements vfs.FilesystemImpl.SymlinkAt. func (fs *filesystem) SymlinkAt(ctx context.Context, rp *vfs.ResolvingPath, target string) error { - return fs.doCreateAt(rp, false /* dir */, func(parent *dentry, name string) error { + return fs.doCreateAt(rp, false /* dir */, func(parentDir *directory, name string) error { child := fs.newDentry(fs.newSymlink(rp.Credentials(), target)) - parent.vfsd.InsertChild(&child.vfsd, name) - parent.inode.impl.(*directory).childList.PushBack(child) + parentDir.insertChildLocked(child, name) return nil }) } @@ -603,22 +657,21 @@ func (fs *filesystem) SymlinkAt(ctx context.Context, rp *vfs.ResolvingPath, targ func (fs *filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error { fs.mu.Lock() defer fs.mu.Unlock() - parent, err := walkParentDirLocked(rp, rp.Start().Impl().(*dentry)) + parentDir, err := walkParentDirLocked(rp, rp.Start().Impl().(*dentry)) if err != nil { return err } - if err := parent.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true /* isDir */); err != nil { + if err := parentDir.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec); err != nil { return err } name := rp.Component() if name == "." || name == ".." { return syserror.EISDIR } - childVFSD := parent.vfsd.Child(name) - if childVFSD == nil { + child, ok := parentDir.childMap[name] + if !ok { return syserror.ENOENT } - child := childVFSD.Impl().(*dentry) if child.inode.isDir() { return syserror.EISDIR } @@ -631,66 +684,81 @@ func (fs *filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error } defer mnt.EndWrite() vfsObj := rp.VirtualFilesystem() - if err := vfsObj.PrepareDeleteDentry(vfs.MountNamespaceFromContext(ctx), childVFSD); err != nil { + mntns := vfs.MountNamespaceFromContext(ctx) + defer mntns.DecRef() + if err := vfsObj.PrepareDeleteDentry(mntns, &child.vfsd); err != nil { return err } - parent.inode.impl.(*directory).childList.Remove(child) + parentDir.removeChildLocked(child) child.inode.decLinksLocked() - vfsObj.CommitDeleteDentry(childVFSD) + vfsObj.CommitDeleteDentry(&child.vfsd) + parentDir.inode.touchCMtime() return nil } +// BoundEndpointAt implements FilesystemImpl.BoundEndpointAt. +func (fs *filesystem) BoundEndpointAt(ctx context.Context, rp *vfs.ResolvingPath) (transport.BoundEndpoint, error) { + fs.mu.RLock() + defer fs.mu.RUnlock() + d, err := resolveLocked(rp) + if err != nil { + return nil, err + } + switch impl := d.inode.impl.(type) { + case *socketFile: + return impl.ep, nil + default: + return nil, syserror.ECONNREFUSED + } +} + // ListxattrAt implements vfs.FilesystemImpl.ListxattrAt. -func (fs *filesystem) ListxattrAt(ctx context.Context, rp *vfs.ResolvingPath) ([]string, error) { +func (fs *filesystem) ListxattrAt(ctx context.Context, rp *vfs.ResolvingPath, size uint64) ([]string, error) { fs.mu.RLock() defer fs.mu.RUnlock() - _, err := resolveLocked(rp) + d, err := resolveLocked(rp) if err != nil { return nil, err } - // TODO(b/127675828): support extended attributes - return nil, syserror.ENOTSUP + return d.inode.listxattr(size) } // GetxattrAt implements vfs.FilesystemImpl.GetxattrAt. -func (fs *filesystem) GetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) (string, error) { +func (fs *filesystem) GetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.GetxattrOptions) (string, error) { fs.mu.RLock() defer fs.mu.RUnlock() - _, err := resolveLocked(rp) + d, err := resolveLocked(rp) if err != nil { return "", err } - // TODO(b/127675828): support extended attributes - return "", syserror.ENOTSUP + return d.inode.getxattr(rp.Credentials(), &opts) } // SetxattrAt implements vfs.FilesystemImpl.SetxattrAt. func (fs *filesystem) SetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetxattrOptions) error { fs.mu.RLock() defer fs.mu.RUnlock() - _, err := resolveLocked(rp) + d, err := resolveLocked(rp) if err != nil { return err } - // TODO(b/127675828): support extended attributes - return syserror.ENOTSUP + return d.inode.setxattr(rp.Credentials(), &opts) } // RemovexattrAt implements vfs.FilesystemImpl.RemovexattrAt. func (fs *filesystem) RemovexattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) error { fs.mu.RLock() defer fs.mu.RUnlock() - _, err := resolveLocked(rp) + d, err := resolveLocked(rp) if err != nil { return err } - // TODO(b/127675828): support extended attributes - return syserror.ENOTSUP + return d.inode.removexattr(rp.Credentials(), name) } // PrependPath implements vfs.FilesystemImpl.PrependPath. func (fs *filesystem) PrependPath(ctx context.Context, vfsroot, vd vfs.VirtualDentry, b *fspath.Builder) error { fs.mu.RLock() defer fs.mu.RUnlock() - return vfs.GenericPrependPath(vfsroot, vd, b) + return genericPrependPath(vfsroot, vd.Mount(), vd.Dentry().Impl().(*dentry), b) } diff --git a/pkg/sentry/fsimpl/tmpfs/named_pipe.go b/pkg/sentry/fsimpl/tmpfs/named_pipe.go index 0c57fdca3..8d77b3fa8 100644 --- a/pkg/sentry/fsimpl/tmpfs/named_pipe.go +++ b/pkg/sentry/fsimpl/tmpfs/named_pipe.go @@ -16,10 +16,8 @@ package tmpfs import ( "gvisor.dev/gvisor/pkg/abi/linux" - "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/kernel/pipe" - "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/usermem" ) @@ -33,27 +31,8 @@ type namedPipe struct { // * fs.mu must be locked. // * rp.Mount().CheckBeginWrite() has been called successfully. func (fs *filesystem) newNamedPipe(creds *auth.Credentials, mode linux.FileMode) *inode { - file := &namedPipe{pipe: pipe.NewVFSPipe(pipe.DefaultPipeSize, usermem.PageSize)} - file.inode.init(file, fs, creds, mode) + file := &namedPipe{pipe: pipe.NewVFSPipe(true /* isNamed */, pipe.DefaultPipeSize, usermem.PageSize)} + file.inode.init(file, fs, creds, linux.S_IFIFO|mode) file.inode.nlink = 1 // Only the parent has a link. return &file.inode } - -// namedPipeFD implements vfs.FileDescriptionImpl. Methods are implemented -// entirely via struct embedding. -type namedPipeFD struct { - fileDescription - - *pipe.VFSPipeFD -} - -func newNamedPipeFD(ctx context.Context, np *namedPipe, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, flags uint32) (*vfs.FileDescription, error) { - var err error - var fd namedPipeFD - fd.VFSPipeFD, err = np.pipe.NewVFSPipeFD(ctx, vfsd, &fd.vfsfd, flags) - if err != nil { - return nil, err - } - fd.vfsfd.Init(&fd, flags, rp.Mount(), vfsd, &vfs.FileDescriptionOptions{}) - return &fd.vfsfd, nil -} diff --git a/pkg/sentry/fsimpl/tmpfs/pipe_test.go b/pkg/sentry/fsimpl/tmpfs/pipe_test.go index 5ee7f2a72..1614f2c39 100644 --- a/pkg/sentry/fsimpl/tmpfs/pipe_test.go +++ b/pkg/sentry/fsimpl/tmpfs/pipe_test.go @@ -151,7 +151,10 @@ func setup(t *testing.T) (context.Context, *auth.Credentials, *vfs.VirtualFilesy creds := auth.CredentialsFromContext(ctx) // Create VFS. - vfsObj := vfs.New() + vfsObj := &vfs.VirtualFilesystem{} + if err := vfsObj.Init(); err != nil { + t.Fatalf("VFS init: %v", err) + } vfsObj.MustRegisterFilesystemType("tmpfs", FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ AllowUserMount: true, }) diff --git a/pkg/sentry/fsimpl/tmpfs/regular_file.go b/pkg/sentry/fsimpl/tmpfs/regular_file.go index dab346a41..57e5e28ec 100644 --- a/pkg/sentry/fsimpl/tmpfs/regular_file.go +++ b/pkg/sentry/fsimpl/tmpfs/regular_file.go @@ -15,6 +15,7 @@ package tmpfs import ( + "fmt" "io" "math" "sync/atomic" @@ -22,6 +23,7 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/safemem" + "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fs/fsutil" "gvisor.dev/gvisor/pkg/sentry/fs/lock" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" @@ -34,71 +36,230 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) +// regularFile is a regular (=S_IFREG) tmpfs file. type regularFile struct { inode inode // memFile is a platform.File used to allocate pages to this regularFile. memFile *pgalloc.MemoryFile - // mu protects the fields below. - mu sync.RWMutex + // mapsMu protects mappings. + mapsMu sync.Mutex `state:"nosave"` + + // mappings tracks mappings of the file into memmap.MappingSpaces. + // + // Protected by mapsMu. + mappings memmap.MappingSet + + // writableMappingPages tracks how many pages of virtual memory are mapped + // as potentially writable from this file. If a page has multiple mappings, + // each mapping is counted separately. + // + // This counter is susceptible to overflow as we can potentially count + // mappings from many VMAs. We count pages rather than bytes to slightly + // mitigate this. + // + // Protected by mapsMu. + writableMappingPages uint64 + + // dataMu protects the fields below. + dataMu sync.RWMutex // data maps offsets into the file to offsets into memFile that store // the file's data. + // + // Protected by dataMu. data fsutil.FileRangeSet - // size is the size of data, but accessed using atomic memory - // operations to avoid locking in inode.stat(). - size uint64 - // seals represents file seals on this inode. + // + // Protected by dataMu. seals uint32 + + // size is the size of data. + // + // Protected by both dataMu and inode.mu; reading it requires holding + // either mutex, while writing requires holding both AND using atomics. + // Readers that do not require consistency (like Stat) may read the + // value atomically without holding either lock. + size uint64 } func (fs *filesystem) newRegularFile(creds *auth.Credentials, mode linux.FileMode) *inode { file := ®ularFile{ memFile: fs.memFile, } - file.inode.init(file, fs, creds, mode) + file.inode.init(file, fs, creds, linux.S_IFREG|mode) file.inode.nlink = 1 // from parent directory return &file.inode } // truncate grows or shrinks the file to the given size. It returns true if the // file size was updated. -func (rf *regularFile) truncate(size uint64) (bool, error) { - rf.mu.Lock() - defer rf.mu.Unlock() +func (rf *regularFile) truncate(newSize uint64) (bool, error) { + rf.inode.mu.Lock() + defer rf.inode.mu.Unlock() + return rf.truncateLocked(newSize) +} - if size == rf.size { +// Preconditions: rf.inode.mu must be held. +func (rf *regularFile) truncateLocked(newSize uint64) (bool, error) { + oldSize := rf.size + if newSize == oldSize { // Nothing to do. return false, nil } - if size > rf.size { - // Growing the file. + // Need to hold inode.mu and dataMu while modifying size. + rf.dataMu.Lock() + if newSize > oldSize { + // Can we grow the file? if rf.seals&linux.F_SEAL_GROW != 0 { - // Seal does not allow growth. + rf.dataMu.Unlock() return false, syserror.EPERM } - rf.size = size + // We only need to update the file size. + atomic.StoreUint64(&rf.size, newSize) + rf.dataMu.Unlock() return true, nil } - // Shrinking the file + // We are shrinking the file. First check if this is allowed. if rf.seals&linux.F_SEAL_SHRINK != 0 { - // Seal does not allow shrink. + rf.dataMu.Unlock() return false, syserror.EPERM } - // TODO(gvisor.dev/issues/1197): Invalidate mappings once we have - // mappings. + // Update the file size. + atomic.StoreUint64(&rf.size, newSize) + rf.dataMu.Unlock() + + // Invalidate past translations of truncated pages. + oldpgend := fs.OffsetPageEnd(int64(oldSize)) + newpgend := fs.OffsetPageEnd(int64(newSize)) + if newpgend < oldpgend { + rf.mapsMu.Lock() + rf.mappings.Invalidate(memmap.MappableRange{newpgend, oldpgend}, memmap.InvalidateOpts{ + // Compare Linux's mm/shmem.c:shmem_setattr() => + // mm/memory.c:unmap_mapping_range(evencows=1). + InvalidatePrivate: true, + }) + rf.mapsMu.Unlock() + } - rf.data.Truncate(size, rf.memFile) - rf.size = size + // We are now guaranteed that there are no translations of truncated pages, + // and can remove them. + rf.dataMu.Lock() + rf.data.Truncate(newSize, rf.memFile) + rf.dataMu.Unlock() return true, nil } +// AddMapping implements memmap.Mappable.AddMapping. +func (rf *regularFile) AddMapping(ctx context.Context, ms memmap.MappingSpace, ar usermem.AddrRange, offset uint64, writable bool) error { + rf.mapsMu.Lock() + defer rf.mapsMu.Unlock() + rf.dataMu.RLock() + defer rf.dataMu.RUnlock() + + // Reject writable mapping if F_SEAL_WRITE is set. + if rf.seals&linux.F_SEAL_WRITE != 0 && writable { + return syserror.EPERM + } + + rf.mappings.AddMapping(ms, ar, offset, writable) + if writable { + pagesBefore := rf.writableMappingPages + + // ar is guaranteed to be page aligned per memmap.Mappable. + rf.writableMappingPages += uint64(ar.Length() / usermem.PageSize) + + if rf.writableMappingPages < pagesBefore { + panic(fmt.Sprintf("Overflow while mapping potentially writable pages pointing to a tmpfs file. Before %v, after %v", pagesBefore, rf.writableMappingPages)) + } + } + + return nil +} + +// RemoveMapping implements memmap.Mappable.RemoveMapping. +func (rf *regularFile) RemoveMapping(ctx context.Context, ms memmap.MappingSpace, ar usermem.AddrRange, offset uint64, writable bool) { + rf.mapsMu.Lock() + defer rf.mapsMu.Unlock() + + rf.mappings.RemoveMapping(ms, ar, offset, writable) + + if writable { + pagesBefore := rf.writableMappingPages + + // ar is guaranteed to be page aligned per memmap.Mappable. + rf.writableMappingPages -= uint64(ar.Length() / usermem.PageSize) + + if rf.writableMappingPages > pagesBefore { + panic(fmt.Sprintf("Underflow while unmapping potentially writable pages pointing to a tmpfs file. Before %v, after %v", pagesBefore, rf.writableMappingPages)) + } + } +} + +// CopyMapping implements memmap.Mappable.CopyMapping. +func (rf *regularFile) CopyMapping(ctx context.Context, ms memmap.MappingSpace, srcAR, dstAR usermem.AddrRange, offset uint64, writable bool) error { + return rf.AddMapping(ctx, ms, dstAR, offset, writable) +} + +// Translate implements memmap.Mappable.Translate. +func (rf *regularFile) Translate(ctx context.Context, required, optional memmap.MappableRange, at usermem.AccessType) ([]memmap.Translation, error) { + rf.dataMu.Lock() + defer rf.dataMu.Unlock() + + // Constrain translations to f.attr.Size (rounded up) to prevent + // translation to pages that may be concurrently truncated. + pgend := fs.OffsetPageEnd(int64(rf.size)) + var beyondEOF bool + if required.End > pgend { + if required.Start >= pgend { + return nil, &memmap.BusError{io.EOF} + } + beyondEOF = true + required.End = pgend + } + if optional.End > pgend { + optional.End = pgend + } + + cerr := rf.data.Fill(ctx, required, optional, rf.memFile, usage.Tmpfs, func(_ context.Context, dsts safemem.BlockSeq, _ uint64) (uint64, error) { + // Newly-allocated pages are zeroed, so we don't need to do anything. + return dsts.NumBytes(), nil + }) + + var ts []memmap.Translation + var translatedEnd uint64 + for seg := rf.data.FindSegment(required.Start); seg.Ok() && seg.Start() < required.End; seg, _ = seg.NextNonEmpty() { + segMR := seg.Range().Intersect(optional) + ts = append(ts, memmap.Translation{ + Source: segMR, + File: rf.memFile, + Offset: seg.FileRangeOf(segMR).Start, + Perms: usermem.AnyAccess, + }) + translatedEnd = segMR.End + } + + // Don't return the error returned by f.data.Fill if it occurred outside of + // required. + if translatedEnd < required.End && cerr != nil { + return ts, &memmap.BusError{cerr} + } + if beyondEOF { + return ts, &memmap.BusError{io.EOF} + } + return ts, nil +} + +// InvalidateUnsavable implements memmap.Mappable.InvalidateUnsavable. +func (*regularFile) InvalidateUnsavable(context.Context) error { + return nil +} + type regularFileFD struct { fileDescription @@ -125,7 +286,8 @@ func (fd *regularFileFD) PRead(ctx context.Context, dst usermem.IOSequence, offs rw := getRegularFileReadWriter(f, offset) n, err := dst.CopyOutFrom(ctx, rw) putRegularFileReadWriter(rw) - return int64(n), err + fd.inode().touchAtime(fd.vfsfd.Mount()) + return n, err } // Read implements vfs.FileDescriptionImpl.Read. @@ -147,13 +309,23 @@ func (fd *regularFileFD) PWrite(ctx context.Context, src usermem.IOSequence, off return 0, nil } f := fd.inode().impl.(*regularFile) - end := offset + srclen - if end < offset { + if end := offset + srclen; end < offset { // Overflow. return 0, syserror.EFBIG } + + var err error + srclen, err = vfs.CheckLimit(ctx, offset, srclen) + if err != nil { + return 0, err + } + src = src.TakeFirst64(srclen) + + f.inode.mu.Lock() rw := getRegularFileReadWriter(f, offset) n, err := src.CopyInTo(ctx, rw) + fd.inode().touchCMtimeLocked() + f.inode.mu.Unlock() putRegularFileReadWriter(rw) return n, err } @@ -215,6 +387,12 @@ func (fd *regularFileFD) UnlockPOSIX(ctx context.Context, uid lock.UniqueID, rng return nil } +// ConfigureMMap implements vfs.FileDescriptionImpl.ConfigureMMap. +func (fd *regularFileFD) ConfigureMMap(ctx context.Context, opts *memmap.MMapOpts) error { + file := fd.inode().impl.(*regularFile) + return vfs.GenericConfigureMMap(&fd.vfsfd, file, opts) +} + // regularFileReadWriter implements safemem.Reader and Safemem.Writer. type regularFileReadWriter struct { file *regularFile @@ -244,14 +422,15 @@ func putRegularFileReadWriter(rw *regularFileReadWriter) { // ReadToBlocks implements safemem.Reader.ReadToBlocks. func (rw *regularFileReadWriter) ReadToBlocks(dsts safemem.BlockSeq) (uint64, error) { - rw.file.mu.RLock() + rw.file.dataMu.RLock() + defer rw.file.dataMu.RUnlock() + size := rw.file.size // Compute the range to read (limited by file size and overflow-checked). - if rw.off >= rw.file.size { - rw.file.mu.RUnlock() + if rw.off >= size { return 0, io.EOF } - end := rw.file.size + end := size if rend := rw.off + dsts.NumBytes(); rend > rw.off && rend < end { end = rend } @@ -265,7 +444,6 @@ func (rw *regularFileReadWriter) ReadToBlocks(dsts safemem.BlockSeq) (uint64, er // Get internal mappings. ims, err := rw.file.memFile.MapInternal(seg.FileRangeOf(seg.Range().Intersect(mr)), usermem.Read) if err != nil { - rw.file.mu.RUnlock() return done, err } @@ -275,7 +453,6 @@ func (rw *regularFileReadWriter) ReadToBlocks(dsts safemem.BlockSeq) (uint64, er rw.off += uint64(n) dsts = dsts.DropFirst64(n) if err != nil { - rw.file.mu.RUnlock() return done, err } @@ -291,7 +468,6 @@ func (rw *regularFileReadWriter) ReadToBlocks(dsts safemem.BlockSeq) (uint64, er rw.off += uint64(n) dsts = dsts.DropFirst64(n) if err != nil { - rw.file.mu.RUnlock() return done, err } @@ -299,13 +475,16 @@ func (rw *regularFileReadWriter) ReadToBlocks(dsts safemem.BlockSeq) (uint64, er seg, gap = gap.NextSegment(), fsutil.FileRangeGapIterator{} } } - rw.file.mu.RUnlock() return done, nil } // WriteFromBlocks implements safemem.Writer.WriteFromBlocks. +// +// Preconditions: inode.mu must be held. func (rw *regularFileReadWriter) WriteFromBlocks(srcs safemem.BlockSeq) (uint64, error) { - rw.file.mu.Lock() + // Hold dataMu so we can modify size. + rw.file.dataMu.Lock() + defer rw.file.dataMu.Unlock() // Compute the range to write (overflow-checked). end := rw.off + srcs.NumBytes() @@ -316,7 +495,6 @@ func (rw *regularFileReadWriter) WriteFromBlocks(srcs safemem.BlockSeq) (uint64, // Check if seals prevent either file growth or all writes. switch { case rw.file.seals&linux.F_SEAL_WRITE != 0: // Write sealed - rw.file.mu.Unlock() return 0, syserror.EPERM case end > rw.file.size && rw.file.seals&linux.F_SEAL_GROW != 0: // Grow sealed // When growth is sealed, Linux effectively allows writes which would @@ -338,7 +516,6 @@ func (rw *regularFileReadWriter) WriteFromBlocks(srcs safemem.BlockSeq) (uint64, } if end <= rw.off { // Truncation would result in no data being written. - rw.file.mu.Unlock() return 0, syserror.EPERM } } @@ -395,9 +572,8 @@ exitLoop: // If the write ends beyond the file's previous size, it causes the // file to grow. if rw.off > rw.file.size { - atomic.StoreUint64(&rw.file.size, rw.off) + rw.file.size = rw.off } - rw.file.mu.Unlock() return done, retErr } diff --git a/pkg/sentry/fsimpl/tmpfs/regular_file_test.go b/pkg/sentry/fsimpl/tmpfs/regular_file_test.go index e9f71e334..0399725cf 100644 --- a/pkg/sentry/fsimpl/tmpfs/regular_file_test.go +++ b/pkg/sentry/fsimpl/tmpfs/regular_file_test.go @@ -40,7 +40,11 @@ var nextFileID int64 func newTmpfsRoot(ctx context.Context) (*vfs.VirtualFilesystem, vfs.VirtualDentry, func(), error) { creds := auth.CredentialsFromContext(ctx) - vfsObj := vfs.New() + vfsObj := &vfs.VirtualFilesystem{} + if err := vfsObj.Init(); err != nil { + return nil, vfs.VirtualDentry{}, nil, fmt.Errorf("VFS init: %v", err) + } + vfsObj.MustRegisterFilesystemType("tmpfs", FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ AllowUserMount: true, }) diff --git a/pkg/tcpip/packet_buffer_state.go b/pkg/sentry/fsimpl/tmpfs/socket_file.go index ad3cc24fa..25c2321af 100644 --- a/pkg/tcpip/packet_buffer_state.go +++ b/pkg/sentry/fsimpl/tmpfs/socket_file.go @@ -12,16 +12,23 @@ // See the License for the specific language governing permissions and // limitations under the License. -package tcpip +package tmpfs -import "gvisor.dev/gvisor/pkg/tcpip/buffer" +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" +) -// beforeSave is invoked by stateify. -func (pk *PacketBuffer) beforeSave() { - // Non-Data fields may be slices of the Data field. This causes - // problems for SR, so during save we make each header independent. - pk.Header = pk.Header.DeepCopy() - pk.LinkHeader = append(buffer.View(nil), pk.LinkHeader...) - pk.NetworkHeader = append(buffer.View(nil), pk.NetworkHeader...) - pk.TransportHeader = append(buffer.View(nil), pk.TransportHeader...) +// socketFile is a socket (=S_IFSOCK) tmpfs file. +type socketFile struct { + inode inode + ep transport.BoundEndpoint +} + +func (fs *filesystem) newSocketFile(creds *auth.Credentials, mode linux.FileMode, ep transport.BoundEndpoint) *inode { + file := &socketFile{ep: ep} + file.inode.init(file, fs, creds, mode) + file.inode.nlink = 1 // from parent directory + return &file.inode } diff --git a/pkg/sentry/fsimpl/tmpfs/stat_test.go b/pkg/sentry/fsimpl/tmpfs/stat_test.go index ebe035dee..60c2c980e 100644 --- a/pkg/sentry/fsimpl/tmpfs/stat_test.go +++ b/pkg/sentry/fsimpl/tmpfs/stat_test.go @@ -29,7 +29,7 @@ func TestStatAfterCreate(t *testing.T) { mode := linux.FileMode(0644) // Run with different file types. - // TODO(gvisor.dev/issues/1197): Also test symlinks and sockets. + // TODO(gvisor.dev/issue/1197): Also test symlinks and sockets. for _, typ := range []string{"file", "dir", "pipe"} { t.Run(fmt.Sprintf("type=%q", typ), func(t *testing.T) { var ( @@ -71,9 +71,15 @@ func TestStatAfterCreate(t *testing.T) { t.Errorf("got btime %d, want 0", got.Btime.ToNsec()) } - // Size should be 0. - if got.Size != 0 { - t.Errorf("got size %d, want 0", got.Size) + // Size should be 0 (except for directories, which make up a size + // of 20 per entry, including the "." and ".." entries present in + // otherwise-empty directories). + wantSize := uint64(0) + if typ == "dir" { + wantSize = 40 + } + if got.Size != wantSize { + t.Errorf("got size %d, want %d", got.Size, wantSize) } // Nlink should be 1 for files, 2 for dirs. @@ -140,7 +146,7 @@ func TestSetStatAtime(t *testing.T) { Mask: 0, Atime: linux.NsecToStatxTimestamp(100), }}); err != nil { - t.Errorf("SetStat atime without mask failed: %v") + t.Errorf("SetStat atime without mask failed: %v", err) } // Atime should be unchanged. if gotStat, err := fd.Stat(ctx, allStatOptions); err != nil { @@ -155,7 +161,7 @@ func TestSetStatAtime(t *testing.T) { Atime: linux.NsecToStatxTimestamp(100), } if err := fd.SetStat(ctx, vfs.SetStatOptions{Stat: setStat}); err != nil { - t.Errorf("SetStat atime with mask failed: %v") + t.Errorf("SetStat atime with mask failed: %v", err) } if gotStat, err := fd.Stat(ctx, allStatOptions); err != nil { t.Errorf("Stat got error: %v", err) @@ -169,7 +175,7 @@ func TestSetStat(t *testing.T) { mode := linux.FileMode(0644) // Run with different file types. - // TODO(gvisor.dev/issues/1197): Also test symlinks and sockets. + // TODO(gvisor.dev/issue/1197): Also test symlinks and sockets. for _, typ := range []string{"file", "dir", "pipe"} { t.Run(fmt.Sprintf("type=%q", typ), func(t *testing.T) { var ( @@ -205,7 +211,7 @@ func TestSetStat(t *testing.T) { Mask: 0, Atime: linux.NsecToStatxTimestamp(100), }}); err != nil { - t.Errorf("SetStat atime without mask failed: %v") + t.Errorf("SetStat atime without mask failed: %v", err) } // Atime should be unchanged. if gotStat, err := fd.Stat(ctx, allStatOptions); err != nil { @@ -220,7 +226,7 @@ func TestSetStat(t *testing.T) { Atime: linux.NsecToStatxTimestamp(100), } if err := fd.SetStat(ctx, vfs.SetStatOptions{Stat: setStat}); err != nil { - t.Errorf("SetStat atime with mask failed: %v") + t.Errorf("SetStat atime with mask failed: %v", err) } if gotStat, err := fd.Stat(ctx, allStatOptions); err != nil { t.Errorf("Stat got error: %v", err) diff --git a/pkg/sentry/fsimpl/tmpfs/symlink.go b/pkg/sentry/fsimpl/tmpfs/symlink.go index 5246aca84..47e075ed4 100644 --- a/pkg/sentry/fsimpl/tmpfs/symlink.go +++ b/pkg/sentry/fsimpl/tmpfs/symlink.go @@ -15,6 +15,7 @@ package tmpfs import ( + "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" ) @@ -27,7 +28,7 @@ func (fs *filesystem) newSymlink(creds *auth.Credentials, target string) *inode link := &symlink{ target: target, } - link.inode.init(link, fs, creds, 0777) + link.inode.init(link, fs, creds, linux.S_IFLNK|0777) link.inode.nlink = 1 // from parent directory return &link.inode } diff --git a/pkg/sentry/fsimpl/tmpfs/tmpfs.go b/pkg/sentry/fsimpl/tmpfs/tmpfs.go index 2108d0f4d..efc931468 100644 --- a/pkg/sentry/fsimpl/tmpfs/tmpfs.go +++ b/pkg/sentry/fsimpl/tmpfs/tmpfs.go @@ -12,20 +12,25 @@ // See the License for the specific language governing permissions and // limitations under the License. -// Package tmpfs provides a filesystem implementation that behaves like tmpfs: -// the Dentry tree is the sole source of truth for the state of the filesystem. +// Package tmpfs provides an in-memory filesystem whose contents are +// application-mutable, consistent with Linux's tmpfs. // // Lock order: // // filesystem.mu -// regularFileFD.offMu -// regularFile.mu // inode.mu +// regularFileFD.offMu +// *** "memmap.Mappable locks" below this point +// regularFile.mapsMu +// *** "memmap.Mappable locks taken by Translate" below this point +// regularFile.dataMu +// directory.iterMu package tmpfs import ( "fmt" "math" + "strings" "sync/atomic" "gvisor.dev/gvisor/pkg/abi/linux" @@ -36,10 +41,15 @@ import ( "gvisor.dev/gvisor/pkg/sentry/pgalloc" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/sentry/vfs/lock" + "gvisor.dev/gvisor/pkg/sentry/vfs/memxattr" "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" ) +// Name is the default filesystem name. +const Name = "tmpfs" + // FilesystemType implements vfs.FilesystemType. type FilesystemType struct{} @@ -59,6 +69,27 @@ type filesystem struct { nextInoMinusOne uint64 // accessed using atomic memory operations } +// Name implements vfs.FilesystemType.Name. +func (FilesystemType) Name() string { + return Name +} + +// FilesystemOpts is used to pass configuration data to tmpfs. +type FilesystemOpts struct { + // RootFileType is the FileType of the filesystem root. Valid values + // are: S_IFDIR, S_IFREG, and S_IFLNK. Defaults to S_IFDIR. + RootFileType uint16 + + // RootSymlinkTarget is the target of the root symlink. Only valid if + // RootFileType == S_IFLNK. + RootSymlinkTarget string + + // FilesystemType allows setting a different FilesystemType for this + // tmpfs filesystem. This allows tmpfs to "impersonate" other + // filesystems, like ramdiskfs and cgroupfs. + FilesystemType vfs.FilesystemType +} + // GetFilesystem implements vfs.FilesystemType.GetFilesystem. func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { memFileProvider := pgalloc.MemoryFileProviderFromContext(ctx) @@ -70,8 +101,32 @@ func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virt memFile: memFileProvider.MemoryFile(), clock: clock, } - fs.vfsfs.Init(vfsObj, &fs) - root := fs.newDentry(fs.newDirectory(creds, 01777)) + + rootFileType := uint16(linux.S_IFDIR) + newFSType := vfs.FilesystemType(&fstype) + tmpfsOpts, ok := opts.InternalData.(FilesystemOpts) + if ok { + if tmpfsOpts.RootFileType != 0 { + rootFileType = tmpfsOpts.RootFileType + } + if tmpfsOpts.FilesystemType != nil { + newFSType = tmpfsOpts.FilesystemType + } + } + + fs.vfsfs.Init(vfsObj, newFSType, &fs) + + var root *dentry + switch rootFileType { + case linux.S_IFREG: + root = fs.newDentry(fs.newRegularFile(creds, 0777)) + case linux.S_IFLNK: + root = fs.newDentry(fs.newSymlink(creds, tmpfsOpts.RootSymlinkTarget)) + case linux.S_IFDIR: + root = &fs.newDirectory(creds, 01777).dentry + default: + return nil, nil, fmt.Errorf("invalid tmpfs root file type: %#o", rootFileType) + } return &fs.vfsfs, &root.vfsd, nil } @@ -83,20 +138,29 @@ func (fs *filesystem) Release() { type dentry struct { vfsd vfs.Dentry + // parent is this dentry's parent directory. Each referenced dentry holds a + // reference on parent.dentry. If this dentry is a filesystem root, parent + // is nil. parent is protected by filesystem.mu. + parent *dentry + + // name is the name of this dentry in its parent. If this dentry is a + // filesystem root, name is the empty string. name is protected by + // filesystem.mu. + name string + + // dentryEntry (ugh) links dentries into their parent directory.childList. + dentryEntry + // inode is the inode represented by this dentry. Multiple Dentries may // share a single non-directory inode (with hard links). inode is // immutable. - inode *inode - + // // tmpfs doesn't count references on dentries; because the dentry tree is // the sole source of truth, it is by definition always consistent with the // state of the filesystem. However, it does count references on inodes, // because inode resources are released when all references are dropped. - // (tmpfs doesn't really have resources to release, but we implement - // reference counting because tmpfs regular files will.) - - // dentryEntry (ugh) links dentries into their parent directory.childList. - dentryEntry + // dentry therefore forwards reference counting directly to inode. + inode *inode } func (fs *filesystem) newDentry(inode *inode) *dentry { @@ -137,10 +201,15 @@ type inode struct { // filesystem.RmdirAt() drops the reference. refs int64 + // xattrs implements extended attributes. + // + // TODO(b/148380782): Support xattrs other than user.* + xattrs memxattr.SimpleExtendedAttributes + // Inode metadata. Writing multiple fields atomically requires holding // mu, othewise atomic operations can be used. mu sync.Mutex - mode uint32 // excluding file type bits, which are based on impl + mode uint32 // file type and mode nlink uint32 // protected by filesystem.mu instead of inode.mu uid uint32 // auth.KUID, but stored as raw uint32 for sync/atomic gid uint32 // auth.KGID, but ... @@ -151,10 +220,6 @@ type inode struct { ctime int64 // nanoseconds mtime int64 // nanoseconds - // Only meaningful for device special files. - rdevMajor uint32 - rdevMinor uint32 - // Advisory file locks, which lock at the inode level. locks lock.FileLocks @@ -164,6 +229,9 @@ type inode struct { const maxLinks = math.MaxUint32 func (i *inode) init(impl interface{}, fs *filesystem, creds *auth.Credentials, mode linux.FileMode) { + if mode.FileType() == 0 { + panic("file type is required in FileMode") + } i.clock = fs.clock i.refs = 1 i.mode = uint32(mode) @@ -171,7 +239,7 @@ func (i *inode) init(impl interface{}, fs *filesystem, creds *auth.Credentials, i.gid = uint32(creds.EffectiveKGID) i.ino = atomic.AddUint64(&fs.nextInoMinusOne, 1) // Tmpfs creation sets atime, ctime, and mtime to current time. - now := i.clock.Now().Nanoseconds() + now := fs.clock.Now().Nanoseconds() i.atime = now i.ctime = now i.mtime = now @@ -188,7 +256,7 @@ func (i *inode) incLinksLocked() { panic("tmpfs.inode.incLinksLocked() called with no existing links") } if i.nlink == maxLinks { - panic("memfs.inode.incLinksLocked() called with maximum link count") + panic("tmpfs.inode.incLinksLocked() called with maximum link count") } atomic.AddUint32(&i.nlink, 1) } @@ -223,20 +291,20 @@ func (i *inode) tryIncRef() bool { func (i *inode) decRef() { if refs := atomic.AddInt64(&i.refs, -1); refs == 0 { - // This is unnecessary; it's mostly to simulate what tmpfs would do. if regFile, ok := i.impl.(*regularFile); ok { - regFile.mu.Lock() + // Release memory used by regFile to store data. Since regFile is + // no longer usable, we don't need to grab any locks or update any + // metadata. regFile.data.DropAll(regFile.memFile) - atomic.StoreUint64(®File.size, 0) - regFile.mu.Unlock() } } else if refs < 0 { panic("tmpfs.inode.decRef() called without holding a reference") } } -func (i *inode) checkPermissions(creds *auth.Credentials, ats vfs.AccessTypes, isDir bool) error { - return vfs.GenericCheckPermissions(creds, ats, isDir, uint16(atomic.LoadUint32(&i.mode)), auth.KUID(atomic.LoadUint32(&i.uid)), auth.KGID(atomic.LoadUint32(&i.gid))) +func (i *inode) checkPermissions(creds *auth.Credentials, ats vfs.AccessTypes) error { + mode := linux.FileMode(atomic.LoadUint32(&i.mode)) + return vfs.GenericCheckPermissions(creds, ats, mode, auth.KUID(atomic.LoadUint32(&i.uid)), auth.KGID(atomic.LoadUint32(&i.gid))) } // Go won't inline this function, and returning linux.Statx (which is quite @@ -247,44 +315,37 @@ func (i *inode) checkPermissions(creds *auth.Credentials, ats vfs.AccessTypes, i // a concurrent modification), so we do not require holding inode.mu. func (i *inode) statTo(stat *linux.Statx) { stat.Mask = linux.STATX_TYPE | linux.STATX_MODE | linux.STATX_NLINK | - linux.STATX_UID | linux.STATX_GID | linux.STATX_INO | linux.STATX_ATIME | - linux.STATX_BTIME | linux.STATX_CTIME | linux.STATX_MTIME - stat.Blksize = 1 // usermem.PageSize in tmpfs + linux.STATX_UID | linux.STATX_GID | linux.STATX_INO | linux.STATX_SIZE | + linux.STATX_BLOCKS | linux.STATX_ATIME | linux.STATX_CTIME | + linux.STATX_MTIME + stat.Blksize = usermem.PageSize stat.Nlink = atomic.LoadUint32(&i.nlink) stat.UID = atomic.LoadUint32(&i.uid) stat.GID = atomic.LoadUint32(&i.gid) stat.Mode = uint16(atomic.LoadUint32(&i.mode)) stat.Ino = i.ino - // Linux's tmpfs has no concept of btime, so zero-value is returned. stat.Atime = linux.NsecToStatxTimestamp(i.atime) stat.Ctime = linux.NsecToStatxTimestamp(i.ctime) stat.Mtime = linux.NsecToStatxTimestamp(i.mtime) - // TODO(gvisor.dev/issues/1197): Device number. + // TODO(gvisor.dev/issue/1197): Device number. switch impl := i.impl.(type) { case *regularFile: - stat.Mode |= linux.S_IFREG stat.Mask |= linux.STATX_SIZE | linux.STATX_BLOCKS stat.Size = uint64(atomic.LoadUint64(&impl.size)) - // In tmpfs, this will be FileRangeSet.Span() / 512 (but also cached in - // a uint64 accessed using atomic memory operations to avoid taking - // locks). + // TODO(jamieliu): This should be impl.data.Span() / 512, but this is + // too expensive to compute here. Cache it in regularFile. stat.Blocks = allocatedBlocksForSize(stat.Size) case *directory: - stat.Mode |= linux.S_IFDIR + // "20" is mm/shmem.c:BOGO_DIRENT_SIZE. + stat.Size = 20 * (2 + uint64(atomic.LoadInt64(&impl.numChildren))) + // stat.Blocks is 0. case *symlink: - stat.Mode |= linux.S_IFLNK - stat.Mask |= linux.STATX_SIZE | linux.STATX_BLOCKS stat.Size = uint64(len(impl.target)) - stat.Blocks = allocatedBlocksForSize(stat.Size) - case *namedPipe: - stat.Mode |= linux.S_IFIFO + // stat.Blocks is 0. + case *namedPipe, *socketFile: + // stat.Size and stat.Blocks are 0. case *deviceFile: - switch impl.kind { - case vfs.BlockDevice: - stat.Mode |= linux.S_IFBLK - case vfs.CharDevice: - stat.Mode |= linux.S_IFCHR - } + // stat.Size and stat.Blocks are 0. stat.RdevMajor = impl.major stat.RdevMinor = impl.minor default: @@ -292,18 +353,27 @@ func (i *inode) statTo(stat *linux.Statx) { } } -func (i *inode) setStat(stat linux.Statx) error { +func (i *inode) setStat(ctx context.Context, creds *auth.Credentials, stat *linux.Statx) error { if stat.Mask == 0 { return nil } + if stat.Mask&^(linux.STATX_MODE|linux.STATX_UID|linux.STATX_GID|linux.STATX_ATIME|linux.STATX_MTIME|linux.STATX_CTIME|linux.STATX_SIZE) != 0 { + return syserror.EPERM + } + mode := linux.FileMode(atomic.LoadUint32(&i.mode)) + if err := vfs.CheckSetStat(ctx, creds, stat, mode, auth.KUID(atomic.LoadUint32(&i.uid)), auth.KGID(atomic.LoadUint32(&i.gid))); err != nil { + return err + } i.mu.Lock() + defer i.mu.Unlock() var ( needsMtimeBump bool needsCtimeBump bool ) mask := stat.Mask if mask&linux.STATX_MODE != 0 { - atomic.StoreUint32(&i.mode, uint32(stat.Mode)) + ft := atomic.LoadUint32(&i.mode) & linux.S_IFMT + atomic.StoreUint32(&i.mode, ft|uint32(stat.Mode&^linux.S_IFMT)) needsCtimeBump = true } if mask&linux.STATX_UID != 0 { @@ -317,7 +387,7 @@ func (i *inode) setStat(stat linux.Statx) error { if mask&linux.STATX_SIZE != 0 { switch impl := i.impl.(type) { case *regularFile: - updated, err := impl.truncate(stat.Size) + updated, err := impl.truncateLocked(stat.Size) if err != nil { return err } @@ -331,29 +401,41 @@ func (i *inode) setStat(stat linux.Statx) error { return syserror.EINVAL } } + now := i.clock.Now().Nanoseconds() if mask&linux.STATX_ATIME != 0 { - atomic.StoreInt64(&i.atime, stat.Atime.ToNsecCapped()) + if stat.Atime.Nsec == linux.UTIME_NOW { + atomic.StoreInt64(&i.atime, now) + } else { + atomic.StoreInt64(&i.atime, stat.Atime.ToNsecCapped()) + } needsCtimeBump = true } if mask&linux.STATX_MTIME != 0 { - atomic.StoreInt64(&i.mtime, stat.Mtime.ToNsecCapped()) + if stat.Mtime.Nsec == linux.UTIME_NOW { + atomic.StoreInt64(&i.mtime, now) + } else { + atomic.StoreInt64(&i.mtime, stat.Mtime.ToNsecCapped()) + } needsCtimeBump = true // Ignore the mtime bump, since we just set it ourselves. needsMtimeBump = false } if mask&linux.STATX_CTIME != 0 { - atomic.StoreInt64(&i.ctime, stat.Ctime.ToNsecCapped()) + if stat.Ctime.Nsec == linux.UTIME_NOW { + atomic.StoreInt64(&i.ctime, now) + } else { + atomic.StoreInt64(&i.ctime, stat.Ctime.ToNsecCapped()) + } // Ignore the ctime bump, since we just set it ourselves. needsCtimeBump = false } - now := i.clock.Now().Nanoseconds() if needsMtimeBump { atomic.StoreInt64(&i.mtime, now) } if needsCtimeBump { atomic.StoreInt64(&i.ctime, now) } - i.mu.Unlock() + return nil } @@ -412,6 +494,8 @@ func (i *inode) direntType() uint8 { return linux.DT_DIR case *symlink: return linux.DT_LNK + case *socketFile: + return linux.DT_SOCK case *deviceFile: switch impl.kind { case vfs.BlockDevice: @@ -426,6 +510,96 @@ func (i *inode) direntType() uint8 { } } +func (i *inode) isDir() bool { + return linux.FileMode(i.mode).FileType() == linux.S_IFDIR +} + +func (i *inode) touchAtime(mnt *vfs.Mount) { + if err := mnt.CheckBeginWrite(); err != nil { + return + } + now := i.clock.Now().Nanoseconds() + i.mu.Lock() + atomic.StoreInt64(&i.atime, now) + i.mu.Unlock() + mnt.EndWrite() +} + +// Preconditions: The caller has called vfs.Mount.CheckBeginWrite(). +func (i *inode) touchCtime() { + now := i.clock.Now().Nanoseconds() + i.mu.Lock() + atomic.StoreInt64(&i.ctime, now) + i.mu.Unlock() +} + +// Preconditions: The caller has called vfs.Mount.CheckBeginWrite(). +func (i *inode) touchCMtime() { + now := i.clock.Now().Nanoseconds() + i.mu.Lock() + atomic.StoreInt64(&i.mtime, now) + atomic.StoreInt64(&i.ctime, now) + i.mu.Unlock() +} + +// Preconditions: The caller has called vfs.Mount.CheckBeginWrite() and holds +// inode.mu. +func (i *inode) touchCMtimeLocked() { + now := i.clock.Now().Nanoseconds() + atomic.StoreInt64(&i.mtime, now) + atomic.StoreInt64(&i.ctime, now) +} + +func (i *inode) listxattr(size uint64) ([]string, error) { + return i.xattrs.Listxattr(size) +} + +func (i *inode) getxattr(creds *auth.Credentials, opts *vfs.GetxattrOptions) (string, error) { + if err := i.checkPermissions(creds, vfs.MayRead); err != nil { + return "", err + } + if !strings.HasPrefix(opts.Name, linux.XATTR_USER_PREFIX) { + return "", syserror.EOPNOTSUPP + } + if !i.userXattrSupported() { + return "", syserror.ENODATA + } + return i.xattrs.Getxattr(opts) +} + +func (i *inode) setxattr(creds *auth.Credentials, opts *vfs.SetxattrOptions) error { + if err := i.checkPermissions(creds, vfs.MayWrite); err != nil { + return err + } + if !strings.HasPrefix(opts.Name, linux.XATTR_USER_PREFIX) { + return syserror.EOPNOTSUPP + } + if !i.userXattrSupported() { + return syserror.EPERM + } + return i.xattrs.Setxattr(opts) +} + +func (i *inode) removexattr(creds *auth.Credentials, name string) error { + if err := i.checkPermissions(creds, vfs.MayWrite); err != nil { + return err + } + if !strings.HasPrefix(name, linux.XATTR_USER_PREFIX) { + return syserror.EOPNOTSUPP + } + if !i.userXattrSupported() { + return syserror.EPERM + } + return i.xattrs.Removexattr(name) +} + +// Extended attributes in the user.* namespace are only supported for regular +// files and directories. +func (i *inode) userXattrSupported() bool { + filetype := linux.S_IFMT & atomic.LoadUint32(&i.mode) + return filetype == linux.S_IFREG || filetype == linux.S_IFDIR +} + // fileDescription is embedded by tmpfs implementations of // vfs.FileDescriptionImpl. type fileDescription struct { @@ -450,5 +624,26 @@ func (fd *fileDescription) Stat(ctx context.Context, opts vfs.StatOptions) (linu // SetStat implements vfs.FileDescriptionImpl.SetStat. func (fd *fileDescription) SetStat(ctx context.Context, opts vfs.SetStatOptions) error { - return fd.inode().setStat(opts.Stat) + creds := auth.CredentialsFromContext(ctx) + return fd.inode().setStat(ctx, creds, &opts.Stat) +} + +// Listxattr implements vfs.FileDescriptionImpl.Listxattr. +func (fd *fileDescription) Listxattr(ctx context.Context, size uint64) ([]string, error) { + return fd.inode().listxattr(size) +} + +// Getxattr implements vfs.FileDescriptionImpl.Getxattr. +func (fd *fileDescription) Getxattr(ctx context.Context, opts vfs.GetxattrOptions) (string, error) { + return fd.inode().getxattr(auth.CredentialsFromContext(ctx), &opts) +} + +// Setxattr implements vfs.FileDescriptionImpl.Setxattr. +func (fd *fileDescription) Setxattr(ctx context.Context, opts vfs.SetxattrOptions) error { + return fd.inode().setxattr(auth.CredentialsFromContext(ctx), &opts) +} + +// Removexattr implements vfs.FileDescriptionImpl.Removexattr. +func (fd *fileDescription) Removexattr(ctx context.Context, name string) error { + return fd.inode().removexattr(auth.CredentialsFromContext(ctx), name) } diff --git a/pkg/sentry/inet/BUILD b/pkg/sentry/inet/BUILD index 334432abf..07bf39fed 100644 --- a/pkg/sentry/inet/BUILD +++ b/pkg/sentry/inet/BUILD @@ -10,6 +10,7 @@ go_library( srcs = [ "context.go", "inet.go", + "namespace.go", "test_stack.go", ], deps = [ diff --git a/pkg/sentry/inet/namespace.go b/pkg/sentry/inet/namespace.go new file mode 100644 index 000000000..029af3025 --- /dev/null +++ b/pkg/sentry/inet/namespace.go @@ -0,0 +1,102 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package inet + +// Namespace represents a network namespace. See network_namespaces(7). +// +// +stateify savable +type Namespace struct { + // stack is the network stack implementation of this network namespace. + stack Stack `state:"nosave"` + + // creator allows kernel to create new network stack for network namespaces. + // If nil, no networking will function if network is namespaced. + // + // At afterLoad(), creator will be used to create network stack. Stateify + // needs to wait for this field to be loaded before calling afterLoad(). + creator NetworkStackCreator `state:"wait"` + + // isRoot indicates whether this is the root network namespace. + isRoot bool +} + +// NewRootNamespace creates the root network namespace, with creator +// allowing new network namespaces to be created. If creator is nil, no +// networking will function if the network is namespaced. +func NewRootNamespace(stack Stack, creator NetworkStackCreator) *Namespace { + return &Namespace{ + stack: stack, + creator: creator, + isRoot: true, + } +} + +// NewNamespace creates a new network namespace from the root. +func NewNamespace(root *Namespace) *Namespace { + n := &Namespace{ + creator: root.creator, + } + n.init() + return n +} + +// Stack returns the network stack of n. Stack may return nil if no network +// stack is configured. +func (n *Namespace) Stack() Stack { + return n.stack +} + +// IsRoot returns whether n is the root network namespace. +func (n *Namespace) IsRoot() bool { + return n.isRoot +} + +// RestoreRootStack restores the root network namespace with stack. This should +// only be called when restoring kernel. +func (n *Namespace) RestoreRootStack(stack Stack) { + if !n.isRoot { + panic("RestoreRootStack can only be called on root network namespace") + } + if n.stack != nil { + panic("RestoreRootStack called after a stack has already been set") + } + n.stack = stack +} + +func (n *Namespace) init() { + // Root network namespace will have stack assigned later. + if n.isRoot { + return + } + if n.creator != nil { + var err error + n.stack, err = n.creator.CreateStack() + if err != nil { + panic(err) + } + } +} + +// afterLoad is invoked by stateify. +func (n *Namespace) afterLoad() { + n.init() +} + +// NetworkStackCreator allows new instances of a network stack to be created. It +// is used by the kernel to create new network namespaces when requested. +type NetworkStackCreator interface { + // CreateStack creates a new network stack for a network namespace. + CreateStack() (Stack, error) +} diff --git a/pkg/sentry/kernel/BUILD b/pkg/sentry/kernel/BUILD index 2231d6973..e47af66d6 100644 --- a/pkg/sentry/kernel/BUILD +++ b/pkg/sentry/kernel/BUILD @@ -157,6 +157,7 @@ go_library( "//pkg/context", "//pkg/cpuid", "//pkg/eventchannel", + "//pkg/fspath", "//pkg/log", "//pkg/metric", "//pkg/refs", @@ -167,6 +168,10 @@ go_library( "//pkg/sentry/fs", "//pkg/sentry/fs/lock", "//pkg/sentry/fs/timerfd", + "//pkg/sentry/fsbridge", + "//pkg/sentry/fsimpl/kernfs", + "//pkg/sentry/fsimpl/pipefs", + "//pkg/sentry/fsimpl/sockfs", "//pkg/sentry/hostcpu", "//pkg/sentry/inet", "//pkg/sentry/kernel/auth", @@ -199,6 +204,7 @@ go_library( "//pkg/tcpip/stack", "//pkg/usermem", "//pkg/waiter", + "//tools/go_marshal/marshal", ], ) diff --git a/pkg/sentry/kernel/epoll/epoll.go b/pkg/sentry/kernel/epoll/epoll.go index 8bffb78fc..592650923 100644 --- a/pkg/sentry/kernel/epoll/epoll.go +++ b/pkg/sentry/kernel/epoll/epoll.go @@ -296,8 +296,10 @@ func (*readyCallback) Callback(w *waiter.Entry) { e.waitingList.Remove(entry) e.readyList.PushBack(entry) entry.curList = &e.readyList + e.listsMu.Unlock() e.Notify(waiter.EventIn) + return } e.listsMu.Unlock() diff --git a/pkg/sentry/kernel/epoll/epoll_state.go b/pkg/sentry/kernel/epoll/epoll_state.go index a0d35d350..8e9f200d0 100644 --- a/pkg/sentry/kernel/epoll/epoll_state.go +++ b/pkg/sentry/kernel/epoll/epoll_state.go @@ -38,11 +38,14 @@ func (e *EventPoll) afterLoad() { } } - for it := e.waitingList.Front(); it != nil; it = it.Next() { - if it.id.File.Readiness(it.mask) != 0 { - e.waitingList.Remove(it) - e.readyList.PushBack(it) - it.curList = &e.readyList + for it := e.waitingList.Front(); it != nil; { + entry := it + it = it.Next() + + if entry.id.File.Readiness(entry.mask) != 0 { + e.waitingList.Remove(entry) + e.readyList.PushBack(entry) + entry.curList = &e.readyList e.Notify(waiter.EventIn) } } diff --git a/pkg/sentry/kernel/fd_table.go b/pkg/sentry/kernel/fd_table.go index 23b88f7a6..ed40b5303 100644 --- a/pkg/sentry/kernel/fd_table.go +++ b/pkg/sentry/kernel/fd_table.go @@ -191,10 +191,12 @@ func (f *FDTable) Size() int { return int(size) } -// forEach iterates over all non-nil files. +// forEach iterates over all non-nil files in sorted order. // // It is the caller's responsibility to acquire an appropriate lock. func (f *FDTable) forEach(fn func(fd int32, file *fs.File, fileVFS2 *vfs.FileDescription, flags FDFlags)) { + // retries tracks the number of failed TryIncRef attempts for the same FD. + retries := 0 fd := int32(0) for { file, fileVFS2, flags, ok := f.getAll(fd) @@ -204,17 +206,26 @@ func (f *FDTable) forEach(fn func(fd int32, file *fs.File, fileVFS2 *vfs.FileDes switch { case file != nil: if !file.TryIncRef() { + retries++ + if retries > 1000 { + panic(fmt.Sprintf("File in FD table has been destroyed. FD: %d, File: %+v, FileOps: %+v", fd, file, file.FileOperations)) + } continue // Race caught. } fn(fd, file, nil, flags) file.DecRef() case fileVFS2 != nil: if !fileVFS2.TryIncRef() { + retries++ + if retries > 1000 { + panic(fmt.Sprintf("File in FD table has been destroyed. FD: %d, File: %+v, Impl: %+v", fd, fileVFS2, fileVFS2.Impl())) + } continue // Race caught. } fn(fd, nil, fileVFS2, flags) fileVFS2.DecRef() } + retries = 0 fd++ } } @@ -296,6 +307,105 @@ func (f *FDTable) NewFDs(ctx context.Context, fd int32, files []*fs.File, flags return fds, nil } +// NewFDsVFS2 allocates new FDs guaranteed to be the lowest number available +// greater than or equal to the fd parameter. All files will share the set +// flags. Success is guaranteed to be all or none. +func (f *FDTable) NewFDsVFS2(ctx context.Context, fd int32, files []*vfs.FileDescription, flags FDFlags) (fds []int32, err error) { + if fd < 0 { + // Don't accept negative FDs. + return nil, syscall.EINVAL + } + + // Default limit. + end := int32(math.MaxInt32) + + // Ensure we don't get past the provided limit. + if limitSet := limits.FromContext(ctx); limitSet != nil { + lim := limitSet.Get(limits.NumberOfFiles) + if lim.Cur != limits.Infinity { + end = int32(lim.Cur) + } + if fd >= end { + return nil, syscall.EMFILE + } + } + + f.mu.Lock() + defer f.mu.Unlock() + + // From f.next to find available fd. + if fd < f.next { + fd = f.next + } + + // Install all entries. + for i := fd; i < end && len(fds) < len(files); i++ { + if d, _, _ := f.getVFS2(i); d == nil { + f.setVFS2(i, files[len(fds)], flags) // Set the descriptor. + fds = append(fds, i) // Record the file descriptor. + } + } + + // Failure? Unwind existing FDs. + if len(fds) < len(files) { + for _, i := range fds { + f.setVFS2(i, nil, FDFlags{}) // Zap entry. + } + return nil, syscall.EMFILE + } + + if fd == f.next { + // Update next search start position. + f.next = fds[len(fds)-1] + 1 + } + + return fds, nil +} + +// NewFDVFS2 allocates a file descriptor greater than or equal to minfd for +// the given file description. If it succeeds, it takes a reference on file. +func (f *FDTable) NewFDVFS2(ctx context.Context, minfd int32, file *vfs.FileDescription, flags FDFlags) (int32, error) { + if minfd < 0 { + // Don't accept negative FDs. + return -1, syscall.EINVAL + } + + // Default limit. + end := int32(math.MaxInt32) + + // Ensure we don't get past the provided limit. + if limitSet := limits.FromContext(ctx); limitSet != nil { + lim := limitSet.Get(limits.NumberOfFiles) + if lim.Cur != limits.Infinity { + end = int32(lim.Cur) + } + if minfd >= end { + return -1, syscall.EMFILE + } + } + + f.mu.Lock() + defer f.mu.Unlock() + + // From f.next to find available fd. + fd := minfd + if fd < f.next { + fd = f.next + } + for fd < end { + if d, _, _ := f.getVFS2(fd); d == nil { + f.setVFS2(fd, file, flags) + if fd == f.next { + // Update next search start position. + f.next = fd + 1 + } + return fd, nil + } + fd++ + } + return -1, syscall.EMFILE +} + // NewFDAt sets the file reference for the given FD. If there is an active // reference for that FD, the ref count for that existing reference is // decremented. @@ -316,9 +426,6 @@ func (f *FDTable) newFDAt(ctx context.Context, fd int32, file *fs.File, fileVFS2 return syscall.EBADF } - f.mu.Lock() - defer f.mu.Unlock() - // Check the limit for the provided file. if limitSet := limits.FromContext(ctx); limitSet != nil { if lim := limitSet.Get(limits.NumberOfFiles); lim.Cur != limits.Infinity && uint64(fd) >= lim.Cur { @@ -327,6 +434,8 @@ func (f *FDTable) newFDAt(ctx context.Context, fd int32, file *fs.File, fileVFS2 } // Install the entry. + f.mu.Lock() + defer f.mu.Unlock() f.setAll(fd, file, fileVFS2, flags) return nil } @@ -404,7 +513,10 @@ func (f *FDTable) GetVFS2(fd int32) (*vfs.FileDescription, FDFlags) { } } -// GetFDs returns a list of valid fds. +// GetFDs returns a sorted list of valid fds. +// +// Precondition: The caller must be running on the task goroutine, or Task.mu +// must be locked. func (f *FDTable) GetFDs() []int32 { fds := make([]int32, 0, int(atomic.LoadInt32(&f.used))) f.forEach(func(fd int32, _ *fs.File, _ *vfs.FileDescription, _ FDFlags) { @@ -479,7 +591,9 @@ func (f *FDTable) Remove(fd int32) (*fs.File, *vfs.FileDescription) { case orig2 != nil: orig2.IncRef() } - f.setAll(fd, nil, nil, FDFlags{}) // Zap entry. + if orig != nil || orig2 != nil { + f.setAll(fd, nil, nil, FDFlags{}) // Zap entry. + } return orig, orig2 } diff --git a/pkg/sentry/kernel/fs_context.go b/pkg/sentry/kernel/fs_context.go index 2448c1d99..47f78df9a 100644 --- a/pkg/sentry/kernel/fs_context.go +++ b/pkg/sentry/kernel/fs_context.go @@ -19,6 +19,7 @@ import ( "gvisor.dev/gvisor/pkg/refs" "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/sync" ) @@ -37,10 +38,16 @@ type FSContext struct { // destroyed. root *fs.Dirent + // rootVFS2 is the filesystem root. + rootVFS2 vfs.VirtualDentry + // cwd is the current working directory. Will be nil iff the FSContext // has been destroyed. cwd *fs.Dirent + // cwdVFS2 is the current working directory. + cwdVFS2 vfs.VirtualDentry + // umask is the current file mode creation mask. When a thread using this // context invokes a syscall that creates a file, bits set in umask are // removed from the permissions that the file is created with. @@ -60,6 +67,19 @@ func newFSContext(root, cwd *fs.Dirent, umask uint) *FSContext { return &f } +// NewFSContextVFS2 returns a new filesystem context. +func NewFSContextVFS2(root, cwd vfs.VirtualDentry, umask uint) *FSContext { + root.IncRef() + cwd.IncRef() + f := FSContext{ + rootVFS2: root, + cwdVFS2: cwd, + umask: umask, + } + f.EnableLeakCheck("kernel.FSContext") + return &f +} + // destroy is the destructor for an FSContext. // // This will call DecRef on both root and cwd Dirents. If either call to @@ -75,11 +95,17 @@ func (f *FSContext) destroy() { f.mu.Lock() defer f.mu.Unlock() - f.root.DecRef() - f.root = nil - - f.cwd.DecRef() - f.cwd = nil + if VFS2Enabled { + f.rootVFS2.DecRef() + f.rootVFS2 = vfs.VirtualDentry{} + f.cwdVFS2.DecRef() + f.cwdVFS2 = vfs.VirtualDentry{} + } else { + f.root.DecRef() + f.root = nil + f.cwd.DecRef() + f.cwd = nil + } } // DecRef implements RefCounter.DecRef with destructor f.destroy. @@ -93,12 +119,21 @@ func (f *FSContext) DecRef() { func (f *FSContext) Fork() *FSContext { f.mu.Lock() defer f.mu.Unlock() - f.cwd.IncRef() - f.root.IncRef() + + if VFS2Enabled { + f.cwdVFS2.IncRef() + f.rootVFS2.IncRef() + } else { + f.cwd.IncRef() + f.root.IncRef() + } + return &FSContext{ - cwd: f.cwd, - root: f.root, - umask: f.umask, + cwd: f.cwd, + root: f.root, + cwdVFS2: f.cwdVFS2, + rootVFS2: f.rootVFS2, + umask: f.umask, } } @@ -109,12 +144,23 @@ func (f *FSContext) Fork() *FSContext { func (f *FSContext) WorkingDirectory() *fs.Dirent { f.mu.Lock() defer f.mu.Unlock() - if f.cwd != nil { - f.cwd.IncRef() - } + + f.cwd.IncRef() return f.cwd } +// WorkingDirectoryVFS2 returns the current working directory. +// +// This will return nil if called after destroy(), otherwise it will return a +// Dirent with a reference taken. +func (f *FSContext) WorkingDirectoryVFS2() vfs.VirtualDentry { + f.mu.Lock() + defer f.mu.Unlock() + + f.cwdVFS2.IncRef() + return f.cwdVFS2 +} + // SetWorkingDirectory sets the current working directory. // This will take an extra reference on the Dirent. // @@ -137,6 +183,20 @@ func (f *FSContext) SetWorkingDirectory(d *fs.Dirent) { old.DecRef() } +// SetWorkingDirectoryVFS2 sets the current working directory. +// This will take an extra reference on the VirtualDentry. +// +// This is not a valid call after destroy. +func (f *FSContext) SetWorkingDirectoryVFS2(d vfs.VirtualDentry) { + f.mu.Lock() + defer f.mu.Unlock() + + old := f.cwdVFS2 + f.cwdVFS2 = d + d.IncRef() + old.DecRef() +} + // RootDirectory returns the current filesystem root. // // This will return nil if called after destroy(), otherwise it will return a @@ -150,6 +210,18 @@ func (f *FSContext) RootDirectory() *fs.Dirent { return f.root } +// RootDirectoryVFS2 returns the current filesystem root. +// +// This will return nil if called after destroy(), otherwise it will return a +// Dirent with a reference taken. +func (f *FSContext) RootDirectoryVFS2() vfs.VirtualDentry { + f.mu.Lock() + defer f.mu.Unlock() + + f.rootVFS2.IncRef() + return f.rootVFS2 +} + // SetRootDirectory sets the root directory. // This will take an extra reference on the Dirent. // @@ -172,6 +244,28 @@ func (f *FSContext) SetRootDirectory(d *fs.Dirent) { old.DecRef() } +// SetRootDirectoryVFS2 sets the root directory. It takes a reference on vd. +// +// This is not a valid call after free. +func (f *FSContext) SetRootDirectoryVFS2(vd vfs.VirtualDentry) { + if !vd.Ok() { + panic("FSContext.SetRootDirectoryVFS2 called with zero-value VirtualDentry") + } + + f.mu.Lock() + + if !f.rootVFS2.Ok() { + f.mu.Unlock() + panic(fmt.Sprintf("FSContext.SetRootDirectoryVFS2(%v)) called after destroy", vd)) + } + + old := f.rootVFS2 + vd.IncRef() + f.rootVFS2 = vd + f.mu.Unlock() + old.DecRef() +} + // Umask returns the current umask. func (f *FSContext) Umask() uint { f.mu.Lock() diff --git a/pkg/sentry/kernel/kernel.go b/pkg/sentry/kernel/kernel.go index 3ee760ba2..fef60e636 100644 --- a/pkg/sentry/kernel/kernel.go +++ b/pkg/sentry/kernel/kernel.go @@ -43,11 +43,15 @@ import ( "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/cpuid" "gvisor.dev/gvisor/pkg/eventchannel" + "gvisor.dev/gvisor/pkg/fspath" "gvisor.dev/gvisor/pkg/log" "gvisor.dev/gvisor/pkg/refs" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fs/timerfd" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/pipefs" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/sockfs" "gvisor.dev/gvisor/pkg/sentry/hostcpu" "gvisor.dev/gvisor/pkg/sentry/inet" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" @@ -71,6 +75,10 @@ import ( "gvisor.dev/gvisor/pkg/tcpip" ) +// VFS2Enabled is set to true when VFS2 is enabled. Added as a global for allow +// easy access everywhere. To be removed once VFS2 becomes the default. +var VFS2Enabled = false + // Kernel represents an emulated Linux kernel. It must be initialized by calling // Init() or LoadFrom(). // @@ -105,7 +113,7 @@ type Kernel struct { timekeeper *Timekeeper tasks *TaskSet rootUserNamespace *auth.UserNamespace - networkStack inet.Stack `state:"nosave"` + rootNetworkNamespace *inet.Namespace applicationCores uint useHostCores bool extraAuxv []arch.AuxEntry @@ -219,6 +227,11 @@ type Kernel struct { // by extMu. nextSocketEntry uint64 + // socketMount is a disconnected vfs.Mount, not included in k.vfs, + // representing a sockfs.filesystem. socketMount is used to back + // VirtualDentries representing anonymous sockets. + socketMount *vfs.Mount + // deviceRegistry is used to save/restore device.SimpleDevices. deviceRegistry struct{} `state:".(*device.Registry)"` @@ -238,6 +251,17 @@ type Kernel struct { // SpecialOpts contains special kernel options. SpecialOpts + + // VFS keeps the filesystem state used across the kernel. + vfs vfs.VirtualFilesystem + + // pipeMount is the Mount used for pipes created by the pipe() and pipe2() + // syscalls (as opposed to named pipes created by mknod()). + pipeMount *vfs.Mount + + // If set to true, report address space activation waits as if the task is in + // external wait so that the watchdog doesn't report the task stuck. + SleepForAddressSpaceActivation bool } // InitKernelArgs holds arguments to Init. @@ -251,8 +275,9 @@ type InitKernelArgs struct { // RootUserNamespace is the root user namespace. RootUserNamespace *auth.UserNamespace - // NetworkStack is the TCP/IP network stack. NetworkStack may be nil. - NetworkStack inet.Stack + // RootNetworkNamespace is the root network namespace. If nil, no networking + // will be available. + RootNetworkNamespace *inet.Namespace // ApplicationCores is the number of logical CPUs visible to sandboxed // applications. The set of logical CPU IDs is [0, ApplicationCores); thus @@ -311,7 +336,10 @@ func (k *Kernel) Init(args InitKernelArgs) error { k.rootUTSNamespace = args.RootUTSNamespace k.rootIPCNamespace = args.RootIPCNamespace k.rootAbstractSocketNamespace = args.RootAbstractSocketNamespace - k.networkStack = args.NetworkStack + k.rootNetworkNamespace = args.RootNetworkNamespace + if k.rootNetworkNamespace == nil { + k.rootNetworkNamespace = inet.NewRootNamespace(nil, nil) + } k.applicationCores = args.ApplicationCores if args.UseHostCores { k.useHostCores = true @@ -331,6 +359,29 @@ func (k *Kernel) Init(args InitKernelArgs) error { k.monotonicClock = &timekeeperClock{tk: args.Timekeeper, c: sentrytime.Monotonic} k.futexes = futex.NewManager() k.netlinkPorts = port.New() + + if VFS2Enabled { + if err := k.vfs.Init(); err != nil { + return fmt.Errorf("failed to initialize VFS: %v", err) + } + + pipeFilesystem := pipefs.NewFilesystem(&k.vfs) + defer pipeFilesystem.DecRef() + pipeMount, err := k.vfs.NewDisconnectedMount(pipeFilesystem, nil, &vfs.MountOptions{}) + if err != nil { + return fmt.Errorf("failed to create pipefs mount: %v", err) + } + k.pipeMount = pipeMount + + socketFilesystem := sockfs.NewFilesystem(&k.vfs) + defer socketFilesystem.DecRef() + socketMount, err := k.vfs.NewDisconnectedMount(socketFilesystem, nil, &vfs.MountOptions{}) + if err != nil { + return fmt.Errorf("failed to initialize socket mount: %v", err) + } + k.socketMount = socketMount + } + return nil } @@ -450,6 +501,11 @@ func (k *Kernel) flushMountSourceRefs() error { // // Precondition: Must be called with the kernel paused. func (ts *TaskSet) forEachFDPaused(f func(*fs.File, *vfs.FileDescription) error) (err error) { + // TODO(gvisor.dev/issue/1663): Add save support for VFS2. + if VFS2Enabled { + return nil + } + ts.mu.RLock() defer ts.mu.RUnlock() for t := range ts.Root.tids { @@ -467,7 +523,7 @@ func (ts *TaskSet) forEachFDPaused(f func(*fs.File, *vfs.FileDescription) error) } func (ts *TaskSet) flushWritesToFiles(ctx context.Context) error { - // TODO(gvisor.dev/issues/1663): Add save support for VFS2. + // TODO(gvisor.dev/issue/1663): Add save support for VFS2. return ts.forEachFDPaused(func(file *fs.File, _ *vfs.FileDescription) error { if flags := file.Flags(); !flags.Write { return nil @@ -516,17 +572,32 @@ func (k *Kernel) invalidateUnsavableMappings(ctx context.Context) error { } func (ts *TaskSet) unregisterEpollWaiters() { + // TODO(gvisor.dev/issue/1663): Add save support for VFS2. + if VFS2Enabled { + return + } + ts.mu.RLock() defer ts.mu.RUnlock() + + // Tasks that belong to the same process could potentially point to the + // same FDTable. So we retain a map of processed ones to avoid + // processing the same FDTable multiple times. + processed := make(map[*FDTable]struct{}) for t := range ts.Root.tids { // We can skip locking Task.mu here since the kernel is paused. - if t.fdTable != nil { - t.fdTable.forEach(func(_ int32, file *fs.File, _ *vfs.FileDescription, _ FDFlags) { - if e, ok := file.FileOperations.(*epoll.EventPoll); ok { - e.UnregisterEpollWaiters() - } - }) + if t.fdTable == nil { + continue + } + if _, ok := processed[t.fdTable]; ok { + continue } + t.fdTable.forEach(func(_ int32, file *fs.File, _ *vfs.FileDescription, _ FDFlags) { + if e, ok := file.FileOperations.(*epoll.EventPoll); ok { + e.UnregisterEpollWaiters() + } + }) + processed[t.fdTable] = struct{}{} } } @@ -534,8 +605,6 @@ func (ts *TaskSet) unregisterEpollWaiters() { func (k *Kernel) LoadFrom(r io.Reader, net inet.Stack, clocks sentrytime.Clocks) error { loadStart := time.Now() - k.networkStack = net - initAppCores := k.applicationCores // Load the pre-saved CPUID FeatureSet. @@ -566,6 +635,10 @@ func (k *Kernel) LoadFrom(r io.Reader, net inet.Stack, clocks sentrytime.Clocks) log.Infof("Kernel load stats: %s", &stats) log.Infof("Kernel load took [%s].", time.Since(kernelStart)) + // rootNetworkNamespace should be populated after loading the state file. + // Restore the root network stack. + k.rootNetworkNamespace.RestoreRootStack(net) + // Load the memory file's state. memoryStart := time.Now() if err := k.mf.LoadFrom(k.SupervisorContext(), r); err != nil { @@ -624,7 +697,7 @@ type CreateProcessArgs struct { // File is a passed host FD pointing to a file to load as the init binary. // // This is checked if and only if Filename is "". - File *fs.File + File fsbridge.File // Argvv is a list of arguments. Argv []string @@ -673,6 +746,13 @@ type CreateProcessArgs struct { // increment it). MountNamespace *fs.MountNamespace + // MountNamespaceVFS2 optionally contains the mount namespace for this + // process. If nil, the init process's mount namespace is used. + // + // Anyone setting MountNamespaceVFS2 must donate a reference (i.e. + // increment it). + MountNamespaceVFS2 *vfs.MountNamespace + // ContainerID is the container that the process belongs to. ContainerID string } @@ -711,13 +791,26 @@ func (ctx *createProcessContext) Value(key interface{}) interface{} { return ctx.args.Credentials case fs.CtxRoot: if ctx.args.MountNamespace != nil { - // MountNamespace.Root() will take a reference on the root - // dirent for us. + // MountNamespace.Root() will take a reference on the root dirent for us. return ctx.args.MountNamespace.Root() } return nil + case vfs.CtxRoot: + if ctx.args.MountNamespaceVFS2 == nil { + return nil + } + // MountNamespaceVFS2.Root() takes a reference on the root dirent for us. + return ctx.args.MountNamespaceVFS2.Root() + case vfs.CtxMountNamespace: + if ctx.k.globalInit == nil { + return nil + } + // MountNamespaceVFS2 takes a reference for us. + return ctx.k.GlobalInit().Leader().MountNamespaceVFS2() case fs.CtxDirentCacheLimiter: return ctx.k.DirentCacheLimiter + case inet.CtxStack: + return ctx.k.RootNetworkNamespace().Stack() case ktime.CtxRealtimeClock: return ctx.k.RealtimeClock() case limits.CtxLimits: @@ -757,34 +850,77 @@ func (k *Kernel) CreateProcess(args CreateProcessArgs) (*ThreadGroup, ThreadID, defer k.extMu.Unlock() log.Infof("EXEC: %v", args.Argv) - // Grab the mount namespace. - mounts := args.MountNamespace - if mounts == nil { - mounts = k.GlobalInit().Leader().MountNamespace() - mounts.IncRef() - } - - tg := k.NewThreadGroup(mounts, args.PIDNamespace, NewSignalHandlers(), linux.SIGCHLD, args.Limits) ctx := args.NewContext(k) - // Get the root directory from the MountNamespace. - root := mounts.Root() - // The call to newFSContext below will take a reference on root, so we - // don't need to hold this one. - defer root.DecRef() - - // Grab the working directory. - remainingTraversals := uint(args.MaxSymlinkTraversals) - wd := root // Default. - if args.WorkingDirectory != "" { - var err error - wd, err = mounts.FindInode(ctx, root, nil, args.WorkingDirectory, &remainingTraversals) - if err != nil { - return nil, 0, fmt.Errorf("failed to find initial working directory %q: %v", args.WorkingDirectory, err) + var ( + opener fsbridge.Lookup + fsContext *FSContext + mntns *fs.MountNamespace + ) + + if VFS2Enabled { + mntnsVFS2 := args.MountNamespaceVFS2 + if mntnsVFS2 == nil { + // MountNamespaceVFS2 adds a reference to the namespace, which is + // transferred to the new process. + mntnsVFS2 = k.GlobalInit().Leader().MountNamespaceVFS2() + } + // Get the root directory from the MountNamespace. + root := args.MountNamespaceVFS2.Root() + // The call to newFSContext below will take a reference on root, so we + // don't need to hold this one. + defer root.DecRef() + + // Grab the working directory. + wd := root // Default. + if args.WorkingDirectory != "" { + pop := vfs.PathOperation{ + Root: root, + Start: wd, + Path: fspath.Parse(args.WorkingDirectory), + FollowFinalSymlink: true, + } + var err error + wd, err = k.VFS().GetDentryAt(ctx, args.Credentials, &pop, &vfs.GetDentryOptions{ + CheckSearchable: true, + }) + if err != nil { + return nil, 0, fmt.Errorf("failed to find initial working directory %q: %v", args.WorkingDirectory, err) + } + defer wd.DecRef() + } + opener = fsbridge.NewVFSLookup(mntnsVFS2, root, wd) + fsContext = NewFSContextVFS2(root, wd, args.Umask) + + } else { + mntns = args.MountNamespace + if mntns == nil { + mntns = k.GlobalInit().Leader().MountNamespace() + mntns.IncRef() } - defer wd.DecRef() + // Get the root directory from the MountNamespace. + root := mntns.Root() + // The call to newFSContext below will take a reference on root, so we + // don't need to hold this one. + defer root.DecRef() + + // Grab the working directory. + remainingTraversals := args.MaxSymlinkTraversals + wd := root // Default. + if args.WorkingDirectory != "" { + var err error + wd, err = mntns.FindInode(ctx, root, nil, args.WorkingDirectory, &remainingTraversals) + if err != nil { + return nil, 0, fmt.Errorf("failed to find initial working directory %q: %v", args.WorkingDirectory, err) + } + defer wd.DecRef() + } + opener = fsbridge.NewFSLookup(mntns, root, wd) + fsContext = newFSContext(root, wd, args.Umask) } + tg := k.NewThreadGroup(mntns, args.PIDNamespace, NewSignalHandlers(), linux.SIGCHLD, args.Limits) + // Check which file to start from. switch { case args.Filename != "": @@ -805,11 +941,9 @@ func (k *Kernel) CreateProcess(args CreateProcessArgs) (*ThreadGroup, ThreadID, } // Create a fresh task context. - remainingTraversals = uint(args.MaxSymlinkTraversals) + remainingTraversals := args.MaxSymlinkTraversals loadArgs := loader.LoadArgs{ - Mounts: mounts, - Root: root, - WorkingDirectory: wd, + Opener: opener, RemainingTraversals: &remainingTraversals, ResolveFinal: true, Filename: args.Filename, @@ -834,13 +968,15 @@ func (k *Kernel) CreateProcess(args CreateProcessArgs) (*ThreadGroup, ThreadID, Kernel: k, ThreadGroup: tg, TaskContext: tc, - FSContext: newFSContext(root, wd, args.Umask), + FSContext: fsContext, FDTable: args.FDTable, Credentials: args.Credentials, + NetworkNamespace: k.RootNetworkNamespace(), AllowedCPUMask: sched.NewFullCPUSet(k.applicationCores), UTSNamespace: args.UTSNamespace, IPCNamespace: args.IPCNamespace, AbstractSocketNamespace: args.AbstractSocketNamespace, + MountNamespaceVFS2: args.MountNamespaceVFS2, ContainerID: args.ContainerID, } t, err := k.tasks.NewTask(config) @@ -923,9 +1059,15 @@ func (k *Kernel) pauseTimeLocked() { // This means we'll iterate FDTables shared by multiple tasks repeatedly, // but ktime.Timer.Pause is idempotent so this is harmless. if t.fdTable != nil { - t.fdTable.forEach(func(_ int32, file *fs.File, _ *vfs.FileDescription, _ FDFlags) { - if tfd, ok := file.FileOperations.(*timerfd.TimerOperations); ok { - tfd.PauseTimer() + t.fdTable.forEach(func(_ int32, file *fs.File, fd *vfs.FileDescription, _ FDFlags) { + if VFS2Enabled { + if tfd, ok := fd.Impl().(*vfs.TimerFileDescription); ok { + tfd.PauseTimer() + } + } else { + if tfd, ok := file.FileOperations.(*timerfd.TimerOperations); ok { + tfd.PauseTimer() + } } }) } @@ -953,9 +1095,15 @@ func (k *Kernel) resumeTimeLocked() { } } if t.fdTable != nil { - t.fdTable.forEach(func(_ int32, file *fs.File, _ *vfs.FileDescription, _ FDFlags) { - if tfd, ok := file.FileOperations.(*timerfd.TimerOperations); ok { - tfd.ResumeTimer() + t.fdTable.forEach(func(_ int32, file *fs.File, fd *vfs.FileDescription, _ FDFlags) { + if VFS2Enabled { + if tfd, ok := fd.Impl().(*vfs.TimerFileDescription); ok { + tfd.ResumeTimer() + } + } else { + if tfd, ok := file.FileOperations.(*timerfd.TimerOperations); ok { + tfd.ResumeTimer() + } } }) } @@ -1100,6 +1248,14 @@ func (k *Kernel) SendExternalSignal(info *arch.SignalInfo, context string) { k.sendExternalSignal(info, context) } +// SendExternalSignalThreadGroup injects a signal into an specific ThreadGroup. +// This function doesn't skip signals like SendExternalSignal does. +func (k *Kernel) SendExternalSignalThreadGroup(tg *ThreadGroup, info *arch.SignalInfo) error { + k.extMu.Lock() + defer k.extMu.Unlock() + return tg.SendSignal(info) +} + // SendContainerSignal sends the given signal to all processes inside the // namespace that match the given container ID. func (k *Kernel) SendContainerSignal(cid string, info *arch.SignalInfo) error { @@ -1178,10 +1334,9 @@ func (k *Kernel) RootAbstractSocketNamespace() *AbstractSocketNamespace { return k.rootAbstractSocketNamespace } -// NetworkStack returns the network stack. NetworkStack may return nil if no -// network stack is available. -func (k *Kernel) NetworkStack() inet.Stack { - return k.networkStack +// RootNetworkNamespace returns the root network namespace, always non-nil. +func (k *Kernel) RootNetworkNamespace() *inet.Namespace { + return k.rootNetworkNamespace } // GlobalInit returns the thread group with ID 1 in the root PID namespace, or @@ -1311,9 +1466,10 @@ func (k *Kernel) SupervisorContext() context.Context { // +stateify savable type SocketEntry struct { socketEntry - k *Kernel - Sock *refs.WeakRef - ID uint64 // Socket table entry number. + k *Kernel + Sock *refs.WeakRef + SockVFS2 *vfs.FileDescription + ID uint64 // Socket table entry number. } // WeakRefGone implements refs.WeakRefUser.WeakRefGone. @@ -1336,7 +1492,30 @@ func (k *Kernel) RecordSocket(sock *fs.File) { k.extMu.Unlock() } +// RecordSocketVFS2 adds a VFS2 socket to the system-wide socket table for +// tracking. +// +// Precondition: Caller must hold a reference to sock. +// +// Note that the socket table will not hold a reference on the +// vfs.FileDescription, because we do not support weak refs on VFS2 files. +func (k *Kernel) RecordSocketVFS2(sock *vfs.FileDescription) { + k.extMu.Lock() + id := k.nextSocketEntry + k.nextSocketEntry++ + s := &SocketEntry{ + k: k, + ID: id, + SockVFS2: sock, + } + k.sockets.PushBack(s) + k.extMu.Unlock() +} + // ListSockets returns a snapshot of all sockets. +// +// Callers of ListSockets() in VFS2 should use SocketEntry.SockVFS2.TryIncRef() +// to get a reference on a socket in the table. func (k *Kernel) ListSockets() []*SocketEntry { k.extMu.Lock() var socks []*SocketEntry @@ -1347,6 +1526,11 @@ func (k *Kernel) ListSockets() []*SocketEntry { return socks } +// SocketMount returns the global socket mount. +func (k *Kernel) SocketMount() *vfs.Mount { + return k.socketMount +} + // supervisorContext is a privileged context. type supervisorContext struct { context.NoopSleeper @@ -1378,8 +1562,24 @@ func (ctx supervisorContext) Value(key interface{}) interface{} { return ctx.k.globalInit.mounts.Root() } return nil + case vfs.CtxRoot: + if ctx.k.globalInit == nil { + return vfs.VirtualDentry{} + } + mntns := ctx.k.GlobalInit().Leader().MountNamespaceVFS2() + defer mntns.DecRef() + // Root() takes a reference on the root dirent for us. + return mntns.Root() + case vfs.CtxMountNamespace: + if ctx.k.globalInit == nil { + return nil + } + // MountNamespaceVFS2() takes a reference for us. + return ctx.k.GlobalInit().Leader().MountNamespaceVFS2() case fs.CtxDirentCacheLimiter: return ctx.k.DirentCacheLimiter + case inet.CtxStack: + return ctx.k.RootNetworkNamespace().Stack() case ktime.CtxRealtimeClock: return ctx.k.RealtimeClock() case limits.CtxLimits: @@ -1423,3 +1623,13 @@ func (k *Kernel) EmitUnimplementedEvent(ctx context.Context) { Registers: t.Arch().StateData().Proto(), }) } + +// VFS returns the virtual filesystem for the kernel. +func (k *Kernel) VFS() *vfs.VirtualFilesystem { + return &k.vfs +} + +// PipeMount returns the pipefs mount. +func (k *Kernel) PipeMount() *vfs.Mount { + return k.pipeMount +} diff --git a/pkg/sentry/kernel/pipe/BUILD b/pkg/sentry/kernel/pipe/BUILD index 4c049d5b4..f29dc0472 100644 --- a/pkg/sentry/kernel/pipe/BUILD +++ b/pkg/sentry/kernel/pipe/BUILD @@ -1,25 +1,10 @@ load("//tools:defs.bzl", "go_library", "go_test") -load("//tools/go_generics:defs.bzl", "go_template_instance") package(licenses = ["notice"]) -go_template_instance( - name = "buffer_list", - out = "buffer_list.go", - package = "pipe", - prefix = "buffer", - template = "//pkg/ilist:generic_list", - types = { - "Element": "*buffer", - "Linker": "*buffer", - }, -) - go_library( name = "pipe", srcs = [ - "buffer.go", - "buffer_list.go", "device.go", "node.go", "pipe.go", @@ -33,8 +18,8 @@ go_library( deps = [ "//pkg/abi/linux", "//pkg/amutex", + "//pkg/buffer", "//pkg/context", - "//pkg/safemem", "//pkg/sentry/arch", "//pkg/sentry/device", "//pkg/sentry/fs", @@ -51,7 +36,6 @@ go_test( name = "pipe_test", size = "small", srcs = [ - "buffer_test.go", "node_test.go", "pipe_test.go", ], diff --git a/pkg/sentry/kernel/pipe/buffer.go b/pkg/sentry/kernel/pipe/buffer.go deleted file mode 100644 index fe3be5dbd..000000000 --- a/pkg/sentry/kernel/pipe/buffer.go +++ /dev/null @@ -1,115 +0,0 @@ -// Copyright 2018 The gVisor Authors. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -package pipe - -import ( - "io" - - "gvisor.dev/gvisor/pkg/safemem" - "gvisor.dev/gvisor/pkg/sync" -) - -// buffer encapsulates a queueable byte buffer. -// -// Note that the total size is slightly less than two pages. This -// is done intentionally to ensure that the buffer object aligns -// with runtime internals. We have no hard size or alignment -// requirements. This two page size will effectively minimize -// internal fragmentation, but still have a large enough chunk -// to limit excessive segmentation. -// -// +stateify savable -type buffer struct { - data [8144]byte - read int - write int - bufferEntry -} - -// Reset resets internal data. -// -// This must be called before use. -func (b *buffer) Reset() { - b.read = 0 - b.write = 0 -} - -// Empty indicates the buffer is empty. -// -// This indicates there is no data left to read. -func (b *buffer) Empty() bool { - return b.read == b.write -} - -// Full indicates the buffer is full. -// -// This indicates there is no capacity left to write. -func (b *buffer) Full() bool { - return b.write == len(b.data) -} - -// WriteFromBlocks implements safemem.Writer.WriteFromBlocks. -func (b *buffer) WriteFromBlocks(srcs safemem.BlockSeq) (uint64, error) { - dst := safemem.BlockSeqOf(safemem.BlockFromSafeSlice(b.data[b.write:])) - n, err := safemem.CopySeq(dst, srcs) - b.write += int(n) - return n, err -} - -// WriteFromReader writes to the buffer from an io.Reader. -func (b *buffer) WriteFromReader(r io.Reader, count int64) (int64, error) { - dst := b.data[b.write:] - if count < int64(len(dst)) { - dst = b.data[b.write:][:count] - } - n, err := r.Read(dst) - b.write += n - return int64(n), err -} - -// ReadToBlocks implements safemem.Reader.ReadToBlocks. -func (b *buffer) ReadToBlocks(dsts safemem.BlockSeq) (uint64, error) { - src := safemem.BlockSeqOf(safemem.BlockFromSafeSlice(b.data[b.read:b.write])) - n, err := safemem.CopySeq(dsts, src) - b.read += int(n) - return n, err -} - -// ReadToWriter reads from the buffer into an io.Writer. -func (b *buffer) ReadToWriter(w io.Writer, count int64, dup bool) (int64, error) { - src := b.data[b.read:b.write] - if count < int64(len(src)) { - src = b.data[b.read:][:count] - } - n, err := w.Write(src) - if !dup { - b.read += n - } - return int64(n), err -} - -// bufferPool is a pool for buffers. -var bufferPool = sync.Pool{ - New: func() interface{} { - return new(buffer) - }, -} - -// newBuffer grabs a new buffer from the pool. -func newBuffer() *buffer { - b := bufferPool.Get().(*buffer) - b.Reset() - return b -} diff --git a/pkg/sentry/kernel/pipe/pipe.go b/pkg/sentry/kernel/pipe/pipe.go index 08410283f..62c8691f1 100644 --- a/pkg/sentry/kernel/pipe/pipe.go +++ b/pkg/sentry/kernel/pipe/pipe.go @@ -20,6 +20,7 @@ import ( "sync/atomic" "syscall" + "gvisor.dev/gvisor/pkg/buffer" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sync" @@ -70,10 +71,10 @@ type Pipe struct { // mu protects all pipe internal state below. mu sync.Mutex `state:"nosave"` - // data is the buffer queue of pipe contents. + // view is the underlying set of buffers. // // This is protected by mu. - data bufferList + view buffer.View // max is the maximum size of the pipe in bytes. When this max has been // reached, writers will get EWOULDBLOCK. @@ -81,11 +82,6 @@ type Pipe struct { // This is protected by mu. max int64 - // size is the current size of the pipe in bytes. - // - // This is protected by mu. - size int64 - // hadWriter indicates if this pipe ever had a writer. Note that this // does not necessarily indicate there is *currently* a writer, just // that there has been a writer at some point since the pipe was @@ -196,7 +192,7 @@ type readOps struct { limit func(int64) // read performs the actual read operation. - read func(*buffer) (int64, error) + read func(*buffer.View) (int64, error) } // read reads data from the pipe into dst and returns the number of bytes @@ -213,7 +209,7 @@ func (p *Pipe) read(ctx context.Context, ops readOps) (int64, error) { defer p.mu.Unlock() // Is the pipe empty? - if p.size == 0 { + if p.view.Size() == 0 { if !p.HasWriters() { // There are no writers, return EOF. return 0, nil @@ -222,71 +218,13 @@ func (p *Pipe) read(ctx context.Context, ops readOps) (int64, error) { } // Limit how much we consume. - if ops.left() > p.size { - ops.limit(p.size) - } - - done := int64(0) - for ops.left() > 0 { - // Pop the first buffer. - first := p.data.Front() - if first == nil { - break - } - - // Copy user data. - n, err := ops.read(first) - done += int64(n) - p.size -= n - - // Empty buffer? - if first.Empty() { - // Push to the free list. - p.data.Remove(first) - bufferPool.Put(first) - } - - // Handle errors. - if err != nil { - return done, err - } - } - - return done, nil -} - -// dup duplicates all data from this pipe into the given writer. -// -// There is no blocking behavior implemented here. The writer may propagate -// some blocking error. All the writes must be complete writes. -func (p *Pipe) dup(ctx context.Context, ops readOps) (int64, error) { - p.mu.Lock() - defer p.mu.Unlock() - - // Is the pipe empty? - if p.size == 0 { - if !p.HasWriters() { - // See above. - return 0, nil - } - return 0, syserror.ErrWouldBlock - } - - // Limit how much we consume. - if ops.left() > p.size { - ops.limit(p.size) + if ops.left() > p.view.Size() { + ops.limit(p.view.Size()) } - done := int64(0) - for buf := p.data.Front(); buf != nil; buf = buf.Next() { - n, err := ops.read(buf) - done += n - if err != nil { - return done, err - } - } - - return done, nil + // Copy user data; the read op is responsible for trimming. + done, err := ops.read(&p.view) + return done, err } type writeOps struct { @@ -297,7 +235,7 @@ type writeOps struct { limit func(int64) // write should write to the provided buffer. - write func(*buffer) (int64, error) + write func(*buffer.View) (int64, error) } // write writes data from sv into the pipe and returns the number of bytes @@ -317,35 +255,28 @@ func (p *Pipe) write(ctx context.Context, ops writeOps) (int64, error) { // POSIX requires that a write smaller than atomicIOBytes (PIPE_BUF) be // atomic, but requires no atomicity for writes larger than this. wanted := ops.left() - if avail := p.max - p.size; wanted > avail { + avail := p.max - p.view.Size() + if wanted > avail { if wanted <= p.atomicIOBytes { return 0, syserror.ErrWouldBlock } ops.limit(avail) } - done := int64(0) - for ops.left() > 0 { - // Need a new buffer? - last := p.data.Back() - if last == nil || last.Full() { - // Add a new buffer to the data list. - last = newBuffer() - p.data.PushBack(last) - } - - // Copy user data. - n, err := ops.write(last) - done += int64(n) - p.size += n + // Copy user data. + done, err := ops.write(&p.view) + if err != nil { + return done, err + } - // Handle errors. - if err != nil { - return done, err - } + if done < avail { + // Non-failure, but short write. + return done, nil } - if wanted > done { - // Partial write due to full pipe. + if done < wanted { + // Partial write due to full pipe. Note that this could also be + // the short write case above, we would expect a second call + // and the write to return zero bytes in this case. return done, syserror.ErrWouldBlock } @@ -396,7 +327,7 @@ func (p *Pipe) HasWriters() bool { // Precondition: mu must be held. func (p *Pipe) rReadinessLocked() waiter.EventMask { ready := waiter.EventMask(0) - if p.HasReaders() && p.data.Front() != nil { + if p.HasReaders() && p.view.Size() != 0 { ready |= waiter.EventIn } if !p.HasWriters() && p.hadWriter { @@ -422,7 +353,7 @@ func (p *Pipe) rReadiness() waiter.EventMask { // Precondition: mu must be held. func (p *Pipe) wReadinessLocked() waiter.EventMask { ready := waiter.EventMask(0) - if p.HasWriters() && p.size < p.max { + if p.HasWriters() && p.view.Size() < p.max { ready |= waiter.EventOut } if !p.HasReaders() { @@ -451,7 +382,7 @@ func (p *Pipe) rwReadiness() waiter.EventMask { func (p *Pipe) queued() int64 { p.mu.Lock() defer p.mu.Unlock() - return p.size + return p.view.Size() } // FifoSize implements fs.FifoSizer.FifoSize. @@ -474,7 +405,7 @@ func (p *Pipe) SetFifoSize(size int64) (int64, error) { } p.mu.Lock() defer p.mu.Unlock() - if size < p.size { + if size < p.view.Size() { return 0, syserror.EBUSY } p.max = size diff --git a/pkg/sentry/kernel/pipe/pipe_util.go b/pkg/sentry/kernel/pipe/pipe_util.go index 80158239e..5a1d4fd57 100644 --- a/pkg/sentry/kernel/pipe/pipe_util.go +++ b/pkg/sentry/kernel/pipe/pipe_util.go @@ -21,6 +21,7 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/amutex" + "gvisor.dev/gvisor/pkg/buffer" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sync" @@ -49,9 +50,10 @@ func (p *Pipe) Read(ctx context.Context, dst usermem.IOSequence) (int64, error) limit: func(l int64) { dst = dst.TakeFirst64(l) }, - read: func(buf *buffer) (int64, error) { - n, err := dst.CopyOutFrom(ctx, buf) + read: func(view *buffer.View) (int64, error) { + n, err := dst.CopyOutFrom(ctx, view) dst = dst.DropFirst64(n) + view.TrimFront(n) return n, err }, }) @@ -70,16 +72,15 @@ func (p *Pipe) WriteTo(ctx context.Context, w io.Writer, count int64, dup bool) limit: func(l int64) { count = l }, - read: func(buf *buffer) (int64, error) { - n, err := buf.ReadToWriter(w, count, dup) + read: func(view *buffer.View) (int64, error) { + n, err := view.ReadToWriter(w, count) + if !dup { + view.TrimFront(n) + } count -= n return n, err }, } - if dup { - // There is no notification for dup operations. - return p.dup(ctx, ops) - } n, err := p.read(ctx, ops) if n > 0 { p.Notify(waiter.EventOut) @@ -96,8 +97,8 @@ func (p *Pipe) Write(ctx context.Context, src usermem.IOSequence) (int64, error) limit: func(l int64) { src = src.TakeFirst64(l) }, - write: func(buf *buffer) (int64, error) { - n, err := src.CopyInTo(ctx, buf) + write: func(view *buffer.View) (int64, error) { + n, err := src.CopyInTo(ctx, view) src = src.DropFirst64(n) return n, err }, @@ -117,8 +118,8 @@ func (p *Pipe) ReadFrom(ctx context.Context, r io.Reader, count int64) (int64, e limit: func(l int64) { count = l }, - write: func(buf *buffer) (int64, error) { - n, err := buf.WriteFromReader(r, count) + write: func(view *buffer.View) (int64, error) { + n, err := view.WriteFromReader(r, count) count -= n return n, err }, diff --git a/pkg/sentry/kernel/pipe/vfs.go b/pkg/sentry/kernel/pipe/vfs.go index a5675bd70..b54f08a30 100644 --- a/pkg/sentry/kernel/pipe/vfs.go +++ b/pkg/sentry/kernel/pipe/vfs.go @@ -49,38 +49,42 @@ type VFSPipe struct { } // NewVFSPipe returns an initialized VFSPipe. -func NewVFSPipe(sizeBytes, atomicIOBytes int64) *VFSPipe { +func NewVFSPipe(isNamed bool, sizeBytes, atomicIOBytes int64) *VFSPipe { var vp VFSPipe - initPipe(&vp.pipe, true /* isNamed */, sizeBytes, atomicIOBytes) + initPipe(&vp.pipe, isNamed, sizeBytes, atomicIOBytes) return &vp } -// NewVFSPipeFD opens a named pipe. Named pipes have special blocking semantics -// during open: +// ReaderWriterPair returns read-only and write-only FDs for vp. // -// "Normally, opening the FIFO blocks until the other end is opened also. A -// process can open a FIFO in nonblocking mode. In this case, opening for -// read-only will succeed even if no-one has opened on the write side yet, -// opening for write-only will fail with ENXIO (no such device or address) -// unless the other end has already been opened. Under Linux, opening a FIFO -// for read and write will succeed both in blocking and nonblocking mode. POSIX -// leaves this behavior undefined. This can be used to open a FIFO for writing -// while there are no readers available." - fifo(7) -func (vp *VFSPipe) NewVFSPipeFD(ctx context.Context, vfsd *vfs.Dentry, vfsfd *vfs.FileDescription, flags uint32) (*VFSPipeFD, error) { +// Preconditions: statusFlags should not contain an open access mode. +func (vp *VFSPipe) ReaderWriterPair(mnt *vfs.Mount, vfsd *vfs.Dentry, statusFlags uint32) (*vfs.FileDescription, *vfs.FileDescription) { + return vp.newFD(mnt, vfsd, linux.O_RDONLY|statusFlags), vp.newFD(mnt, vfsd, linux.O_WRONLY|statusFlags) +} + +// Open opens the pipe represented by vp. +func (vp *VFSPipe) Open(ctx context.Context, mnt *vfs.Mount, vfsd *vfs.Dentry, statusFlags uint32) (*vfs.FileDescription, error) { vp.mu.Lock() defer vp.mu.Unlock() - readable := vfs.MayReadFileWithOpenFlags(flags) - writable := vfs.MayWriteFileWithOpenFlags(flags) + readable := vfs.MayReadFileWithOpenFlags(statusFlags) + writable := vfs.MayWriteFileWithOpenFlags(statusFlags) if !readable && !writable { return nil, syserror.EINVAL } - vfd, err := vp.open(vfsd, vfsfd, flags) - if err != nil { - return nil, err - } + fd := vp.newFD(mnt, vfsd, statusFlags) + // Named pipes have special blocking semantics during open: + // + // "Normally, opening the FIFO blocks until the other end is opened also. A + // process can open a FIFO in nonblocking mode. In this case, opening for + // read-only will succeed even if no-one has opened on the write side yet, + // opening for write-only will fail with ENXIO (no such device or address) + // unless the other end has already been opened. Under Linux, opening a + // FIFO for read and write will succeed both in blocking and nonblocking + // mode. POSIX leaves this behavior undefined. This can be used to open a + // FIFO for writing while there are no readers available." - fifo(7) switch { case readable && writable: // Pipes opened for read-write always succeed without blocking. @@ -89,23 +93,26 @@ func (vp *VFSPipe) NewVFSPipeFD(ctx context.Context, vfsd *vfs.Dentry, vfsfd *vf case readable: newHandleLocked(&vp.rWakeup) - // If this pipe is being opened as nonblocking and there's no + // If this pipe is being opened as blocking and there's no // writer, we have to wait for a writer to open the other end. - if flags&linux.O_NONBLOCK == 0 && !vp.pipe.HasWriters() && !waitFor(&vp.mu, &vp.wWakeup, ctx) { + if vp.pipe.isNamed && statusFlags&linux.O_NONBLOCK == 0 && !vp.pipe.HasWriters() && !waitFor(&vp.mu, &vp.wWakeup, ctx) { + fd.DecRef() return nil, syserror.EINTR } case writable: newHandleLocked(&vp.wWakeup) - if !vp.pipe.HasReaders() { - // Nonblocking, write-only opens fail with ENXIO when - // the read side isn't open yet. - if flags&linux.O_NONBLOCK != 0 { + if vp.pipe.isNamed && !vp.pipe.HasReaders() { + // Non-blocking, write-only opens fail with ENXIO when the read + // side isn't open yet. + if statusFlags&linux.O_NONBLOCK != 0 { + fd.DecRef() return nil, syserror.ENXIO } // Wait for a reader to open the other end. if !waitFor(&vp.mu, &vp.rWakeup, ctx) { + fd.DecRef() return nil, syserror.EINTR } } @@ -114,96 +121,93 @@ func (vp *VFSPipe) NewVFSPipeFD(ctx context.Context, vfsd *vfs.Dentry, vfsfd *vf panic("invalid pipe flags: must be readable, writable, or both") } - return vfd, nil + return fd, nil } // Preconditions: vp.mu must be held. -func (vp *VFSPipe) open(vfsd *vfs.Dentry, vfsfd *vfs.FileDescription, flags uint32) (*VFSPipeFD, error) { - var fd VFSPipeFD - fd.flags = flags - fd.readable = vfs.MayReadFileWithOpenFlags(flags) - fd.writable = vfs.MayWriteFileWithOpenFlags(flags) - fd.vfsfd = vfsfd - fd.pipe = &vp.pipe +func (vp *VFSPipe) newFD(mnt *vfs.Mount, vfsd *vfs.Dentry, statusFlags uint32) *vfs.FileDescription { + fd := &VFSPipeFD{ + pipe: &vp.pipe, + } + fd.vfsfd.Init(fd, statusFlags, mnt, vfsd, &vfs.FileDescriptionOptions{ + DenyPRead: true, + DenyPWrite: true, + UseDentryMetadata: true, + }) switch { - case fd.readable && fd.writable: + case fd.vfsfd.IsReadable() && fd.vfsfd.IsWritable(): vp.pipe.rOpen() vp.pipe.wOpen() - case fd.readable: + case fd.vfsfd.IsReadable(): vp.pipe.rOpen() - case fd.writable: + case fd.vfsfd.IsWritable(): vp.pipe.wOpen() default: panic("invalid pipe flags: must be readable, writable, or both") } - return &fd, nil + return &fd.vfsfd } -// VFSPipeFD implements a subset of vfs.FileDescriptionImpl for pipes. It is -// expected that filesystesm will use this in a struct implementing -// vfs.FileDescriptionImpl. +// VFSPipeFD implements vfs.FileDescriptionImpl for pipes. type VFSPipeFD struct { - pipe *Pipe - flags uint32 - readable bool - writable bool - vfsfd *vfs.FileDescription + vfsfd vfs.FileDescription + vfs.FileDescriptionDefaultImpl + vfs.DentryMetadataFileDescriptionImpl + + pipe *Pipe } // Release implements vfs.FileDescriptionImpl.Release. func (fd *VFSPipeFD) Release() { var event waiter.EventMask - if fd.readable { + if fd.vfsfd.IsReadable() { fd.pipe.rClose() - event |= waiter.EventIn + event |= waiter.EventOut } - if fd.writable { + if fd.vfsfd.IsWritable() { fd.pipe.wClose() - event |= waiter.EventOut + event |= waiter.EventIn | waiter.EventHUp } if event == 0 { panic("invalid pipe flags: must be readable, writable, or both") } - if fd.writable { - fd.vfsfd.VirtualDentry().Mount().EndWrite() - } - fd.pipe.Notify(event) } -// OnClose implements vfs.FileDescriptionImpl.OnClose. -func (fd *VFSPipeFD) OnClose(_ context.Context) error { - return nil +// Readiness implements waiter.Waitable.Readiness. +func (fd *VFSPipeFD) Readiness(mask waiter.EventMask) waiter.EventMask { + switch { + case fd.vfsfd.IsReadable() && fd.vfsfd.IsWritable(): + return fd.pipe.rwReadiness() + case fd.vfsfd.IsReadable(): + return fd.pipe.rReadiness() + case fd.vfsfd.IsWritable(): + return fd.pipe.wReadiness() + default: + panic("pipe FD is neither readable nor writable") + } } -// PRead implements vfs.FileDescriptionImpl.PRead. -func (fd *VFSPipeFD) PRead(_ context.Context, _ usermem.IOSequence, _ int64, _ vfs.ReadOptions) (int64, error) { - return 0, syserror.ESPIPE +// EventRegister implements waiter.Waitable.EventRegister. +func (fd *VFSPipeFD) EventRegister(e *waiter.Entry, mask waiter.EventMask) { + fd.pipe.EventRegister(e, mask) +} + +// EventUnregister implements waiter.Waitable.EventUnregister. +func (fd *VFSPipeFD) EventUnregister(e *waiter.Entry) { + fd.pipe.EventUnregister(e) } // Read implements vfs.FileDescriptionImpl.Read. func (fd *VFSPipeFD) Read(ctx context.Context, dst usermem.IOSequence, _ vfs.ReadOptions) (int64, error) { - if !fd.readable { - return 0, syserror.EINVAL - } - return fd.pipe.Read(ctx, dst) } -// PWrite implements vfs.FileDescriptionImpl.PWrite. -func (fd *VFSPipeFD) PWrite(_ context.Context, _ usermem.IOSequence, _ int64, _ vfs.WriteOptions) (int64, error) { - return 0, syserror.ESPIPE -} - // Write implements vfs.FileDescriptionImpl.Write. func (fd *VFSPipeFD) Write(ctx context.Context, src usermem.IOSequence, _ vfs.WriteOptions) (int64, error) { - if !fd.writable { - return 0, syserror.EINVAL - } - return fd.pipe.Write(ctx, src) } @@ -211,3 +215,17 @@ func (fd *VFSPipeFD) Write(ctx context.Context, src usermem.IOSequence, _ vfs.Wr func (fd *VFSPipeFD) Ioctl(ctx context.Context, uio usermem.IO, args arch.SyscallArguments) (uintptr, error) { return fd.pipe.Ioctl(ctx, uio, args) } + +// PipeSize implements fcntl(F_GETPIPE_SZ). +func (fd *VFSPipeFD) PipeSize() int64 { + // Inline Pipe.FifoSize() rather than calling it with nil Context and + // fs.File and ignoring the returned error (which is always nil). + fd.pipe.mu.Lock() + defer fd.pipe.mu.Unlock() + return fd.pipe.max +} + +// SetPipeSize implements fcntl(F_SETPIPE_SZ). +func (fd *VFSPipeFD) SetPipeSize(size int64) (int64, error) { + return fd.pipe.SetFifoSize(size) +} diff --git a/pkg/sentry/kernel/ptrace.go b/pkg/sentry/kernel/ptrace.go index 35ad97d5d..e23e796ef 100644 --- a/pkg/sentry/kernel/ptrace.go +++ b/pkg/sentry/kernel/ptrace.go @@ -184,7 +184,6 @@ func (t *Task) CanTrace(target *Task, attach bool) bool { if targetCreds.PermittedCaps&^callerCreds.PermittedCaps != 0 { return false } - // TODO: Yama LSM return true } diff --git a/pkg/sentry/kernel/rseq.go b/pkg/sentry/kernel/rseq.go index efebfd872..18416643b 100644 --- a/pkg/sentry/kernel/rseq.go +++ b/pkg/sentry/kernel/rseq.go @@ -303,26 +303,14 @@ func (t *Task) rseqAddrInterrupt() { return } - buf = t.CopyScratchBuffer(linux.SizeOfRSeqCriticalSection) - if _, err := t.CopyInBytes(critAddr, buf); err != nil { + var cs linux.RSeqCriticalSection + if _, err := cs.CopyIn(t, critAddr); err != nil { t.Debugf("Failed to copy critical section from %#x for rseq: %v", critAddr, err) t.forceSignal(linux.SIGSEGV, false /* unconditional */) t.SendSignal(SignalInfoPriv(linux.SIGSEGV)) return } - // Manually marshal RSeqCriticalSection as this is in the hot path when - // rseq is enabled. It must be as fast as possible. - // - // TODO(b/130243041): Replace with go_marshal. - cs := linux.RSeqCriticalSection{ - Version: usermem.ByteOrder.Uint32(buf[0:4]), - Flags: usermem.ByteOrder.Uint32(buf[4:8]), - Start: usermem.ByteOrder.Uint64(buf[8:16]), - PostCommitOffset: usermem.ByteOrder.Uint64(buf[16:24]), - Abort: usermem.ByteOrder.Uint64(buf[24:32]), - } - if cs.Version != 0 { t.Debugf("Unknown version in %+v", cs) t.forceSignal(linux.SIGSEGV, false /* unconditional */) diff --git a/pkg/sentry/kernel/semaphore/semaphore.go b/pkg/sentry/kernel/semaphore/semaphore.go index 1000f3287..c00fa1138 100644 --- a/pkg/sentry/kernel/semaphore/semaphore.go +++ b/pkg/sentry/kernel/semaphore/semaphore.go @@ -554,6 +554,7 @@ func (s *sem) wakeWaiters() { for w := s.waiters.Front(); w != nil; { if s.value < w.value { // Still blocked, skip it. + w = w.Next() continue } w.ch <- struct{}{} diff --git a/pkg/sentry/kernel/sessions.go b/pkg/sentry/kernel/sessions.go index 047b5214d..0e19286de 100644 --- a/pkg/sentry/kernel/sessions.go +++ b/pkg/sentry/kernel/sessions.go @@ -246,7 +246,7 @@ func (pg *ProcessGroup) SendSignal(info *arch.SignalInfo) error { var lastErr error for tg := range tasks.Root.tgids { - if tg.ProcessGroup() == pg { + if tg.processGroup == pg { tg.signalHandlers.mu.Lock() infoCopy := *info if err := tg.leader.sendSignalLocked(&infoCopy, true /*group*/); err != nil { diff --git a/pkg/sentry/kernel/shm/shm.go b/pkg/sentry/kernel/shm/shm.go index 208569057..f66cfcc7f 100644 --- a/pkg/sentry/kernel/shm/shm.go +++ b/pkg/sentry/kernel/shm/shm.go @@ -461,7 +461,7 @@ func (s *Shm) AddMapping(ctx context.Context, _ memmap.MappingSpace, _ usermem.A func (s *Shm) RemoveMapping(ctx context.Context, _ memmap.MappingSpace, _ usermem.AddrRange, _ uint64, _ bool) { s.mu.Lock() defer s.mu.Unlock() - // TODO(b/38173783): RemoveMapping may be called during task exit, when ctx + // RemoveMapping may be called during task exit, when ctx // is context.Background. Gracefully handle missing clocks. Failing to // update the detach time in these cases is ok, since no one can observe the // omission. diff --git a/pkg/sentry/kernel/syscalls.go b/pkg/sentry/kernel/syscalls.go index 93c4fe969..84156d5a1 100644 --- a/pkg/sentry/kernel/syscalls.go +++ b/pkg/sentry/kernel/syscalls.go @@ -209,65 +209,61 @@ type Stracer interface { // SyscallEnter is called on syscall entry. // // The returned private data is passed to SyscallExit. - // - // TODO(gvisor.dev/issue/155): remove kernel imports from the strace - // package so that the type can be used directly. SyscallEnter(t *Task, sysno uintptr, args arch.SyscallArguments, flags uint32) interface{} // SyscallExit is called on syscall exit. SyscallExit(context interface{}, t *Task, sysno, rval uintptr, err error) } -// SyscallTable is a lookup table of system calls. Critically, a SyscallTable -// is *immutable*. In order to make supporting suspend and resume sane, they -// must be uniquely registered and may not change during operation. +// SyscallTable is a lookup table of system calls. // -// +stateify savable +// Note that a SyscallTable is not savable directly. Instead, they are saved as +// an OS/Arch pair and lookup happens again on restore. type SyscallTable struct { // OS is the operating system that this syscall table implements. - OS abi.OS `state:"wait"` + OS abi.OS // Arch is the architecture that this syscall table targets. - Arch arch.Arch `state:"wait"` + Arch arch.Arch // The OS version that this syscall table implements. - Version Version `state:"manual"` + Version Version // AuditNumber is a numeric constant that represents the syscall table. If // non-zero, auditNumber must be one of the AUDIT_ARCH_* values defined by // linux/audit.h. - AuditNumber uint32 `state:"manual"` + AuditNumber uint32 // Table is the collection of functions. - Table map[uintptr]Syscall `state:"manual"` + Table map[uintptr]Syscall // lookup is a fixed-size array that holds the syscalls (indexed by // their numbers). It is used for fast look ups. - lookup []SyscallFn `state:"manual"` + lookup []SyscallFn // Emulate is a collection of instruction addresses to emulate. The // keys are addresses, and the values are system call numbers. - Emulate map[usermem.Addr]uintptr `state:"manual"` + Emulate map[usermem.Addr]uintptr // The function to call in case of a missing system call. - Missing MissingFn `state:"manual"` + Missing MissingFn // Stracer traces this syscall table. - Stracer Stracer `state:"manual"` + Stracer Stracer // External is used to handle an external callback. - External func(*Kernel) `state:"manual"` + External func(*Kernel) // ExternalFilterBefore is called before External is called before the syscall is executed. // External is not called if it returns false. - ExternalFilterBefore func(*Task, uintptr, arch.SyscallArguments) bool `state:"manual"` + ExternalFilterBefore func(*Task, uintptr, arch.SyscallArguments) bool // ExternalFilterAfter is called before External is called after the syscall is executed. // External is not called if it returns false. - ExternalFilterAfter func(*Task, uintptr, arch.SyscallArguments) bool `state:"manual"` + ExternalFilterAfter func(*Task, uintptr, arch.SyscallArguments) bool // FeatureEnable stores the strace and one-shot enable bits. - FeatureEnable SyscallFlagsTable `state:"manual"` + FeatureEnable SyscallFlagsTable } // allSyscallTables contains all known tables. @@ -330,6 +326,13 @@ func RegisterSyscallTable(s *SyscallTable) { allSyscallTables = append(allSyscallTables, s) } +// FlushSyscallTablesTestOnly flushes the syscall tables for tests. Used for +// parameterized VFSv2 tests. +// TODO(gvisor.dv/issue/1624): Remove when VFS1 is no longer supported. +func FlushSyscallTablesTestOnly() { + allSyscallTables = nil +} + // Lookup returns the syscall implementation, if one exists. func (s *SyscallTable) Lookup(sysno uintptr) SyscallFn { if sysno < uintptr(len(s.lookup)) { diff --git a/pkg/sentry/kernel/syscalls_state.go b/pkg/sentry/kernel/syscalls_state.go index 00358326b..90f890495 100644 --- a/pkg/sentry/kernel/syscalls_state.go +++ b/pkg/sentry/kernel/syscalls_state.go @@ -14,16 +14,34 @@ package kernel -import "fmt" +import ( + "fmt" -// afterLoad is invoked by stateify. -func (s *SyscallTable) afterLoad() { - otherTable, ok := LookupSyscallTable(s.OS, s.Arch) - if !ok { - // Couldn't find a reference? - panic(fmt.Sprintf("syscall table not found for OS %v Arch %v", s.OS, s.Arch)) + "gvisor.dev/gvisor/pkg/abi" + "gvisor.dev/gvisor/pkg/sentry/arch" +) + +// syscallTableInfo is used to reload the SyscallTable. +// +// +stateify savable +type syscallTableInfo struct { + OS abi.OS + Arch arch.Arch +} + +// saveSt saves the SyscallTable. +func (tc *TaskContext) saveSt() syscallTableInfo { + return syscallTableInfo{ + OS: tc.st.OS, + Arch: tc.st.Arch, } +} - // Copy the table. - *s = *otherTable +// loadSt loads the SyscallTable. +func (tc *TaskContext) loadSt(sti syscallTableInfo) { + st, ok := LookupSyscallTable(sti.OS, sti.Arch) + if !ok { + panic(fmt.Sprintf("syscall table not found for OS %v, Arch %v", sti.OS, sti.Arch)) + } + tc.st = st // Save the table reference. } diff --git a/pkg/sentry/kernel/task.go b/pkg/sentry/kernel/task.go index 981e8c7fe..e5d133d6c 100644 --- a/pkg/sentry/kernel/task.go +++ b/pkg/sentry/kernel/task.go @@ -37,6 +37,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/usage" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" "gvisor.dev/gvisor/pkg/waiter" ) @@ -424,6 +425,11 @@ type Task struct { // abstractSockets is protected by mu. abstractSockets *AbstractSocketNamespace + // mountNamespaceVFS2 is the task's mount namespace. + // + // It is protected by mu. It is owned by the task goroutine. + mountNamespaceVFS2 *vfs.MountNamespace + // parentDeathSignal is sent to this task's thread group when its parent exits. // // parentDeathSignal is protected by mu. @@ -481,13 +487,10 @@ type Task struct { numaPolicy int32 numaNodeMask uint64 - // If netns is true, the task is in a non-root network namespace. Network - // namespaces aren't currently implemented in full; being in a network - // namespace simply prevents the task from observing any network devices - // (including loopback) or using abstract socket addresses (see unix(7)). + // netns is the task's network namespace. netns is never nil. // - // netns is protected by mu. netns is owned by the task goroutine. - netns bool + // netns is protected by mu. + netns *inet.Namespace // If rseqPreempted is true, before the next call to p.Switch(), // interrupt rseq critical regions as defined by rseqAddr and @@ -638,6 +641,11 @@ func (t *Task) Value(key interface{}) interface{} { return int32(t.ThreadGroup().ID()) case fs.CtxRoot: return t.fsContext.RootDirectory() + case vfs.CtxRoot: + return t.fsContext.RootDirectoryVFS2() + case vfs.CtxMountNamespace: + t.mountNamespaceVFS2.IncRef() + return t.mountNamespaceVFS2 case fs.CtxDirentCacheLimiter: return t.k.DirentCacheLimiter case inet.CtxStack: @@ -701,6 +709,14 @@ func (t *Task) SyscallRestartBlock() SyscallRestartBlock { // Preconditions: The caller must be running on the task goroutine, or t.mu // must be locked. func (t *Task) IsChrooted() bool { + if VFS2Enabled { + realRoot := t.mountNamespaceVFS2.Root() + defer realRoot.DecRef() + root := t.fsContext.RootDirectoryVFS2() + defer root.DecRef() + return root != realRoot + } + realRoot := t.tg.mounts.Root() defer realRoot.DecRef() root := t.fsContext.RootDirectory() @@ -761,6 +777,15 @@ func (t *Task) NewFDs(fd int32, files []*fs.File, flags FDFlags) ([]int32, error return t.fdTable.NewFDs(t, fd, files, flags) } +// NewFDsVFS2 is a convenience wrapper for t.FDTable().NewFDsVFS2. +// +// This automatically passes the task as the context. +// +// Precondition: same as FDTable. +func (t *Task) NewFDsVFS2(fd int32, files []*vfs.FileDescription, flags FDFlags) ([]int32, error) { + return t.fdTable.NewFDsVFS2(t, fd, files, flags) +} + // NewFDFrom is a convenience wrapper for t.FDTable().NewFDs with a single file. // // This automatically passes the task as the context. @@ -774,6 +799,15 @@ func (t *Task) NewFDFrom(fd int32, file *fs.File, flags FDFlags) (int32, error) return fds[0], nil } +// NewFDFromVFS2 is a convenience wrapper for t.FDTable().NewFDVFS2. +// +// This automatically passes the task as the context. +// +// Precondition: same as FDTable.Get. +func (t *Task) NewFDFromVFS2(fd int32, file *vfs.FileDescription, flags FDFlags) (int32, error) { + return t.fdTable.NewFDVFS2(t, fd, file, flags) +} + // NewFDAt is a convenience wrapper for t.FDTable().NewFDAt. // // This automatically passes the task as the context. @@ -783,6 +817,15 @@ func (t *Task) NewFDAt(fd int32, file *fs.File, flags FDFlags) error { return t.fdTable.NewFDAt(t, fd, file, flags) } +// NewFDAtVFS2 is a convenience wrapper for t.FDTable().NewFDAtVFS2. +// +// This automatically passes the task as the context. +// +// Precondition: same as FDTable. +func (t *Task) NewFDAtVFS2(fd int32, file *vfs.FileDescription, flags FDFlags) error { + return t.fdTable.NewFDAtVFS2(t, fd, file, flags) +} + // WithMuLocked executes f with t.mu locked. func (t *Task) WithMuLocked(f func(*Task)) { t.mu.Lock() @@ -796,6 +839,15 @@ func (t *Task) MountNamespace() *fs.MountNamespace { return t.tg.mounts } +// MountNamespaceVFS2 returns t's MountNamespace. A reference is taken on the +// returned mount namespace. +func (t *Task) MountNamespaceVFS2() *vfs.MountNamespace { + t.mu.Lock() + defer t.mu.Unlock() + t.mountNamespaceVFS2.IncRef() + return t.mountNamespaceVFS2 +} + // AbstractSockets returns t's AbstractSocketNamespace. func (t *Task) AbstractSockets() *AbstractSocketNamespace { return t.abstractSockets @@ -805,3 +857,30 @@ func (t *Task) AbstractSockets() *AbstractSocketNamespace { func (t *Task) ContainerID() string { return t.containerID } + +// OOMScoreAdj gets the task's thread group's OOM score adjustment. +func (t *Task) OOMScoreAdj() int32 { + return atomic.LoadInt32(&t.tg.oomScoreAdj) +} + +// SetOOMScoreAdj sets the task's thread group's OOM score adjustment. The +// value should be between -1000 and 1000 inclusive. +func (t *Task) SetOOMScoreAdj(adj int32) error { + if adj > 1000 || adj < -1000 { + return syserror.EINVAL + } + atomic.StoreInt32(&t.tg.oomScoreAdj, adj) + return nil +} + +// UID returns t's uid. +// TODO(gvisor.dev/issue/170): This method is not namespaced yet. +func (t *Task) UID() uint32 { + return uint32(t.Credentials().EffectiveKUID) +} + +// GID returns t's gid. +// TODO(gvisor.dev/issue/170): This method is not namespaced yet. +func (t *Task) GID() uint32 { + return uint32(t.Credentials().EffectiveKGID) +} diff --git a/pkg/sentry/kernel/task_clone.go b/pkg/sentry/kernel/task_clone.go index 53d4d211b..e1ecca99e 100644 --- a/pkg/sentry/kernel/task_clone.go +++ b/pkg/sentry/kernel/task_clone.go @@ -15,8 +15,11 @@ package kernel import ( + "sync/atomic" + "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/bpf" + "gvisor.dev/gvisor/pkg/sentry/inet" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" ) @@ -54,8 +57,7 @@ type SharingOptions struct { NewUserNamespace bool // If NewNetworkNamespace is true, the task should have an independent - // network namespace. (Note that network namespaces are not really - // implemented; see comment on Task.netns for details.) + // network namespace. NewNetworkNamespace bool // If NewFiles is true, the task should use an independent file descriptor @@ -199,6 +201,17 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) { ipcns = NewIPCNamespace(userns) } + netns := t.NetworkNamespace() + if opts.NewNetworkNamespace { + netns = inet.NewNamespace(netns) + } + + // TODO(b/63601033): Implement CLONE_NEWNS. + mntnsVFS2 := t.mountNamespaceVFS2 + if mntnsVFS2 != nil { + mntnsVFS2.IncRef() + } + tc, err := t.tc.Fork(t, t.k, !opts.NewAddressSpace) if err != nil { return 0, nil, err @@ -241,12 +254,15 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) { rseqAddr := usermem.Addr(0) rseqSignature := uint32(0) if opts.NewThreadGroup { - tg.mounts.IncRef() + if tg.mounts != nil { + tg.mounts.IncRef() + } sh := t.tg.signalHandlers if opts.NewSignalHandlers { sh = sh.Fork() } tg = t.k.NewThreadGroup(tg.mounts, pidns, sh, opts.TerminationSignal, tg.limits.GetCopy()) + tg.oomScoreAdj = atomic.LoadInt32(&t.tg.oomScoreAdj) rseqAddr = t.rseqAddr rseqSignature = t.rseqSignature } @@ -260,11 +276,12 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) { FDTable: fdTable, Credentials: creds, Niceness: t.Niceness(), - NetworkNamespaced: t.netns, + NetworkNamespace: netns, AllowedCPUMask: t.CPUMask(), UTSNamespace: utsns, IPCNamespace: ipcns, AbstractSocketNamespace: t.abstractSockets, + MountNamespaceVFS2: mntnsVFS2, RSeqAddr: rseqAddr, RSeqSignature: rseqSignature, ContainerID: t.ContainerID(), @@ -274,9 +291,6 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) { } else { cfg.InheritParent = t } - if opts.NewNetworkNamespace { - cfg.NetworkNamespaced = true - } nt, err := t.tg.pidns.owner.NewTask(cfg) if err != nil { if opts.NewThreadGroup { @@ -473,7 +487,7 @@ func (t *Task) Unshare(opts *SharingOptions) error { t.mu.Unlock() return syserror.EPERM } - t.netns = true + t.netns = inet.NewNamespace(t.netns) } if opts.NewUTSNamespace { if !haveCapSysAdmin { diff --git a/pkg/sentry/kernel/task_context.go b/pkg/sentry/kernel/task_context.go index 2d6e7733c..9fa528384 100644 --- a/pkg/sentry/kernel/task_context.go +++ b/pkg/sentry/kernel/task_context.go @@ -49,7 +49,7 @@ type TaskContext struct { fu *futex.Manager // st is the task's syscall table. - st *SyscallTable + st *SyscallTable `state:".(syscallTableInfo)"` } // release releases all resources held by the TaskContext. release is called by @@ -58,7 +58,6 @@ func (tc *TaskContext) release() { // Nil out pointers so that if the task is saved after release, it doesn't // follow the pointers to possibly now-invalid objects. if tc.MemoryManager != nil { - // TODO(b/38173783) tc.MemoryManager.DecUsers(context.Background()) tc.MemoryManager = nil } @@ -136,11 +135,11 @@ func (t *Task) Stack() *arch.Stack { func (k *Kernel) LoadTaskImage(ctx context.Context, args loader.LoadArgs) (*TaskContext, *syserr.Error) { // If File is not nil, we should load that instead of resolving Filename. if args.File != nil { - args.Filename = args.File.MappedName(ctx) + args.Filename = args.File.PathnameWithDeleted(ctx) } // Prepare a new user address space to load into. - m := mm.NewMemoryManager(k, k) + m := mm.NewMemoryManager(k, k, k.SleepForAddressSpaceActivation) defer m.DecUsers(ctx) args.MemoryManager = m diff --git a/pkg/sentry/kernel/task_exec.go b/pkg/sentry/kernel/task_exec.go index 8f57a34a6..00c425cca 100644 --- a/pkg/sentry/kernel/task_exec.go +++ b/pkg/sentry/kernel/task_exec.go @@ -220,7 +220,7 @@ func (r *runSyscallAfterExecStop) execute(t *Task) taskRunState { t.mu.Unlock() t.unstopVforkParent() // NOTE(b/30316266): All locks must be dropped prior to calling Activate. - t.MemoryManager().Activate() + t.MemoryManager().Activate(t) t.ptraceExec(oldTID) return (*runSyscallExit)(nil) diff --git a/pkg/sentry/kernel/task_exit.go b/pkg/sentry/kernel/task_exit.go index 435761e5a..c4ade6e8e 100644 --- a/pkg/sentry/kernel/task_exit.go +++ b/pkg/sentry/kernel/task_exit.go @@ -269,6 +269,13 @@ func (*runExitMain) execute(t *Task) taskRunState { t.fsContext.DecRef() t.fdTable.DecRef() + t.mu.Lock() + if t.mountNamespaceVFS2 != nil { + t.mountNamespaceVFS2.DecRef() + t.mountNamespaceVFS2 = nil + } + t.mu.Unlock() + // If this is the last task to exit from the thread group, release the // thread group's resources. if lastExiter { diff --git a/pkg/sentry/kernel/task_identity.go b/pkg/sentry/kernel/task_identity.go index ce3e6ef28..0325967e4 100644 --- a/pkg/sentry/kernel/task_identity.go +++ b/pkg/sentry/kernel/task_identity.go @@ -455,7 +455,7 @@ func (t *Task) SetKeepCaps(k bool) { t.creds.Store(creds) } -// updateCredsForExec updates t.creds to reflect an execve(). +// updateCredsForExecLocked updates t.creds to reflect an execve(). // // NOTE(b/30815691): We currently do not implement privileged executables // (set-user/group-ID bits and file capabilities). This allows us to make a lot diff --git a/pkg/sentry/kernel/task_log.go b/pkg/sentry/kernel/task_log.go index 41259210c..eeccaa197 100644 --- a/pkg/sentry/kernel/task_log.go +++ b/pkg/sentry/kernel/task_log.go @@ -32,21 +32,21 @@ const ( // Infof logs an formatted info message by calling log.Infof. func (t *Task) Infof(fmt string, v ...interface{}) { if log.IsLogging(log.Info) { - log.Infof(t.logPrefix.Load().(string)+fmt, v...) + log.InfofAtDepth(1, t.logPrefix.Load().(string)+fmt, v...) } } // Warningf logs a warning string by calling log.Warningf. func (t *Task) Warningf(fmt string, v ...interface{}) { if log.IsLogging(log.Warning) { - log.Warningf(t.logPrefix.Load().(string)+fmt, v...) + log.WarningfAtDepth(1, t.logPrefix.Load().(string)+fmt, v...) } } // Debugf creates a debug string that includes the task ID. func (t *Task) Debugf(fmt string, v ...interface{}) { if log.IsLogging(log.Debug) { - log.Debugf(t.logPrefix.Load().(string)+fmt, v...) + log.DebugfAtDepth(1, t.logPrefix.Load().(string)+fmt, v...) } } @@ -198,18 +198,11 @@ func (t *Task) traceExecEvent(tc *TaskContext) { if !trace.IsEnabled() { return } - d := tc.MemoryManager.Executable() - if d == nil { + file := tc.MemoryManager.Executable() + if file == nil { trace.Logf(t.traceContext, traceCategory, "exec: << unknown >>") return } - defer d.DecRef() - root := t.fsContext.RootDirectory() - if root == nil { - trace.Logf(t.traceContext, traceCategory, "exec: << no root directory >>") - return - } - defer root.DecRef() - n, _ := d.FullName(root) - trace.Logf(t.traceContext, traceCategory, "exec: %s", n) + defer file.DecRef() + trace.Logf(t.traceContext, traceCategory, "exec: %s", file.PathnameWithDeleted(t)) } diff --git a/pkg/sentry/kernel/task_net.go b/pkg/sentry/kernel/task_net.go index 172a31e1d..f7711232c 100644 --- a/pkg/sentry/kernel/task_net.go +++ b/pkg/sentry/kernel/task_net.go @@ -22,14 +22,23 @@ import ( func (t *Task) IsNetworkNamespaced() bool { t.mu.Lock() defer t.mu.Unlock() - return t.netns + return !t.netns.IsRoot() } // NetworkContext returns the network stack used by the task. NetworkContext // may return nil if no network stack is available. +// +// TODO(gvisor.dev/issue/1833): Migrate callers of this method to +// NetworkNamespace(). func (t *Task) NetworkContext() inet.Stack { - if t.IsNetworkNamespaced() { - return nil - } - return t.k.networkStack + t.mu.Lock() + defer t.mu.Unlock() + return t.netns.Stack() +} + +// NetworkNamespace returns the network namespace observed by the task. +func (t *Task) NetworkNamespace() *inet.Namespace { + t.mu.Lock() + defer t.mu.Unlock() + return t.netns } diff --git a/pkg/sentry/kernel/task_run.go b/pkg/sentry/kernel/task_run.go index 5568c91bc..2ba8d7e63 100644 --- a/pkg/sentry/kernel/task_run.go +++ b/pkg/sentry/kernel/task_run.go @@ -126,13 +126,39 @@ func (t *Task) doStop() { } } +func (*runApp) handleCPUIDInstruction(t *Task) error { + if len(arch.CPUIDInstruction) == 0 { + // CPUID emulation isn't supported, but this code can be + // executed, because the ptrace platform returns + // ErrContextSignalCPUID on page faults too. Look at + // pkg/sentry/platform/ptrace/ptrace.go:context.Switch for more + // details. + return platform.ErrContextSignal + } + // Is this a CPUID instruction? + region := trace.StartRegion(t.traceContext, cpuidRegion) + expected := arch.CPUIDInstruction[:] + found := make([]byte, len(expected)) + _, err := t.CopyIn(usermem.Addr(t.Arch().IP()), &found) + if err == nil && bytes.Equal(expected, found) { + // Skip the cpuid instruction. + t.Arch().CPUIDEmulate(t) + t.Arch().SetIP(t.Arch().IP() + uintptr(len(expected))) + region.End() + + return nil + } + region.End() // Not an actual CPUID, but required copy-in. + return platform.ErrContextSignal +} + // The runApp state checks for interrupts before executing untrusted // application code. // // +stateify savable type runApp struct{} -func (*runApp) execute(t *Task) taskRunState { +func (app *runApp) execute(t *Task) taskRunState { if t.interrupted() { // Checkpointing instructs tasks to stop by sending an interrupt, so we // must check for stops before entering runInterrupt (instead of @@ -237,21 +263,10 @@ func (*runApp) execute(t *Task) taskRunState { return (*runApp)(nil) case platform.ErrContextSignalCPUID: - // Is this a CPUID instruction? - region := trace.StartRegion(t.traceContext, cpuidRegion) - expected := arch.CPUIDInstruction[:] - found := make([]byte, len(expected)) - _, err := t.CopyIn(usermem.Addr(t.Arch().IP()), &found) - if err == nil && bytes.Equal(expected, found) { - // Skip the cpuid instruction. - t.Arch().CPUIDEmulate(t) - t.Arch().SetIP(t.Arch().IP() + uintptr(len(expected))) - region.End() - + if err := app.handleCPUIDInstruction(t); err == nil { // Resume execution. return (*runApp)(nil) } - region.End() // Not an actual CPUID, but required copy-in. // The instruction at the given RIP was not a CPUID, and we // fallthrough to the default signal deliver behavior below. @@ -338,7 +353,7 @@ func (*runApp) execute(t *Task) taskRunState { default: // What happened? Can't continue. t.Warningf("Unexpected SwitchToApp error: %v", err) - t.PrepareExit(ExitStatus{Code: t.ExtractErrno(err, -1)}) + t.PrepareExit(ExitStatus{Code: ExtractErrno(err, -1)}) return (*runExit)(nil) } } diff --git a/pkg/sentry/kernel/task_signals.go b/pkg/sentry/kernel/task_signals.go index 8802db142..f07de2089 100644 --- a/pkg/sentry/kernel/task_signals.go +++ b/pkg/sentry/kernel/task_signals.go @@ -174,7 +174,7 @@ func (t *Task) deliverSignal(info *arch.SignalInfo, act arch.SignalAct) taskRunS fallthrough case (sre == ERESTARTSYS && !act.IsRestart()): t.Debugf("Not restarting syscall %d after errno %d: interrupted by signal %d", t.Arch().SyscallNo(), sre, info.Signo) - t.Arch().SetReturn(uintptr(-t.ExtractErrno(syserror.EINTR, -1))) + t.Arch().SetReturn(uintptr(-ExtractErrno(syserror.EINTR, -1))) default: t.Debugf("Restarting syscall %d after errno %d: interrupted by signal %d", t.Arch().SyscallNo(), sre, info.Signo) t.Arch().RestartSyscall() @@ -513,8 +513,6 @@ func (t *Task) canReceiveSignalLocked(sig linux.Signal) bool { if t.stop != nil { return false } - // - TODO(b/38173783): No special case for when t is also the sending task, - // because the identity of the sender is unknown. // - Do not choose tasks that have already been interrupted, as they may be // busy handling another signal. if len(t.interruptChan) != 0 { diff --git a/pkg/sentry/kernel/task_start.go b/pkg/sentry/kernel/task_start.go index de838beef..a5035bb7f 100644 --- a/pkg/sentry/kernel/task_start.go +++ b/pkg/sentry/kernel/task_start.go @@ -17,10 +17,12 @@ package kernel import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/inet" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/kernel/futex" "gvisor.dev/gvisor/pkg/sentry/kernel/sched" "gvisor.dev/gvisor/pkg/sentry/usage" + "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" ) @@ -64,9 +66,8 @@ type TaskConfig struct { // Niceness is the niceness of the new task. Niceness int - // If NetworkNamespaced is true, the new task should observe a non-root - // network namespace. - NetworkNamespaced bool + // NetworkNamespace is the network namespace to be used for the new task. + NetworkNamespace *inet.Namespace // AllowedCPUMask contains the cpus that this task can run on. AllowedCPUMask sched.CPUSet @@ -80,6 +81,9 @@ type TaskConfig struct { // AbstractSocketNamespace is the AbstractSocketNamespace of the new task. AbstractSocketNamespace *AbstractSocketNamespace + // MountNamespaceVFS2 is the MountNamespace of the new task. + MountNamespaceVFS2 *vfs.MountNamespace + // RSeqAddr is a pointer to the the userspace linux.RSeq structure. RSeqAddr usermem.Addr @@ -116,28 +120,29 @@ func (ts *TaskSet) newTask(cfg *TaskConfig) (*Task, error) { parent: cfg.Parent, children: make(map[*Task]struct{}), }, - runState: (*runApp)(nil), - interruptChan: make(chan struct{}, 1), - signalMask: cfg.SignalMask, - signalStack: arch.SignalStack{Flags: arch.SignalStackFlagDisable}, - tc: *tc, - fsContext: cfg.FSContext, - fdTable: cfg.FDTable, - p: cfg.Kernel.Platform.NewContext(), - k: cfg.Kernel, - ptraceTracees: make(map[*Task]struct{}), - allowedCPUMask: cfg.AllowedCPUMask.Copy(), - ioUsage: &usage.IO{}, - niceness: cfg.Niceness, - netns: cfg.NetworkNamespaced, - utsns: cfg.UTSNamespace, - ipcns: cfg.IPCNamespace, - abstractSockets: cfg.AbstractSocketNamespace, - rseqCPU: -1, - rseqAddr: cfg.RSeqAddr, - rseqSignature: cfg.RSeqSignature, - futexWaiter: futex.NewWaiter(), - containerID: cfg.ContainerID, + runState: (*runApp)(nil), + interruptChan: make(chan struct{}, 1), + signalMask: cfg.SignalMask, + signalStack: arch.SignalStack{Flags: arch.SignalStackFlagDisable}, + tc: *tc, + fsContext: cfg.FSContext, + fdTable: cfg.FDTable, + p: cfg.Kernel.Platform.NewContext(), + k: cfg.Kernel, + ptraceTracees: make(map[*Task]struct{}), + allowedCPUMask: cfg.AllowedCPUMask.Copy(), + ioUsage: &usage.IO{}, + niceness: cfg.Niceness, + netns: cfg.NetworkNamespace, + utsns: cfg.UTSNamespace, + ipcns: cfg.IPCNamespace, + abstractSockets: cfg.AbstractSocketNamespace, + mountNamespaceVFS2: cfg.MountNamespaceVFS2, + rseqCPU: -1, + rseqAddr: cfg.RSeqAddr, + rseqSignature: cfg.RSeqSignature, + futexWaiter: futex.NewWaiter(), + containerID: cfg.ContainerID, } t.creds.Store(cfg.Credentials) t.endStopCond.L = &t.tg.signalHandlers.mu diff --git a/pkg/sentry/kernel/task_syscall.go b/pkg/sentry/kernel/task_syscall.go index d555d69a8..c9db78e06 100644 --- a/pkg/sentry/kernel/task_syscall.go +++ b/pkg/sentry/kernel/task_syscall.go @@ -194,6 +194,19 @@ func (t *Task) executeSyscall(sysno uintptr, args arch.SyscallArguments) (rval u // // The syscall path is very hot; avoid defer. func (t *Task) doSyscall() taskRunState { + // Save value of the register which is clobbered in the following + // t.Arch().SetReturn(-ENOSYS) operation. This is dedicated to arm64. + // + // On x86, register rax was shared by syscall number and return + // value, and at the entry of the syscall handler, the rax was + // saved to regs.orig_rax which was exposed to user space. + // But on arm64, syscall number was passed through X8, and the X0 + // was shared by the first syscall argument and return value. The + // X0 was saved to regs.orig_x0 which was not exposed to user space. + // So we have to do the same operation here to save the X0 value + // into the task context. + t.Arch().SyscallSaveOrig() + sysno := t.Arch().SyscallNo() args := t.Arch().SyscallArgs() @@ -269,6 +282,7 @@ func (*runSyscallAfterSyscallEnterStop) execute(t *Task) taskRunState { return (*runSyscallExit)(nil) } args := t.Arch().SyscallArgs() + return t.doSyscallInvoke(sysno, args) } @@ -298,7 +312,7 @@ func (t *Task) doSyscallInvoke(sysno uintptr, args arch.SyscallArguments) taskRu return ctrl.next } } else if err != nil { - t.Arch().SetReturn(uintptr(-t.ExtractErrno(err, int(sysno)))) + t.Arch().SetReturn(uintptr(-ExtractErrno(err, int(sysno)))) t.haveSyscallReturn = true } else { t.Arch().SetReturn(rval) @@ -417,7 +431,7 @@ func (t *Task) doVsyscallInvoke(sysno uintptr, args arch.SyscallArguments, calle // A return is not emulated in this case. return (*runApp)(nil) } - t.Arch().SetReturn(uintptr(-t.ExtractErrno(err, int(sysno)))) + t.Arch().SetReturn(uintptr(-ExtractErrno(err, int(sysno)))) } t.Arch().SetIP(t.Arch().Value(caller)) t.Arch().SetStack(t.Arch().Stack() + uintptr(t.Arch().Width())) @@ -427,7 +441,7 @@ func (t *Task) doVsyscallInvoke(sysno uintptr, args arch.SyscallArguments, calle // ExtractErrno extracts an integer error number from the error. // The syscall number is purely for context in the error case. Use -1 if // syscall number is unknown. -func (t *Task) ExtractErrno(err error, sysno int) int { +func ExtractErrno(err error, sysno int) int { switch err := err.(type) { case nil: return 0 @@ -441,11 +455,11 @@ func (t *Task) ExtractErrno(err error, sysno int) int { // handled (and the SIGBUS is delivered). return int(syscall.EFAULT) case *os.PathError: - return t.ExtractErrno(err.Err, sysno) + return ExtractErrno(err.Err, sysno) case *os.LinkError: - return t.ExtractErrno(err.Err, sysno) + return ExtractErrno(err.Err, sysno) case *os.SyscallError: - return t.ExtractErrno(err.Err, sysno) + return ExtractErrno(err.Err, sysno) default: if errno, ok := syserror.TranslateError(err); ok { return int(errno) diff --git a/pkg/sentry/kernel/task_usermem.go b/pkg/sentry/kernel/task_usermem.go index 2bf3ce8a8..b02044ad2 100644 --- a/pkg/sentry/kernel/task_usermem.go +++ b/pkg/sentry/kernel/task_usermem.go @@ -30,7 +30,7 @@ var MAX_RW_COUNT = int(usermem.Addr(math.MaxInt32).RoundDown()) // Activate ensures that the task has an active address space. func (t *Task) Activate() { if mm := t.MemoryManager(); mm != nil { - if err := mm.Activate(); err != nil { + if err := mm.Activate(t); err != nil { panic("unable to activate mm: " + err.Error()) } } diff --git a/pkg/sentry/kernel/thread_group.go b/pkg/sentry/kernel/thread_group.go index 768e958d2..52849f5b3 100644 --- a/pkg/sentry/kernel/thread_group.go +++ b/pkg/sentry/kernel/thread_group.go @@ -254,9 +254,16 @@ type ThreadGroup struct { // // tty is protected by the signal mutex. tty *TTY + + // oomScoreAdj is the thread group's OOM score adjustment. This is + // currently not used but is maintained for consistency. + // TODO(gvisor.dev/issue/1967) + // + // oomScoreAdj is accessed using atomic memory operations. + oomScoreAdj int32 } -// NewThreadGroup returns a new, empty thread group in PID namespace ns. The +// NewThreadGroup returns a new, empty thread group in PID namespace pidns. The // thread group leader will send its parent terminationSignal when it exits. // The new thread group isn't visible to the system until a task has been // created inside of it by a successful call to TaskSet.NewTask. @@ -317,7 +324,9 @@ func (tg *ThreadGroup) release() { for _, it := range its { it.DestroyTimer() } - tg.mounts.DecRef() + if tg.mounts != nil { + tg.mounts.DecRef() + } } // forEachChildThreadGroupLocked indicates over all child ThreadGroups. diff --git a/pkg/sentry/kernel/time/time.go b/pkg/sentry/kernel/time/time.go index 706de83ef..e959700f2 100644 --- a/pkg/sentry/kernel/time/time.go +++ b/pkg/sentry/kernel/time/time.go @@ -245,7 +245,7 @@ type Clock interface { type WallRateClock struct{} // WallTimeUntil implements Clock.WallTimeUntil. -func (WallRateClock) WallTimeUntil(t, now Time) time.Duration { +func (*WallRateClock) WallTimeUntil(t, now Time) time.Duration { return t.Sub(now) } @@ -254,16 +254,16 @@ func (WallRateClock) WallTimeUntil(t, now Time) time.Duration { type NoClockEvents struct{} // Readiness implements waiter.Waitable.Readiness. -func (NoClockEvents) Readiness(mask waiter.EventMask) waiter.EventMask { +func (*NoClockEvents) Readiness(mask waiter.EventMask) waiter.EventMask { return 0 } // EventRegister implements waiter.Waitable.EventRegister. -func (NoClockEvents) EventRegister(e *waiter.Entry, mask waiter.EventMask) { +func (*NoClockEvents) EventRegister(e *waiter.Entry, mask waiter.EventMask) { } // EventUnregister implements waiter.Waitable.EventUnregister. -func (NoClockEvents) EventUnregister(e *waiter.Entry) { +func (*NoClockEvents) EventUnregister(e *waiter.Entry) { } // ClockEventsQueue implements waiter.Waitable by wrapping waiter.Queue and @@ -273,7 +273,7 @@ type ClockEventsQueue struct { } // Readiness implements waiter.Waitable.Readiness. -func (ClockEventsQueue) Readiness(mask waiter.EventMask) waiter.EventMask { +func (*ClockEventsQueue) Readiness(mask waiter.EventMask) waiter.EventMask { return 0 } diff --git a/pkg/sentry/loader/BUILD b/pkg/sentry/loader/BUILD index 23790378a..c6aa65f28 100644 --- a/pkg/sentry/loader/BUILD +++ b/pkg/sentry/loader/BUILD @@ -33,6 +33,7 @@ go_library( "//pkg/sentry/fs", "//pkg/sentry/fs/anon", "//pkg/sentry/fs/fsutil", + "//pkg/sentry/fsbridge", "//pkg/sentry/kernel/auth", "//pkg/sentry/limits", "//pkg/sentry/memmap", @@ -40,6 +41,7 @@ go_library( "//pkg/sentry/pgalloc", "//pkg/sentry/uniqueid", "//pkg/sentry/usage", + "//pkg/sentry/vfs", "//pkg/syserr", "//pkg/syserror", "//pkg/usermem", diff --git a/pkg/sentry/loader/elf.go b/pkg/sentry/loader/elf.go index 122ed05c2..616fafa2c 100644 --- a/pkg/sentry/loader/elf.go +++ b/pkg/sentry/loader/elf.go @@ -27,7 +27,7 @@ import ( "gvisor.dev/gvisor/pkg/cpuid" "gvisor.dev/gvisor/pkg/log" "gvisor.dev/gvisor/pkg/sentry/arch" - "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" "gvisor.dev/gvisor/pkg/sentry/limits" "gvisor.dev/gvisor/pkg/sentry/memmap" "gvisor.dev/gvisor/pkg/sentry/mm" @@ -97,11 +97,11 @@ type elfInfo struct { // accepts from the ELF, and it doesn't parse unnecessary parts of the file. // // ctx may be nil if f does not need it. -func parseHeader(ctx context.Context, f *fs.File) (elfInfo, error) { +func parseHeader(ctx context.Context, f fsbridge.File) (elfInfo, error) { // Check ident first; it will tell us the endianness of the rest of the // structs. var ident [elf.EI_NIDENT]byte - _, err := readFull(ctx, f, usermem.BytesIOSequence(ident[:]), 0) + _, err := f.ReadFull(ctx, usermem.BytesIOSequence(ident[:]), 0) if err != nil { log.Infof("Error reading ELF ident: %v", err) // The entire ident array always exists. @@ -137,7 +137,7 @@ func parseHeader(ctx context.Context, f *fs.File) (elfInfo, error) { var hdr elf.Header64 hdrBuf := make([]byte, header64Size) - _, err = readFull(ctx, f, usermem.BytesIOSequence(hdrBuf), 0) + _, err = f.ReadFull(ctx, usermem.BytesIOSequence(hdrBuf), 0) if err != nil { log.Infof("Error reading ELF header: %v", err) // The entire header always exists. @@ -187,7 +187,7 @@ func parseHeader(ctx context.Context, f *fs.File) (elfInfo, error) { } phdrBuf := make([]byte, totalPhdrSize) - _, err = readFull(ctx, f, usermem.BytesIOSequence(phdrBuf), int64(hdr.Phoff)) + _, err = f.ReadFull(ctx, usermem.BytesIOSequence(phdrBuf), int64(hdr.Phoff)) if err != nil { log.Infof("Error reading ELF phdrs: %v", err) // If phdrs were specified, they should all exist. @@ -227,7 +227,7 @@ func parseHeader(ctx context.Context, f *fs.File) (elfInfo, error) { // mapSegment maps a phdr into the Task. offset is the offset to apply to // phdr.Vaddr. -func mapSegment(ctx context.Context, m *mm.MemoryManager, f *fs.File, phdr *elf.ProgHeader, offset usermem.Addr) error { +func mapSegment(ctx context.Context, m *mm.MemoryManager, f fsbridge.File, phdr *elf.ProgHeader, offset usermem.Addr) error { // We must make a page-aligned mapping. adjust := usermem.Addr(phdr.Vaddr).PageOffset() @@ -395,7 +395,7 @@ type loadedELF struct { // // Preconditions: // * f is an ELF file -func loadParsedELF(ctx context.Context, m *mm.MemoryManager, f *fs.File, info elfInfo, sharedLoadOffset usermem.Addr) (loadedELF, error) { +func loadParsedELF(ctx context.Context, m *mm.MemoryManager, f fsbridge.File, info elfInfo, sharedLoadOffset usermem.Addr) (loadedELF, error) { first := true var start, end usermem.Addr var interpreter string @@ -431,7 +431,7 @@ func loadParsedELF(ctx context.Context, m *mm.MemoryManager, f *fs.File, info el } path := make([]byte, phdr.Filesz) - _, err := readFull(ctx, f, usermem.BytesIOSequence(path), int64(phdr.Off)) + _, err := f.ReadFull(ctx, usermem.BytesIOSequence(path), int64(phdr.Off)) if err != nil { // If an interpreter was specified, it should exist. ctx.Infof("Error reading PT_INTERP path: %v", err) @@ -564,7 +564,7 @@ func loadParsedELF(ctx context.Context, m *mm.MemoryManager, f *fs.File, info el // Preconditions: // * f is an ELF file // * f is the first ELF loaded into m -func loadInitialELF(ctx context.Context, m *mm.MemoryManager, fs *cpuid.FeatureSet, f *fs.File) (loadedELF, arch.Context, error) { +func loadInitialELF(ctx context.Context, m *mm.MemoryManager, fs *cpuid.FeatureSet, f fsbridge.File) (loadedELF, arch.Context, error) { info, err := parseHeader(ctx, f) if err != nil { ctx.Infof("Failed to parse initial ELF: %v", err) @@ -602,7 +602,7 @@ func loadInitialELF(ctx context.Context, m *mm.MemoryManager, fs *cpuid.FeatureS // // Preconditions: // * f is an ELF file -func loadInterpreterELF(ctx context.Context, m *mm.MemoryManager, f *fs.File, initial loadedELF) (loadedELF, error) { +func loadInterpreterELF(ctx context.Context, m *mm.MemoryManager, f fsbridge.File, initial loadedELF) (loadedELF, error) { info, err := parseHeader(ctx, f) if err != nil { if err == syserror.ENOEXEC { @@ -649,16 +649,14 @@ func loadELF(ctx context.Context, args LoadArgs) (loadedELF, arch.Context, error // Refresh the traversal limit. *args.RemainingTraversals = linux.MaxSymlinkTraversals args.Filename = bin.interpreter - d, i, err := openPath(ctx, args) + intFile, err := openPath(ctx, args) if err != nil { ctx.Infof("Error opening interpreter %s: %v", bin.interpreter, err) return loadedELF{}, nil, err } - defer i.DecRef() - // We don't need the Dirent. - d.DecRef() + defer intFile.DecRef() - interp, err = loadInterpreterELF(ctx, args.MemoryManager, i, bin) + interp, err = loadInterpreterELF(ctx, args.MemoryManager, intFile, bin) if err != nil { ctx.Infof("Error loading interpreter: %v", err) return loadedELF{}, nil, err diff --git a/pkg/sentry/loader/interpreter.go b/pkg/sentry/loader/interpreter.go index 098a45d36..3886b4d33 100644 --- a/pkg/sentry/loader/interpreter.go +++ b/pkg/sentry/loader/interpreter.go @@ -19,7 +19,7 @@ import ( "io" "gvisor.dev/gvisor/pkg/context" - "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" ) @@ -37,9 +37,9 @@ const ( ) // parseInterpreterScript returns the interpreter path and argv. -func parseInterpreterScript(ctx context.Context, filename string, f *fs.File, argv []string) (newpath string, newargv []string, err error) { +func parseInterpreterScript(ctx context.Context, filename string, f fsbridge.File, argv []string) (newpath string, newargv []string, err error) { line := make([]byte, interpMaxLineLength) - n, err := readFull(ctx, f, usermem.BytesIOSequence(line), 0) + n, err := f.ReadFull(ctx, usermem.BytesIOSequence(line), 0) // Short read is OK. if err != nil && err != io.ErrUnexpectedEOF { if err == io.EOF { diff --git a/pkg/sentry/loader/loader.go b/pkg/sentry/loader/loader.go index 57422706d..88449fe95 100644 --- a/pkg/sentry/loader/loader.go +++ b/pkg/sentry/loader/loader.go @@ -20,7 +20,6 @@ import ( "fmt" "io" "path" - "strings" "gvisor.dev/gvisor/pkg/abi" "gvisor.dev/gvisor/pkg/abi/linux" @@ -29,8 +28,10 @@ import ( "gvisor.dev/gvisor/pkg/rand" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/mm" + "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserr" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" @@ -41,16 +42,6 @@ type LoadArgs struct { // MemoryManager is the memory manager to load the executable into. MemoryManager *mm.MemoryManager - // Mounts is the mount namespace in which to look up Filename. - Mounts *fs.MountNamespace - - // Root is the root directory under which to look up Filename. - Root *fs.Dirent - - // WorkingDirectory is the working directory under which to look up - // Filename. - WorkingDirectory *fs.Dirent - // RemainingTraversals is the maximum number of symlinks to follow to // resolve Filename. This counter is passed by reference to keep it // updated throughout the call stack. @@ -65,7 +56,12 @@ type LoadArgs struct { // File is an open fs.File object of the executable. If File is not // nil, then File will be loaded and Filename will be ignored. - File *fs.File + // + // The caller is responsible for checking that the user can execute this file. + File fsbridge.File + + // Opener is used to open the executable file when 'File' is nil. + Opener fsbridge.Lookup // CloseOnExec indicates that the executable (or one of its parent // directories) was opened with O_CLOEXEC. If the executable is an @@ -106,103 +102,32 @@ func readFull(ctx context.Context, f *fs.File, dst usermem.IOSequence, offset in // installed in the Task FDTable. The caller takes ownership of both. // // args.Filename must be a readable, executable, regular file. -func openPath(ctx context.Context, args LoadArgs) (*fs.Dirent, *fs.File, error) { +func openPath(ctx context.Context, args LoadArgs) (fsbridge.File, error) { if args.Filename == "" { ctx.Infof("cannot open empty name") - return nil, nil, syserror.ENOENT - } - - var d *fs.Dirent - var err error - if args.ResolveFinal { - d, err = args.Mounts.FindInode(ctx, args.Root, args.WorkingDirectory, args.Filename, args.RemainingTraversals) - } else { - d, err = args.Mounts.FindLink(ctx, args.Root, args.WorkingDirectory, args.Filename, args.RemainingTraversals) - } - if err != nil { - return nil, nil, err - } - // Defer a DecRef for the sake of failure cases. - defer d.DecRef() - - if !args.ResolveFinal && fs.IsSymlink(d.Inode.StableAttr) { - return nil, nil, syserror.ELOOP - } - - if err := checkPermission(ctx, d); err != nil { - return nil, nil, err - } - - // If they claim it's a directory, then make sure. - // - // N.B. we reject directories below, but we must first reject - // non-directories passed as directories. - if strings.HasSuffix(args.Filename, "/") && !fs.IsDir(d.Inode.StableAttr) { - return nil, nil, syserror.ENOTDIR - } - - if err := checkIsRegularFile(ctx, d, args.Filename); err != nil { - return nil, nil, err - } - - f, err := d.Inode.GetFile(ctx, d, fs.FileFlags{Read: true}) - if err != nil { - return nil, nil, err - } - // Defer a DecRef for the sake of failure cases. - defer f.DecRef() - - if err := checkPread(ctx, f, args.Filename); err != nil { - return nil, nil, err - } - - d.IncRef() - f.IncRef() - return d, f, err -} - -// checkFile performs checks on a file to be executed. -func checkFile(ctx context.Context, f *fs.File, filename string) error { - if err := checkPermission(ctx, f.Dirent); err != nil { - return err - } - - if err := checkIsRegularFile(ctx, f.Dirent, filename); err != nil { - return err + return nil, syserror.ENOENT } - return checkPread(ctx, f, filename) -} - -// checkPermission checks whether the file is readable and executable. -func checkPermission(ctx context.Context, d *fs.Dirent) error { - perms := fs.PermMask{ - // TODO(gvisor.dev/issue/160): Linux requires only execute - // permission, not read. However, our backing filesystems may - // prevent us from reading the file without read permission. - // - // Additionally, a task with a non-readable executable has - // additional constraints on access via ptrace and procfs. - Read: true, - Execute: true, + // TODO(gvisor.dev/issue/160): Linux requires only execute permission, + // not read. However, our backing filesystems may prevent us from reading + // the file without read permission. Additionally, a task with a + // non-readable executable has additional constraints on access via + // ptrace and procfs. + opts := vfs.OpenOptions{ + Flags: linux.O_RDONLY, + FileExec: true, } - return d.Inode.CheckPermission(ctx, perms) + return args.Opener.OpenPath(ctx, args.Filename, opts, args.RemainingTraversals, args.ResolveFinal) } // checkIsRegularFile prevents us from trying to execute a directory, pipe, etc. -func checkIsRegularFile(ctx context.Context, d *fs.Dirent, filename string) error { - attr := d.Inode.StableAttr - if !fs.IsRegular(attr) { - ctx.Infof("%s is not regular: %v", filename, attr) - return syserror.EACCES +func checkIsRegularFile(ctx context.Context, file fsbridge.File, filename string) error { + t, err := file.Type(ctx) + if err != nil { + return err } - return nil -} - -// checkPread checks whether we can read the file at arbitrary offsets. -func checkPread(ctx context.Context, f *fs.File, filename string) error { - if !f.Flags().Pread { - ctx.Infof("%s cannot be read at an offset: %+v", filename, f.Flags()) + if t != linux.ModeRegular { + ctx.Infof("%q is not a regular file: %v", filename, t) return syserror.EACCES } return nil @@ -224,8 +149,10 @@ const ( maxLoaderAttempts = 6 ) -// loadExecutable loads an executable that is pointed to by args.File. If nil, -// the path args.Filename is resolved and loaded. If the executable is an +// loadExecutable loads an executable that is pointed to by args.File. The +// caller is responsible for checking that the user can execute this file. +// If nil, the path args.Filename is resolved and loaded (check that the user +// can execute this file is done here in this case). If the executable is an // interpreter script rather than an ELF, the binary of the corresponding // interpreter will be loaded. // @@ -234,37 +161,27 @@ const ( // * arch.Context matching the binary arch // * fs.Dirent of the binary file // * Possibly updated args.Argv -func loadExecutable(ctx context.Context, args LoadArgs) (loadedELF, arch.Context, *fs.Dirent, []string, error) { +func loadExecutable(ctx context.Context, args LoadArgs) (loadedELF, arch.Context, fsbridge.File, []string, error) { for i := 0; i < maxLoaderAttempts; i++ { - var ( - d *fs.Dirent - err error - ) if args.File == nil { - d, args.File, err = openPath(ctx, args) - // We will return d in the successful case, but defer a DecRef for the - // sake of intermediate loops and failure cases. - if d != nil { - defer d.DecRef() - } - if args.File != nil { - defer args.File.DecRef() + var err error + args.File, err = openPath(ctx, args) + if err != nil { + ctx.Infof("Error opening %s: %v", args.Filename, err) + return loadedELF{}, nil, nil, nil, err } + // Ensure file is release in case the code loops or errors out. + defer args.File.DecRef() } else { - d = args.File.Dirent - d.IncRef() - defer d.DecRef() - err = checkFile(ctx, args.File, args.Filename) - } - if err != nil { - ctx.Infof("Error opening %s: %v", args.Filename, err) - return loadedELF{}, nil, nil, nil, err + if err := checkIsRegularFile(ctx, args.File, args.Filename); err != nil { + return loadedELF{}, nil, nil, nil, err + } } // Check the header. Is this an ELF or interpreter script? var hdr [4]uint8 // N.B. We assume that reading from a regular file cannot block. - _, err = readFull(ctx, args.File, usermem.BytesIOSequence(hdr[:]), 0) + _, err := args.File.ReadFull(ctx, usermem.BytesIOSequence(hdr[:]), 0) // Allow unexpected EOF, as a valid executable could be only three bytes // (e.g., #!a). if err != nil && err != io.ErrUnexpectedEOF { @@ -281,9 +198,10 @@ func loadExecutable(ctx context.Context, args LoadArgs) (loadedELF, arch.Context ctx.Infof("Error loading ELF: %v", err) return loadedELF{}, nil, nil, nil, err } - // An ELF is always terminal. Hold on to d. - d.IncRef() - return loaded, ac, d, args.Argv, err + // An ELF is always terminal. Hold on to file. + args.File.IncRef() + return loaded, ac, args.File, args.Argv, err + case bytes.Equal(hdr[:2], []byte(interpreterScriptMagic)): if args.CloseOnExec { return loadedELF{}, nil, nil, nil, syserror.ENOENT @@ -295,6 +213,7 @@ func loadExecutable(ctx context.Context, args LoadArgs) (loadedELF, arch.Context } // Refresh the traversal limit for the interpreter. *args.RemainingTraversals = linux.MaxSymlinkTraversals + default: ctx.Infof("Unknown magic: %v", hdr) return loadedELF{}, nil, nil, nil, syserror.ENOEXEC @@ -317,11 +236,11 @@ func loadExecutable(ctx context.Context, args LoadArgs) (loadedELF, arch.Context // * Load is called on the Task goroutine. func Load(ctx context.Context, args LoadArgs, extraAuxv []arch.AuxEntry, vdso *VDSO) (abi.OS, arch.Context, string, *syserr.Error) { // Load the executable itself. - loaded, ac, d, newArgv, err := loadExecutable(ctx, args) + loaded, ac, file, newArgv, err := loadExecutable(ctx, args) if err != nil { return 0, nil, "", syserr.NewDynamic(fmt.Sprintf("Failed to load %s: %v", args.Filename, err), syserr.FromError(err).ToLinux()) } - defer d.DecRef() + defer file.DecRef() // Load the VDSO. vdsoAddr, err := loadVDSO(ctx, args.MemoryManager, vdso, loaded) @@ -390,7 +309,7 @@ func Load(ctx context.Context, args LoadArgs, extraAuxv []arch.AuxEntry, vdso *V m.SetEnvvStart(sl.EnvvStart) m.SetEnvvEnd(sl.EnvvEnd) m.SetAuxv(auxv) - m.SetExecutable(d) + m.SetExecutable(file) symbolValue, err := getSymbolValueFromVDSO("rt_sigreturn") if err != nil { diff --git a/pkg/sentry/loader/vdso.go b/pkg/sentry/loader/vdso.go index 01eaefd26..00977fc08 100644 --- a/pkg/sentry/loader/vdso.go +++ b/pkg/sentry/loader/vdso.go @@ -29,6 +29,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fs/anon" "gvisor.dev/gvisor/pkg/sentry/fs/fsutil" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" "gvisor.dev/gvisor/pkg/sentry/memmap" "gvisor.dev/gvisor/pkg/sentry/mm" "gvisor.dev/gvisor/pkg/sentry/pgalloc" @@ -73,6 +74,8 @@ type byteReader struct { var _ fs.FileOperations = (*byteReader)(nil) // newByteReaderFile creates a fake file to read data from. +// +// TODO(gvisor.dev/issue/1623): Convert to VFS2. func newByteReaderFile(ctx context.Context, data []byte) *fs.File { // Create a fake inode. inode := fs.NewInode( @@ -127,7 +130,7 @@ func (b *byteReader) Write(ctx context.Context, file *fs.File, src usermem.IOSeq // * PT_LOAD segments don't extend beyond the end of the file. // // ctx may be nil if f does not need it. -func validateVDSO(ctx context.Context, f *fs.File, size uint64) (elfInfo, error) { +func validateVDSO(ctx context.Context, f fsbridge.File, size uint64) (elfInfo, error) { info, err := parseHeader(ctx, f) if err != nil { log.Infof("Unable to parse VDSO header: %v", err) @@ -246,7 +249,7 @@ func getSymbolValueFromVDSO(symbol string) (uint64, error) { // PrepareVDSO validates the system VDSO and returns a VDSO, containing the // param page for updating by the kernel. func PrepareVDSO(ctx context.Context, mfp pgalloc.MemoryFileProvider) (*VDSO, error) { - vdsoFile := newByteReaderFile(ctx, vdsoBin) + vdsoFile := fsbridge.NewFSFile(newByteReaderFile(ctx, vdsoBin)) // First make sure the VDSO is valid. vdsoFile does not use ctx, so a // nil context can be passed. diff --git a/pkg/sentry/mm/BUILD b/pkg/sentry/mm/BUILD index e5729ced5..73591dab7 100644 --- a/pkg/sentry/mm/BUILD +++ b/pkg/sentry/mm/BUILD @@ -105,8 +105,8 @@ go_library( "//pkg/safecopy", "//pkg/safemem", "//pkg/sentry/arch", - "//pkg/sentry/fs", "//pkg/sentry/fs/proc/seqfile", + "//pkg/sentry/fsbridge", "//pkg/sentry/kernel/auth", "//pkg/sentry/kernel/futex", "//pkg/sentry/kernel/shm", diff --git a/pkg/sentry/mm/address_space.go b/pkg/sentry/mm/address_space.go index e58a63deb..5c667117c 100644 --- a/pkg/sentry/mm/address_space.go +++ b/pkg/sentry/mm/address_space.go @@ -18,7 +18,7 @@ import ( "fmt" "sync/atomic" - "gvisor.dev/gvisor/pkg/atomicbitops" + "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/platform" "gvisor.dev/gvisor/pkg/usermem" ) @@ -39,11 +39,18 @@ func (mm *MemoryManager) AddressSpace() platform.AddressSpace { // // When this MemoryManager is no longer needed by a task, it should call // Deactivate to release the reference. -func (mm *MemoryManager) Activate() error { +func (mm *MemoryManager) Activate(ctx context.Context) error { // Fast path: the MemoryManager already has an active // platform.AddressSpace, and we just need to indicate that we need it too. - if atomicbitops.IncUnlessZeroInt32(&mm.active) { - return nil + for { + active := atomic.LoadInt32(&mm.active) + if active == 0 { + // Fall back to the slow path. + break + } + if atomic.CompareAndSwapInt32(&mm.active, active, active+1) { + return nil + } } for { @@ -85,16 +92,20 @@ func (mm *MemoryManager) Activate() error { if as == nil { // AddressSpace is unavailable, we must wait. // - // activeMu must not be held while waiting, as the user - // of the address space we are waiting on may attempt - // to take activeMu. - // - // Don't call UninterruptibleSleepStart to register the - // wait to allow the watchdog stuck task to trigger in - // case a process is starved waiting for the address - // space. + // activeMu must not be held while waiting, as the user of the address + // space we are waiting on may attempt to take activeMu. mm.activeMu.Unlock() + + sleep := mm.p.CooperativelySchedulesAddressSpace() && mm.sleepForActivation + if sleep { + // Mark this task sleeping while waiting for the address space to + // prevent the watchdog from reporting it as a stuck task. + ctx.UninterruptibleSleepStart(false) + } <-c + if sleep { + ctx.UninterruptibleSleepFinish(false) + } continue } @@ -118,8 +129,15 @@ func (mm *MemoryManager) Activate() error { func (mm *MemoryManager) Deactivate() { // Fast path: this is not the last goroutine to deactivate the // MemoryManager. - if atomicbitops.DecUnlessOneInt32(&mm.active) { - return + for { + active := atomic.LoadInt32(&mm.active) + if active == 1 { + // Fall back to the slow path. + break + } + if atomic.CompareAndSwapInt32(&mm.active, active, active-1) { + return + } } mm.activeMu.Lock() @@ -183,8 +201,10 @@ func (mm *MemoryManager) mapASLocked(pseg pmaIterator, ar usermem.AddrRange, pre if pma.needCOW { perms.Write = false } - if err := mm.as.MapFile(pmaMapAR.Start, pma.file, pseg.fileRangeOf(pmaMapAR), perms, precommit); err != nil { - return err + if perms.Any() { // MapFile precondition + if err := mm.as.MapFile(pmaMapAR.Start, pma.file, pseg.fileRangeOf(pmaMapAR), perms, precommit); err != nil { + return err + } } pseg = pseg.NextSegment() } diff --git a/pkg/sentry/mm/aio_context.go b/pkg/sentry/mm/aio_context.go index cb29d94b0..379148903 100644 --- a/pkg/sentry/mm/aio_context.go +++ b/pkg/sentry/mm/aio_context.go @@ -59,25 +59,27 @@ func (a *aioManager) newAIOContext(events uint32, id uint64) bool { } a.contexts[id] = &AIOContext{ - done: make(chan struct{}, 1), + requestReady: make(chan struct{}, 1), maxOutstanding: events, } return true } -// destroyAIOContext destroys an asynchronous I/O context. +// destroyAIOContext destroys an asynchronous I/O context. It doesn't wait for +// for pending requests to complete. Returns the destroyed AIOContext so it can +// be drained. // -// False is returned if the context does not exist. -func (a *aioManager) destroyAIOContext(id uint64) bool { +// Nil is returned if the context does not exist. +func (a *aioManager) destroyAIOContext(id uint64) *AIOContext { a.mu.Lock() defer a.mu.Unlock() ctx, ok := a.contexts[id] if !ok { - return false + return nil } delete(a.contexts, id) ctx.destroy() - return true + return ctx } // lookupAIOContext looks up the given context. @@ -102,8 +104,8 @@ type ioResult struct { // // +stateify savable type AIOContext struct { - // done is the notification channel used for all requests. - done chan struct{} `state:"nosave"` + // requestReady is the notification channel used for all requests. + requestReady chan struct{} `state:"nosave"` // mu protects below. mu sync.Mutex `state:"nosave"` @@ -129,8 +131,14 @@ func (ctx *AIOContext) destroy() { ctx.mu.Lock() defer ctx.mu.Unlock() ctx.dead = true - if ctx.outstanding == 0 { - close(ctx.done) + ctx.checkForDone() +} + +// Preconditions: ctx.mu must be held by caller. +func (ctx *AIOContext) checkForDone() { + if ctx.dead && ctx.outstanding == 0 { + close(ctx.requestReady) + ctx.requestReady = nil } } @@ -154,11 +162,12 @@ func (ctx *AIOContext) PopRequest() (interface{}, bool) { // Is there anything ready? if e := ctx.results.Front(); e != nil { - ctx.results.Remove(e) - ctx.outstanding-- - if ctx.outstanding == 0 && ctx.dead { - close(ctx.done) + if ctx.outstanding == 0 { + panic("AIOContext outstanding is going negative") } + ctx.outstanding-- + ctx.results.Remove(e) + ctx.checkForDone() return e.data, true } return nil, false @@ -172,26 +181,58 @@ func (ctx *AIOContext) FinishRequest(data interface{}) { // Push to the list and notify opportunistically. The channel notify // here is guaranteed to be safe because outstanding must be non-zero. - // The done channel is only closed when outstanding reaches zero. + // The requestReady channel is only closed when outstanding reaches zero. ctx.results.PushBack(&ioResult{data: data}) select { - case ctx.done <- struct{}{}: + case ctx.requestReady <- struct{}{}: default: } } // WaitChannel returns a channel that is notified when an AIO request is -// completed. -// -// The boolean return value indicates whether or not the context is active. -func (ctx *AIOContext) WaitChannel() (chan struct{}, bool) { +// completed. Returns nil if the context is destroyed and there are no more +// outstanding requests. +func (ctx *AIOContext) WaitChannel() chan struct{} { ctx.mu.Lock() defer ctx.mu.Unlock() - if ctx.outstanding == 0 && ctx.dead { - return nil, false + return ctx.requestReady +} + +// Dead returns true if the context has been destroyed. +func (ctx *AIOContext) Dead() bool { + ctx.mu.Lock() + defer ctx.mu.Unlock() + return ctx.dead +} + +// CancelPendingRequest forgets about a request that hasn't yet completed. +func (ctx *AIOContext) CancelPendingRequest() { + ctx.mu.Lock() + defer ctx.mu.Unlock() + + if ctx.outstanding == 0 { + panic("AIOContext outstanding is going negative") } - return ctx.done, true + ctx.outstanding-- + ctx.checkForDone() +} + +// Drain drops all completed requests. Pending requests remain untouched. +func (ctx *AIOContext) Drain() { + ctx.mu.Lock() + defer ctx.mu.Unlock() + + if ctx.outstanding == 0 { + return + } + size := uint32(ctx.results.Len()) + if ctx.outstanding < size { + panic("AIOContext outstanding is going negative") + } + ctx.outstanding -= size + ctx.results.Reset() + ctx.checkForDone() } // aioMappable implements memmap.MappingIdentity and memmap.Mappable for AIO @@ -332,9 +373,9 @@ func (mm *MemoryManager) NewAIOContext(ctx context.Context, events uint32) (uint Length: aioRingBufferSize, MappingIdentity: m, Mappable: m, - // TODO(fvoznika): Linux does "do_mmap_pgoff(..., PROT_READ | - // PROT_WRITE, ...)" in fs/aio.c:aio_setup_ring(); why do we make this - // mapping read-only? + // Linux uses "do_mmap_pgoff(..., PROT_READ | PROT_WRITE, ...)" in + // fs/aio.c:aio_setup_ring(). Since we don't implement AIO_RING_MAGIC, + // user mode should not write to this page. Perms: usermem.Read, MaxPerms: usermem.Read, }) @@ -349,11 +390,11 @@ func (mm *MemoryManager) NewAIOContext(ctx context.Context, events uint32) (uint return id, nil } -// DestroyAIOContext destroys an asynchronous I/O context. It returns false if -// the context does not exist. -func (mm *MemoryManager) DestroyAIOContext(ctx context.Context, id uint64) bool { +// DestroyAIOContext destroys an asynchronous I/O context. It returns the +// destroyed context. nil if the context does not exist. +func (mm *MemoryManager) DestroyAIOContext(ctx context.Context, id uint64) *AIOContext { if _, ok := mm.LookupAIOContext(ctx, id); !ok { - return false + return nil } // Only unmaps after it assured that the address is a valid aio context to diff --git a/pkg/sentry/mm/aio_context_state.go b/pkg/sentry/mm/aio_context_state.go index c37fc9f7b..3dabac1af 100644 --- a/pkg/sentry/mm/aio_context_state.go +++ b/pkg/sentry/mm/aio_context_state.go @@ -16,5 +16,5 @@ package mm // afterLoad is invoked by stateify. func (a *AIOContext) afterLoad() { - a.done = make(chan struct{}, 1) + a.requestReady = make(chan struct{}, 1) } diff --git a/pkg/sentry/mm/lifecycle.go b/pkg/sentry/mm/lifecycle.go index 47b8fbf43..aac56679b 100644 --- a/pkg/sentry/mm/lifecycle.go +++ b/pkg/sentry/mm/lifecycle.go @@ -18,7 +18,6 @@ import ( "fmt" "sync/atomic" - "gvisor.dev/gvisor/pkg/atomicbitops" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/limits" @@ -29,16 +28,17 @@ import ( ) // NewMemoryManager returns a new MemoryManager with no mappings and 1 user. -func NewMemoryManager(p platform.Platform, mfp pgalloc.MemoryFileProvider) *MemoryManager { +func NewMemoryManager(p platform.Platform, mfp pgalloc.MemoryFileProvider, sleepForActivation bool) *MemoryManager { return &MemoryManager{ - p: p, - mfp: mfp, - haveASIO: p.SupportsAddressSpaceIO(), - privateRefs: &privateRefs{}, - users: 1, - auxv: arch.Auxv{}, - dumpability: UserDumpable, - aioManager: aioManager{contexts: make(map[uint64]*AIOContext)}, + p: p, + mfp: mfp, + haveASIO: p.SupportsAddressSpaceIO(), + privateRefs: &privateRefs{}, + users: 1, + auxv: arch.Auxv{}, + dumpability: UserDumpable, + aioManager: aioManager{contexts: make(map[uint64]*AIOContext)}, + sleepForActivation: sleepForActivation, } } @@ -80,9 +80,11 @@ func (mm *MemoryManager) Fork(ctx context.Context) (*MemoryManager, error) { envv: mm.envv, auxv: append(arch.Auxv(nil), mm.auxv...), // IncRef'd below, once we know that there isn't an error. - executable: mm.executable, - dumpability: mm.dumpability, - aioManager: aioManager{contexts: make(map[uint64]*AIOContext)}, + executable: mm.executable, + dumpability: mm.dumpability, + aioManager: aioManager{contexts: make(map[uint64]*AIOContext)}, + sleepForActivation: mm.sleepForActivation, + vdsoSigReturnAddr: mm.vdsoSigReturnAddr, } // Copy vmas. @@ -229,7 +231,15 @@ func (mm *MemoryManager) Fork(ctx context.Context) (*MemoryManager, error) { // IncUsers increments mm's user count and returns true. If the user count is // already 0, IncUsers does nothing and returns false. func (mm *MemoryManager) IncUsers() bool { - return atomicbitops.IncUnlessZeroInt32(&mm.users) + for { + users := atomic.LoadInt32(&mm.users) + if users == 0 { + return false + } + if atomic.CompareAndSwapInt32(&mm.users, users, users+1) { + return true + } + } } // DecUsers decrements mm's user count. If the user count reaches 0, all diff --git a/pkg/sentry/mm/metadata.go b/pkg/sentry/mm/metadata.go index f550acae0..28e5057f7 100644 --- a/pkg/sentry/mm/metadata.go +++ b/pkg/sentry/mm/metadata.go @@ -16,7 +16,7 @@ package mm import ( "gvisor.dev/gvisor/pkg/sentry/arch" - "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" "gvisor.dev/gvisor/pkg/usermem" ) @@ -132,7 +132,7 @@ func (mm *MemoryManager) SetAuxv(auxv arch.Auxv) { // // An additional reference will be taken in the case of a non-nil executable, // which must be released by the caller. -func (mm *MemoryManager) Executable() *fs.Dirent { +func (mm *MemoryManager) Executable() fsbridge.File { mm.metadataMu.Lock() defer mm.metadataMu.Unlock() @@ -147,15 +147,15 @@ func (mm *MemoryManager) Executable() *fs.Dirent { // SetExecutable sets the executable. // // This takes a reference on d. -func (mm *MemoryManager) SetExecutable(d *fs.Dirent) { +func (mm *MemoryManager) SetExecutable(file fsbridge.File) { mm.metadataMu.Lock() // Grab a new reference. - d.IncRef() + file.IncRef() // Set the executable. orig := mm.executable - mm.executable = d + mm.executable = file mm.metadataMu.Unlock() @@ -167,3 +167,17 @@ func (mm *MemoryManager) SetExecutable(d *fs.Dirent) { orig.DecRef() } } + +// VDSOSigReturn returns the address of vdso_sigreturn. +func (mm *MemoryManager) VDSOSigReturn() uint64 { + mm.metadataMu.Lock() + defer mm.metadataMu.Unlock() + return mm.vdsoSigReturnAddr +} + +// SetVDSOSigReturn sets the address of vdso_sigreturn. +func (mm *MemoryManager) SetVDSOSigReturn(addr uint64) { + mm.metadataMu.Lock() + defer mm.metadataMu.Unlock() + mm.vdsoSigReturnAddr = addr +} diff --git a/pkg/sentry/mm/mm.go b/pkg/sentry/mm/mm.go index 09e582dd3..34d3bde7a 100644 --- a/pkg/sentry/mm/mm.go +++ b/pkg/sentry/mm/mm.go @@ -37,7 +37,7 @@ package mm import ( "gvisor.dev/gvisor/pkg/safemem" "gvisor.dev/gvisor/pkg/sentry/arch" - "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" "gvisor.dev/gvisor/pkg/sentry/memmap" "gvisor.dev/gvisor/pkg/sentry/pgalloc" "gvisor.dev/gvisor/pkg/sentry/platform" @@ -215,7 +215,7 @@ type MemoryManager struct { // is not nil, it holds a reference on the Dirent. // // executable is protected by metadataMu. - executable *fs.Dirent + executable fsbridge.File // dumpability describes if and how this MemoryManager may be dumped to // userspace. @@ -226,6 +226,14 @@ type MemoryManager struct { // aioManager keeps track of AIOContexts used for async IOs. AIOManager // must be cloned when CLONE_VM is used. aioManager aioManager + + // sleepForActivation indicates whether the task should report to be sleeping + // before trying to activate the address space. When set to true, delays in + // activation are not reported as stuck tasks by the watchdog. + sleepForActivation bool + + // vdsoSigReturnAddr is the address of 'vdso_sigreturn'. + vdsoSigReturnAddr uint64 } // vma represents a virtual memory area. diff --git a/pkg/sentry/mm/mm_test.go b/pkg/sentry/mm/mm_test.go index edacca741..fdc308542 100644 --- a/pkg/sentry/mm/mm_test.go +++ b/pkg/sentry/mm/mm_test.go @@ -31,7 +31,7 @@ import ( func testMemoryManager(ctx context.Context) *MemoryManager { p := platform.FromContext(ctx) mfp := pgalloc.MemoryFileProviderFromContext(ctx) - mm := NewMemoryManager(p, mfp) + mm := NewMemoryManager(p, mfp, false) mm.layout = arch.MmapLayout{ MinAddr: p.MinUserAddress(), MaxAddr: p.MaxUserAddress(), diff --git a/pkg/sentry/platform/kvm/bluepill.go b/pkg/sentry/platform/kvm/bluepill.go index 35cd55fef..4b23f7803 100644 --- a/pkg/sentry/platform/kvm/bluepill.go +++ b/pkg/sentry/platform/kvm/bluepill.go @@ -81,12 +81,6 @@ func (c *vCPU) die(context *arch.SignalContext64, msg string) { // Save the death message, which will be thrown. c.dieState.message = msg - // Reload all registers to have an accurate stack trace when we return - // to host mode. This means that the stack should be unwound correctly. - if errno := c.getUserRegisters(&c.dieState.guestRegs); errno != 0 { - throw(msg) - } - // Setup the trampoline. dieArchSetup(c, context, &c.dieState.guestRegs) } diff --git a/pkg/sentry/platform/kvm/bluepill_amd64_unsafe.go b/pkg/sentry/platform/kvm/bluepill_amd64_unsafe.go index a63a6a071..99cac665d 100644 --- a/pkg/sentry/platform/kvm/bluepill_amd64_unsafe.go +++ b/pkg/sentry/platform/kvm/bluepill_amd64_unsafe.go @@ -31,6 +31,12 @@ import ( // //go:nosplit func dieArchSetup(c *vCPU, context *arch.SignalContext64, guestRegs *userRegs) { + // Reload all registers to have an accurate stack trace when we return + // to host mode. This means that the stack should be unwound correctly. + if errno := c.getUserRegisters(&c.dieState.guestRegs); errno != 0 { + throw(c.dieState.message) + } + // If the vCPU is in user mode, we set the stack to the stored stack // value in the vCPU itself. We don't want to unwind the user stack. if guestRegs.RFLAGS&ring0.UserFlagsSet == ring0.UserFlagsSet { diff --git a/pkg/sentry/platform/kvm/bluepill_arm64.s b/pkg/sentry/platform/kvm/bluepill_arm64.s index c61700892..04efa0147 100644 --- a/pkg/sentry/platform/kvm/bluepill_arm64.s +++ b/pkg/sentry/platform/kvm/bluepill_arm64.s @@ -82,6 +82,8 @@ fallback: // dieTrampoline: see bluepill.go, bluepill_arm64_unsafe.go for documentation. TEXT ·dieTrampoline(SB),NOSPLIT,$0 - // TODO(gvisor.dev/issue/1249): dieTrampoline supporting for Arm64. - MOVD R9, 8(RSP) - BL ·dieHandler(SB) + // R0: Fake the old PC as caller + // R1: First argument (vCPU) + MOVD.P R1, 8(RSP) // R1: First argument (vCPU) + MOVD.P R0, 8(RSP) // R0: Fake the old PC as caller + B ·dieHandler(SB) diff --git a/pkg/sentry/platform/kvm/bluepill_arm64_unsafe.go b/pkg/sentry/platform/kvm/bluepill_arm64_unsafe.go index 2f02c03cf..eb5ed574e 100644 --- a/pkg/sentry/platform/kvm/bluepill_arm64_unsafe.go +++ b/pkg/sentry/platform/kvm/bluepill_arm64_unsafe.go @@ -17,10 +17,33 @@ package kvm import ( + "unsafe" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/platform/ring0" ) +// dieArchSetup initialies the state for dieTrampoline. +// +// The arm64 dieTrampoline requires the vCPU to be set in R1, and the last PC +// to be in R0. The trampoline then simulates a call to dieHandler from the +// provided PC. +// //go:nosplit func dieArchSetup(c *vCPU, context *arch.SignalContext64, guestRegs *userRegs) { - // TODO(gvisor.dev/issue/1249): dieTrampoline supporting for Arm64. + // If the vCPU is in user mode, we set the stack to the stored stack + // value in the vCPU itself. We don't want to unwind the user stack. + if guestRegs.Regs.Pstate&ring0.PSR_MODE_MASK == ring0.PSR_MODE_EL0t { + regs := c.CPU.Registers() + context.Regs[0] = regs.Regs[0] + context.Sp = regs.Sp + context.Regs[29] = regs.Regs[29] // stack base address + } else { + context.Regs[0] = guestRegs.Regs.Pc + context.Sp = guestRegs.Regs.Sp + context.Regs[29] = guestRegs.Regs.Regs[29] + context.Pstate = guestRegs.Regs.Pstate + } + context.Regs[1] = uint64(uintptr(unsafe.Pointer(c))) + context.Pc = uint64(dieTrampolineAddr) } diff --git a/pkg/sentry/platform/kvm/kvm.go b/pkg/sentry/platform/kvm/kvm.go index 972ba85c3..a9b4af43e 100644 --- a/pkg/sentry/platform/kvm/kvm.go +++ b/pkg/sentry/platform/kvm/kvm.go @@ -27,6 +27,38 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) +// userMemoryRegion is a region of physical memory. +// +// This mirrors kvm_memory_region. +type userMemoryRegion struct { + slot uint32 + flags uint32 + guestPhysAddr uint64 + memorySize uint64 + userspaceAddr uint64 +} + +// runData is the run structure. This may be mapped for synchronous register +// access (although that doesn't appear to be supported by my kernel at least). +// +// This mirrors kvm_run. +type runData struct { + requestInterruptWindow uint8 + _ [7]uint8 + + exitReason uint32 + readyForInterruptInjection uint8 + ifFlag uint8 + _ [2]uint8 + + cr8 uint64 + apicBase uint64 + + // This is the union data for exits. Interpretation depends entirely on + // the exitReason above (see vCPU code for more information). + data [32]uint64 +} + // KVM represents a lightweight VM context. type KVM struct { platform.NoCPUPreemptionDetection diff --git a/pkg/sentry/platform/kvm/kvm_amd64.go b/pkg/sentry/platform/kvm/kvm_amd64.go index c5a6f9c7d..093497bc4 100644 --- a/pkg/sentry/platform/kvm/kvm_amd64.go +++ b/pkg/sentry/platform/kvm/kvm_amd64.go @@ -21,17 +21,6 @@ import ( "gvisor.dev/gvisor/pkg/sentry/platform/ring0" ) -// userMemoryRegion is a region of physical memory. -// -// This mirrors kvm_memory_region. -type userMemoryRegion struct { - slot uint32 - flags uint32 - guestPhysAddr uint64 - memorySize uint64 - userspaceAddr uint64 -} - // userRegs represents KVM user registers. // // This mirrors kvm_regs. @@ -169,27 +158,6 @@ type modelControlRegisters struct { entries [16]modelControlRegister } -// runData is the run structure. This may be mapped for synchronous register -// access (although that doesn't appear to be supported by my kernel at least). -// -// This mirrors kvm_run. -type runData struct { - requestInterruptWindow uint8 - _ [7]uint8 - - exitReason uint32 - readyForInterruptInjection uint8 - ifFlag uint8 - _ [2]uint8 - - cr8 uint64 - apicBase uint64 - - // This is the union data for exits. Interpretation depends entirely on - // the exitReason above (see vCPU code for more information). - data [32]uint64 -} - // cpuidEntry is a single CPUID entry. // // This mirrors kvm_cpuid_entry2. diff --git a/pkg/sentry/platform/kvm/kvm_arm64.go b/pkg/sentry/platform/kvm/kvm_arm64.go index 2319c86d3..716198712 100644 --- a/pkg/sentry/platform/kvm/kvm_arm64.go +++ b/pkg/sentry/platform/kvm/kvm_arm64.go @@ -18,18 +18,9 @@ package kvm import ( "syscall" -) -// userMemoryRegion is a region of physical memory. -// -// This mirrors kvm_memory_region. -type userMemoryRegion struct { - slot uint32 - flags uint32 - guestPhysAddr uint64 - memorySize uint64 - userspaceAddr uint64 -} + "gvisor.dev/gvisor/pkg/sentry/platform/ring0" +) type kvmOneReg struct { id uint64 @@ -53,31 +44,10 @@ type userRegs struct { fpRegs userFpsimdState } -// runData is the run structure. This may be mapped for synchronous register -// access (although that doesn't appear to be supported by my kernel at least). -// -// This mirrors kvm_run. -type runData struct { - requestInterruptWindow uint8 - _ [7]uint8 - - exitReason uint32 - readyForInterruptInjection uint8 - ifFlag uint8 - _ [2]uint8 - - cr8 uint64 - apicBase uint64 - - // This is the union data for exits. Interpretation depends entirely on - // the exitReason above (see vCPU code for more information). - data [32]uint64 -} - // updateGlobalOnce does global initialization. It has to be called only once. func updateGlobalOnce(fd int) error { physicalInit() err := updateSystemValues(int(fd)) - updateVectorTable() + ring0.Init() return err } diff --git a/pkg/sentry/platform/kvm/machine.go b/pkg/sentry/platform/kvm/machine.go index 8076c7529..f1afc74dc 100644 --- a/pkg/sentry/platform/kvm/machine.go +++ b/pkg/sentry/platform/kvm/machine.go @@ -329,10 +329,12 @@ func (m *machine) Destroy() { } // Get gets an available vCPU. +// +// This will return with the OS thread locked. func (m *machine) Get() *vCPU { + m.mu.RLock() runtime.LockOSThread() tid := procid.Current() - m.mu.RLock() // Check for an exact match. if c := m.vCPUs[tid]; c != nil { @@ -343,8 +345,22 @@ func (m *machine) Get() *vCPU { // The happy path failed. We now proceed to acquire an exclusive lock // (because the vCPU map may change), and scan all available vCPUs. + // In this case, we first unlock the OS thread. Otherwise, if mu is + // not available, the current system thread will be parked and a new + // system thread spawned. We avoid this situation by simply refreshing + // tid after relocking the system thread. m.mu.RUnlock() + runtime.UnlockOSThread() m.mu.Lock() + runtime.LockOSThread() + tid = procid.Current() + + // Recheck for an exact match. + if c := m.vCPUs[tid]; c != nil { + c.lock() + m.mu.Unlock() + return c + } for { // Scan for an available vCPU. diff --git a/pkg/sentry/platform/kvm/machine_arm64_unsafe.go b/pkg/sentry/platform/kvm/machine_arm64_unsafe.go index 1c8384e6b..3b35858ae 100644 --- a/pkg/sentry/platform/kvm/machine_arm64_unsafe.go +++ b/pkg/sentry/platform/kvm/machine_arm64_unsafe.go @@ -29,30 +29,6 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) -// setMemoryRegion initializes a region. -// -// This may be called from bluepillHandler, and therefore returns an errno -// directly (instead of wrapping in an error) to avoid allocations. -// -//go:nosplit -func (m *machine) setMemoryRegion(slot int, physical, length, virtual uintptr) syscall.Errno { - userRegion := userMemoryRegion{ - slot: uint32(slot), - flags: 0, - guestPhysAddr: uint64(physical), - memorySize: uint64(length), - userspaceAddr: uint64(virtual), - } - - // Set the region. - _, _, errno := syscall.RawSyscall( - syscall.SYS_IOCTL, - uintptr(m.fd), - _KVM_SET_USER_MEMORY_REGION, - uintptr(unsafe.Pointer(&userRegion))) - return errno -} - type kvmVcpuInit struct { target uint32 features [7]uint32 @@ -72,69 +48,6 @@ func (m *machine) initArchState() error { return nil } -func getPageWithReflect(p uintptr) []byte { - return (*(*[0xFFFFFF]byte)(unsafe.Pointer(p & ^uintptr(syscall.Getpagesize()-1))))[:syscall.Getpagesize()] -} - -// Work around: move ring0.Vectors() into a specific address with 11-bits alignment. -// -// According to the design documentation of Arm64, -// the start address of exception vector table should be 11-bits aligned. -// Please see the code in linux kernel as reference: arch/arm64/kernel/entry.S -// But, we can't align a function's start address to a specific address by using golang. -// We have raised this question in golang community: -// https://groups.google.com/forum/m/#!topic/golang-dev/RPj90l5x86I -// This function will be removed when golang supports this feature. -// -// There are 2 jobs were implemented in this function: -// 1, move the start address of exception vector table into the specific address. -// 2, modify the offset of each instruction. -func updateVectorTable() { - fromLocation := reflect.ValueOf(ring0.Vectors).Pointer() - offset := fromLocation & (1<<11 - 1) - if offset != 0 { - offset = 1<<11 - offset - } - - toLocation := fromLocation + offset - page := getPageWithReflect(toLocation) - if err := syscall.Mprotect(page, syscall.PROT_READ|syscall.PROT_WRITE|syscall.PROT_EXEC); err != nil { - panic(err) - } - - page = getPageWithReflect(toLocation + 4096) - if err := syscall.Mprotect(page, syscall.PROT_READ|syscall.PROT_WRITE|syscall.PROT_EXEC); err != nil { - panic(err) - } - - // Move exception-vector-table into the specific address. - var entry *uint32 - var entryFrom *uint32 - for i := 1; i <= 0x800; i++ { - entry = (*uint32)(unsafe.Pointer(toLocation + 0x800 - uintptr(i))) - entryFrom = (*uint32)(unsafe.Pointer(fromLocation + 0x800 - uintptr(i))) - *entry = *entryFrom - } - - // The offset from the address of each unconditionally branch is changed. - // We should modify the offset of each instruction. - nums := []uint32{0x0, 0x80, 0x100, 0x180, 0x200, 0x280, 0x300, 0x380, 0x400, 0x480, 0x500, 0x580, 0x600, 0x680, 0x700, 0x780} - for _, num := range nums { - entry = (*uint32)(unsafe.Pointer(toLocation + uintptr(num))) - *entry = *entry - (uint32)(offset/4) - } - - page = getPageWithReflect(toLocation) - if err := syscall.Mprotect(page, syscall.PROT_READ|syscall.PROT_EXEC); err != nil { - panic(err) - } - - page = getPageWithReflect(toLocation + 4096) - if err := syscall.Mprotect(page, syscall.PROT_READ|syscall.PROT_EXEC); err != nil { - panic(err) - } -} - // initArchState initializes architecture-specific state. func (c *vCPU) initArchState() error { var ( diff --git a/pkg/sentry/platform/ptrace/BUILD b/pkg/sentry/platform/ptrace/BUILD index 95abd321e..30402c2df 100644 --- a/pkg/sentry/platform/ptrace/BUILD +++ b/pkg/sentry/platform/ptrace/BUILD @@ -9,6 +9,7 @@ go_library( "ptrace.go", "ptrace_amd64.go", "ptrace_arm64.go", + "ptrace_arm64_unsafe.go", "ptrace_unsafe.go", "stub_amd64.s", "stub_arm64.s", diff --git a/pkg/sentry/platform/ptrace/ptrace_amd64.go b/pkg/sentry/platform/ptrace/ptrace_amd64.go index db0212538..24fc5dc62 100644 --- a/pkg/sentry/platform/ptrace/ptrace_amd64.go +++ b/pkg/sentry/platform/ptrace/ptrace_amd64.go @@ -31,3 +31,17 @@ func fpRegSet(useXsave bool) uintptr { func stackPointer(r *syscall.PtraceRegs) uintptr { return uintptr(r.Rsp) } + +// x86 use the fs_base register to store the TLS pointer which can be +// get/set in "func (t *thread) get/setRegs(regs *syscall.PtraceRegs)". +// So both of the get/setTLS() operations are noop here. + +// getTLS gets the thread local storage register. +func (t *thread) getTLS(tls *uint64) error { + return nil +} + +// setTLS sets the thread local storage register. +func (t *thread) setTLS(tls *uint64) error { + return nil +} diff --git a/pkg/sentry/platform/ptrace/ptrace_arm64_unsafe.go b/pkg/sentry/platform/ptrace/ptrace_arm64_unsafe.go new file mode 100644 index 000000000..32b8a6be9 --- /dev/null +++ b/pkg/sentry/platform/ptrace/ptrace_arm64_unsafe.go @@ -0,0 +1,62 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// +build arm64 + +package ptrace + +import ( + "syscall" + "unsafe" + + "gvisor.dev/gvisor/pkg/abi/linux" +) + +// getTLS gets the thread local storage register. +func (t *thread) getTLS(tls *uint64) error { + iovec := syscall.Iovec{ + Base: (*byte)(unsafe.Pointer(tls)), + Len: uint64(unsafe.Sizeof(*tls)), + } + _, _, errno := syscall.RawSyscall6( + syscall.SYS_PTRACE, + syscall.PTRACE_GETREGSET, + uintptr(t.tid), + linux.NT_ARM_TLS, + uintptr(unsafe.Pointer(&iovec)), + 0, 0) + if errno != 0 { + return errno + } + return nil +} + +// setTLS sets the thread local storage register. +func (t *thread) setTLS(tls *uint64) error { + iovec := syscall.Iovec{ + Base: (*byte)(unsafe.Pointer(tls)), + Len: uint64(unsafe.Sizeof(*tls)), + } + _, _, errno := syscall.RawSyscall6( + syscall.SYS_PTRACE, + syscall.PTRACE_SETREGSET, + uintptr(t.tid), + linux.NT_ARM_TLS, + uintptr(unsafe.Pointer(&iovec)), + 0, 0) + if errno != 0 { + return errno + } + return nil +} diff --git a/pkg/sentry/platform/ptrace/subprocess.go b/pkg/sentry/platform/ptrace/subprocess.go index 31b7cec53..a644609ef 100644 --- a/pkg/sentry/platform/ptrace/subprocess.go +++ b/pkg/sentry/platform/ptrace/subprocess.go @@ -506,6 +506,9 @@ func (s *subprocess) switchToApp(c *context, ac arch.Context) bool { regs := &ac.StateData().Regs t.resetSysemuRegs(regs) + // Extract TLS register + tls := uint64(ac.TLS()) + // Check for interrupts, and ensure that future interrupts will signal t. if !c.interrupt.Enable(t) { // Pending interrupt; simulate. @@ -526,6 +529,9 @@ func (s *subprocess) switchToApp(c *context, ac arch.Context) bool { if err := t.setFPRegs(fpState, uint64(fpLen), useXsave); err != nil { panic(fmt.Sprintf("ptrace set fpregs (%+v) failed: %v", fpState, err)) } + if err := t.setTLS(&tls); err != nil { + panic(fmt.Sprintf("ptrace set tls (%+v) failed: %v", tls, err)) + } for { // Start running until the next system call. @@ -555,6 +561,12 @@ func (s *subprocess) switchToApp(c *context, ac arch.Context) bool { if err := t.getFPRegs(fpState, uint64(fpLen), useXsave); err != nil { panic(fmt.Sprintf("ptrace get fpregs failed: %v", err)) } + if err := t.getTLS(&tls); err != nil { + panic(fmt.Sprintf("ptrace get tls failed: %v", err)) + } + if !ac.SetTLS(uintptr(tls)) { + panic(fmt.Sprintf("tls value %v is invalid", tls)) + } // Is it a system call? if sig == (syscallEvent | syscall.SIGTRAP) { diff --git a/pkg/sentry/platform/ptrace/subprocess_amd64.go b/pkg/sentry/platform/ptrace/subprocess_amd64.go index e99798c56..cd74945e7 100644 --- a/pkg/sentry/platform/ptrace/subprocess_amd64.go +++ b/pkg/sentry/platform/ptrace/subprocess_amd64.go @@ -21,6 +21,7 @@ import ( "strings" "syscall" + "golang.org/x/sys/unix" "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/seccomp" "gvisor.dev/gvisor/pkg/sentry/arch" @@ -183,13 +184,76 @@ func enableCpuidFault() { // appendArchSeccompRules append architecture specific seccomp rules when creating BPF program. // Ref attachedThread() for more detail. -func appendArchSeccompRules(rules []seccomp.RuleSet) []seccomp.RuleSet { - return append(rules, seccomp.RuleSet{ - Rules: seccomp.SyscallRules{ - syscall.SYS_ARCH_PRCTL: []seccomp.Rule{ - {seccomp.AllowValue(linux.ARCH_SET_CPUID), seccomp.AllowValue(0)}, +func appendArchSeccompRules(rules []seccomp.RuleSet, defaultAction linux.BPFAction) []seccomp.RuleSet { + rules = append(rules, + // Rules for trapping vsyscall access. + seccomp.RuleSet{ + Rules: seccomp.SyscallRules{ + syscall.SYS_GETTIMEOFDAY: {}, + syscall.SYS_TIME: {}, + unix.SYS_GETCPU: {}, // SYS_GETCPU was not defined in package syscall on amd64. }, - }, - Action: linux.SECCOMP_RET_ALLOW, - }) + Action: linux.SECCOMP_RET_TRAP, + Vsyscall: true, + }) + if defaultAction != linux.SECCOMP_RET_ALLOW { + rules = append(rules, + seccomp.RuleSet{ + Rules: seccomp.SyscallRules{ + syscall.SYS_ARCH_PRCTL: []seccomp.Rule{ + {seccomp.AllowValue(linux.ARCH_SET_CPUID), seccomp.AllowValue(0)}, + }, + }, + Action: linux.SECCOMP_RET_ALLOW, + }) + } + return rules +} + +// probeSeccomp returns true iff seccomp is run after ptrace notifications, +// which is generally the case for kernel version >= 4.8. This check is dynamic +// because kernels have be backported behavior. +// +// See createStub for more information. +// +// Precondition: the runtime OS thread must be locked. +func probeSeccomp() bool { + // Create a completely new, destroyable process. + t, err := attachedThread(0, linux.SECCOMP_RET_ERRNO) + if err != nil { + panic(fmt.Sprintf("seccomp probe failed: %v", err)) + } + defer t.destroy() + + // Set registers to the yield system call. This call is not allowed + // by the filters specified in the attachThread function. + regs := createSyscallRegs(&t.initRegs, syscall.SYS_SCHED_YIELD) + if err := t.setRegs(®s); err != nil { + panic(fmt.Sprintf("ptrace set regs failed: %v", err)) + } + + for { + // Attempt an emulation. + if _, _, errno := syscall.RawSyscall6(syscall.SYS_PTRACE, unix.PTRACE_SYSEMU, uintptr(t.tid), 0, 0, 0, 0); errno != 0 { + panic(fmt.Sprintf("ptrace syscall-enter failed: %v", errno)) + } + + sig := t.wait(stopped) + if sig == (syscallEvent | syscall.SIGTRAP) { + // Did the seccomp errno hook already run? This would + // indicate that seccomp is first in line and we're + // less than 4.8. + if err := t.getRegs(®s); err != nil { + panic(fmt.Sprintf("ptrace get-regs failed: %v", err)) + } + if _, err := syscallReturnValue(®s); err == nil { + // The seccomp errno mode ran first, and reset + // the error in the registers. + return false + } + // The seccomp hook did not run yet, and therefore it + // is safe to use RET_KILL mode for dispatched calls. + return true + } + } } diff --git a/pkg/sentry/platform/ptrace/subprocess_arm64.go b/pkg/sentry/platform/ptrace/subprocess_arm64.go index 7b975137f..7f5c393f0 100644 --- a/pkg/sentry/platform/ptrace/subprocess_arm64.go +++ b/pkg/sentry/platform/ptrace/subprocess_arm64.go @@ -160,6 +160,15 @@ func enableCpuidFault() { // appendArchSeccompRules append architecture specific seccomp rules when creating BPF program. // Ref attachedThread() for more detail. -func appendArchSeccompRules(rules []seccomp.RuleSet) []seccomp.RuleSet { +func appendArchSeccompRules(rules []seccomp.RuleSet, defaultAction linux.BPFAction) []seccomp.RuleSet { return rules } + +// probeSeccomp returns true if seccomp is run after ptrace notifications, +// which is generally the case for kernel version >= 4.8. +// +// On arm64, the support of PTRACE_SYSEMU was added in the 5.3 kernel, so +// probeSeccomp can always return true. +func probeSeccomp() bool { + return true +} diff --git a/pkg/sentry/platform/ptrace/subprocess_linux.go b/pkg/sentry/platform/ptrace/subprocess_linux.go index 74968dfdf..2ce528601 100644 --- a/pkg/sentry/platform/ptrace/subprocess_linux.go +++ b/pkg/sentry/platform/ptrace/subprocess_linux.go @@ -20,7 +20,6 @@ import ( "fmt" "syscall" - "golang.org/x/sys/unix" "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/log" "gvisor.dev/gvisor/pkg/procid" @@ -30,54 +29,6 @@ import ( const syscallEvent syscall.Signal = 0x80 -// probeSeccomp returns true iff seccomp is run after ptrace notifications, -// which is generally the case for kernel version >= 4.8. This check is dynamic -// because kernels have be backported behavior. -// -// See createStub for more information. -// -// Precondition: the runtime OS thread must be locked. -func probeSeccomp() bool { - // Create a completely new, destroyable process. - t, err := attachedThread(0, linux.SECCOMP_RET_ERRNO) - if err != nil { - panic(fmt.Sprintf("seccomp probe failed: %v", err)) - } - defer t.destroy() - - // Set registers to the yield system call. This call is not allowed - // by the filters specified in the attachThread function. - regs := createSyscallRegs(&t.initRegs, syscall.SYS_SCHED_YIELD) - if err := t.setRegs(®s); err != nil { - panic(fmt.Sprintf("ptrace set regs failed: %v", err)) - } - - for { - // Attempt an emulation. - if _, _, errno := syscall.RawSyscall6(syscall.SYS_PTRACE, unix.PTRACE_SYSEMU, uintptr(t.tid), 0, 0, 0, 0); errno != 0 { - panic(fmt.Sprintf("ptrace syscall-enter failed: %v", errno)) - } - - sig := t.wait(stopped) - if sig == (syscallEvent | syscall.SIGTRAP) { - // Did the seccomp errno hook already run? This would - // indicate that seccomp is first in line and we're - // less than 4.8. - if err := t.getRegs(®s); err != nil { - panic(fmt.Sprintf("ptrace get-regs failed: %v", err)) - } - if _, err := syscallReturnValue(®s); err == nil { - // The seccomp errno mode ran first, and reset - // the error in the registers. - return false - } - // The seccomp hook did not run yet, and therefore it - // is safe to use RET_KILL mode for dispatched calls. - return true - } - } -} - // createStub creates a fresh stub processes. // // Precondition: the runtime OS thread must be locked. @@ -123,18 +74,7 @@ func attachedThread(flags uintptr, defaultAction linux.BPFAction) (*thread, erro // stub and all its children. This is used to create child stubs // (below), so we must include the ability to fork, but otherwise lock // down available calls only to what is needed. - rules := []seccomp.RuleSet{ - // Rules for trapping vsyscall access. - { - Rules: seccomp.SyscallRules{ - syscall.SYS_GETTIMEOFDAY: {}, - syscall.SYS_TIME: {}, - unix.SYS_GETCPU: {}, // SYS_GETCPU was not defined in package syscall on amd64. - }, - Action: linux.SECCOMP_RET_TRAP, - Vsyscall: true, - }, - } + rules := []seccomp.RuleSet{} if defaultAction != linux.SECCOMP_RET_ALLOW { rules = append(rules, seccomp.RuleSet{ Rules: seccomp.SyscallRules{ @@ -173,9 +113,8 @@ func attachedThread(flags uintptr, defaultAction linux.BPFAction) (*thread, erro }, Action: linux.SECCOMP_RET_ALLOW, }) - - rules = appendArchSeccompRules(rules) } + rules = appendArchSeccompRules(rules, defaultAction) instrs, err := seccomp.BuildProgram(rules, defaultAction) if err != nil { return nil, err diff --git a/pkg/sentry/platform/ring0/BUILD b/pkg/sentry/platform/ring0/BUILD index 934b6fbcd..b69520030 100644 --- a/pkg/sentry/platform/ring0/BUILD +++ b/pkg/sentry/platform/ring0/BUILD @@ -72,11 +72,13 @@ go_library( "lib_amd64.s", "lib_arm64.go", "lib_arm64.s", + "lib_arm64_unsafe.go", "ring0.go", ], visibility = ["//pkg/sentry:internal"], deps = [ "//pkg/cpuid", + "//pkg/safecopy", "//pkg/sentry/platform/ring0/pagetables", "//pkg/usermem", ], diff --git a/pkg/sentry/platform/ring0/aarch64.go b/pkg/sentry/platform/ring0/aarch64.go index f6da41c27..8122ac6e2 100644 --- a/pkg/sentry/platform/ring0/aarch64.go +++ b/pkg/sentry/platform/ring0/aarch64.go @@ -27,26 +27,27 @@ const ( _PTE_PGT_BASE = 0x7000 _PTE_PGT_SIZE = 0x1000 - _PSR_MODE_EL0t = 0x0 - _PSR_MODE_EL1t = 0x4 - _PSR_MODE_EL1h = 0x5 - _PSR_EL_MASK = 0xf - - _PSR_D_BIT = 0x200 - _PSR_A_BIT = 0x100 - _PSR_I_BIT = 0x80 - _PSR_F_BIT = 0x40 + _PSR_D_BIT = 0x00000200 + _PSR_A_BIT = 0x00000100 + _PSR_I_BIT = 0x00000080 + _PSR_F_BIT = 0x00000040 ) const ( + // PSR bits + PSR_MODE_EL0t = 0x00000000 + PSR_MODE_EL1t = 0x00000004 + PSR_MODE_EL1h = 0x00000005 + PSR_MODE_MASK = 0x0000000f + // KernelFlagsSet should always be set in the kernel. - KernelFlagsSet = _PSR_MODE_EL1h + KernelFlagsSet = PSR_MODE_EL1h // UserFlagsSet are always set in userspace. - UserFlagsSet = _PSR_MODE_EL0t + UserFlagsSet = PSR_MODE_EL0t - KernelFlagsClear = _PSR_EL_MASK - UserFlagsClear = _PSR_EL_MASK + KernelFlagsClear = PSR_MODE_MASK + UserFlagsClear = PSR_MODE_MASK PsrDefaultSet = _PSR_D_BIT | _PSR_A_BIT | _PSR_I_BIT | _PSR_F_BIT ) diff --git a/pkg/sentry/platform/ring0/entry_arm64.s b/pkg/sentry/platform/ring0/entry_arm64.s index baa6c4910..db6465663 100644 --- a/pkg/sentry/platform/ring0/entry_arm64.s +++ b/pkg/sentry/platform/ring0/entry_arm64.s @@ -25,10 +25,14 @@ // not available for calls. // +// ERET returns using the ELR and SPSR for the current exception level. #define ERET() \ WORD $0xd69f03e0 +// RSV_REG is a register that holds el1 information temporarily. #define RSV_REG R18_PLATFORM + +// RSV_REG_APP is a register that holds el0 information temporarily. #define RSV_REG_APP R9 #define FPEN_NOTRAP 0x3 @@ -36,6 +40,12 @@ #define FPEN_ENABLE (FPEN_NOTRAP << FPEN_SHIFT) +// Saves a register set. +// +// This is a macro because it may need to executed in contents where a stack is +// not available for calls. +// +// The following registers are not saved: R9, R18. #define REGISTERS_SAVE(reg, offset) \ MOVD R0, offset+PTRACE_R0(reg); \ MOVD R1, offset+PTRACE_R1(reg); \ @@ -67,6 +77,12 @@ MOVD R29, offset+PTRACE_R29(reg); \ MOVD R30, offset+PTRACE_R30(reg); +// Loads a register set. +// +// This is a macro because it may need to executed in contents where a stack is +// not available for calls. +// +// The following registers are not loaded: R9, R18. #define REGISTERS_LOAD(reg, offset) \ MOVD offset+PTRACE_R0(reg), R0; \ MOVD offset+PTRACE_R1(reg), R1; \ @@ -98,7 +114,7 @@ MOVD offset+PTRACE_R29(reg), R29; \ MOVD offset+PTRACE_R30(reg), R30; -//NOP +// NOP-s #define nop31Instructions() \ WORD $0xd503201f; \ WORD $0xd503201f; \ @@ -254,6 +270,7 @@ #define ESR_ELx_WFx_ISS_WFE (UL(1) << 0) #define ESR_ELx_xVC_IMM_MASK ((1UL << 16) - 1) +// LOAD_KERNEL_ADDRESS loads a kernel address. #define LOAD_KERNEL_ADDRESS(from, to) \ MOVD from, to; \ ORR $0xffff000000000000, to, to; @@ -263,15 +280,18 @@ LOAD_KERNEL_ADDRESS(CPU_SELF(from), RSV_REG); \ MOVD $CPU_STACK_TOP(RSV_REG), RSV_REG; \ MOVD RSV_REG, RSP; \ + WORD $0xd538d092; \ //MRS TPIDR_EL1, R18 ISB $15; \ DSB $15; +// SWITCH_TO_APP_PAGETABLE sets a new pagetable for a container application. #define SWITCH_TO_APP_PAGETABLE(from) \ MOVD CPU_TTBR0_APP(from), RSV_REG; \ WORD $0xd5182012; \ // MSR R18, TTBR0_EL1 ISB $15; \ DSB $15; +// SWITCH_TO_KVM_PAGETABLE sets the kvm pagetable. #define SWITCH_TO_KVM_PAGETABLE(from) \ MOVD CPU_TTBR0_KVM(from), RSV_REG; \ WORD $0xd5182012; \ // MSR R18, TTBR0_EL1 @@ -294,6 +314,7 @@ WORD $0xd5181040; \ //MSR R0, CPACR_EL1 ISB $15; +// KERNEL_ENTRY_FROM_EL0 is the entry code of the vcpu from el0 to el1. #define KERNEL_ENTRY_FROM_EL0 \ SUB $16, RSP, RSP; \ // step1, save r18, r9 into kernel temporary stack. STP (RSV_REG, RSV_REG_APP), 16*0(RSP); \ @@ -315,19 +336,22 @@ WORD $0xd5384103; \ // MRS SP_EL0, R3 MOVD R3, PTRACE_SP(RSV_REG_APP); +// KERNEL_ENTRY_FROM_EL1 is the entry code of the vcpu from el1 to el1. #define KERNEL_ENTRY_FROM_EL1 \ WORD $0xd538d092; \ //MRS TPIDR_EL1, R18 - REGISTERS_SAVE(RSV_REG, CPU_REGISTERS); \ // save sentry context + REGISTERS_SAVE(RSV_REG, CPU_REGISTERS); \ // Save sentry context. MOVD RSV_REG_APP, CPU_REGISTERS+PTRACE_R9(RSV_REG); \ WORD $0xd5384004; \ // MRS SPSR_EL1, R4 MOVD R4, CPU_REGISTERS+PTRACE_PSTATE(RSV_REG); \ MRS ELR_EL1, R4; \ MOVD R4, CPU_REGISTERS+PTRACE_PC(RSV_REG); \ MOVD RSP, R4; \ - MOVD R4, CPU_REGISTERS+PTRACE_SP(RSV_REG); + MOVD R4, CPU_REGISTERS+PTRACE_SP(RSV_REG); \ + LOAD_KERNEL_STACK(RSV_REG); // Load the temporary stack. +// Halt halts execution. TEXT ·Halt(SB),NOSPLIT,$0 - // clear bluepill. + // Clear bluepill. WORD $0xd538d092 //MRS TPIDR_EL1, R18 CMP RSV_REG, R9 BNE mmio_exit @@ -341,8 +365,22 @@ mmio_exit: // MMIO_EXIT. MOVD $0, R9 MOVD R0, 0xffff000000001000(R9) - B ·kernelExitToEl1(SB) + RET + +// HaltAndResume halts execution and point the pointer to the resume function. +TEXT ·HaltAndResume(SB),NOSPLIT,$0 + BL ·Halt(SB) + B ·kernelExitToEl1(SB) // Resume. + +// HaltEl1SvcAndResume calls Hooks.KernelSyscall and resume. +TEXT ·HaltEl1SvcAndResume(SB),NOSPLIT,$0 + WORD $0xd538d092 // MRS TPIDR_EL1, R18 + MOVD CPU_SELF(RSV_REG), R3 // Load vCPU. + MOVD R3, 8(RSP) // First argument (vCPU). + CALL ·kernelSyscall(SB) // Call the trampoline. + B ·kernelExitToEl1(SB) // Resume. +// Shutdown stops the guest. TEXT ·Shutdown(SB),NOSPLIT,$0 // PSCI EVENT. MOVD $0x84000009, R0 @@ -356,6 +394,8 @@ TEXT ·Current(SB),NOSPLIT,$0-8 #define STACK_FRAME_SIZE 16 +// kernelExitToEl0 is the entrypoint for application in guest_el0. +// Prepare the vcpu environment for container application. TEXT ·kernelExitToEl0(SB),NOSPLIT,$0 // Step1, save sentry context into memory. REGISTERS_SAVE(RSV_REG, CPU_REGISTERS) @@ -426,9 +466,26 @@ TEXT ·kernelExitToEl0(SB),NOSPLIT,$0 ERET() +// kernelExitToEl1 is the entrypoint for sentry in guest_el1. +// Prepare the vcpu environment for sentry. TEXT ·kernelExitToEl1(SB),NOSPLIT,$0 + WORD $0xd538d092 //MRS TPIDR_EL1, R18 + + MOVD CPU_REGISTERS+PTRACE_PSTATE(RSV_REG), R1 + WORD $0xd5184001 //MSR R1, SPSR_EL1 + + MOVD CPU_REGISTERS+PTRACE_PC(RSV_REG), R1 + MSR R1, ELR_EL1 + + MOVD CPU_REGISTERS+PTRACE_SP(RSV_REG), R1 + MOVD R1, RSP + + REGISTERS_LOAD(RSV_REG, CPU_REGISTERS) + MOVD CPU_REGISTERS+PTRACE_R9(RSV_REG), RSV_REG_APP + ERET() +// Start is the CPU entrypoint. TEXT ·Start(SB),NOSPLIT,$0 IRQ_DISABLE MOVD R8, RSV_REG @@ -437,18 +494,23 @@ TEXT ·Start(SB),NOSPLIT,$0 B ·kernelExitToEl1(SB) +// El1_sync_invalid is the handler for an invalid EL1_sync. TEXT ·El1_sync_invalid(SB),NOSPLIT,$0 B ·Shutdown(SB) +// El1_irq_invalid is the handler for an invalid El1_irq. TEXT ·El1_irq_invalid(SB),NOSPLIT,$0 B ·Shutdown(SB) +// El1_fiq_invalid is the handler for an invalid El1_fiq. TEXT ·El1_fiq_invalid(SB),NOSPLIT,$0 B ·Shutdown(SB) +// El1_error_invalid is the handler for an invalid El1_error. TEXT ·El1_error_invalid(SB),NOSPLIT,$0 B ·Shutdown(SB) +// El1_sync is the handler for El1_sync. TEXT ·El1_sync(SB),NOSPLIT,$0 KERNEL_ENTRY_FROM_EL1 WORD $0xd5385219 // MRS ESR_EL1, R25 @@ -484,10 +546,10 @@ el1_da: MOVD $PageFault, R3 MOVD R3, CPU_VECTOR_CODE(RSV_REG) - B ·Halt(SB) + B ·HaltAndResume(SB) el1_ia: - B ·Halt(SB) + B ·HaltAndResume(SB) el1_sp_pc: B ·Shutdown(SB) @@ -496,7 +558,9 @@ el1_undef: B ·Shutdown(SB) el1_svc: - B ·Halt(SB) + MOVD $0, CPU_ERROR_CODE(RSV_REG) + MOVD $0, CPU_ERROR_TYPE(RSV_REG) + B ·HaltEl1SvcAndResume(SB) el1_dbg: B ·Shutdown(SB) @@ -508,15 +572,19 @@ el1_fpsimd_acc: el1_invalid: B ·Shutdown(SB) +// El1_irq is the handler for El1_irq. TEXT ·El1_irq(SB),NOSPLIT,$0 B ·Shutdown(SB) +// El1_fiq is the handler for El1_fiq. TEXT ·El1_fiq(SB),NOSPLIT,$0 B ·Shutdown(SB) +// El1_error is the handler for El1_error. TEXT ·El1_error(SB),NOSPLIT,$0 B ·Shutdown(SB) +// El0_sync is the handler for El0_sync. TEXT ·El0_sync(SB),NOSPLIT,$0 KERNEL_ENTRY_FROM_EL0 WORD $0xd5385219 // MRS ESR_EL1, R25 @@ -554,7 +622,7 @@ el0_svc: MOVD $Syscall, R3 MOVD R3, CPU_VECTOR_CODE(RSV_REG) - B ·Halt(SB) + B ·HaltAndResume(SB) el0_da: WORD $0xd538d092 //MRS TPIDR_EL1, R18 @@ -568,7 +636,7 @@ el0_da: MOVD $PageFault, R3 MOVD R3, CPU_VECTOR_CODE(RSV_REG) - B ·Halt(SB) + B ·HaltAndResume(SB) el0_ia: B ·Shutdown(SB) @@ -613,7 +681,7 @@ TEXT ·El0_error(SB),NOSPLIT,$0 MOVD $VirtualizationException, R3 MOVD R3, CPU_VECTOR_CODE(RSV_REG) - B ·Halt(SB) + B ·HaltAndResume(SB) TEXT ·El0_sync_invalid(SB),NOSPLIT,$0 B ·Shutdown(SB) @@ -627,6 +695,7 @@ TEXT ·El0_fiq_invalid(SB),NOSPLIT,$0 TEXT ·El0_error_invalid(SB),NOSPLIT,$0 B ·Shutdown(SB) +// Vectors implements exception vector table. TEXT ·Vectors(SB),NOSPLIT,$0 B ·El1_sync_invalid(SB) nop31Instructions() diff --git a/pkg/sentry/platform/ring0/kernel_arm64.go b/pkg/sentry/platform/ring0/kernel_arm64.go index c3d341998..ccacaea6b 100644 --- a/pkg/sentry/platform/ring0/kernel_arm64.go +++ b/pkg/sentry/platform/ring0/kernel_arm64.go @@ -16,6 +16,14 @@ package ring0 +// HaltAndResume halts execution and point the pointer to the resume function. +//go:nosplit +func HaltAndResume() + +// HaltEl1SvcAndResume calls Hooks.KernelSyscall and resume. +//go:nosplit +func HaltEl1SvcAndResume() + // init initializes architecture-specific state. func (k *Kernel) init(opts KernelOpts) { // Save the root page tables. diff --git a/pkg/sentry/platform/ring0/lib_arm64.go b/pkg/sentry/platform/ring0/lib_arm64.go index af075aae4..242b9305c 100644 --- a/pkg/sentry/platform/ring0/lib_arm64.go +++ b/pkg/sentry/platform/ring0/lib_arm64.go @@ -37,3 +37,10 @@ func SaveVRegs(*byte) // LoadVRegs loads V0-V31 registers. func LoadVRegs(*byte) + +// Init sets function pointers based on architectural features. +// +// This must be called prior to using ring0. +func Init() { + rewriteVectors() +} diff --git a/pkg/sentry/platform/ring0/lib_arm64_unsafe.go b/pkg/sentry/platform/ring0/lib_arm64_unsafe.go new file mode 100644 index 000000000..c05166fea --- /dev/null +++ b/pkg/sentry/platform/ring0/lib_arm64_unsafe.go @@ -0,0 +1,108 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// +build arm64 + +package ring0 + +import ( + "reflect" + "syscall" + "unsafe" + + "gvisor.dev/gvisor/pkg/safecopy" + "gvisor.dev/gvisor/pkg/usermem" +) + +const ( + nopInstruction = 0xd503201f + instSize = unsafe.Sizeof(uint32(0)) + vectorsRawLen = 0x800 +) + +func unsafeSlice(addr uintptr, length int) (slice []uint32) { + hdr := (*reflect.SliceHeader)(unsafe.Pointer(&slice)) + hdr.Data = addr + hdr.Len = length / int(instSize) + hdr.Cap = length / int(instSize) + return slice +} + +// Work around: move ring0.Vectors() into a specific address with 11-bits alignment. +// +// According to the design documentation of Arm64, +// the start address of exception vector table should be 11-bits aligned. +// Please see the code in linux kernel as reference: arch/arm64/kernel/entry.S +// But, we can't align a function's start address to a specific address by using golang. +// We have raised this question in golang community: +// https://groups.google.com/forum/m/#!topic/golang-dev/RPj90l5x86I +// This function will be removed when golang supports this feature. +// +// There are 2 jobs were implemented in this function: +// 1, move the start address of exception vector table into the specific address. +// 2, modify the offset of each instruction. +func rewriteVectors() { + vectorsBegin := reflect.ValueOf(Vectors).Pointer() + + // The exception-vector-table is required to be 11-bits aligned. + // And the size is 0x800. + // Please see the documentation as reference: + // https://developer.arm.com/docs/100933/0100/aarch64-exception-vector-table + // + // But, golang does not allow to set a function's address to a specific value. + // So, for gvisor, I defined the size of exception-vector-table as 4K, + // filled the 2nd 2K part with NOP-s. + // So that, I can safely move the 1st 2K part into the address with 11-bits alignment. + // + // So, the prerequisite for this function to work correctly is: + // vectorsSafeLen >= 0x1000 + // vectorsRawLen = 0x800 + vectorsSafeLen := int(safecopy.FindEndAddress(vectorsBegin) - vectorsBegin) + if vectorsSafeLen < 2*vectorsRawLen { + panic("Can't update vectors") + } + + vectorsSafeTable := unsafeSlice(vectorsBegin, vectorsSafeLen) // Now a []uint32 + vectorsRawLen32 := vectorsRawLen / int(instSize) + + offset := vectorsBegin & (1<<11 - 1) + if offset != 0 { + offset = 1<<11 - offset + } + + pageBegin := (vectorsBegin + offset) & ^uintptr(usermem.PageSize-1) + + _, _, errno := syscall.Syscall(syscall.SYS_MPROTECT, uintptr(pageBegin), uintptr(usermem.PageSize), uintptr(syscall.PROT_READ|syscall.PROT_WRITE|syscall.PROT_EXEC)) + if errno != 0 { + panic(errno.Error()) + } + + offset = offset / instSize // By index, not bytes. + // Move exception-vector-table into the specific address, should uses memmove here. + for i := 1; i <= vectorsRawLen32; i++ { + vectorsSafeTable[int(offset)+vectorsRawLen32-i] = vectorsSafeTable[vectorsRawLen32-i] + } + + // Adjust branch since instruction was moved forward. + for i := 0; i < vectorsRawLen32; i++ { + if vectorsSafeTable[int(offset)+i] != nopInstruction { + vectorsSafeTable[int(offset)+i] -= uint32(offset) + } + } + + _, _, errno = syscall.Syscall(syscall.SYS_MPROTECT, uintptr(pageBegin), uintptr(usermem.PageSize), uintptr(syscall.PROT_READ|syscall.PROT_EXEC)) + if errno != 0 { + panic(errno.Error()) + } +} diff --git a/pkg/sentry/platform/ring0/pagetables/BUILD b/pkg/sentry/platform/ring0/pagetables/BUILD index 971eed7fa..16d5f478b 100644 --- a/pkg/sentry/platform/ring0/pagetables/BUILD +++ b/pkg/sentry/platform/ring0/pagetables/BUILD @@ -7,7 +7,7 @@ go_template( name = "generic_walker", srcs = select_arch( amd64 = ["walker_amd64.go"], - arm64 = ["walker_amd64.go"], + arm64 = ["walker_arm64.go"], ), opt_types = [ "Visitor", @@ -80,6 +80,9 @@ go_library( "pagetables_amd64.go", "pagetables_arm64.go", "pagetables_x86.go", + "pcids.go", + "pcids_aarch64.go", + "pcids_aarch64.s", "pcids_x86.go", "walker_amd64.go", "walker_arm64.go", diff --git a/pkg/sentry/platform/ring0/pagetables/pagetables_x86.go b/pkg/sentry/platform/ring0/pagetables/pagetables_x86.go index dcf061df9..157438d9b 100644 --- a/pkg/sentry/platform/ring0/pagetables/pagetables_x86.go +++ b/pkg/sentry/platform/ring0/pagetables/pagetables_x86.go @@ -12,7 +12,7 @@ // See the License for the specific language governing permissions and // limitations under the License. -// +build i386 amd64 +// +build 386 amd64 package pagetables diff --git a/pkg/sentry/platform/ring0/pagetables/pcids.go b/pkg/sentry/platform/ring0/pagetables/pcids.go new file mode 100644 index 000000000..964496aac --- /dev/null +++ b/pkg/sentry/platform/ring0/pagetables/pcids.go @@ -0,0 +1,104 @@ +// Copyright 2018 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package pagetables + +import ( + "gvisor.dev/gvisor/pkg/sync" +) + +// PCIDs is a simple PCID database. +// +// This is not protected by locks and is thus suitable for use only with a +// single CPU at a time. +type PCIDs struct { + // mu protects below. + mu sync.Mutex + + // cache are the assigned page tables. + cache map[*PageTables]uint16 + + // avail are available PCIDs. + avail []uint16 +} + +// NewPCIDs returns a new PCID database. +// +// start is the first index to assign. Typically this will be one, as the zero +// pcid will always be flushed on transition (see pagetables_x86.go). This may +// be more than one if specific PCIDs are reserved. +// +// Nil is returned iff the start and size are out of range. +func NewPCIDs(start, size uint16) *PCIDs { + if start+uint16(size) > limitPCID { + return nil // See comment. + } + p := &PCIDs{ + cache: make(map[*PageTables]uint16), + } + for pcid := start; pcid < start+size; pcid++ { + p.avail = append(p.avail, pcid) + } + return p +} + +// Assign assigns a PCID to the given PageTables. +// +// This may overwrite any previous assignment provided. If this in the case, +// true is returned to indicate that the PCID should be flushed. +func (p *PCIDs) Assign(pt *PageTables) (uint16, bool) { + p.mu.Lock() + if pcid, ok := p.cache[pt]; ok { + p.mu.Unlock() + return pcid, false // No flush. + } + + // Is there something available? + if len(p.avail) > 0 { + pcid := p.avail[len(p.avail)-1] + p.avail = p.avail[:len(p.avail)-1] + p.cache[pt] = pcid + + // We need to flush because while this is in the available + // pool, it may have been used previously. + p.mu.Unlock() + return pcid, true + } + + // Evict an existing table. + for old, pcid := range p.cache { + delete(p.cache, old) + p.cache[pt] = pcid + + // A flush is definitely required in this case, these page + // tables may still be active. (They will just be assigned some + // other PCID if and when they hit the given CPU again.) + p.mu.Unlock() + return pcid, true + } + + // No PCID. + p.mu.Unlock() + return 0, false +} + +// Drop drops references to a set of page tables. +func (p *PCIDs) Drop(pt *PageTables) { + p.mu.Lock() + if pcid, ok := p.cache[pt]; ok { + delete(p.cache, pt) + p.avail = append(p.avail, pcid) + } + p.mu.Unlock() +} diff --git a/pkg/sentry/platform/ring0/pagetables/pcids_aarch64.go b/pkg/sentry/platform/ring0/pagetables/pcids_aarch64.go new file mode 100644 index 000000000..fbfd41d83 --- /dev/null +++ b/pkg/sentry/platform/ring0/pagetables/pcids_aarch64.go @@ -0,0 +1,32 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// +build arm64 + +package pagetables + +// limitPCID is the maximum value of PCIDs. +// +// In VMSAv8-64, the PCID(ASID) size is an IMPLEMENTATION DEFINED choice +// of 8 bits or 16 bits, and ID_AA64MMFR0_EL1.ASIDBits identifies the +// supported size. When an implementation supports a 16-bit ASID, TCR_ELx.AS +// selects whether the top 8 bits of the ASID are used. +var limitPCID uint16 + +// GetASIDBits return the system ASID bits, 8 or 16 bits. +func GetASIDBits() uint8 + +func init() { + limitPCID = uint16(1)<<GetASIDBits() - 1 +} diff --git a/pkg/sentry/platform/ring0/pagetables/pcids_aarch64.s b/pkg/sentry/platform/ring0/pagetables/pcids_aarch64.s new file mode 100644 index 000000000..e9d62d768 --- /dev/null +++ b/pkg/sentry/platform/ring0/pagetables/pcids_aarch64.s @@ -0,0 +1,45 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// +build arm64 + +#include "funcdata.h" +#include "textflag.h" + +#define ID_AA64MMFR0_ASIDBITS_SHIFT 4 +#define ID_AA64MMFR0_ASIDBITS_16 2 +#define TCR_EL1_AS_BIT 36 + +// GetASIDBits return the system ASID bits, 8 or 16 bits. +// +// func GetASIDBits() uint8 +TEXT ·GetASIDBits(SB),NOSPLIT,$0-1 + // First, check whether 16bits ASID is supported. + // ID_AA64MMFR0_EL1.ASIDBITS[7:4] == 0010. + WORD $0xd5380700 // MRS ID_AA64MMFR0_EL1, R0 + UBFX $ID_AA64MMFR0_ASIDBITS_SHIFT, R0, $4, R0 + CMPW $ID_AA64MMFR0_ASIDBITS_16, R0 + BNE bits_8 + + // Second, check whether 16bits ASID is enabled. + // TCR_EL1.AS[36] == 1. + WORD $0xd5382040 // MRS TCR_EL1, R0 + TBZ $TCR_EL1_AS_BIT, R0, bits_8 + MOVD $16, R0 + B done +bits_8: + MOVD $8, R0 +done: + MOVB R0, ret+0(FP) + RET diff --git a/pkg/sentry/platform/ring0/pagetables/pcids_x86.go b/pkg/sentry/platform/ring0/pagetables/pcids_x86.go index e199bae18..91fc5e8dd 100644 --- a/pkg/sentry/platform/ring0/pagetables/pcids_x86.go +++ b/pkg/sentry/platform/ring0/pagetables/pcids_x86.go @@ -1,4 +1,4 @@ -// Copyright 2018 The gVisor Authors. +// Copyright 2020 The gVisor Authors. // // Licensed under the Apache License, Version 2.0 (the "License"); // you may not use this file except in compliance with the License. @@ -16,94 +16,5 @@ package pagetables -import ( - "gvisor.dev/gvisor/pkg/sync" -) - -// limitPCID is the number of valid PCIDs. -const limitPCID = 4096 - -// PCIDs is a simple PCID database. -// -// This is not protected by locks and is thus suitable for use only with a -// single CPU at a time. -type PCIDs struct { - // mu protects below. - mu sync.Mutex - - // cache are the assigned page tables. - cache map[*PageTables]uint16 - - // avail are available PCIDs. - avail []uint16 -} - -// NewPCIDs returns a new PCID database. -// -// start is the first index to assign. Typically this will be one, as the zero -// pcid will always be flushed on transition (see pagetables_x86.go). This may -// be more than one if specific PCIDs are reserved. -// -// Nil is returned iff the start and size are out of range. -func NewPCIDs(start, size uint16) *PCIDs { - if start+uint16(size) >= limitPCID { - return nil // See comment. - } - p := &PCIDs{ - cache: make(map[*PageTables]uint16), - } - for pcid := start; pcid < start+size; pcid++ { - p.avail = append(p.avail, pcid) - } - return p -} - -// Assign assigns a PCID to the given PageTables. -// -// This may overwrite any previous assignment provided. If this in the case, -// true is returned to indicate that the PCID should be flushed. -func (p *PCIDs) Assign(pt *PageTables) (uint16, bool) { - p.mu.Lock() - if pcid, ok := p.cache[pt]; ok { - p.mu.Unlock() - return pcid, false // No flush. - } - - // Is there something available? - if len(p.avail) > 0 { - pcid := p.avail[len(p.avail)-1] - p.avail = p.avail[:len(p.avail)-1] - p.cache[pt] = pcid - - // We need to flush because while this is in the available - // pool, it may have been used previously. - p.mu.Unlock() - return pcid, true - } - - // Evict an existing table. - for old, pcid := range p.cache { - delete(p.cache, old) - p.cache[pt] = pcid - - // A flush is definitely required in this case, these page - // tables may still be active. (They will just be assigned some - // other PCID if and when they hit the given CPU again.) - p.mu.Unlock() - return pcid, true - } - - // No PCID. - p.mu.Unlock() - return 0, false -} - -// Drop drops references to a set of page tables. -func (p *PCIDs) Drop(pt *PageTables) { - p.mu.Lock() - if pcid, ok := p.cache[pt]; ok { - delete(p.cache, pt) - p.avail = append(p.avail, pcid) - } - p.mu.Unlock() -} +// limitPCID is the maximum value of valid PCIDs. +const limitPCID = 4095 diff --git a/pkg/sentry/platform/ring0/x86.go b/pkg/sentry/platform/ring0/x86.go index 5f80d64e8..9da0ea685 100644 --- a/pkg/sentry/platform/ring0/x86.go +++ b/pkg/sentry/platform/ring0/x86.go @@ -12,7 +12,7 @@ // See the License for the specific language governing permissions and // limitations under the License. -// +build i386 amd64 +// +build 386 amd64 package ring0 diff --git a/pkg/sentry/sighandling/sighandling.go b/pkg/sentry/sighandling/sighandling.go index ba1f9043d..83195d5a1 100644 --- a/pkg/sentry/sighandling/sighandling.go +++ b/pkg/sentry/sighandling/sighandling.go @@ -85,6 +85,11 @@ func StartSignalForwarding(handler func(linux.Signal)) func() { for sig := 1; sig <= numSignals+1; sig++ { sigchan := make(chan os.Signal, 1) sigchans = append(sigchans, sigchan) + + // SIGURG is used by Go's runtime scheduler. + if sig == int(linux.SIGURG) { + continue + } signal.Notify(sigchan, syscall.Signal(sig)) } // Start up our listener. diff --git a/pkg/sentry/socket/BUILD b/pkg/sentry/socket/BUILD index 611fa22c3..c40c6d673 100644 --- a/pkg/sentry/socket/BUILD +++ b/pkg/sentry/socket/BUILD @@ -16,6 +16,7 @@ go_library( "//pkg/sentry/kernel", "//pkg/sentry/kernel/time", "//pkg/sentry/socket/unix/transport", + "//pkg/sentry/vfs", "//pkg/syserr", "//pkg/tcpip", "//pkg/usermem", diff --git a/pkg/sentry/socket/control/BUILD b/pkg/sentry/socket/control/BUILD index 79e16d6e8..4d42d29cb 100644 --- a/pkg/sentry/socket/control/BUILD +++ b/pkg/sentry/socket/control/BUILD @@ -19,6 +19,7 @@ go_library( "//pkg/sentry/socket", "//pkg/sentry/socket/unix/transport", "//pkg/syserror", + "//pkg/tcpip", "//pkg/usermem", ], ) diff --git a/pkg/sentry/socket/control/control.go b/pkg/sentry/socket/control/control.go index 00265f15b..8834a1e1a 100644 --- a/pkg/sentry/socket/control/control.go +++ b/pkg/sentry/socket/control/control.go @@ -26,6 +26,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/socket" "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/tcpip" "gvisor.dev/gvisor/pkg/usermem" ) @@ -189,7 +190,7 @@ func putUint32(buf []byte, n uint32) []byte { // putCmsg writes a control message header and as much data as will fit into // the unused capacity of a buffer. func putCmsg(buf []byte, flags int, msgType uint32, align uint, data []int32) ([]byte, int) { - space := AlignDown(cap(buf)-len(buf), 4) + space := binary.AlignDown(cap(buf)-len(buf), 4) // We can't write to space that doesn't exist, so if we are going to align // the available space, we must align down. @@ -282,19 +283,9 @@ func PackCredentials(t *kernel.Task, creds SCMCredentials, buf []byte, flags int return putCmsg(buf, flags, linux.SCM_CREDENTIALS, align, c) } -// AlignUp rounds a length up to an alignment. align must be a power of 2. -func AlignUp(length int, align uint) int { - return (length + int(align) - 1) & ^(int(align) - 1) -} - -// AlignDown rounds a down to an alignment. align must be a power of 2. -func AlignDown(length int, align uint) int { - return length & ^(int(align) - 1) -} - // alignSlice extends a slice's length (up to the capacity) to align it. func alignSlice(buf []byte, align uint) []byte { - aligned := AlignUp(len(buf), align) + aligned := binary.AlignUp(len(buf), align) if aligned > cap(buf) { // Linux allows unaligned data if there isn't room for alignment. // Since there isn't room for alignment, there isn't room for any @@ -338,7 +329,7 @@ func PackTOS(t *kernel.Task, tos uint8, buf []byte) []byte { } // PackTClass packs an IPV6_TCLASS socket control message. -func PackTClass(t *kernel.Task, tClass int32, buf []byte) []byte { +func PackTClass(t *kernel.Task, tClass uint32, buf []byte) []byte { return putCmsgStruct( buf, linux.SOL_IPV6, @@ -348,6 +339,22 @@ func PackTClass(t *kernel.Task, tClass int32, buf []byte) []byte { ) } +// PackIPPacketInfo packs an IP_PKTINFO socket control message. +func PackIPPacketInfo(t *kernel.Task, packetInfo tcpip.IPPacketInfo, buf []byte) []byte { + var p linux.ControlMessageIPPacketInfo + p.NIC = int32(packetInfo.NIC) + copy(p.LocalAddr[:], []byte(packetInfo.LocalAddr)) + copy(p.DestinationAddr[:], []byte(packetInfo.DestinationAddr)) + + return putCmsgStruct( + buf, + linux.SOL_IP, + linux.IP_PKTINFO, + t.Arch().Width(), + p, + ) +} + // PackControlMessages packs control messages into the given buffer. // // We skip control messages specific to Unix domain sockets. @@ -372,12 +379,16 @@ func PackControlMessages(t *kernel.Task, cmsgs socket.ControlMessages, buf []byt buf = PackTClass(t, cmsgs.IP.TClass, buf) } + if cmsgs.IP.HasIPPacketInfo { + buf = PackIPPacketInfo(t, cmsgs.IP.PacketInfo, buf) + } + return buf } // cmsgSpace is equivalent to CMSG_SPACE in Linux. func cmsgSpace(t *kernel.Task, dataLen int) int { - return linux.SizeOfControlMessageHeader + AlignUp(dataLen, t.Arch().Width()) + return linux.SizeOfControlMessageHeader + binary.AlignUp(dataLen, t.Arch().Width()) } // CmsgsSpace returns the number of bytes needed to fit the control messages @@ -404,6 +415,16 @@ func CmsgsSpace(t *kernel.Task, cmsgs socket.ControlMessages) int { return space } +// NewIPPacketInfo returns the IPPacketInfo struct. +func NewIPPacketInfo(packetInfo linux.ControlMessageIPPacketInfo) tcpip.IPPacketInfo { + var p tcpip.IPPacketInfo + p.NIC = tcpip.NICID(packetInfo.NIC) + copy([]byte(p.LocalAddr), packetInfo.LocalAddr[:]) + copy([]byte(p.DestinationAddr), packetInfo.DestinationAddr[:]) + + return p +} + // Parse parses a raw socket control message into portable objects. func Parse(t *kernel.Task, socketOrEndpoint interface{}, buf []byte) (socket.ControlMessages, error) { var ( @@ -437,7 +458,7 @@ func Parse(t *kernel.Task, socketOrEndpoint interface{}, buf []byte) (socket.Con case linux.SOL_SOCKET: switch h.Type { case linux.SCM_RIGHTS: - rightsSize := AlignDown(length, linux.SizeOfControlMessageRight) + rightsSize := binary.AlignDown(length, linux.SizeOfControlMessageRight) numRights := rightsSize / linux.SizeOfControlMessageRight if len(fds)+numRights > linux.SCM_MAX_FD { @@ -448,7 +469,7 @@ func Parse(t *kernel.Task, socketOrEndpoint interface{}, buf []byte) (socket.Con fds = append(fds, int32(usermem.ByteOrder.Uint32(buf[j:j+linux.SizeOfControlMessageRight]))) } - i += AlignUp(length, width) + i += binary.AlignUp(length, width) case linux.SCM_CREDENTIALS: if length < linux.SizeOfControlMessageCredentials { @@ -462,7 +483,7 @@ func Parse(t *kernel.Task, socketOrEndpoint interface{}, buf []byte) (socket.Con return socket.ControlMessages{}, err } cmsgs.Unix.Credentials = scmCreds - i += AlignUp(length, width) + i += binary.AlignUp(length, width) default: // Unknown message type. @@ -476,7 +497,19 @@ func Parse(t *kernel.Task, socketOrEndpoint interface{}, buf []byte) (socket.Con } cmsgs.IP.HasTOS = true binary.Unmarshal(buf[i:i+linux.SizeOfControlMessageTOS], usermem.ByteOrder, &cmsgs.IP.TOS) - i += AlignUp(length, width) + i += binary.AlignUp(length, width) + + case linux.IP_PKTINFO: + if length < linux.SizeOfControlMessageIPPacketInfo { + return socket.ControlMessages{}, syserror.EINVAL + } + + cmsgs.IP.HasIPPacketInfo = true + var packetInfo linux.ControlMessageIPPacketInfo + binary.Unmarshal(buf[i:i+linux.SizeOfControlMessageIPPacketInfo], usermem.ByteOrder, &packetInfo) + + cmsgs.IP.PacketInfo = NewIPPacketInfo(packetInfo) + i += binary.AlignUp(length, width) default: return socket.ControlMessages{}, syserror.EINVAL @@ -489,7 +522,7 @@ func Parse(t *kernel.Task, socketOrEndpoint interface{}, buf []byte) (socket.Con } cmsgs.IP.HasTClass = true binary.Unmarshal(buf[i:i+linux.SizeOfControlMessageTClass], usermem.ByteOrder, &cmsgs.IP.TClass) - i += AlignUp(length, width) + i += binary.AlignUp(length, width) default: return socket.ControlMessages{}, syserror.EINVAL diff --git a/pkg/sentry/socket/hostinet/socket.go b/pkg/sentry/socket/hostinet/socket.go index de76388ac..22f78d2e2 100644 --- a/pkg/sentry/socket/hostinet/socket.go +++ b/pkg/sentry/socket/hostinet/socket.go @@ -289,7 +289,7 @@ func (s *socketOperations) GetSockOpt(t *kernel.Task, level int, name int, outPt switch level { case linux.SOL_IP: switch name { - case linux.IP_TOS, linux.IP_RECVTOS: + case linux.IP_TOS, linux.IP_RECVTOS, linux.IP_PKTINFO: optlen = sizeofInt32 } case linux.SOL_IPV6: @@ -336,6 +336,8 @@ func (s *socketOperations) SetSockOpt(t *kernel.Task, level int, name int, opt [ switch name { case linux.IP_TOS, linux.IP_RECVTOS: optlen = sizeofInt32 + case linux.IP_PKTINFO: + optlen = linux.SizeOfControlMessageIPPacketInfo } case linux.SOL_IPV6: switch name { @@ -473,7 +475,14 @@ func (s *socketOperations) RecvMsg(t *kernel.Task, dst usermem.IOSequence, flags case syscall.IP_TOS: controlMessages.IP.HasTOS = true binary.Unmarshal(unixCmsg.Data[:linux.SizeOfControlMessageTOS], usermem.ByteOrder, &controlMessages.IP.TOS) + + case syscall.IP_PKTINFO: + controlMessages.IP.HasIPPacketInfo = true + var packetInfo linux.ControlMessageIPPacketInfo + binary.Unmarshal(unixCmsg.Data[:linux.SizeOfControlMessageIPPacketInfo], usermem.ByteOrder, &packetInfo) + controlMessages.IP.PacketInfo = control.NewIPPacketInfo(packetInfo) } + case syscall.SOL_IPV6: switch unixCmsg.Header.Type { case syscall.IPV6_TCLASS: diff --git a/pkg/sentry/socket/netfilter/BUILD b/pkg/sentry/socket/netfilter/BUILD index c91ec7494..721094bbf 100644 --- a/pkg/sentry/socket/netfilter/BUILD +++ b/pkg/sentry/socket/netfilter/BUILD @@ -7,6 +7,8 @@ go_library( srcs = [ "extensions.go", "netfilter.go", + "owner_matcher.go", + "targets.go", "tcp_matcher.go", "udp_matcher.go", ], @@ -21,7 +23,6 @@ go_library( "//pkg/syserr", "//pkg/tcpip", "//pkg/tcpip/header", - "//pkg/tcpip/iptables", "//pkg/tcpip/stack", "//pkg/usermem", ], diff --git a/pkg/sentry/socket/netfilter/extensions.go b/pkg/sentry/socket/netfilter/extensions.go index 22fd0ebe7..0336a32d8 100644 --- a/pkg/sentry/socket/netfilter/extensions.go +++ b/pkg/sentry/socket/netfilter/extensions.go @@ -19,7 +19,7 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/binary" - "gvisor.dev/gvisor/pkg/tcpip/iptables" + "gvisor.dev/gvisor/pkg/tcpip/stack" "gvisor.dev/gvisor/pkg/usermem" ) @@ -37,12 +37,12 @@ type matchMaker interface { // name is the matcher name as stored in the xt_entry_match struct. name() string - // marshal converts from an iptables.Matcher to an ABI struct. - marshal(matcher iptables.Matcher) []byte + // marshal converts from an stack.Matcher to an ABI struct. + marshal(matcher stack.Matcher) []byte // unmarshal converts from the ABI matcher struct to an - // iptables.Matcher. - unmarshal(buf []byte, filter iptables.IPHeaderFilter) (iptables.Matcher, error) + // stack.Matcher. + unmarshal(buf []byte, filter stack.IPHeaderFilter) (stack.Matcher, error) } // matchMakers maps the name of supported matchers to the matchMaker that @@ -58,7 +58,7 @@ func registerMatchMaker(mm matchMaker) { matchMakers[mm.name()] = mm } -func marshalMatcher(matcher iptables.Matcher) []byte { +func marshalMatcher(matcher stack.Matcher) []byte { matchMaker, ok := matchMakers[matcher.Name()] if !ok { panic(fmt.Sprintf("Unknown matcher of type %T.", matcher)) @@ -72,7 +72,7 @@ func marshalEntryMatch(name string, data []byte) []byte { nflog("marshaling matcher %q", name) // We have to pad this struct size to a multiple of 8 bytes. - size := alignUp(linux.SizeOfXTEntryMatch+len(data), 8) + size := binary.AlignUp(linux.SizeOfXTEntryMatch+len(data), 8) matcher := linux.KernelXTEntryMatch{ XTEntryMatch: linux.XTEntryMatch{ MatchSize: uint16(size), @@ -86,15 +86,10 @@ func marshalEntryMatch(name string, data []byte) []byte { return append(buf, make([]byte, size-len(buf))...) } -func unmarshalMatcher(match linux.XTEntryMatch, filter iptables.IPHeaderFilter, buf []byte) (iptables.Matcher, error) { +func unmarshalMatcher(match linux.XTEntryMatch, filter stack.IPHeaderFilter, buf []byte) (stack.Matcher, error) { matchMaker, ok := matchMakers[match.Name.String()] if !ok { return nil, fmt.Errorf("unsupported matcher with name %q", match.Name.String()) } return matchMaker.unmarshal(buf, filter) } - -// alignUp rounds a length up to an alignment. align must be a power of 2. -func alignUp(length int, align uint) int { - return (length + int(align) - 1) & ^(int(align) - 1) -} diff --git a/pkg/sentry/socket/netfilter/netfilter.go b/pkg/sentry/socket/netfilter/netfilter.go index ea02627de..878f81fd5 100644 --- a/pkg/sentry/socket/netfilter/netfilter.go +++ b/pkg/sentry/socket/netfilter/netfilter.go @@ -26,7 +26,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/syserr" "gvisor.dev/gvisor/pkg/tcpip" - "gvisor.dev/gvisor/pkg/tcpip/iptables" + "gvisor.dev/gvisor/pkg/tcpip/header" "gvisor.dev/gvisor/pkg/tcpip/stack" "gvisor.dev/gvisor/pkg/usermem" ) @@ -35,6 +35,11 @@ import ( // shouldn't be reached - an error has occurred if we fall through to one. const errorTargetName = "ERROR" +// redirectTargetName is used to mark targets as redirect targets. Redirect +// targets should be reached for only NAT and Mangle tables. These targets will +// change the destination port/destination IP for packets. +const redirectTargetName = "REDIRECT" + // Metadata is used to verify that we are correctly serializing and // deserializing iptables into structs consumable by the iptables tool. We save // a metadata struct when the tables are written, and when they are read out we @@ -50,7 +55,9 @@ type metadata struct { // nflog logs messages related to the writing and reading of iptables. func nflog(format string, args ...interface{}) { - log.Infof("netfilter: "+format, args...) + if log.IsLogging(log.Debug) { + log.Debugf("netfilter: "+format, args...) + } } // GetInfo returns information about iptables. @@ -121,19 +128,19 @@ func GetEntries(t *kernel.Task, stack *stack.Stack, outPtr usermem.Addr, outLen return entries, nil } -func findTable(stack *stack.Stack, tablename linux.TableName) (iptables.Table, error) { - ipt := stack.IPTables() +func findTable(stk *stack.Stack, tablename linux.TableName) (stack.Table, error) { + ipt := stk.IPTables() table, ok := ipt.Tables[tablename.String()] if !ok { - return iptables.Table{}, fmt.Errorf("couldn't find table %q", tablename) + return stack.Table{}, fmt.Errorf("couldn't find table %q", tablename) } return table, nil } // FillDefaultIPTables sets stack's IPTables to the default tables and // populates them with metadata. -func FillDefaultIPTables(stack *stack.Stack) { - ipt := iptables.DefaultTables() +func FillDefaultIPTables(stk *stack.Stack) { + ipt := stack.DefaultTables() // In order to fill in the metadata, we have to translate ipt from its // netstack format to Linux's giant-binary-blob format. @@ -146,14 +153,14 @@ func FillDefaultIPTables(stack *stack.Stack) { ipt.Tables[name] = table } - stack.SetIPTables(ipt) + stk.SetIPTables(ipt) } // convertNetstackToBinary converts the iptables as stored in netstack to the // format expected by the iptables tool. Linux stores each table as a binary // blob that can only be traversed by parsing a bit, reading some offsets, // jumping to those offsets, parsing again, etc. -func convertNetstackToBinary(tablename string, table iptables.Table) (linux.KernelIPTGetEntries, metadata, error) { +func convertNetstackToBinary(tablename string, table stack.Table) (linux.KernelIPTGetEntries, metadata, error) { // Return values. var entries linux.KernelIPTGetEntries var meta metadata @@ -226,21 +233,29 @@ func convertNetstackToBinary(tablename string, table iptables.Table) (linux.Kern return entries, meta, nil } -func marshalTarget(target iptables.Target) []byte { - switch target.(type) { - case iptables.UnconditionalAcceptTarget: - return marshalStandardTarget(iptables.Accept) - case iptables.UnconditionalDropTarget: - return marshalStandardTarget(iptables.Drop) - case iptables.ErrorTarget: - return marshalErrorTarget() +func marshalTarget(target stack.Target) []byte { + switch tg := target.(type) { + case stack.AcceptTarget: + return marshalStandardTarget(stack.RuleAccept) + case stack.DropTarget: + return marshalStandardTarget(stack.RuleDrop) + case stack.ErrorTarget: + return marshalErrorTarget(errorTargetName) + case stack.UserChainTarget: + return marshalErrorTarget(tg.Name) + case stack.ReturnTarget: + return marshalStandardTarget(stack.RuleReturn) + case stack.RedirectTarget: + return marshalRedirectTarget() + case JumpTarget: + return marshalJumpTarget(tg) default: panic(fmt.Errorf("unknown target of type %T", target)) } } -func marshalStandardTarget(verdict iptables.Verdict) []byte { - nflog("convert to binary: marshalling standard target with size %d", linux.SizeOfXTStandardTarget) +func marshalStandardTarget(verdict stack.RuleVerdict) []byte { + nflog("convert to binary: marshalling standard target") // The target's name will be the empty string. target := linux.XTStandardTarget{ @@ -254,60 +269,87 @@ func marshalStandardTarget(verdict iptables.Verdict) []byte { return binary.Marshal(ret, usermem.ByteOrder, target) } -func marshalErrorTarget() []byte { +func marshalErrorTarget(errorName string) []byte { // This is an error target named error target := linux.XTErrorTarget{ Target: linux.XTEntryTarget{ TargetSize: linux.SizeOfXTErrorTarget, }, } - copy(target.Name[:], errorTargetName) + copy(target.Name[:], errorName) copy(target.Target.Name[:], errorTargetName) ret := make([]byte, 0, linux.SizeOfXTErrorTarget) return binary.Marshal(ret, usermem.ByteOrder, target) } +func marshalRedirectTarget() []byte { + // This is a redirect target named redirect + target := linux.XTRedirectTarget{ + Target: linux.XTEntryTarget{ + TargetSize: linux.SizeOfXTRedirectTarget, + }, + } + copy(target.Target.Name[:], redirectTargetName) + + ret := make([]byte, 0, linux.SizeOfXTRedirectTarget) + return binary.Marshal(ret, usermem.ByteOrder, target) +} + +func marshalJumpTarget(jt JumpTarget) []byte { + nflog("convert to binary: marshalling jump target") + + // The target's name will be the empty string. + target := linux.XTStandardTarget{ + Target: linux.XTEntryTarget{ + TargetSize: linux.SizeOfXTStandardTarget, + }, + // Verdict is overloaded by the ABI. When positive, it holds + // the jump offset from the start of the table. + Verdict: int32(jt.Offset), + } + + ret := make([]byte, 0, linux.SizeOfXTStandardTarget) + return binary.Marshal(ret, usermem.ByteOrder, target) +} + // translateFromStandardVerdict translates verdicts the same way as the iptables // tool. -func translateFromStandardVerdict(verdict iptables.Verdict) int32 { +func translateFromStandardVerdict(verdict stack.RuleVerdict) int32 { switch verdict { - case iptables.Accept: + case stack.RuleAccept: return -linux.NF_ACCEPT - 1 - case iptables.Drop: + case stack.RuleDrop: return -linux.NF_DROP - 1 - case iptables.Queue: - return -linux.NF_QUEUE - 1 - case iptables.Return: + case stack.RuleReturn: return linux.NF_RETURN - case iptables.Jump: + default: // TODO(gvisor.dev/issue/170): Support Jump. - panic("Jump isn't supported yet") + panic(fmt.Sprintf("unknown standard verdict: %d", verdict)) } - panic(fmt.Sprintf("unknown standard verdict: %d", verdict)) } -// translateToStandardVerdict translates from the value in a -// linux.XTStandardTarget to an iptables.Verdict. -func translateToStandardVerdict(val int32) (iptables.Verdict, error) { +// translateToStandardTarget translates from the value in a +// linux.XTStandardTarget to an stack.Verdict. +func translateToStandardTarget(val int32) (stack.Target, error) { // TODO(gvisor.dev/issue/170): Support other verdicts. switch val { case -linux.NF_ACCEPT - 1: - return iptables.Accept, nil + return stack.AcceptTarget{}, nil case -linux.NF_DROP - 1: - return iptables.Drop, nil + return stack.DropTarget{}, nil case -linux.NF_QUEUE - 1: - return iptables.Invalid, errors.New("unsupported iptables verdict QUEUE") + return nil, errors.New("unsupported iptables verdict QUEUE") case linux.NF_RETURN: - return iptables.Invalid, errors.New("unsupported iptables verdict RETURN") + return stack.ReturnTarget{}, nil default: - return iptables.Invalid, fmt.Errorf("unknown iptables verdict %d", val) + return nil, fmt.Errorf("unknown iptables verdict %d", val) } } // SetEntries sets iptables rules for a single table. See // net/ipv4/netfilter/ip_tables.c:translate_table for reference. -func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { +func SetEntries(stk *stack.Stack, optVal []byte) *syserr.Error { // Get the basic rules data (struct ipt_replace). if len(optVal) < linux.SizeOfIPTReplace { nflog("optVal has insufficient size for replace %d", len(optVal)) @@ -319,10 +361,12 @@ func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { binary.Unmarshal(replaceBuf, usermem.ByteOrder, &replace) // TODO(gvisor.dev/issue/170): Support other tables. - var table iptables.Table + var table stack.Table switch replace.Name.String() { - case iptables.TablenameFilter: - table = iptables.EmptyFilterTable() + case stack.TablenameFilter: + table = stack.EmptyFilterTable() + case stack.TablenameNat: + table = stack.EmptyNatTable() default: nflog("we don't yet support writing to the %q table (gvisor.dev/issue/170)", replace.Name.String()) return syserr.ErrInvalidArgument @@ -332,7 +376,8 @@ func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { // Convert input into a list of rules and their offsets. var offset uint32 - var offsets []uint32 + // offsets maps rule byte offsets to their position in table.Rules. + offsets := map[uint32]int{} for entryIdx := uint32(0); entryIdx < replace.NumEntries; entryIdx++ { nflog("set entries: processing entry at offset %d", offset) @@ -381,23 +426,24 @@ func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { nflog("entry doesn't have enough room for its target (only %d bytes remain)", len(optVal)) return syserr.ErrInvalidArgument } - target, err := parseTarget(optVal[:targetSize]) + target, err := parseTarget(filter, optVal[:targetSize]) if err != nil { nflog("failed to parse target: %v", err) return syserr.ErrInvalidArgument } optVal = optVal[targetSize:] - table.Rules = append(table.Rules, iptables.Rule{ + table.Rules = append(table.Rules, stack.Rule{ Filter: filter, Target: target, Matchers: matchers, }) - offsets = append(offsets, offset) + offsets[offset] = int(entryIdx) offset += uint32(entry.NextOffset) if initialOptValLen-len(optVal) != int(entry.NextOffset) { nflog("entry NextOffset is %d, but entry took up %d bytes", entry.NextOffset, initialOptValLen-len(optVal)) + return syserr.ErrInvalidArgument } } @@ -406,31 +452,76 @@ func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { for hook, _ := range replace.HookEntry { if table.ValidHooks()&(1<<hook) != 0 { hk := hookFromLinux(hook) - for ruleIdx, offset := range offsets { + for offset, ruleIdx := range offsets { if offset == replace.HookEntry[hook] { table.BuiltinChains[hk] = ruleIdx } if offset == replace.Underflow[hook] { + if !validUnderflow(table.Rules[ruleIdx]) { + nflog("underflow for hook %d isn't an unconditional ACCEPT or DROP") + return syserr.ErrInvalidArgument + } table.Underflows[hk] = ruleIdx } } - if ruleIdx := table.BuiltinChains[hk]; ruleIdx == iptables.HookUnset { + if ruleIdx := table.BuiltinChains[hk]; ruleIdx == stack.HookUnset { nflog("hook %v is unset.", hk) return syserr.ErrInvalidArgument } - if ruleIdx := table.Underflows[hk]; ruleIdx == iptables.HookUnset { + if ruleIdx := table.Underflows[hk]; ruleIdx == stack.HookUnset { nflog("underflow %v is unset.", hk) return syserr.ErrInvalidArgument } } } + // Add the user chains. + for ruleIdx, rule := range table.Rules { + target, ok := rule.Target.(stack.UserChainTarget) + if !ok { + continue + } + + // We found a user chain. Before inserting it into the table, + // check that: + // - There's some other rule after it. + // - There are no matchers. + if ruleIdx == len(table.Rules)-1 { + nflog("user chain must have a rule or default policy") + return syserr.ErrInvalidArgument + } + if len(table.Rules[ruleIdx].Matchers) != 0 { + nflog("user chain's first node must have no matchers") + return syserr.ErrInvalidArgument + } + table.UserChains[target.Name] = ruleIdx + 1 + } + + // Set each jump to point to the appropriate rule. Right now they hold byte + // offsets. + for ruleIdx, rule := range table.Rules { + jump, ok := rule.Target.(JumpTarget) + if !ok { + continue + } + + // Find the rule corresponding to the jump rule offset. + jumpTo, ok := offsets[jump.Offset] + if !ok { + nflog("failed to find a rule to jump to") + return syserr.ErrInvalidArgument + } + jump.RuleNum = jumpTo + rule.Target = jump + table.Rules[ruleIdx] = rule + } + // TODO(gvisor.dev/issue/170): Support other chains. - // Since we only support modifying the INPUT chain right now, make sure - // all other chains point to ACCEPT rules. + // Since we only support modifying the INPUT, PREROUTING and OUTPUT chain right now, + // make sure all other chains point to ACCEPT rules. for hook, ruleIdx := range table.BuiltinChains { - if hook != iptables.Input { - if _, ok := table.Rules[ruleIdx].Target.(iptables.UnconditionalAcceptTarget); !ok { + if hook == stack.Forward || hook == stack.Postrouting { + if _, ok := table.Rules[ruleIdx].Target.(stack.AcceptTarget); !ok { nflog("hook %d is unsupported.", hook) return syserr.ErrInvalidArgument } @@ -442,7 +533,7 @@ func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { // - There are no chains without an unconditional final rule. // - There are no chains without an unconditional underflow rule. - ipt := stack.IPTables() + ipt := stk.IPTables() table.SetMetadata(metadata{ HookEntry: replace.HookEntry, Underflow: replace.Underflow, @@ -450,16 +541,16 @@ func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { Size: replace.Size, }) ipt.Tables[replace.Name.String()] = table - stack.SetIPTables(ipt) + stk.SetIPTables(ipt) return nil } // parseMatchers parses 0 or more matchers from optVal. optVal should contain // only the matchers. -func parseMatchers(filter iptables.IPHeaderFilter, optVal []byte) ([]iptables.Matcher, error) { +func parseMatchers(filter stack.IPHeaderFilter, optVal []byte) ([]stack.Matcher, error) { nflog("set entries: parsing matchers of size %d", len(optVal)) - var matchers []iptables.Matcher + var matchers []stack.Matcher for len(optVal) > 0 { nflog("set entries: optVal has len %d", len(optVal)) @@ -501,7 +592,7 @@ func parseMatchers(filter iptables.IPHeaderFilter, optVal []byte) ([]iptables.Ma // parseTarget parses a target from optVal. optVal should contain only the // target. -func parseTarget(optVal []byte) (iptables.Target, error) { +func parseTarget(filter stack.IPHeaderFilter, optVal []byte) (stack.Target, error) { nflog("set entries: parsing target of size %d", len(optVal)) if len(optVal) < linux.SizeOfXTEntryTarget { return nil, fmt.Errorf("optVal has insufficient size for entry target %d", len(optVal)) @@ -519,18 +610,12 @@ func parseTarget(optVal []byte) (iptables.Target, error) { buf = optVal[:linux.SizeOfXTStandardTarget] binary.Unmarshal(buf, usermem.ByteOrder, &standardTarget) - verdict, err := translateToStandardVerdict(standardTarget.Verdict) - if err != nil { - return nil, err - } - switch verdict { - case iptables.Accept: - return iptables.UnconditionalAcceptTarget{}, nil - case iptables.Drop: - return iptables.UnconditionalDropTarget{}, nil - default: - return nil, fmt.Errorf("Unknown verdict: %v", verdict) + if standardTarget.Verdict < 0 { + // A Verdict < 0 indicates a non-jump verdict. + return translateToStandardTarget(standardTarget.Verdict) } + // A verdict >= 0 indicates a jump. + return JumpTarget{Offset: uint32(standardTarget.Verdict)}, nil case errorTargetName: // Error target. @@ -548,55 +633,128 @@ func parseTarget(optVal []byte) (iptables.Target, error) { // somehow fall through every rule. // * To mark the start of a user defined chain. These // rules have an error with the name of the chain. - switch errorTarget.Name.String() { + switch name := errorTarget.Name.String(); name { case errorTargetName: - return iptables.ErrorTarget{}, nil + nflog("set entries: error target") + return stack.ErrorTarget{}, nil default: - return nil, fmt.Errorf("unknown error target %q doesn't exist or isn't supported yet.", errorTarget.Name.String()) + // User defined chain. + nflog("set entries: user-defined target %q", name) + return stack.UserChainTarget{Name: name}, nil + } + + case redirectTargetName: + // Redirect target. + if len(optVal) < linux.SizeOfXTRedirectTarget { + return nil, fmt.Errorf("netfilter.SetEntries: optVal has insufficient size for redirect target %d", len(optVal)) } + + if filter.Protocol != header.TCPProtocolNumber && filter.Protocol != header.UDPProtocolNumber { + return nil, fmt.Errorf("netfilter.SetEntries: invalid argument") + } + + var redirectTarget linux.XTRedirectTarget + buf = optVal[:linux.SizeOfXTRedirectTarget] + binary.Unmarshal(buf, usermem.ByteOrder, &redirectTarget) + + // Copy linux.XTRedirectTarget to stack.RedirectTarget. + var target stack.RedirectTarget + nfRange := redirectTarget.NfRange + + // RangeSize should be 1. + if nfRange.RangeSize != 1 { + return nil, fmt.Errorf("netfilter.SetEntries: invalid argument") + } + + // TODO(gvisor.dev/issue/170): Check if the flags are valid. + // Also check if we need to map ports or IP. + // For now, redirect target only supports destination port change. + // Port range and IP range are not supported yet. + if nfRange.RangeIPV4.Flags&linux.NF_NAT_RANGE_PROTO_SPECIFIED == 0 { + return nil, fmt.Errorf("netfilter.SetEntries: invalid argument") + } + target.RangeProtoSpecified = true + + target.MinIP = tcpip.Address(nfRange.RangeIPV4.MinIP[:]) + target.MaxIP = tcpip.Address(nfRange.RangeIPV4.MaxIP[:]) + + // TODO(gvisor.dev/issue/170): Port range is not supported yet. + if nfRange.RangeIPV4.MinPort != nfRange.RangeIPV4.MaxPort { + return nil, fmt.Errorf("netfilter.SetEntries: invalid argument") + } + + // Convert port from big endian to little endian. + port := make([]byte, 2) + binary.BigEndian.PutUint16(port, nfRange.RangeIPV4.MinPort) + target.MinPort = binary.LittleEndian.Uint16(port) + + binary.BigEndian.PutUint16(port, nfRange.RangeIPV4.MaxPort) + target.MaxPort = binary.LittleEndian.Uint16(port) + return target, nil } // Unknown target. return nil, fmt.Errorf("unknown target %q doesn't exist or isn't supported yet.", target.Name.String()) } -func filterFromIPTIP(iptip linux.IPTIP) (iptables.IPHeaderFilter, error) { +func filterFromIPTIP(iptip linux.IPTIP) (stack.IPHeaderFilter, error) { if containsUnsupportedFields(iptip) { - return iptables.IPHeaderFilter{}, fmt.Errorf("unsupported fields in struct iptip: %+v", iptip) + return stack.IPHeaderFilter{}, fmt.Errorf("unsupported fields in struct iptip: %+v", iptip) + } + if len(iptip.Dst) != header.IPv4AddressSize || len(iptip.DstMask) != header.IPv4AddressSize { + return stack.IPHeaderFilter{}, fmt.Errorf("incorrect length of destination (%d) and/or destination mask (%d) fields", len(iptip.Dst), len(iptip.DstMask)) } - return iptables.IPHeaderFilter{ - Protocol: tcpip.TransportProtocolNumber(iptip.Protocol), + return stack.IPHeaderFilter{ + Protocol: tcpip.TransportProtocolNumber(iptip.Protocol), + Dst: tcpip.Address(iptip.Dst[:]), + DstMask: tcpip.Address(iptip.DstMask[:]), + DstInvert: iptip.InverseFlags&linux.IPT_INV_DSTIP != 0, }, nil } func containsUnsupportedFields(iptip linux.IPTIP) bool { - // Currently we check that everything except protocol is zeroed. + // The following features are supported: + // - Protocol + // - Dst and DstMask + // - The inverse destination IP check flag var emptyInetAddr = linux.InetAddr{} var emptyInterface = [linux.IFNAMSIZ]byte{} - return iptip.Dst != emptyInetAddr || - iptip.Src != emptyInetAddr || + // Disable any supported inverse flags. + inverseMask := uint8(linux.IPT_INV_DSTIP) + return iptip.Src != emptyInetAddr || iptip.SrcMask != emptyInetAddr || - iptip.DstMask != emptyInetAddr || iptip.InputInterface != emptyInterface || iptip.OutputInterface != emptyInterface || iptip.InputInterfaceMask != emptyInterface || iptip.OutputInterfaceMask != emptyInterface || iptip.Flags != 0 || - iptip.InverseFlags != 0 + iptip.InverseFlags&^inverseMask != 0 +} + +func validUnderflow(rule stack.Rule) bool { + if len(rule.Matchers) != 0 { + return false + } + switch rule.Target.(type) { + case stack.AcceptTarget, stack.DropTarget: + return true + default: + return false + } } -func hookFromLinux(hook int) iptables.Hook { +func hookFromLinux(hook int) stack.Hook { switch hook { case linux.NF_INET_PRE_ROUTING: - return iptables.Prerouting + return stack.Prerouting case linux.NF_INET_LOCAL_IN: - return iptables.Input + return stack.Input case linux.NF_INET_FORWARD: - return iptables.Forward + return stack.Forward case linux.NF_INET_LOCAL_OUT: - return iptables.Output + return stack.Output case linux.NF_INET_POST_ROUTING: - return iptables.Postrouting + return stack.Postrouting } panic(fmt.Sprintf("Unknown hook %d does not correspond to a builtin chain", hook)) } diff --git a/pkg/sentry/socket/netfilter/owner_matcher.go b/pkg/sentry/socket/netfilter/owner_matcher.go new file mode 100644 index 000000000..5949a7c29 --- /dev/null +++ b/pkg/sentry/socket/netfilter/owner_matcher.go @@ -0,0 +1,128 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package netfilter + +import ( + "fmt" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/binary" + "gvisor.dev/gvisor/pkg/tcpip/stack" + "gvisor.dev/gvisor/pkg/usermem" +) + +const matcherNameOwner = "owner" + +func init() { + registerMatchMaker(ownerMarshaler{}) +} + +// ownerMarshaler implements matchMaker for owner matching. +type ownerMarshaler struct{} + +// name implements matchMaker.name. +func (ownerMarshaler) name() string { + return matcherNameOwner +} + +// marshal implements matchMaker.marshal. +func (ownerMarshaler) marshal(mr stack.Matcher) []byte { + matcher := mr.(*OwnerMatcher) + iptOwnerInfo := linux.IPTOwnerInfo{ + UID: matcher.uid, + GID: matcher.gid, + } + + // Support for UID match. + // TODO(gvisor.dev/issue/170): Need to support gid match. + if matcher.matchUID { + iptOwnerInfo.Match = linux.XT_OWNER_UID + } else if matcher.matchGID { + panic("GID match is not supported.") + } else { + panic("UID match is not set.") + } + + buf := make([]byte, 0, linux.SizeOfIPTOwnerInfo) + return marshalEntryMatch(matcherNameOwner, binary.Marshal(buf, usermem.ByteOrder, iptOwnerInfo)) +} + +// unmarshal implements matchMaker.unmarshal. +func (ownerMarshaler) unmarshal(buf []byte, filter stack.IPHeaderFilter) (stack.Matcher, error) { + if len(buf) < linux.SizeOfIPTOwnerInfo { + return nil, fmt.Errorf("buf has insufficient size for owner match: %d", len(buf)) + } + + // For alignment reasons, the match's total size may + // exceed what's strictly necessary to hold matchData. + var matchData linux.IPTOwnerInfo + binary.Unmarshal(buf[:linux.SizeOfIPTOwnerInfo], usermem.ByteOrder, &matchData) + nflog("parseMatchers: parsed IPTOwnerInfo: %+v", matchData) + + if matchData.Invert != 0 { + return nil, fmt.Errorf("invert flag is not supported for owner match") + } + + // Support for UID match. + // TODO(gvisor.dev/issue/170): Need to support gid match. + if matchData.Match&linux.XT_OWNER_UID != linux.XT_OWNER_UID { + return nil, fmt.Errorf("owner match is only supported for uid") + } + + // Check Flags. + var owner OwnerMatcher + owner.uid = matchData.UID + owner.gid = matchData.GID + owner.matchUID = true + + return &owner, nil +} + +type OwnerMatcher struct { + uid uint32 + gid uint32 + matchUID bool + matchGID bool + invert uint8 +} + +// Name implements Matcher.Name. +func (*OwnerMatcher) Name() string { + return matcherNameOwner +} + +// Match implements Matcher.Match. +func (om *OwnerMatcher) Match(hook stack.Hook, pkt stack.PacketBuffer, interfaceName string) (bool, bool) { + // Support only for OUTPUT chain. + // TODO(gvisor.dev/issue/170): Need to support for POSTROUTING chain also. + if hook != stack.Output { + return false, true + } + + // If the packet owner is not set, drop the packet. + // Support for uid match. + // TODO(gvisor.dev/issue/170): Need to support gid match. + if pkt.Owner == nil || !om.matchUID { + return false, true + } + + // TODO(gvisor.dev/issue/170): Need to add tests to verify + // drop rule when packet UID does not match owner matcher UID. + if pkt.Owner.UID() != om.uid { + return false, false + } + + return true, false +} diff --git a/pkg/sentry/socket/netfilter/targets.go b/pkg/sentry/socket/netfilter/targets.go new file mode 100644 index 000000000..c948de876 --- /dev/null +++ b/pkg/sentry/socket/netfilter/targets.go @@ -0,0 +1,34 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package netfilter + +import ( + "gvisor.dev/gvisor/pkg/tcpip/stack" +) + +// JumpTarget implements stack.Target. +type JumpTarget struct { + // Offset is the byte offset of the rule to jump to. It is used for + // marshaling and unmarshaling. + Offset uint32 + + // RuleNum is the rule to jump to. + RuleNum int +} + +// Action implements stack.Target.Action. +func (jt JumpTarget) Action(stack.PacketBuffer) (stack.RuleVerdict, int) { + return stack.RuleJump, jt.RuleNum +} diff --git a/pkg/sentry/socket/netfilter/tcp_matcher.go b/pkg/sentry/socket/netfilter/tcp_matcher.go index f9945e214..ff1cfd8f6 100644 --- a/pkg/sentry/socket/netfilter/tcp_matcher.go +++ b/pkg/sentry/socket/netfilter/tcp_matcher.go @@ -19,9 +19,8 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/binary" - "gvisor.dev/gvisor/pkg/tcpip" "gvisor.dev/gvisor/pkg/tcpip/header" - "gvisor.dev/gvisor/pkg/tcpip/iptables" + "gvisor.dev/gvisor/pkg/tcpip/stack" "gvisor.dev/gvisor/pkg/usermem" ) @@ -40,7 +39,7 @@ func (tcpMarshaler) name() string { } // marshal implements matchMaker.marshal. -func (tcpMarshaler) marshal(mr iptables.Matcher) []byte { +func (tcpMarshaler) marshal(mr stack.Matcher) []byte { matcher := mr.(*TCPMatcher) xttcp := linux.XTTCP{ SourcePortStart: matcher.sourcePortStart, @@ -53,7 +52,7 @@ func (tcpMarshaler) marshal(mr iptables.Matcher) []byte { } // unmarshal implements matchMaker.unmarshal. -func (tcpMarshaler) unmarshal(buf []byte, filter iptables.IPHeaderFilter) (iptables.Matcher, error) { +func (tcpMarshaler) unmarshal(buf []byte, filter stack.IPHeaderFilter) (stack.Matcher, error) { if len(buf) < linux.SizeOfXTTCP { return nil, fmt.Errorf("buf has insufficient size for TCP match: %d", len(buf)) } @@ -97,7 +96,7 @@ func (*TCPMatcher) Name() string { } // Match implements Matcher.Match. -func (tm *TCPMatcher) Match(hook iptables.Hook, pkt tcpip.PacketBuffer, interfaceName string) (bool, bool) { +func (tm *TCPMatcher) Match(hook stack.Hook, pkt stack.PacketBuffer, interfaceName string) (bool, bool) { netHeader := header.IPv4(pkt.NetworkHeader) if netHeader.TransportProtocol() != header.TCPProtocolNumber { @@ -115,7 +114,7 @@ func (tm *TCPMatcher) Match(hook iptables.Hook, pkt tcpip.PacketBuffer, interfac // Now we need the transport header. However, this may not have been set // yet. // TODO(gvisor.dev/issue/170): Parsing the transport header should - // ultimately be moved into the iptables.Check codepath as matchers are + // ultimately be moved into the stack.Check codepath as matchers are // added. var tcpHeader header.TCP if pkt.TransportHeader != nil { diff --git a/pkg/sentry/socket/netfilter/udp_matcher.go b/pkg/sentry/socket/netfilter/udp_matcher.go index 86aa11696..3359418c1 100644 --- a/pkg/sentry/socket/netfilter/udp_matcher.go +++ b/pkg/sentry/socket/netfilter/udp_matcher.go @@ -19,9 +19,8 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/binary" - "gvisor.dev/gvisor/pkg/tcpip" "gvisor.dev/gvisor/pkg/tcpip/header" - "gvisor.dev/gvisor/pkg/tcpip/iptables" + "gvisor.dev/gvisor/pkg/tcpip/stack" "gvisor.dev/gvisor/pkg/usermem" ) @@ -40,7 +39,7 @@ func (udpMarshaler) name() string { } // marshal implements matchMaker.marshal. -func (udpMarshaler) marshal(mr iptables.Matcher) []byte { +func (udpMarshaler) marshal(mr stack.Matcher) []byte { matcher := mr.(*UDPMatcher) xtudp := linux.XTUDP{ SourcePortStart: matcher.sourcePortStart, @@ -53,7 +52,7 @@ func (udpMarshaler) marshal(mr iptables.Matcher) []byte { } // unmarshal implements matchMaker.unmarshal. -func (udpMarshaler) unmarshal(buf []byte, filter iptables.IPHeaderFilter) (iptables.Matcher, error) { +func (udpMarshaler) unmarshal(buf []byte, filter stack.IPHeaderFilter) (stack.Matcher, error) { if len(buf) < linux.SizeOfXTUDP { return nil, fmt.Errorf("buf has insufficient size for UDP match: %d", len(buf)) } @@ -94,11 +93,11 @@ func (*UDPMatcher) Name() string { } // Match implements Matcher.Match. -func (um *UDPMatcher) Match(hook iptables.Hook, pkt tcpip.PacketBuffer, interfaceName string) (bool, bool) { +func (um *UDPMatcher) Match(hook stack.Hook, pkt stack.PacketBuffer, interfaceName string) (bool, bool) { netHeader := header.IPv4(pkt.NetworkHeader) // TODO(gvisor.dev/issue/170): Proto checks should ultimately be moved - // into the iptables.Check codepath as matchers are added. + // into the stack.Check codepath as matchers are added. if netHeader.TransportProtocol() != header.UDPProtocolNumber { return false, false } @@ -114,7 +113,7 @@ func (um *UDPMatcher) Match(hook iptables.Hook, pkt tcpip.PacketBuffer, interfac // Now we need the transport header. However, this may not have been set // yet. // TODO(gvisor.dev/issue/170): Parsing the transport header should - // ultimately be moved into the iptables.Check codepath as matchers are + // ultimately be moved into the stack.Check codepath as matchers are // added. var udpHeader header.UDP if pkt.TransportHeader != nil { diff --git a/pkg/sentry/socket/netlink/message.go b/pkg/sentry/socket/netlink/message.go index 4ea252ccb..0899c61d1 100644 --- a/pkg/sentry/socket/netlink/message.go +++ b/pkg/sentry/socket/netlink/message.go @@ -23,18 +23,11 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) -// alignUp rounds a length up to an alignment. -// -// Preconditions: align is a power of two. -func alignUp(length int, align uint) int { - return (length + int(align) - 1) &^ (int(align) - 1) -} - // alignPad returns the length of padding required for alignment. // // Preconditions: align is a power of two. func alignPad(length int, align uint) int { - return alignUp(length, align) - length + return binary.AlignUp(length, align) - length } // Message contains a complete serialized netlink message. @@ -138,7 +131,7 @@ func (m *Message) Finalize() []byte { // Align the message. Note that the message length in the header (set // above) is the useful length of the message, not the total aligned // length. See net/netlink/af_netlink.c:__nlmsg_put. - aligned := alignUp(len(m.buf), linux.NLMSG_ALIGNTO) + aligned := binary.AlignUp(len(m.buf), linux.NLMSG_ALIGNTO) m.putZeros(aligned - len(m.buf)) return m.buf } @@ -173,7 +166,7 @@ func (m *Message) PutAttr(atype uint16, v interface{}) { m.Put(v) // Align the attribute. - aligned := alignUp(l, linux.NLA_ALIGNTO) + aligned := binary.AlignUp(l, linux.NLA_ALIGNTO) m.putZeros(aligned - l) } @@ -190,7 +183,7 @@ func (m *Message) PutAttrString(atype uint16, s string) { m.putZeros(1) // Align the attribute. - aligned := alignUp(l, linux.NLA_ALIGNTO) + aligned := binary.AlignUp(l, linux.NLA_ALIGNTO) m.putZeros(aligned - l) } diff --git a/pkg/sentry/socket/netstack/BUILD b/pkg/sentry/socket/netstack/BUILD index ab01cb4fa..cbf46b1e9 100644 --- a/pkg/sentry/socket/netstack/BUILD +++ b/pkg/sentry/socket/netstack/BUILD @@ -38,7 +38,6 @@ go_library( "//pkg/tcpip", "//pkg/tcpip/buffer", "//pkg/tcpip/header", - "//pkg/tcpip/iptables", "//pkg/tcpip/network/ipv4", "//pkg/tcpip/network/ipv6", "//pkg/tcpip/stack", diff --git a/pkg/sentry/socket/netstack/netstack.go b/pkg/sentry/socket/netstack/netstack.go index ed2fbcceb..d5879c10f 100644 --- a/pkg/sentry/socket/netstack/netstack.go +++ b/pkg/sentry/socket/netstack/netstack.go @@ -29,6 +29,7 @@ import ( "io" "math" "reflect" + "sync/atomic" "syscall" "time" @@ -62,7 +63,13 @@ import ( func mustCreateMetric(name, description string) *tcpip.StatCounter { var cm tcpip.StatCounter - metric.MustRegisterCustomUint64Metric(name, false /* sync */, description, cm.Value) + metric.MustRegisterCustomUint64Metric(name, true /* cumulative */, false /* sync */, description, cm.Value) + return &cm +} + +func mustCreateGauge(name, description string) *tcpip.StatCounter { + var cm tcpip.StatCounter + metric.MustRegisterCustomUint64Metric(name, false /* cumulative */, false /* sync */, description, cm.Value) return &cm } @@ -150,10 +157,10 @@ var Metrics = tcpip.Stats{ TCP: tcpip.TCPStats{ ActiveConnectionOpenings: mustCreateMetric("/netstack/tcp/active_connection_openings", "Number of connections opened successfully via Connect."), PassiveConnectionOpenings: mustCreateMetric("/netstack/tcp/passive_connection_openings", "Number of connections opened successfully via Listen."), - CurrentEstablished: mustCreateMetric("/netstack/tcp/current_established", "Number of connections in ESTABLISHED state now."), - CurrentConnected: mustCreateMetric("/netstack/tcp/current_open", "Number of connections that are in connected state."), + CurrentEstablished: mustCreateGauge("/netstack/tcp/current_established", "Number of connections in ESTABLISHED state now."), + CurrentConnected: mustCreateGauge("/netstack/tcp/current_open", "Number of connections that are in connected state."), EstablishedResets: mustCreateMetric("/netstack/tcp/established_resets", "Number of times TCP connections have made a direct transition to the CLOSED state from either the ESTABLISHED state or the CLOSE-WAIT state"), - EstablishedClosed: mustCreateMetric("/netstack/tcp/established_closed", "number of times established TCP connections made a transition to CLOSED state."), + EstablishedClosed: mustCreateMetric("/netstack/tcp/established_closed", "Number of times established TCP connections made a transition to CLOSED state."), EstablishedTimedout: mustCreateMetric("/netstack/tcp/established_timedout", "Number of times an established connection was reset because of keep-alive time out."), ListenOverflowSynDrop: mustCreateMetric("/netstack/tcp/listen_overflow_syn_drop", "Number of times the listen queue overflowed and a SYN was dropped."), ListenOverflowAckDrop: mustCreateMetric("/netstack/tcp/listen_overflow_ack_drop", "Number of times the listen queue overflowed and the final ACK in the handshake was dropped."), @@ -264,6 +271,12 @@ type SocketOperations struct { skType linux.SockType protocol int + // readViewHasData is 1 iff readView has data to be read, 0 otherwise. + // Must be accessed using atomic operations. It must only be written + // with readMu held but can be read without holding readMu. The latter + // is required to avoid deadlocks in epoll Readiness checks. + readViewHasData uint32 + // readMu protects access to the below fields. readMu sync.Mutex `state:"nosave"` // readView contains the remaining payload from the last packet. @@ -293,7 +306,7 @@ type SocketOperations struct { // New creates a new endpoint socket. func New(t *kernel.Task, family int, skType linux.SockType, protocol int, queue *waiter.Queue, endpoint tcpip.Endpoint) (*fs.File, *syserr.Error) { if skType == linux.SOCK_STREAM { - if err := endpoint.SetSockOptInt(tcpip.DelayOption, 1); err != nil { + if err := endpoint.SetSockOptBool(tcpip.DelayOption, true); err != nil { return nil, syserr.TranslateNetstackError(err) } } @@ -410,21 +423,24 @@ func (s *SocketOperations) isPacketBased() bool { // fetchReadView updates the readView field of the socket if it's currently // empty. It assumes that the socket is locked. +// +// Precondition: s.readMu must be held. func (s *SocketOperations) fetchReadView() *syserr.Error { if len(s.readView) > 0 { return nil } - s.readView = nil s.sender = tcpip.FullAddress{} v, cms, err := s.Endpoint.Read(&s.sender) if err != nil { + atomic.StoreUint32(&s.readViewHasData, 0) return syserr.TranslateNetstackError(err) } s.readView = v s.readCM = cms + atomic.StoreUint32(&s.readViewHasData, 1) return nil } @@ -525,7 +541,7 @@ func (s *SocketOperations) Write(ctx context.Context, _ *fs.File, src usermem.IO } if resCh != nil { - t := ctx.(*kernel.Task) + t := kernel.TaskFromContext(ctx) if err := t.Block(resCh); err != nil { return 0, syserr.FromError(err).ToError() } @@ -598,7 +614,7 @@ func (s *SocketOperations) ReadFrom(ctx context.Context, _ *fs.File, r io.Reader } if resCh != nil { - t := ctx.(*kernel.Task) + t := kernel.TaskFromContext(ctx) if err := t.Block(resCh); err != nil { return 0, syserr.FromError(err).ToError() } @@ -623,11 +639,9 @@ func (s *SocketOperations) Readiness(mask waiter.EventMask) waiter.EventMask { // Check our cached value iff the caller asked for readability and the // endpoint itself is currently not readable. if (mask & ^r & waiter.EventIn) != 0 { - s.readMu.Lock() - if len(s.readView) > 0 { + if atomic.LoadUint32(&s.readViewHasData) == 1 { r |= waiter.EventIn } - s.readMu.Unlock() } return r @@ -655,7 +669,7 @@ func (s *SocketOperations) checkFamily(family uint16, exact bool) *syserr.Error // This is a hack to work around the fact that both IPv4 and IPv6 ANY are // represented by the empty string. // -// TODO(gvisor.dev/issues/1556): remove this function. +// TODO(gvisor.dev/issue/1556): remove this function. func (s *SocketOperations) mapFamily(addr tcpip.FullAddress, family uint16) tcpip.FullAddress { if len(addr.Addr) == 0 && s.family == linux.AF_INET6 && family == linux.AF_INET { addr.Addr = "\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\xff\xff\x00\x00\x00\x00" @@ -712,14 +726,44 @@ func (s *SocketOperations) Connect(t *kernel.Task, sockaddr []byte, blocking boo // Bind implements the linux syscall bind(2) for sockets backed by // tcpip.Endpoint. func (s *SocketOperations) Bind(t *kernel.Task, sockaddr []byte) *syserr.Error { - addr, family, err := AddressAndFamily(sockaddr) - if err != nil { - return err + if len(sockaddr) < 2 { + return syserr.ErrInvalidArgument } - if err := s.checkFamily(family, true /* exact */); err != nil { - return err + + family := usermem.ByteOrder.Uint16(sockaddr) + var addr tcpip.FullAddress + + // Bind for AF_PACKET requires only family, protocol and ifindex. + // In function AddressAndFamily, we check the address length which is + // not needed for AF_PACKET bind. + if family == linux.AF_PACKET { + var a linux.SockAddrLink + if len(sockaddr) < sockAddrLinkSize { + return syserr.ErrInvalidArgument + } + binary.Unmarshal(sockaddr[:sockAddrLinkSize], usermem.ByteOrder, &a) + + if a.Protocol != uint16(s.protocol) { + return syserr.ErrInvalidArgument + } + + addr = tcpip.FullAddress{ + NIC: tcpip.NICID(a.InterfaceIndex), + Addr: tcpip.Address(a.HardwareAddr[:header.EthernetAddressSize]), + } + } else { + var err *syserr.Error + addr, family, err = AddressAndFamily(sockaddr) + if err != nil { + return err + } + + if err = s.checkFamily(family, true /* exact */); err != nil { + return err + } + + addr = s.mapFamily(addr, family) } - addr = s.mapFamily(addr, family) // Issue the bind request to the endpoint. return syserr.TranslateNetstackError(s.Endpoint.Bind(addr)) @@ -902,7 +946,7 @@ func (s *SocketOperations) GetSockOpt(t *kernel.Task, level, name int, outPtr us // GetSockOpt can be used to implement the linux syscall getsockopt(2) for // sockets backed by a commonEndpoint. -func GetSockOpt(t *kernel.Task, s socket.Socket, ep commonEndpoint, family int, skType linux.SockType, level, name, outLen int) (interface{}, *syserr.Error) { +func GetSockOpt(t *kernel.Task, s socket.SocketOps, ep commonEndpoint, family int, skType linux.SockType, level, name, outLen int) (interface{}, *syserr.Error) { switch level { case linux.SOL_SOCKET: return getSockOptSocket(t, s, ep, family, skType, name, outLen) @@ -927,8 +971,15 @@ func GetSockOpt(t *kernel.Task, s socket.Socket, ep commonEndpoint, family int, return nil, syserr.ErrProtocolNotAvailable } +func boolToInt32(v bool) int32 { + if v { + return 1 + } + return 0 +} + // getSockOptSocket implements GetSockOpt when level is SOL_SOCKET. -func getSockOptSocket(t *kernel.Task, s socket.Socket, ep commonEndpoint, family int, skType linux.SockType, name, outLen int) (interface{}, *syserr.Error) { +func getSockOptSocket(t *kernel.Task, s socket.SocketOps, ep commonEndpoint, family int, skType linux.SockType, name, outLen int) (interface{}, *syserr.Error) { // TODO(b/124056281): Stop rejecting short optLen values in getsockopt. switch name { case linux.SO_ERROR: @@ -960,12 +1011,11 @@ func getSockOptSocket(t *kernel.Task, s socket.Socket, ep commonEndpoint, family return nil, syserr.ErrInvalidArgument } - var v tcpip.PasscredOption - if err := ep.GetSockOpt(&v); err != nil { + v, err := ep.GetSockOptBool(tcpip.PasscredOption) + if err != nil { return nil, syserr.TranslateNetstackError(err) } - - return int32(v), nil + return boolToInt32(v), nil case linux.SO_SNDBUF: if outLen < sizeOfInt32 { @@ -1004,24 +1054,22 @@ func getSockOptSocket(t *kernel.Task, s socket.Socket, ep commonEndpoint, family return nil, syserr.ErrInvalidArgument } - var v tcpip.ReuseAddressOption - if err := ep.GetSockOpt(&v); err != nil { + v, err := ep.GetSockOptBool(tcpip.ReuseAddressOption) + if err != nil { return nil, syserr.TranslateNetstackError(err) } - - return int32(v), nil + return boolToInt32(v), nil case linux.SO_REUSEPORT: if outLen < sizeOfInt32 { return nil, syserr.ErrInvalidArgument } - var v tcpip.ReusePortOption - if err := ep.GetSockOpt(&v); err != nil { + v, err := ep.GetSockOptBool(tcpip.ReusePortOption) + if err != nil { return nil, syserr.TranslateNetstackError(err) } - - return int32(v), nil + return boolToInt32(v), nil case linux.SO_BINDTODEVICE: var v tcpip.BindToDeviceOption @@ -1051,24 +1099,22 @@ func getSockOptSocket(t *kernel.Task, s socket.Socket, ep commonEndpoint, family return nil, syserr.ErrInvalidArgument } - var v tcpip.BroadcastOption - if err := ep.GetSockOpt(&v); err != nil { + v, err := ep.GetSockOptBool(tcpip.BroadcastOption) + if err != nil { return nil, syserr.TranslateNetstackError(err) } - - return int32(v), nil + return boolToInt32(v), nil case linux.SO_KEEPALIVE: if outLen < sizeOfInt32 { return nil, syserr.ErrInvalidArgument } - var v tcpip.KeepaliveEnabledOption - if err := ep.GetSockOpt(&v); err != nil { + v, err := ep.GetSockOptBool(tcpip.KeepaliveEnabledOption) + if err != nil { return nil, syserr.TranslateNetstackError(err) } - - return int32(v), nil + return boolToInt32(v), nil case linux.SO_LINGER: if outLen < linux.SizeOfLinger { @@ -1118,47 +1164,41 @@ func getSockOptTCP(t *kernel.Task, ep commonEndpoint, name, outLen int) (interfa return nil, syserr.ErrInvalidArgument } - v, err := ep.GetSockOptInt(tcpip.DelayOption) + v, err := ep.GetSockOptBool(tcpip.DelayOption) if err != nil { return nil, syserr.TranslateNetstackError(err) } - - if v == 0 { - return int32(1), nil - } - return int32(0), nil + return boolToInt32(!v), nil case linux.TCP_CORK: if outLen < sizeOfInt32 { return nil, syserr.ErrInvalidArgument } - var v tcpip.CorkOption - if err := ep.GetSockOpt(&v); err != nil { + v, err := ep.GetSockOptBool(tcpip.CorkOption) + if err != nil { return nil, syserr.TranslateNetstackError(err) } - - return int32(v), nil + return boolToInt32(v), nil case linux.TCP_QUICKACK: if outLen < sizeOfInt32 { return nil, syserr.ErrInvalidArgument } - var v tcpip.QuickAckOption - if err := ep.GetSockOpt(&v); err != nil { + v, err := ep.GetSockOptBool(tcpip.QuickAckOption) + if err != nil { return nil, syserr.TranslateNetstackError(err) } - - return int32(v), nil + return boolToInt32(v), nil case linux.TCP_MAXSEG: if outLen < sizeOfInt32 { return nil, syserr.ErrInvalidArgument } - var v tcpip.MaxSegOption - if err := ep.GetSockOpt(&v); err != nil { + v, err := ep.GetSockOptInt(tcpip.MaxSegOption) + if err != nil { return nil, syserr.TranslateNetstackError(err) } @@ -1290,11 +1330,7 @@ func getSockOptIPv6(t *kernel.Task, ep commonEndpoint, name, outLen int) (interf if err != nil { return nil, syserr.TranslateNetstackError(err) } - var o int32 - if v { - o = 1 - } - return o, nil + return boolToInt32(v), nil case linux.IPV6_PATHMTU: t.Kernel().EmitUnimplementedEvent(t) @@ -1304,8 +1340,8 @@ func getSockOptIPv6(t *kernel.Task, ep commonEndpoint, name, outLen int) (interf if outLen == 0 { return make([]byte, 0), nil } - var v tcpip.IPv6TrafficClassOption - if err := ep.GetSockOpt(&v); err != nil { + v, err := ep.GetSockOptInt(tcpip.IPv6TrafficClassOption) + if err != nil { return nil, syserr.TranslateNetstackError(err) } @@ -1318,6 +1354,17 @@ func getSockOptIPv6(t *kernel.Task, ep commonEndpoint, name, outLen int) (interf } return ib, nil + case linux.IPV6_RECVTCLASS: + if outLen < sizeOfInt32 { + return nil, syserr.ErrInvalidArgument + } + + v, err := ep.GetSockOptBool(tcpip.ReceiveTClassOption) + if err != nil { + return nil, syserr.TranslateNetstackError(err) + } + return boolToInt32(v), nil + default: emitUnimplementedEventIPv6(t, name) } @@ -1332,8 +1379,8 @@ func getSockOptIP(t *kernel.Task, ep commonEndpoint, name, outLen int, family in return nil, syserr.ErrInvalidArgument } - var v tcpip.TTLOption - if err := ep.GetSockOpt(&v); err != nil { + v, err := ep.GetSockOptInt(tcpip.TTLOption) + if err != nil { return nil, syserr.TranslateNetstackError(err) } @@ -1349,8 +1396,8 @@ func getSockOptIP(t *kernel.Task, ep commonEndpoint, name, outLen int, family in return nil, syserr.ErrInvalidArgument } - var v tcpip.MulticastTTLOption - if err := ep.GetSockOpt(&v); err != nil { + v, err := ep.GetSockOptInt(tcpip.MulticastTTLOption) + if err != nil { return nil, syserr.TranslateNetstackError(err) } @@ -1375,23 +1422,19 @@ func getSockOptIP(t *kernel.Task, ep commonEndpoint, name, outLen int, family in return nil, syserr.ErrInvalidArgument } - var v tcpip.MulticastLoopOption - if err := ep.GetSockOpt(&v); err != nil { + v, err := ep.GetSockOptBool(tcpip.MulticastLoopOption) + if err != nil { return nil, syserr.TranslateNetstackError(err) } - - if v { - return int32(1), nil - } - return int32(0), nil + return boolToInt32(v), nil case linux.IP_TOS: // Length handling for parity with Linux. if outLen == 0 { return []byte(nil), nil } - var v tcpip.IPv4TOSOption - if err := ep.GetSockOpt(&v); err != nil { + v, err := ep.GetSockOptInt(tcpip.IPv4TOSOption) + if err != nil { return nil, syserr.TranslateNetstackError(err) } if outLen < sizeOfInt32 { @@ -1408,11 +1451,18 @@ func getSockOptIP(t *kernel.Task, ep commonEndpoint, name, outLen int, family in if err != nil { return nil, syserr.TranslateNetstackError(err) } - var o int32 - if v { - o = 1 + return boolToInt32(v), nil + + case linux.IP_PKTINFO: + if outLen < sizeOfInt32 { + return nil, syserr.ErrInvalidArgument + } + + v, err := ep.GetSockOptBool(tcpip.ReceiveIPPacketInfoOption) + if err != nil { + return nil, syserr.TranslateNetstackError(err) } - return o, nil + return boolToInt32(v), nil default: emitUnimplementedEventIP(t, name) @@ -1472,7 +1522,7 @@ func (s *SocketOperations) SetSockOpt(t *kernel.Task, level int, name int, optVa // SetSockOpt can be used to implement the linux syscall setsockopt(2) for // sockets backed by a commonEndpoint. -func SetSockOpt(t *kernel.Task, s socket.Socket, ep commonEndpoint, level int, name int, optVal []byte) *syserr.Error { +func SetSockOpt(t *kernel.Task, s socket.SocketOps, ep commonEndpoint, level int, name int, optVal []byte) *syserr.Error { switch level { case linux.SOL_SOCKET: return setSockOptSocket(t, s, ep, name, optVal) @@ -1499,7 +1549,7 @@ func SetSockOpt(t *kernel.Task, s socket.Socket, ep commonEndpoint, level int, n } // setSockOptSocket implements SetSockOpt when level is SOL_SOCKET. -func setSockOptSocket(t *kernel.Task, s socket.Socket, ep commonEndpoint, name int, optVal []byte) *syserr.Error { +func setSockOptSocket(t *kernel.Task, s socket.SocketOps, ep commonEndpoint, name int, optVal []byte) *syserr.Error { switch name { case linux.SO_SNDBUF: if len(optVal) < sizeOfInt32 { @@ -1523,7 +1573,7 @@ func setSockOptSocket(t *kernel.Task, s socket.Socket, ep commonEndpoint, name i } v := usermem.ByteOrder.Uint32(optVal) - return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.ReuseAddressOption(v))) + return syserr.TranslateNetstackError(ep.SetSockOptBool(tcpip.ReuseAddressOption, v != 0)) case linux.SO_REUSEPORT: if len(optVal) < sizeOfInt32 { @@ -1531,7 +1581,7 @@ func setSockOptSocket(t *kernel.Task, s socket.Socket, ep commonEndpoint, name i } v := usermem.ByteOrder.Uint32(optVal) - return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.ReusePortOption(v))) + return syserr.TranslateNetstackError(ep.SetSockOptBool(tcpip.ReusePortOption, v != 0)) case linux.SO_BINDTODEVICE: n := bytes.IndexByte(optVal, 0) @@ -1559,7 +1609,7 @@ func setSockOptSocket(t *kernel.Task, s socket.Socket, ep commonEndpoint, name i } v := usermem.ByteOrder.Uint32(optVal) - return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.BroadcastOption(v))) + return syserr.TranslateNetstackError(ep.SetSockOptBool(tcpip.BroadcastOption, v != 0)) case linux.SO_PASSCRED: if len(optVal) < sizeOfInt32 { @@ -1567,7 +1617,7 @@ func setSockOptSocket(t *kernel.Task, s socket.Socket, ep commonEndpoint, name i } v := usermem.ByteOrder.Uint32(optVal) - return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.PasscredOption(v))) + return syserr.TranslateNetstackError(ep.SetSockOptBool(tcpip.PasscredOption, v != 0)) case linux.SO_KEEPALIVE: if len(optVal) < sizeOfInt32 { @@ -1575,7 +1625,7 @@ func setSockOptSocket(t *kernel.Task, s socket.Socket, ep commonEndpoint, name i } v := usermem.ByteOrder.Uint32(optVal) - return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.KeepaliveEnabledOption(v))) + return syserr.TranslateNetstackError(ep.SetSockOptBool(tcpip.KeepaliveEnabledOption, v != 0)) case linux.SO_SNDTIMEO: if len(optVal) < linux.SizeOfTimeval { @@ -1647,11 +1697,7 @@ func setSockOptTCP(t *kernel.Task, ep commonEndpoint, name int, optVal []byte) * } v := usermem.ByteOrder.Uint32(optVal) - var o int - if v == 0 { - o = 1 - } - return syserr.TranslateNetstackError(ep.SetSockOptInt(tcpip.DelayOption, o)) + return syserr.TranslateNetstackError(ep.SetSockOptBool(tcpip.DelayOption, v == 0)) case linux.TCP_CORK: if len(optVal) < sizeOfInt32 { @@ -1659,7 +1705,7 @@ func setSockOptTCP(t *kernel.Task, ep commonEndpoint, name int, optVal []byte) * } v := usermem.ByteOrder.Uint32(optVal) - return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.CorkOption(v))) + return syserr.TranslateNetstackError(ep.SetSockOptBool(tcpip.CorkOption, v != 0)) case linux.TCP_QUICKACK: if len(optVal) < sizeOfInt32 { @@ -1667,7 +1713,7 @@ func setSockOptTCP(t *kernel.Task, ep commonEndpoint, name int, optVal []byte) * } v := usermem.ByteOrder.Uint32(optVal) - return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.QuickAckOption(v))) + return syserr.TranslateNetstackError(ep.SetSockOptBool(tcpip.QuickAckOption, v != 0)) case linux.TCP_MAXSEG: if len(optVal) < sizeOfInt32 { @@ -1675,7 +1721,7 @@ func setSockOptTCP(t *kernel.Task, ep commonEndpoint, name int, optVal []byte) * } v := usermem.ByteOrder.Uint32(optVal) - return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.MaxSegOption(v))) + return syserr.TranslateNetstackError(ep.SetSockOptInt(tcpip.MaxSegOption, int(v))) case linux.TCP_KEEPIDLE: if len(optVal) < sizeOfInt32 { @@ -1762,6 +1808,7 @@ func setSockOptIPv6(t *kernel.Task, ep commonEndpoint, name int, optVal []byte) linux.IPV6_IPSEC_POLICY, linux.IPV6_JOIN_ANYCAST, linux.IPV6_LEAVE_ANYCAST, + // TODO(b/148887420): Add support for IPV6_PKTINFO. linux.IPV6_PKTINFO, linux.IPV6_ROUTER_ALERT, linux.IPV6_XFRM_POLICY, @@ -1785,7 +1832,15 @@ func setSockOptIPv6(t *kernel.Task, ep commonEndpoint, name int, optVal []byte) if v == -1 { v = 0 } - return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.IPv6TrafficClassOption(v))) + return syserr.TranslateNetstackError(ep.SetSockOptInt(tcpip.IPv6TrafficClassOption, int(v))) + + case linux.IPV6_RECVTCLASS: + v, err := parseIntOrChar(optVal) + if err != nil { + return err + } + + return syserr.TranslateNetstackError(ep.SetSockOptBool(tcpip.ReceiveTClassOption, v != 0)) default: emitUnimplementedEventIPv6(t, name) @@ -1862,7 +1917,7 @@ func setSockOptIP(t *kernel.Task, ep commonEndpoint, name int, optVal []byte) *s if v < 0 || v > 255 { return syserr.ErrInvalidArgument } - return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.MulticastTTLOption(v))) + return syserr.TranslateNetstackError(ep.SetSockOptInt(tcpip.MulticastTTLOption, int(v))) case linux.IP_ADD_MEMBERSHIP: req, err := copyInMulticastRequest(optVal, false /* allowAddr */) @@ -1909,9 +1964,7 @@ func setSockOptIP(t *kernel.Task, ep commonEndpoint, name int, optVal []byte) *s return err } - return syserr.TranslateNetstackError(ep.SetSockOpt( - tcpip.MulticastLoopOption(v != 0), - )) + return syserr.TranslateNetstackError(ep.SetSockOptBool(tcpip.MulticastLoopOption, v != 0)) case linux.MCAST_JOIN_GROUP: // FIXME(b/124219304): Implement MCAST_JOIN_GROUP. @@ -1930,7 +1983,7 @@ func setSockOptIP(t *kernel.Task, ep commonEndpoint, name int, optVal []byte) *s } else if v < 1 || v > 255 { return syserr.ErrInvalidArgument } - return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.TTLOption(v))) + return syserr.TranslateNetstackError(ep.SetSockOptInt(tcpip.TTLOption, int(v))) case linux.IP_TOS: if len(optVal) == 0 { @@ -1940,7 +1993,7 @@ func setSockOptIP(t *kernel.Task, ep commonEndpoint, name int, optVal []byte) *s if err != nil { return err } - return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.IPv4TOSOption(v))) + return syserr.TranslateNetstackError(ep.SetSockOptInt(tcpip.IPv4TOSOption, int(v))) case linux.IP_RECVTOS: v, err := parseIntOrChar(optVal) @@ -1949,6 +2002,16 @@ func setSockOptIP(t *kernel.Task, ep commonEndpoint, name int, optVal []byte) *s } return syserr.TranslateNetstackError(ep.SetSockOptBool(tcpip.ReceiveTOSOption, v != 0)) + case linux.IP_PKTINFO: + if len(optVal) == 0 { + return nil + } + v, err := parseIntOrChar(optVal) + if err != nil { + return err + } + return syserr.TranslateNetstackError(ep.SetSockOptBool(tcpip.ReceiveIPPacketInfoOption, v != 0)) + case linux.IP_ADD_SOURCE_MEMBERSHIP, linux.IP_BIND_ADDRESS_NO_PORT, linux.IP_BLOCK_SOURCE, @@ -1964,7 +2027,6 @@ func setSockOptIP(t *kernel.Task, ep commonEndpoint, name int, optVal []byte) *s linux.IP_NODEFRAG, linux.IP_OPTIONS, linux.IP_PASSSEC, - linux.IP_PKTINFO, linux.IP_RECVERR, linux.IP_RECVFRAGSIZE, linux.IP_RECVOPTS, @@ -2061,7 +2123,6 @@ func emitUnimplementedEventIPv6(t *kernel.Task, name int) { linux.IPV6_RECVPATHMTU, linux.IPV6_RECVPKTINFO, linux.IPV6_RECVRTHDR, - linux.IPV6_RECVTCLASS, linux.IPV6_RTHDR, linux.IPV6_RTHDRDSTOPTS, linux.IPV6_TCLASS, @@ -2256,6 +2317,10 @@ func (s *SocketOperations) coalescingRead(ctx context.Context, dst usermem.IOSeq } copied += n s.readView.TrimFront(n) + if len(s.readView) == 0 { + atomic.StoreUint32(&s.readViewHasData, 0) + } + dst = dst.DropFirst(n) if e != nil { err = syserr.FromError(e) @@ -2302,9 +2367,9 @@ func (s *SocketOperations) nonBlockingRead(ctx context.Context, dst usermem.IOSe // caller-supplied buffer. s.readMu.Lock() n, err := s.coalescingRead(ctx, dst, trunc) - s.readMu.Unlock() cmsg := s.controlMessages() s.fillCmsgInq(&cmsg) + s.readMu.Unlock() return n, 0, nil, 0, cmsg, err } @@ -2378,6 +2443,10 @@ func (s *SocketOperations) nonBlockingRead(ctx context.Context, dst usermem.IOSe s.readView.TrimFront(int(n)) } + if len(s.readView) == 0 { + atomic.StoreUint32(&s.readViewHasData, 0) + } + var flags int if msgLen > int(n) { flags |= linux.MSG_TRUNC @@ -2395,10 +2464,14 @@ func (s *SocketOperations) nonBlockingRead(ctx context.Context, dst usermem.IOSe func (s *SocketOperations) controlMessages() socket.ControlMessages { return socket.ControlMessages{ IP: tcpip.ControlMessages{ - HasTimestamp: s.readCM.HasTimestamp && s.sockOptTimestamp, - Timestamp: s.readCM.Timestamp, - HasTOS: s.readCM.HasTOS, - TOS: s.readCM.TOS, + HasTimestamp: s.readCM.HasTimestamp && s.sockOptTimestamp, + Timestamp: s.readCM.Timestamp, + HasTOS: s.readCM.HasTOS, + TOS: s.readCM.TOS, + HasTClass: s.readCM.HasTClass, + TClass: s.readCM.TClass, + HasIPPacketInfo: s.readCM.HasIPPacketInfo, + PacketInfo: s.readCM.PacketInfo, }, } } @@ -2585,7 +2658,9 @@ func (s *SocketOperations) Ioctl(ctx context.Context, _ *fs.File, io usermem.IO, } // Add bytes removed from the endpoint but not yet sent to the caller. + s.readMu.Lock() v += len(s.readView) + s.readMu.Unlock() if v > math.MaxInt32 { v = math.MaxInt32 diff --git a/pkg/sentry/socket/netstack/provider.go b/pkg/sentry/socket/netstack/provider.go index 5afff2564..c3f04b613 100644 --- a/pkg/sentry/socket/netstack/provider.go +++ b/pkg/sentry/socket/netstack/provider.go @@ -62,10 +62,6 @@ func getTransportProtocol(ctx context.Context, stype linux.SockType, protocol in } case linux.SOCK_RAW: - // TODO(b/142504697): "In order to create a raw socket, a - // process must have the CAP_NET_RAW capability in the user - // namespace that governs its network namespace." - raw(7) - // Raw sockets require CAP_NET_RAW. creds := auth.CredentialsFromContext(ctx) if !creds.HasCapability(linux.CAP_NET_RAW) { @@ -75,6 +71,8 @@ func getTransportProtocol(ctx context.Context, stype linux.SockType, protocol in switch protocol { case syscall.IPPROTO_ICMP: return header.ICMPv4ProtocolNumber, true, nil + case syscall.IPPROTO_ICMPV6: + return header.ICMPv6ProtocolNumber, true, nil case syscall.IPPROTO_UDP: return header.UDPProtocolNumber, true, nil case syscall.IPPROTO_TCP: @@ -124,6 +122,12 @@ func (p *provider) Socket(t *kernel.Task, stype linux.SockType, protocol int) (* ep, e = eps.Stack.NewRawEndpoint(transProto, p.netProto, wq, associated) } else { ep, e = eps.Stack.NewEndpoint(transProto, p.netProto, wq) + + // Assign task to PacketOwner interface to get the UID and GID for + // iptables owner matching. + if e == nil { + ep.SetOwner(t) + } } if e != nil { return nil, syserr.TranslateNetstackError(e) @@ -133,10 +137,6 @@ func (p *provider) Socket(t *kernel.Task, stype linux.SockType, protocol int) (* } func packetSocket(t *kernel.Task, epStack *Stack, stype linux.SockType, protocol int) (*fs.File, *syserr.Error) { - // TODO(b/142504697): "In order to create a packet socket, a process - // must have the CAP_NET_RAW capability in the user namespace that - // governs its network namespace." - packet(7) - // Packet sockets require CAP_NET_RAW. creds := auth.CredentialsFromContext(t) if !creds.HasCapability(linux.CAP_NET_RAW) { diff --git a/pkg/sentry/socket/netstack/stack.go b/pkg/sentry/socket/netstack/stack.go index 0692482e9..f5fa18136 100644 --- a/pkg/sentry/socket/netstack/stack.go +++ b/pkg/sentry/socket/netstack/stack.go @@ -23,7 +23,6 @@ import ( "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/tcpip" "gvisor.dev/gvisor/pkg/tcpip/header" - "gvisor.dev/gvisor/pkg/tcpip/iptables" "gvisor.dev/gvisor/pkg/tcpip/network/ipv4" "gvisor.dev/gvisor/pkg/tcpip/network/ipv6" "gvisor.dev/gvisor/pkg/tcpip/stack" @@ -200,36 +199,66 @@ func (s *Stack) SetTCPSACKEnabled(enabled bool) error { // Statistics implements inet.Stack.Statistics. func (s *Stack) Statistics(stat interface{}, arg string) error { switch stats := stat.(type) { + case *inet.StatDev: + for _, ni := range s.Stack.NICInfo() { + if ni.Name != arg { + continue + } + // TODO(gvisor.dev/issue/2103) Support stubbed stats. + *stats = inet.StatDev{ + // Receive section. + ni.Stats.Rx.Bytes.Value(), // bytes. + ni.Stats.Rx.Packets.Value(), // packets. + 0, // errs. + 0, // drop. + 0, // fifo. + 0, // frame. + 0, // compressed. + 0, // multicast. + // Transmit section. + ni.Stats.Tx.Bytes.Value(), // bytes. + ni.Stats.Tx.Packets.Value(), // packets. + 0, // errs. + 0, // drop. + 0, // fifo. + 0, // colls. + 0, // carrier. + 0, // compressed. + } + break + } case *inet.StatSNMPIP: ip := Metrics.IP + // TODO(gvisor.dev/issue/969) Support stubbed stats. *stats = inet.StatSNMPIP{ - 0, // TODO(gvisor.dev/issue/969): Support Ip/Forwarding. - 0, // TODO(gvisor.dev/issue/969): Support Ip/DefaultTTL. + 0, // Ip/Forwarding. + 0, // Ip/DefaultTTL. ip.PacketsReceived.Value(), // InReceives. - 0, // TODO(gvisor.dev/issue/969): Support Ip/InHdrErrors. + 0, // Ip/InHdrErrors. ip.InvalidDestinationAddressesReceived.Value(), // InAddrErrors. - 0, // TODO(gvisor.dev/issue/969): Support Ip/ForwDatagrams. - 0, // TODO(gvisor.dev/issue/969): Support Ip/InUnknownProtos. - 0, // TODO(gvisor.dev/issue/969): Support Ip/InDiscards. + 0, // Ip/ForwDatagrams. + 0, // Ip/InUnknownProtos. + 0, // Ip/InDiscards. ip.PacketsDelivered.Value(), // InDelivers. ip.PacketsSent.Value(), // OutRequests. ip.OutgoingPacketErrors.Value(), // OutDiscards. - 0, // TODO(gvisor.dev/issue/969): Support Ip/OutNoRoutes. - 0, // TODO(gvisor.dev/issue/969): Support Ip/ReasmTimeout. - 0, // TODO(gvisor.dev/issue/969): Support Ip/ReasmReqds. - 0, // TODO(gvisor.dev/issue/969): Support Ip/ReasmOKs. - 0, // TODO(gvisor.dev/issue/969): Support Ip/ReasmFails. - 0, // TODO(gvisor.dev/issue/969): Support Ip/FragOKs. - 0, // TODO(gvisor.dev/issue/969): Support Ip/FragFails. - 0, // TODO(gvisor.dev/issue/969): Support Ip/FragCreates. + 0, // Ip/OutNoRoutes. + 0, // Support Ip/ReasmTimeout. + 0, // Support Ip/ReasmReqds. + 0, // Support Ip/ReasmOKs. + 0, // Support Ip/ReasmFails. + 0, // Support Ip/FragOKs. + 0, // Support Ip/FragFails. + 0, // Support Ip/FragCreates. } case *inet.StatSNMPICMP: in := Metrics.ICMP.V4PacketsReceived.ICMPv4PacketStats out := Metrics.ICMP.V4PacketsSent.ICMPv4PacketStats + // TODO(gvisor.dev/issue/969) Support stubbed stats. *stats = inet.StatSNMPICMP{ - 0, // TODO(gvisor.dev/issue/969): Support Icmp/InMsgs. + 0, // Icmp/InMsgs. Metrics.ICMP.V4PacketsSent.Dropped.Value(), // InErrors. - 0, // TODO(gvisor.dev/issue/969): Support Icmp/InCsumErrors. + 0, // Icmp/InCsumErrors. in.DstUnreachable.Value(), // InDestUnreachs. in.TimeExceeded.Value(), // InTimeExcds. in.ParamProblem.Value(), // InParmProbs. @@ -241,7 +270,7 @@ func (s *Stack) Statistics(stat interface{}, arg string) error { in.TimestampReply.Value(), // InTimestampReps. in.InfoRequest.Value(), // InAddrMasks. in.InfoReply.Value(), // InAddrMaskReps. - 0, // TODO(gvisor.dev/issue/969): Support Icmp/OutMsgs. + 0, // Icmp/OutMsgs. Metrics.ICMP.V4PacketsReceived.Invalid.Value(), // OutErrors. out.DstUnreachable.Value(), // OutDestUnreachs. out.TimeExceeded.Value(), // OutTimeExcds. @@ -277,15 +306,16 @@ func (s *Stack) Statistics(stat interface{}, arg string) error { } case *inet.StatSNMPUDP: udp := Metrics.UDP + // TODO(gvisor.dev/issue/969) Support stubbed stats. *stats = inet.StatSNMPUDP{ udp.PacketsReceived.Value(), // InDatagrams. udp.UnknownPortErrors.Value(), // NoPorts. - 0, // TODO(gvisor.dev/issue/969): Support Udp/InErrors. + 0, // Udp/InErrors. udp.PacketsSent.Value(), // OutDatagrams. udp.ReceiveBufferErrors.Value(), // RcvbufErrors. - 0, // TODO(gvisor.dev/issue/969): Support Udp/SndbufErrors. - 0, // TODO(gvisor.dev/issue/969): Support Udp/InCsumErrors. - 0, // TODO(gvisor.dev/issue/969): Support Udp/IgnoredMulti. + 0, // Udp/SndbufErrors. + 0, // Udp/InCsumErrors. + 0, // Udp/IgnoredMulti. } default: return syserr.ErrEndpointOperation.ToError() @@ -332,7 +362,7 @@ func (s *Stack) RouteTable() []inet.Route { } // IPTables returns the stack's iptables. -func (s *Stack) IPTables() (iptables.IPTables, error) { +func (s *Stack) IPTables() (stack.IPTables, error) { return s.Stack.IPTables(), nil } diff --git a/pkg/sentry/socket/socket.go b/pkg/sentry/socket/socket.go index 50d9744e6..6580bd6e9 100644 --- a/pkg/sentry/socket/socket.go +++ b/pkg/sentry/socket/socket.go @@ -31,6 +31,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/kernel" ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" + "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserr" "gvisor.dev/gvisor/pkg/tcpip" "gvisor.dev/gvisor/pkg/usermem" @@ -48,11 +49,25 @@ func (c *ControlMessages) Release() { c.Unix.Release() } -// Socket is the interface containing socket syscalls used by the syscall layer -// to redirect them to the appropriate implementation. +// Socket is an interface combining fs.FileOperations and SocketOps, +// representing a VFS1 socket file. type Socket interface { fs.FileOperations + SocketOps +} + +// SocketVFS2 is an interface combining vfs.FileDescription and SocketOps, +// representing a VFS2 socket file. +type SocketVFS2 interface { + vfs.FileDescriptionImpl + SocketOps +} +// SocketOps is the interface containing socket syscalls used by the syscall +// layer to redirect them to the appropriate implementation. +// +// It is implemented by both Socket and SocketVFS2. +type SocketOps interface { // Connect implements the connect(2) linux syscall. Connect(t *kernel.Task, sockaddr []byte, blocking bool) *syserr.Error @@ -153,6 +168,8 @@ var families = make(map[int][]Provider) // RegisterProvider registers the provider of a given address family so that // sockets of that type can be created via socket() and/or socketpair() // syscalls. +// +// This should only be called during the initialization of the address family. func RegisterProvider(family int, provider Provider) { families[family] = append(families[family], provider) } @@ -216,6 +233,74 @@ func NewDirent(ctx context.Context, d *device.Device) *fs.Dirent { return fs.NewDirent(ctx, inode, fmt.Sprintf("socket:[%d]", ino)) } +// ProviderVFS2 is the vfs2 interface implemented by providers of sockets for +// specific address families (e.g., AF_INET). +type ProviderVFS2 interface { + // Socket creates a new socket. + // + // If a nil Socket _and_ a nil error is returned, it means that the + // protocol is not supported. A non-nil error should only be returned + // if the protocol is supported, but an error occurs during creation. + Socket(t *kernel.Task, stype linux.SockType, protocol int) (*vfs.FileDescription, *syserr.Error) + + // Pair creates a pair of connected sockets. + // + // See Socket for error information. + Pair(t *kernel.Task, stype linux.SockType, protocol int) (*vfs.FileDescription, *vfs.FileDescription, *syserr.Error) +} + +// familiesVFS2 holds a map of all known address families and their providers. +var familiesVFS2 = make(map[int][]ProviderVFS2) + +// RegisterProviderVFS2 registers the provider of a given address family so that +// sockets of that type can be created via socket() and/or socketpair() +// syscalls. +// +// This should only be called during the initialization of the address family. +func RegisterProviderVFS2(family int, provider ProviderVFS2) { + familiesVFS2[family] = append(familiesVFS2[family], provider) +} + +// NewVFS2 creates a new socket with the given family, type and protocol. +func NewVFS2(t *kernel.Task, family int, stype linux.SockType, protocol int) (*vfs.FileDescription, *syserr.Error) { + for _, p := range familiesVFS2[family] { + s, err := p.Socket(t, stype, protocol) + if err != nil { + return nil, err + } + if s != nil { + t.Kernel().RecordSocketVFS2(s) + return s, nil + } + } + + return nil, syserr.ErrAddressFamilyNotSupported +} + +// PairVFS2 creates a new connected socket pair with the given family, type and +// protocol. +func PairVFS2(t *kernel.Task, family int, stype linux.SockType, protocol int) (*vfs.FileDescription, *vfs.FileDescription, *syserr.Error) { + providers, ok := familiesVFS2[family] + if !ok { + return nil, nil, syserr.ErrAddressFamilyNotSupported + } + + for _, p := range providers { + s1, s2, err := p.Pair(t, stype, protocol) + if err != nil { + return nil, nil, err + } + if s1 != nil && s2 != nil { + k := t.Kernel() + k.RecordSocketVFS2(s1) + k.RecordSocketVFS2(s2) + return s1, s2, nil + } + } + + return nil, nil, syserr.ErrSocketNotSupported +} + // SendReceiveTimeout stores timeouts for send and receive calls. // // It is meant to be embedded into Socket implementations to help satisfy the diff --git a/pkg/sentry/socket/unix/BUILD b/pkg/sentry/socket/unix/BUILD index 08743deba..de2cc4bdf 100644 --- a/pkg/sentry/socket/unix/BUILD +++ b/pkg/sentry/socket/unix/BUILD @@ -8,23 +8,27 @@ go_library( "device.go", "io.go", "unix.go", + "unix_vfs2.go", ], visibility = ["//pkg/sentry:internal"], deps = [ "//pkg/abi/linux", "//pkg/context", + "//pkg/fspath", "//pkg/refs", "//pkg/safemem", "//pkg/sentry/arch", "//pkg/sentry/device", "//pkg/sentry/fs", "//pkg/sentry/fs/fsutil", + "//pkg/sentry/fsimpl/sockfs", "//pkg/sentry/kernel", "//pkg/sentry/kernel/time", "//pkg/sentry/socket", "//pkg/sentry/socket/control", "//pkg/sentry/socket/netstack", "//pkg/sentry/socket/unix/transport", + "//pkg/sentry/vfs", "//pkg/syserr", "//pkg/syserror", "//pkg/tcpip", diff --git a/pkg/sentry/socket/unix/transport/BUILD b/pkg/sentry/socket/unix/transport/BUILD index 74bcd6300..c708b6030 100644 --- a/pkg/sentry/socket/unix/transport/BUILD +++ b/pkg/sentry/socket/unix/transport/BUILD @@ -30,6 +30,7 @@ go_library( "//pkg/abi/linux", "//pkg/context", "//pkg/ilist", + "//pkg/log", "//pkg/refs", "//pkg/sync", "//pkg/syserr", diff --git a/pkg/sentry/socket/unix/transport/unix.go b/pkg/sentry/socket/unix/transport/unix.go index 2ef654235..2f1b127df 100644 --- a/pkg/sentry/socket/unix/transport/unix.go +++ b/pkg/sentry/socket/unix/transport/unix.go @@ -20,6 +20,7 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/log" "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/syserr" "gvisor.dev/gvisor/pkg/tcpip" @@ -838,24 +839,43 @@ func (e *baseEndpoint) SendMsg(ctx context.Context, data [][]byte, c ControlMess // SetSockOpt sets a socket option. Currently not supported. func (e *baseEndpoint) SetSockOpt(opt interface{}) *tcpip.Error { - switch v := opt.(type) { - case tcpip.PasscredOption: - e.setPasscred(v != 0) - return nil - } return nil } func (e *baseEndpoint) SetSockOptBool(opt tcpip.SockOptBool, v bool) *tcpip.Error { + switch opt { + case tcpip.BroadcastOption: + case tcpip.PasscredOption: + e.setPasscred(v) + case tcpip.ReuseAddressOption: + default: + log.Warningf("Unsupported socket option: %d", opt) + } return nil } func (e *baseEndpoint) SetSockOptInt(opt tcpip.SockOptInt, v int) *tcpip.Error { + switch opt { + case tcpip.SendBufferSizeOption: + case tcpip.ReceiveBufferSizeOption: + default: + log.Warningf("Unsupported socket option: %d", opt) + } return nil } func (e *baseEndpoint) GetSockOptBool(opt tcpip.SockOptBool) (bool, *tcpip.Error) { - return false, tcpip.ErrUnknownProtocolOption + switch opt { + case tcpip.KeepaliveEnabledOption: + return false, nil + + case tcpip.PasscredOption: + return e.Passcred(), nil + + default: + log.Warningf("Unsupported socket option: %d", opt) + return false, tcpip.ErrUnknownProtocolOption + } } func (e *baseEndpoint) GetSockOptInt(opt tcpip.SockOptInt) (int, *tcpip.Error) { @@ -914,29 +934,19 @@ func (e *baseEndpoint) GetSockOptInt(opt tcpip.SockOptInt) (int, *tcpip.Error) { return int(v), nil default: + log.Warningf("Unsupported socket option: %d", opt) return -1, tcpip.ErrUnknownProtocolOption } } // GetSockOpt implements tcpip.Endpoint.GetSockOpt. func (e *baseEndpoint) GetSockOpt(opt interface{}) *tcpip.Error { - switch o := opt.(type) { + switch opt.(type) { case tcpip.ErrorOption: return nil - case *tcpip.PasscredOption: - if e.Passcred() { - *o = tcpip.PasscredOption(1) - } else { - *o = tcpip.PasscredOption(0) - } - return nil - - case *tcpip.KeepaliveEnabledOption: - *o = 0 - return nil - default: + log.Warningf("Unsupported socket option: %T", opt) return tcpip.ErrUnknownProtocolOption } } diff --git a/pkg/sentry/socket/unix/unix.go b/pkg/sentry/socket/unix/unix.go index 4d30aa714..7c64f30fa 100644 --- a/pkg/sentry/socket/unix/unix.go +++ b/pkg/sentry/socket/unix/unix.go @@ -23,6 +23,7 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/fspath" "gvisor.dev/gvisor/pkg/refs" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/fs" @@ -33,6 +34,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/socket/control" "gvisor.dev/gvisor/pkg/sentry/socket/netstack" "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" + "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserr" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/tcpip" @@ -52,11 +54,8 @@ type SocketOperations struct { fsutil.FileNoSplice `state:"nosave"` fsutil.FileNoopFlush `state:"nosave"` fsutil.FileUseInodeUnstableAttr `state:"nosave"` - refs.AtomicRefCount - socket.SendReceiveTimeout - ep transport.Endpoint - stype linux.SockType + socketOpsCommon } // New creates a new unix socket. @@ -75,16 +74,29 @@ func NewWithDirent(ctx context.Context, d *fs.Dirent, ep transport.Endpoint, sty } s := SocketOperations{ - ep: ep, - stype: stype, + socketOpsCommon: socketOpsCommon{ + ep: ep, + stype: stype, + }, } s.EnableLeakCheck("unix.SocketOperations") return fs.NewFile(ctx, d, flags, &s) } +// socketOpsCommon contains the socket operations common to VFS1 and VFS2. +// +// +stateify savable +type socketOpsCommon struct { + refs.AtomicRefCount + socket.SendReceiveTimeout + + ep transport.Endpoint + stype linux.SockType +} + // DecRef implements RefCounter.DecRef. -func (s *SocketOperations) DecRef() { +func (s *socketOpsCommon) DecRef() { s.DecRefWithDestructor(func() { s.ep.Close() }) @@ -97,7 +109,7 @@ func (s *SocketOperations) Release() { s.DecRef() } -func (s *SocketOperations) isPacket() bool { +func (s *socketOpsCommon) isPacket() bool { switch s.stype { case linux.SOCK_DGRAM, linux.SOCK_SEQPACKET: return true @@ -110,7 +122,7 @@ func (s *SocketOperations) isPacket() bool { } // Endpoint extracts the transport.Endpoint. -func (s *SocketOperations) Endpoint() transport.Endpoint { +func (s *socketOpsCommon) Endpoint() transport.Endpoint { return s.ep } @@ -143,7 +155,7 @@ func extractPath(sockaddr []byte) (string, *syserr.Error) { // GetPeerName implements the linux syscall getpeername(2) for sockets backed by // a transport.Endpoint. -func (s *SocketOperations) GetPeerName(t *kernel.Task) (linux.SockAddr, uint32, *syserr.Error) { +func (s *socketOpsCommon) GetPeerName(t *kernel.Task) (linux.SockAddr, uint32, *syserr.Error) { addr, err := s.ep.GetRemoteAddress() if err != nil { return nil, 0, syserr.TranslateNetstackError(err) @@ -155,7 +167,7 @@ func (s *SocketOperations) GetPeerName(t *kernel.Task) (linux.SockAddr, uint32, // GetSockName implements the linux syscall getsockname(2) for sockets backed by // a transport.Endpoint. -func (s *SocketOperations) GetSockName(t *kernel.Task) (linux.SockAddr, uint32, *syserr.Error) { +func (s *socketOpsCommon) GetSockName(t *kernel.Task) (linux.SockAddr, uint32, *syserr.Error) { addr, err := s.ep.GetLocalAddress() if err != nil { return nil, 0, syserr.TranslateNetstackError(err) @@ -178,7 +190,7 @@ func (s *SocketOperations) GetSockOpt(t *kernel.Task, level, name int, outPtr us // Listen implements the linux syscall listen(2) for sockets backed by // a transport.Endpoint. -func (s *SocketOperations) Listen(t *kernel.Task, backlog int) *syserr.Error { +func (s *socketOpsCommon) Listen(t *kernel.Task, backlog int) *syserr.Error { return s.ep.Listen(backlog) } @@ -310,6 +322,8 @@ func (s *SocketOperations) Bind(t *kernel.Task, sockaddr []byte) *syserr.Error { } // Create the socket. + // + // TODO(gvisor.dev/issue/2324): Correctly set file permissions. childDir, err := d.Bind(t, t.FSContext().RootDirectory(), name, bep, fs.FilePermissions{User: fs.PermMask{Read: true}}) if err != nil { return syserr.ErrPortInUse @@ -345,6 +359,31 @@ func extractEndpoint(t *kernel.Task, sockaddr []byte) (transport.BoundEndpoint, return ep, nil } + if kernel.VFS2Enabled { + p := fspath.Parse(path) + root := t.FSContext().RootDirectoryVFS2() + start := root + relPath := !p.Absolute + if relPath { + start = t.FSContext().WorkingDirectoryVFS2() + } + pop := vfs.PathOperation{ + Root: root, + Start: start, + Path: p, + FollowFinalSymlink: true, + } + ep, e := t.Kernel().VFS().BoundEndpointAt(t, t.Credentials(), &pop) + root.DecRef() + if relPath { + start.DecRef() + } + if e != nil { + return nil, syserr.FromError(e) + } + return ep, nil + } + // Find the node in the filesystem. root := t.FSContext().RootDirectory() cwd := t.FSContext().WorkingDirectory() @@ -363,12 +402,11 @@ func extractEndpoint(t *kernel.Task, sockaddr []byte) (transport.BoundEndpoint, // No socket! return nil, syserr.ErrConnectionRefused } - return ep, nil } // Connect implements the linux syscall connect(2) for unix sockets. -func (s *SocketOperations) Connect(t *kernel.Task, sockaddr []byte, blocking bool) *syserr.Error { +func (s *socketOpsCommon) Connect(t *kernel.Task, sockaddr []byte, blocking bool) *syserr.Error { ep, err := extractEndpoint(t, sockaddr) if err != nil { return err @@ -379,7 +417,7 @@ func (s *SocketOperations) Connect(t *kernel.Task, sockaddr []byte, blocking boo return s.ep.Connect(t, ep) } -// Writev implements fs.FileOperations.Write. +// Write implements fs.FileOperations.Write. func (s *SocketOperations) Write(ctx context.Context, _ *fs.File, src usermem.IOSequence, _ int64) (int64, error) { t := kernel.TaskFromContext(ctx) ctrl := control.New(t, s.ep, nil) @@ -399,7 +437,7 @@ func (s *SocketOperations) Write(ctx context.Context, _ *fs.File, src usermem.IO // SendMsg implements the linux syscall sendmsg(2) for unix sockets backed by // a transport.Endpoint. -func (s *SocketOperations) SendMsg(t *kernel.Task, src usermem.IOSequence, to []byte, flags int, haveDeadline bool, deadline ktime.Time, controlMessages socket.ControlMessages) (int, *syserr.Error) { +func (s *socketOpsCommon) SendMsg(t *kernel.Task, src usermem.IOSequence, to []byte, flags int, haveDeadline bool, deadline ktime.Time, controlMessages socket.ControlMessages) (int, *syserr.Error) { w := EndpointWriter{ Ctx: t, Endpoint: s.ep, @@ -453,27 +491,27 @@ func (s *SocketOperations) SendMsg(t *kernel.Task, src usermem.IOSequence, to [] } // Passcred implements transport.Credentialer.Passcred. -func (s *SocketOperations) Passcred() bool { +func (s *socketOpsCommon) Passcred() bool { return s.ep.Passcred() } // ConnectedPasscred implements transport.Credentialer.ConnectedPasscred. -func (s *SocketOperations) ConnectedPasscred() bool { +func (s *socketOpsCommon) ConnectedPasscred() bool { return s.ep.ConnectedPasscred() } // Readiness implements waiter.Waitable.Readiness. -func (s *SocketOperations) Readiness(mask waiter.EventMask) waiter.EventMask { +func (s *socketOpsCommon) Readiness(mask waiter.EventMask) waiter.EventMask { return s.ep.Readiness(mask) } // EventRegister implements waiter.Waitable.EventRegister. -func (s *SocketOperations) EventRegister(e *waiter.Entry, mask waiter.EventMask) { +func (s *socketOpsCommon) EventRegister(e *waiter.Entry, mask waiter.EventMask) { s.ep.EventRegister(e, mask) } // EventUnregister implements waiter.Waitable.EventUnregister. -func (s *SocketOperations) EventUnregister(e *waiter.Entry) { +func (s *socketOpsCommon) EventUnregister(e *waiter.Entry) { s.ep.EventUnregister(e) } @@ -485,7 +523,7 @@ func (s *SocketOperations) SetSockOpt(t *kernel.Task, level int, name int, optVa // Shutdown implements the linux syscall shutdown(2) for sockets backed by // a transport.Endpoint. -func (s *SocketOperations) Shutdown(t *kernel.Task, how int) *syserr.Error { +func (s *socketOpsCommon) Shutdown(t *kernel.Task, how int) *syserr.Error { f, err := netstack.ConvertShutdown(how) if err != nil { return err @@ -511,7 +549,7 @@ func (s *SocketOperations) Read(ctx context.Context, _ *fs.File, dst usermem.IOS // RecvMsg implements the linux syscall recvmsg(2) for sockets backed by // a transport.Endpoint. -func (s *SocketOperations) RecvMsg(t *kernel.Task, dst usermem.IOSequence, flags int, haveDeadline bool, deadline ktime.Time, senderRequested bool, controlDataLen uint64) (n int, msgFlags int, senderAddr linux.SockAddr, senderAddrLen uint32, controlMessages socket.ControlMessages, err *syserr.Error) { +func (s *socketOpsCommon) RecvMsg(t *kernel.Task, dst usermem.IOSequence, flags int, haveDeadline bool, deadline ktime.Time, senderRequested bool, controlDataLen uint64) (n int, msgFlags int, senderAddr linux.SockAddr, senderAddrLen uint32, controlMessages socket.ControlMessages, err *syserr.Error) { trunc := flags&linux.MSG_TRUNC != 0 peek := flags&linux.MSG_PEEK != 0 dontWait := flags&linux.MSG_DONTWAIT != 0 @@ -648,12 +686,12 @@ func (s *SocketOperations) RecvMsg(t *kernel.Task, dst usermem.IOSequence, flags } // State implements socket.Socket.State. -func (s *SocketOperations) State() uint32 { +func (s *socketOpsCommon) State() uint32 { return s.ep.State() } // Type implements socket.Socket.Type. -func (s *SocketOperations) Type() (family int, skType linux.SockType, protocol int) { +func (s *socketOpsCommon) Type() (family int, skType linux.SockType, protocol int) { // Unix domain sockets always have a protocol of 0. return linux.AF_UNIX, s.stype, 0 } @@ -706,4 +744,5 @@ func (*provider) Pair(t *kernel.Task, stype linux.SockType, protocol int) (*fs.F func init() { socket.RegisterProvider(linux.AF_UNIX, &provider{}) + socket.RegisterProviderVFS2(linux.AF_UNIX, &providerVFS2{}) } diff --git a/pkg/sentry/socket/unix/unix_vfs2.go b/pkg/sentry/socket/unix/unix_vfs2.go new file mode 100644 index 000000000..3e54d49c4 --- /dev/null +++ b/pkg/sentry/socket/unix/unix_vfs2.go @@ -0,0 +1,348 @@ +// Copyright 2018 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package unix + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/sockfs" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/socket/control" + "gvisor.dev/gvisor/pkg/sentry/socket/netstack" + "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserr" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/tcpip" + "gvisor.dev/gvisor/pkg/usermem" + "gvisor.dev/gvisor/pkg/waiter" +) + +// SocketVFS2 implements socket.SocketVFS2 (and by extension, +// vfs.FileDescriptionImpl) for Unix sockets. +type SocketVFS2 struct { + vfsfd vfs.FileDescription + vfs.FileDescriptionDefaultImpl + vfs.DentryMetadataFileDescriptionImpl + + socketOpsCommon +} + +// NewVFS2File creates and returns a new vfs.FileDescription for a unix socket. +func NewVFS2File(t *kernel.Task, ep transport.Endpoint, stype linux.SockType) (*vfs.FileDescription, *syserr.Error) { + sock := NewFDImpl(ep, stype) + vfsfd := &sock.vfsfd + if err := sockfs.InitSocket(sock, vfsfd, t.Kernel().SocketMount(), t.Credentials()); err != nil { + return nil, syserr.FromError(err) + } + return vfsfd, nil +} + +// NewFDImpl creates and returns a new SocketVFS2. +func NewFDImpl(ep transport.Endpoint, stype linux.SockType) *SocketVFS2 { + // You can create AF_UNIX, SOCK_RAW sockets. They're the same as + // SOCK_DGRAM and don't require CAP_NET_RAW. + if stype == linux.SOCK_RAW { + stype = linux.SOCK_DGRAM + } + + return &SocketVFS2{ + socketOpsCommon: socketOpsCommon{ + ep: ep, + stype: stype, + }, + } +} + +// GetSockOpt implements the linux syscall getsockopt(2) for sockets backed by +// a transport.Endpoint. +func (s *SocketVFS2) GetSockOpt(t *kernel.Task, level int, name int, outPtr usermem.Addr, outLen int) (interface{}, *syserr.Error) { + return netstack.GetSockOpt(t, s, s.ep, linux.AF_UNIX, s.ep.Type(), level, name, outLen) +} + +// blockingAccept implements a blocking version of accept(2), that is, if no +// connections are ready to be accept, it will block until one becomes ready. +func (s *SocketVFS2) blockingAccept(t *kernel.Task) (transport.Endpoint, *syserr.Error) { + // Register for notifications. + e, ch := waiter.NewChannelEntry(nil) + s.socketOpsCommon.EventRegister(&e, waiter.EventIn) + defer s.socketOpsCommon.EventUnregister(&e) + + // Try to accept the connection; if it fails, then wait until we get a + // notification. + for { + if ep, err := s.ep.Accept(); err != syserr.ErrWouldBlock { + return ep, err + } + + if err := t.Block(ch); err != nil { + return nil, syserr.FromError(err) + } + } +} + +// Accept implements the linux syscall accept(2) for sockets backed by +// a transport.Endpoint. +func (s *SocketVFS2) Accept(t *kernel.Task, peerRequested bool, flags int, blocking bool) (int32, linux.SockAddr, uint32, *syserr.Error) { + // Issue the accept request to get the new endpoint. + ep, err := s.ep.Accept() + if err != nil { + if err != syserr.ErrWouldBlock || !blocking { + return 0, nil, 0, err + } + + var err *syserr.Error + ep, err = s.blockingAccept(t) + if err != nil { + return 0, nil, 0, err + } + } + + // We expect this to be a FileDescription here. + ns, err := NewVFS2File(t, ep, s.stype) + if err != nil { + return 0, nil, 0, err + } + defer ns.DecRef() + + if flags&linux.SOCK_NONBLOCK != 0 { + ns.SetStatusFlags(t, t.Credentials(), linux.SOCK_NONBLOCK) + } + + var addr linux.SockAddr + var addrLen uint32 + if peerRequested { + // Get address of the peer. + var err *syserr.Error + addr, addrLen, err = ns.Impl().(*SocketVFS2).GetPeerName(t) + if err != nil { + return 0, nil, 0, err + } + } + + fd, e := t.NewFDFromVFS2(0, ns, kernel.FDFlags{ + CloseOnExec: flags&linux.SOCK_CLOEXEC != 0, + }) + if e != nil { + return 0, nil, 0, syserr.FromError(e) + } + + t.Kernel().RecordSocketVFS2(ns) + return fd, addr, addrLen, nil +} + +// Bind implements the linux syscall bind(2) for unix sockets. +func (s *SocketVFS2) Bind(t *kernel.Task, sockaddr []byte) *syserr.Error { + p, e := extractPath(sockaddr) + if e != nil { + return e + } + + bep, ok := s.ep.(transport.BoundEndpoint) + if !ok { + // This socket can't be bound. + return syserr.ErrInvalidArgument + } + + return s.ep.Bind(tcpip.FullAddress{Addr: tcpip.Address(p)}, func() *syserr.Error { + // Is it abstract? + if p[0] == 0 { + if t.IsNetworkNamespaced() { + return syserr.ErrInvalidEndpointState + } + if err := t.AbstractSockets().Bind(p[1:], bep, s); err != nil { + // syserr.ErrPortInUse corresponds to EADDRINUSE. + return syserr.ErrPortInUse + } + } else { + path := fspath.Parse(p) + root := t.FSContext().RootDirectoryVFS2() + defer root.DecRef() + start := root + relPath := !path.Absolute + if relPath { + start = t.FSContext().WorkingDirectoryVFS2() + defer start.DecRef() + } + pop := vfs.PathOperation{ + Root: root, + Start: start, + Path: path, + } + err := t.Kernel().VFS().MknodAt(t, t.Credentials(), &pop, &vfs.MknodOptions{ + // TODO(gvisor.dev/issue/2324): The file permissions should be taken + // from s and t.FSContext().Umask() (see net/unix/af_unix.c:unix_bind), + // but VFS1 just always uses 0400. Resolve this inconsistency. + Mode: linux.S_IFSOCK | 0400, + Endpoint: bep, + }) + if err == syserror.EEXIST { + return syserr.ErrAddressInUse + } + return syserr.FromError(err) + } + + return nil + }) +} + +// Ioctl implements vfs.FileDescriptionImpl. +func (s *SocketVFS2) Ioctl(ctx context.Context, uio usermem.IO, args arch.SyscallArguments) (uintptr, error) { + return netstack.Ioctl(ctx, s.ep, uio, args) +} + +// PRead implements vfs.FileDescriptionImpl. +func (s *SocketVFS2) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { + return 0, syserror.ESPIPE +} + +// Read implements vfs.FileDescriptionImpl. +func (s *SocketVFS2) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { + // All flags other than RWF_NOWAIT should be ignored. + // TODO(gvisor.dev/issue/1476): Support RWF_NOWAIT. + if opts.Flags != 0 { + return 0, syserror.EOPNOTSUPP + } + + if dst.NumBytes() == 0 { + return 0, nil + } + return dst.CopyOutFrom(ctx, &EndpointReader{ + Ctx: ctx, + Endpoint: s.ep, + NumRights: 0, + Peek: false, + From: nil, + }) +} + +// PWrite implements vfs.FileDescriptionImpl. +func (s *SocketVFS2) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) { + return 0, syserror.ESPIPE +} + +// Write implements vfs.FileDescriptionImpl. +func (s *SocketVFS2) Write(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) { + // All flags other than RWF_NOWAIT should be ignored. + // TODO(gvisor.dev/issue/1476): Support RWF_NOWAIT. + if opts.Flags != 0 { + return 0, syserror.EOPNOTSUPP + } + + t := kernel.TaskFromContext(ctx) + ctrl := control.New(t, s.ep, nil) + + if src.NumBytes() == 0 { + nInt, err := s.ep.SendMsg(ctx, [][]byte{}, ctrl, nil) + return int64(nInt), err.ToError() + } + + return src.CopyInTo(ctx, &EndpointWriter{ + Ctx: ctx, + Endpoint: s.ep, + Control: ctrl, + To: nil, + }) +} + +// Release implements vfs.FileDescriptionImpl. +func (s *SocketVFS2) Release() { + // Release only decrements a reference on s because s may be referenced in + // the abstract socket namespace. + s.DecRef() +} + +// Readiness implements waiter.Waitable.Readiness. +func (s *SocketVFS2) Readiness(mask waiter.EventMask) waiter.EventMask { + return s.socketOpsCommon.Readiness(mask) +} + +// EventRegister implements waiter.Waitable.EventRegister. +func (s *SocketVFS2) EventRegister(e *waiter.Entry, mask waiter.EventMask) { + s.socketOpsCommon.EventRegister(e, mask) +} + +// EventUnregister implements waiter.Waitable.EventUnregister. +func (s *SocketVFS2) EventUnregister(e *waiter.Entry) { + s.socketOpsCommon.EventUnregister(e) +} + +// SetSockOpt implements the linux syscall setsockopt(2) for sockets backed by +// a transport.Endpoint. +func (s *SocketVFS2) SetSockOpt(t *kernel.Task, level int, name int, optVal []byte) *syserr.Error { + return netstack.SetSockOpt(t, s, s.ep, level, name, optVal) +} + +// providerVFS2 is a unix domain socket provider for VFS2. +type providerVFS2 struct{} + +func (*providerVFS2) Socket(t *kernel.Task, stype linux.SockType, protocol int) (*vfs.FileDescription, *syserr.Error) { + // Check arguments. + if protocol != 0 && protocol != linux.AF_UNIX /* PF_UNIX */ { + return nil, syserr.ErrProtocolNotSupported + } + + // Create the endpoint and socket. + var ep transport.Endpoint + switch stype { + case linux.SOCK_DGRAM, linux.SOCK_RAW: + ep = transport.NewConnectionless(t) + case linux.SOCK_SEQPACKET, linux.SOCK_STREAM: + ep = transport.NewConnectioned(t, stype, t.Kernel()) + default: + return nil, syserr.ErrInvalidArgument + } + + f, err := NewVFS2File(t, ep, stype) + if err != nil { + ep.Close() + return nil, err + } + return f, nil +} + +// Pair creates a new pair of AF_UNIX connected sockets. +func (*providerVFS2) Pair(t *kernel.Task, stype linux.SockType, protocol int) (*vfs.FileDescription, *vfs.FileDescription, *syserr.Error) { + // Check arguments. + if protocol != 0 && protocol != linux.AF_UNIX /* PF_UNIX */ { + return nil, nil, syserr.ErrProtocolNotSupported + } + + switch stype { + case linux.SOCK_STREAM, linux.SOCK_DGRAM, linux.SOCK_SEQPACKET, linux.SOCK_RAW: + // Ok + default: + return nil, nil, syserr.ErrInvalidArgument + } + + // Create the endpoints and sockets. + ep1, ep2 := transport.NewPair(t, stype, t.Kernel()) + s1, err := NewVFS2File(t, ep1, stype) + if err != nil { + ep1.Close() + ep2.Close() + return nil, nil, err + } + s2, err := NewVFS2File(t, ep2, stype) + if err != nil { + s1.DecRef() + ep2.Close() + return nil, nil, err + } + + return s1, s2, nil +} diff --git a/pkg/sentry/strace/BUILD b/pkg/sentry/strace/BUILD index 762a946fe..88d5db9fc 100644 --- a/pkg/sentry/strace/BUILD +++ b/pkg/sentry/strace/BUILD @@ -7,6 +7,7 @@ go_library( srcs = [ "capability.go", "clone.go", + "epoll.go", "futex.go", "linux64_amd64.go", "linux64_arm64.go", @@ -30,7 +31,6 @@ go_library( "//pkg/seccomp", "//pkg/sentry/arch", "//pkg/sentry/kernel", - "//pkg/sentry/socket/control", "//pkg/sentry/socket/netlink", "//pkg/sentry/socket/netstack", "//pkg/sentry/syscalls/linux", diff --git a/pkg/sentry/strace/epoll.go b/pkg/sentry/strace/epoll.go new file mode 100644 index 000000000..a6e48b836 --- /dev/null +++ b/pkg/sentry/strace/epoll.go @@ -0,0 +1,89 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package strace + +import ( + "fmt" + "strings" + + "gvisor.dev/gvisor/pkg/abi" + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/usermem" +) + +func epollEvent(t *kernel.Task, eventAddr usermem.Addr) string { + var e linux.EpollEvent + if _, err := t.CopyIn(eventAddr, &e); err != nil { + return fmt.Sprintf("%#x {error reading event: %v}", eventAddr, err) + } + var sb strings.Builder + fmt.Fprintf(&sb, "%#x ", eventAddr) + writeEpollEvent(&sb, e) + return sb.String() +} + +func epollEvents(t *kernel.Task, eventsAddr usermem.Addr, numEvents, maxBytes uint64) string { + var sb strings.Builder + fmt.Fprintf(&sb, "%#x {", eventsAddr) + addr := eventsAddr + for i := uint64(0); i < numEvents; i++ { + var e linux.EpollEvent + if _, err := t.CopyIn(addr, &e); err != nil { + fmt.Fprintf(&sb, "{error reading event at %#x: %v}", addr, err) + continue + } + writeEpollEvent(&sb, e) + if uint64(sb.Len()) >= maxBytes { + sb.WriteString("...") + break + } + if _, ok := addr.AddLength(uint64(linux.SizeOfEpollEvent)); !ok { + fmt.Fprintf(&sb, "{error reading event at %#x: EFAULT}", addr) + continue + } + } + sb.WriteString("}") + return sb.String() +} + +func writeEpollEvent(sb *strings.Builder, e linux.EpollEvent) { + events := epollEventEvents.Parse(uint64(e.Events)) + fmt.Fprintf(sb, "{events=%s data=[%#x, %#x]}", events, e.Data[0], e.Data[1]) +} + +var epollCtlOps = abi.ValueSet{ + linux.EPOLL_CTL_ADD: "EPOLL_CTL_ADD", + linux.EPOLL_CTL_DEL: "EPOLL_CTL_DEL", + linux.EPOLL_CTL_MOD: "EPOLL_CTL_MOD", +} + +var epollEventEvents = abi.FlagSet{ + {Flag: linux.EPOLLIN, Name: "EPOLLIN"}, + {Flag: linux.EPOLLPRI, Name: "EPOLLPRI"}, + {Flag: linux.EPOLLOUT, Name: "EPOLLOUT"}, + {Flag: linux.EPOLLERR, Name: "EPOLLERR"}, + {Flag: linux.EPOLLHUP, Name: "EPULLHUP"}, + {Flag: linux.EPOLLRDNORM, Name: "EPOLLRDNORM"}, + {Flag: linux.EPOLLRDBAND, Name: "EPOLLRDBAND"}, + {Flag: linux.EPOLLWRNORM, Name: "EPOLLWRNORM"}, + {Flag: linux.EPOLLWRBAND, Name: "EPOLLWRBAND"}, + {Flag: linux.EPOLLMSG, Name: "EPOLLMSG"}, + {Flag: linux.EPOLLRDHUP, Name: "EPOLLRDHUP"}, + {Flag: linux.EPOLLEXCLUSIVE, Name: "EPOLLEXCLUSIVE"}, + {Flag: linux.EPOLLWAKEUP, Name: "EPOLLWAKEUP"}, + {Flag: linux.EPOLLONESHOT, Name: "EPOLLONESHOT"}, + {Flag: linux.EPOLLET, Name: "EPOLLET"}, +} diff --git a/pkg/sentry/strace/linux64_amd64.go b/pkg/sentry/strace/linux64_amd64.go index a4de545e9..71b92eaee 100644 --- a/pkg/sentry/strace/linux64_amd64.go +++ b/pkg/sentry/strace/linux64_amd64.go @@ -256,8 +256,8 @@ var linuxAMD64 = SyscallMap{ 229: makeSyscallInfo("clock_getres", Hex, PostTimespec), 230: makeSyscallInfo("clock_nanosleep", Hex, Hex, Timespec, PostTimespec), 231: makeSyscallInfo("exit_group", Hex), - 232: makeSyscallInfo("epoll_wait", Hex, Hex, Hex, Hex), - 233: makeSyscallInfo("epoll_ctl", Hex, Hex, FD, Hex), + 232: makeSyscallInfo("epoll_wait", FD, EpollEvents, Hex, Hex), + 233: makeSyscallInfo("epoll_ctl", FD, EpollCtlOp, FD, EpollEvent), 234: makeSyscallInfo("tgkill", Hex, Hex, Signal), 235: makeSyscallInfo("utimes", Path, Timeval), // 236: vserver (not implemented in the Linux kernel) @@ -305,7 +305,7 @@ var linuxAMD64 = SyscallMap{ 278: makeSyscallInfo("vmsplice", FD, Hex, Hex, Hex), 279: makeSyscallInfo("move_pages", Hex, Hex, Hex, Hex, Hex, Hex), 280: makeSyscallInfo("utimensat", FD, Path, UTimeTimespec, Hex), - 281: makeSyscallInfo("epoll_pwait", Hex, Hex, Hex, Hex, SigSet, Hex), + 281: makeSyscallInfo("epoll_pwait", FD, EpollEvents, Hex, Hex, SigSet, Hex), 282: makeSyscallInfo("signalfd", Hex, Hex, Hex), 283: makeSyscallInfo("timerfd_create", Hex, Hex), 284: makeSyscallInfo("eventfd", Hex), diff --git a/pkg/sentry/strace/linux64_arm64.go b/pkg/sentry/strace/linux64_arm64.go index 8bc38545f..bd7361a52 100644 --- a/pkg/sentry/strace/linux64_arm64.go +++ b/pkg/sentry/strace/linux64_arm64.go @@ -45,8 +45,8 @@ var linuxARM64 = SyscallMap{ 18: makeSyscallInfo("lookup_dcookie", Hex, Hex, Hex), 19: makeSyscallInfo("eventfd2", Hex, Hex), 20: makeSyscallInfo("epoll_create1", Hex), - 21: makeSyscallInfo("epoll_ctl", Hex, Hex, FD, Hex), - 22: makeSyscallInfo("epoll_pwait", Hex, Hex, Hex, Hex, SigSet, Hex), + 21: makeSyscallInfo("epoll_ctl", FD, EpollCtlOp, FD, EpollEvent), + 22: makeSyscallInfo("epoll_pwait", FD, EpollEvents, Hex, Hex, SigSet, Hex), 23: makeSyscallInfo("dup", FD), 24: makeSyscallInfo("dup3", FD, FD, Hex), 25: makeSyscallInfo("fcntl", FD, Hex, Hex), diff --git a/pkg/sentry/strace/socket.go b/pkg/sentry/strace/socket.go index f7ff4573e..c0512de89 100644 --- a/pkg/sentry/strace/socket.go +++ b/pkg/sentry/strace/socket.go @@ -22,7 +22,6 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/binary" "gvisor.dev/gvisor/pkg/sentry/kernel" - "gvisor.dev/gvisor/pkg/sentry/socket/control" "gvisor.dev/gvisor/pkg/sentry/socket/netlink" "gvisor.dev/gvisor/pkg/sentry/socket/netstack" slinux "gvisor.dev/gvisor/pkg/sentry/syscalls/linux" @@ -220,13 +219,13 @@ func cmsghdr(t *kernel.Task, addr usermem.Addr, length uint64, maxBytes uint64) if skipData { strs = append(strs, fmt.Sprintf("{level=%s, type=%s, length=%d}", level, typ, h.Length)) - i += control.AlignUp(length, width) + i += binary.AlignUp(length, width) continue } switch h.Type { case linux.SCM_RIGHTS: - rightsSize := control.AlignDown(length, linux.SizeOfControlMessageRight) + rightsSize := binary.AlignDown(length, linux.SizeOfControlMessageRight) numRights := rightsSize / linux.SizeOfControlMessageRight fds := make(linux.ControlMessageRights, numRights) @@ -295,7 +294,7 @@ func cmsghdr(t *kernel.Task, addr usermem.Addr, length uint64, maxBytes uint64) default: panic("unreachable") } - i += control.AlignUp(length, width) + i += binary.AlignUp(length, width) } return fmt.Sprintf("%#x %s", addr, strings.Join(strs, ", ")) @@ -633,4 +632,13 @@ var sockOptNames = map[uint64]abi.ValueSet{ linux.MCAST_MSFILTER: "MCAST_MSFILTER", linux.IPV6_ADDRFORM: "IPV6_ADDRFORM", }, + linux.SOL_NETLINK: { + linux.NETLINK_BROADCAST_ERROR: "NETLINK_BROADCAST_ERROR", + linux.NETLINK_CAP_ACK: "NETLINK_CAP_ACK", + linux.NETLINK_DUMP_STRICT_CHK: "NETLINK_DUMP_STRICT_CHK", + linux.NETLINK_EXT_ACK: "NETLINK_EXT_ACK", + linux.NETLINK_LIST_MEMBERSHIPS: "NETLINK_LIST_MEMBERSHIPS", + linux.NETLINK_NO_ENOBUFS: "NETLINK_NO_ENOBUFS", + linux.NETLINK_PKTINFO: "NETLINK_PKTINFO", + }, } diff --git a/pkg/sentry/strace/strace.go b/pkg/sentry/strace/strace.go index a796b2396..68ca537c8 100644 --- a/pkg/sentry/strace/strace.go +++ b/pkg/sentry/strace/strace.go @@ -141,6 +141,10 @@ func path(t *kernel.Task, addr usermem.Addr) string { } func fd(t *kernel.Task, fd int32) string { + if kernel.VFS2Enabled { + return fdVFS2(t, fd) + } + root := t.FSContext().RootDirectory() if root != nil { defer root.DecRef() @@ -169,6 +173,30 @@ func fd(t *kernel.Task, fd int32) string { return fmt.Sprintf("%#x %s", fd, name) } +func fdVFS2(t *kernel.Task, fd int32) string { + root := t.FSContext().RootDirectoryVFS2() + defer root.DecRef() + + vfsObj := root.Mount().Filesystem().VirtualFilesystem() + if fd == linux.AT_FDCWD { + wd := t.FSContext().WorkingDirectoryVFS2() + defer wd.DecRef() + + name, _ := vfsObj.PathnameWithDeleted(t, root, wd) + return fmt.Sprintf("AT_FDCWD %s", name) + } + + file := t.GetFileVFS2(fd) + if file == nil { + // Cast FD to uint64 to avoid printing negative hex. + return fmt.Sprintf("%#x (bad FD)", uint64(fd)) + } + defer file.DecRef() + + name, _ := vfsObj.PathnameWithDeleted(t, root, file.VirtualDentry()) + return fmt.Sprintf("%#x %s", fd, name) +} + func fdpair(t *kernel.Task, addr usermem.Addr) string { var fds [2]int32 _, err := t.CopyIn(addr, &fds) @@ -453,6 +481,12 @@ func (i *SyscallInfo) pre(t *kernel.Task, args arch.SyscallArguments, maximumBlo output = append(output, capData(t, args[arg-1].Pointer(), args[arg].Pointer())) case PollFDs: output = append(output, pollFDs(t, args[arg].Pointer(), uint(args[arg+1].Uint()), false)) + case EpollCtlOp: + output = append(output, epollCtlOps.Parse(uint64(args[arg].Int()))) + case EpollEvent: + output = append(output, epollEvent(t, args[arg].Pointer())) + case EpollEvents: + output = append(output, epollEvents(t, args[arg].Pointer(), 0 /* numEvents */, uint64(maximumBlobSize))) case SelectFDSet: output = append(output, fdSet(t, int(args[0].Int()), args[arg].Pointer())) case Oct: @@ -521,6 +555,8 @@ func (i *SyscallInfo) post(t *kernel.Task, args arch.SyscallArguments, rval uint output[arg] = capData(t, args[arg-1].Pointer(), args[arg].Pointer()) case PollFDs: output[arg] = pollFDs(t, args[arg].Pointer(), uint(args[arg+1].Uint()), true) + case EpollEvents: + output[arg] = epollEvents(t, args[arg].Pointer(), uint64(rval), uint64(maximumBlobSize)) case GetSockOptVal: output[arg] = getSockOptVal(t, args[arg-2].Uint64() /* level */, args[arg-1].Uint64() /* optName */, args[arg].Pointer() /* optVal */, args[arg+1].Pointer() /* optLen */, maximumBlobSize, rval) case SetSockOptVal: @@ -683,7 +719,7 @@ func (s SyscallMap) SyscallEnter(t *kernel.Task, sysno uintptr, args arch.Syscal // SyscallExit implements kernel.Stracer.SyscallExit. It logs the syscall // exit trace. func (s SyscallMap) SyscallExit(context interface{}, t *kernel.Task, sysno, rval uintptr, err error) { - errno := t.ExtractErrno(err, int(sysno)) + errno := kernel.ExtractErrno(err, int(sysno)) c := context.(*syscallContext) elapsed := time.Since(c.start) @@ -742,9 +778,6 @@ func (s SyscallMap) Name(sysno uintptr) string { // // N.B. This is not in an init function because we can't be sure all syscall // tables are registered with the kernel when init runs. -// -// TODO(gvisor.dev/issue/155): remove kernel package dependencies from this -// package and have the kernel package self-initialize all syscall tables. func Initialize() { for _, table := range kernel.SyscallTables() { // Is this known? diff --git a/pkg/sentry/strace/syscalls.go b/pkg/sentry/strace/syscalls.go index 446d1e0f6..7e69b9279 100644 --- a/pkg/sentry/strace/syscalls.go +++ b/pkg/sentry/strace/syscalls.go @@ -228,6 +228,16 @@ const ( // SockOptLevel is the optname argument in getsockopt(2) and // setsockopt(2). SockOptName + + // EpollCtlOp is the op argument to epoll_ctl(2). + EpollCtlOp + + // EpollEvent is the event argument in epoll_ctl(2). + EpollEvent + + // EpollEvents is an array of struct epoll_event. It is the events + // argument in epoll_wait(2)/epoll_pwait(2). + EpollEvents ) // defaultFormat is the syscall argument format to use if the actual format is diff --git a/pkg/sentry/syscalls/linux/BUILD b/pkg/sentry/syscalls/linux/BUILD index be16ee686..0d24fd3c4 100644 --- a/pkg/sentry/syscalls/linux/BUILD +++ b/pkg/sentry/syscalls/linux/BUILD @@ -74,6 +74,7 @@ go_library( "//pkg/sentry/fs/lock", "//pkg/sentry/fs/timerfd", "//pkg/sentry/fs/tmpfs", + "//pkg/sentry/fsbridge", "//pkg/sentry/kernel", "//pkg/sentry/kernel/auth", "//pkg/sentry/kernel/epoll", diff --git a/pkg/sentry/syscalls/linux/sys_aio.go b/pkg/sentry/syscalls/linux/sys_aio.go index b401978db..d781d6a04 100644 --- a/pkg/sentry/syscalls/linux/sys_aio.go +++ b/pkg/sentry/syscalls/linux/sys_aio.go @@ -114,14 +114,28 @@ func IoSetup(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysca func IoDestroy(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { id := args[0].Uint64() - // Destroy the given context. - if !t.MemoryManager().DestroyAIOContext(t, id) { + ctx := t.MemoryManager().DestroyAIOContext(t, id) + if ctx == nil { // Does not exist. return 0, nil, syserror.EINVAL } - // FIXME(fvoznika): Linux blocks until all AIO to the destroyed context is - // done. - return 0, nil, nil + + // Drain completed requests amd wait for pending requests until there are no + // more. + for { + ctx.Drain() + + ch := ctx.WaitChannel() + if ch == nil { + // No more requests, we're done. + return 0, nil, nil + } + // The task cannot be interrupted during the wait. Equivalent to + // TASK_UNINTERRUPTIBLE in Linux. + t.UninterruptibleSleepStart(true /* deactivate */) + <-ch + t.UninterruptibleSleepFinish(true /* activate */) + } } // IoGetevents implements linux syscall io_getevents(2). @@ -200,13 +214,13 @@ func IoGetevents(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.S func waitForRequest(ctx *mm.AIOContext, t *kernel.Task, haveDeadline bool, deadline ktime.Time) (interface{}, error) { for { if v, ok := ctx.PopRequest(); ok { - // Request was readly available. Just return it. + // Request was readily available. Just return it. return v, nil } // Need to wait for request completion. - done, active := ctx.WaitChannel() - if !active { + done := ctx.WaitChannel() + if done == nil { // Context has been destroyed. return nil, syserror.EINVAL } @@ -248,6 +262,10 @@ func memoryFor(t *kernel.Task, cb *ioCallback) (usermem.IOSequence, error) { } func performCallback(t *kernel.Task, file *fs.File, cbAddr usermem.Addr, cb *ioCallback, ioseq usermem.IOSequence, ctx *mm.AIOContext, eventFile *fs.File) { + if ctx.Dead() { + ctx.CancelPendingRequest() + return + } ev := &ioEvent{ Data: cb.Data, Obj: uint64(cbAddr), @@ -272,7 +290,7 @@ func performCallback(t *kernel.Task, file *fs.File, cbAddr usermem.Addr, cb *ioC // Update the result. if err != nil { err = handleIOError(t, ev.Result != 0 /* partial */, err, nil /* never interrupted */, "aio", file) - ev.Result = -int64(t.ExtractErrno(err, 0)) + ev.Result = -int64(kernel.ExtractErrno(err, 0)) } file.DecRef() diff --git a/pkg/sentry/syscalls/linux/sys_epoll.go b/pkg/sentry/syscalls/linux/sys_epoll.go index 5f11b496c..3ab93fbde 100644 --- a/pkg/sentry/syscalls/linux/sys_epoll.go +++ b/pkg/sentry/syscalls/linux/sys_epoll.go @@ -25,6 +25,8 @@ import ( "gvisor.dev/gvisor/pkg/waiter" ) +// LINT.IfChange + // EpollCreate1 implements the epoll_create1(2) linux syscall. func EpollCreate1(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { flags := args[0].Int() @@ -83,8 +85,7 @@ func EpollCtl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysc } mask = waiter.EventMaskFromLinux(e.Events) - data[0] = e.Fd - data[1] = e.Data + data = e.Data } // Perform the requested operations. @@ -165,3 +166,5 @@ func EpollPwait(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sy return EpollWait(t, args) } + +// LINT.ThenChange(vfs2/epoll.go) diff --git a/pkg/sentry/syscalls/linux/sys_file.go b/pkg/sentry/syscalls/linux/sys_file.go index 421845ebb..35a98212a 100644 --- a/pkg/sentry/syscalls/linux/sys_file.go +++ b/pkg/sentry/syscalls/linux/sys_file.go @@ -130,6 +130,8 @@ func copyInPath(t *kernel.Task, addr usermem.Addr, allowEmpty bool) (path string return path, dirPath, nil } +// LINT.IfChange + func openAt(t *kernel.Task, dirFD int32, addr usermem.Addr, flags uint) (fd uintptr, err error) { path, dirPath, err := copyInPath(t, addr, false /* allowEmpty */) if err != nil { @@ -512,7 +514,7 @@ func (ac accessContext) Value(key interface{}) interface{} { } } -func accessAt(t *kernel.Task, dirFD int32, addr usermem.Addr, resolve bool, mode uint) error { +func accessAt(t *kernel.Task, dirFD int32, addr usermem.Addr, mode uint) error { const rOK = 4 const wOK = 2 const xOK = 1 @@ -527,7 +529,7 @@ func accessAt(t *kernel.Task, dirFD int32, addr usermem.Addr, resolve bool, mode return syserror.EINVAL } - return fileOpOn(t, dirFD, path, resolve, func(root *fs.Dirent, d *fs.Dirent, _ uint) error { + return fileOpOn(t, dirFD, path, true /* resolve */, func(root *fs.Dirent, d *fs.Dirent, _ uint) error { // access(2) and faccessat(2) check permissions using real // UID/GID, not effective UID/GID. // @@ -562,19 +564,29 @@ func Access(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal addr := args[0].Pointer() mode := args[1].ModeT() - return 0, nil, accessAt(t, linux.AT_FDCWD, addr, true, mode) + return 0, nil, accessAt(t, linux.AT_FDCWD, addr, mode) } // Faccessat implements linux syscall faccessat(2). +// +// Note that the faccessat() system call does not take a flags argument: +// "The raw faccessat() system call takes only the first three arguments. The +// AT_EACCESS and AT_SYMLINK_NOFOLLOW flags are actually implemented within +// the glibc wrapper function for faccessat(). If either of these flags is +// specified, then the wrapper function employs fstatat(2) to determine access +// permissions." - faccessat(2) func Faccessat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { dirFD := args[0].Int() addr := args[1].Pointer() mode := args[2].ModeT() - flags := args[3].Int() - return 0, nil, accessAt(t, dirFD, addr, flags&linux.AT_SYMLINK_NOFOLLOW == 0, mode) + return 0, nil, accessAt(t, dirFD, addr, mode) } +// LINT.ThenChange(vfs2/filesystem.go) + +// LINT.IfChange + // Ioctl implements linux syscall ioctl(2). func Ioctl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { fd := args[0].Int() @@ -650,6 +662,10 @@ func Ioctl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall } } +// LINT.ThenChange(vfs2/ioctl.go) + +// LINT.IfChange + // Getcwd implements the linux syscall getcwd(2). func Getcwd(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { addr := args[0].Pointer() @@ -760,6 +776,10 @@ func Fchdir(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal return 0, nil, nil } +// LINT.ThenChange(vfs2/fscontext.go) + +// LINT.IfChange + // Close implements linux syscall close(2). func Close(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { fd := args[0].Int() @@ -1094,6 +1114,8 @@ func Fcntl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall } } +// LINT.ThenChange(vfs2/fd.go) + const ( _FADV_NORMAL = 0 _FADV_RANDOM = 1 @@ -1141,6 +1163,8 @@ func Fadvise64(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sys return 0, nil, nil } +// LINT.IfChange + func mkdirAt(t *kernel.Task, dirFD int32, addr usermem.Addr, mode linux.FileMode) error { path, _, err := copyInPath(t, addr, false /* allowEmpty */) if err != nil { @@ -1218,7 +1242,7 @@ func rmdirAt(t *kernel.Task, dirFD int32, addr usermem.Addr) error { return syserror.ENOTEMPTY } - if err := fs.MayDelete(t, root, d, name); err != nil { + if err := d.MayDelete(t, root, name); err != nil { return err } @@ -1421,6 +1445,10 @@ func Linkat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal return 0, nil, linkAt(t, oldDirFD, oldAddr, newDirFD, newAddr, resolve, allowEmpty) } +// LINT.ThenChange(vfs2/filesystem.go) + +// LINT.IfChange + func readlinkAt(t *kernel.Task, dirFD int32, addr usermem.Addr, bufAddr usermem.Addr, size uint) (copied uintptr, err error) { path, dirPath, err := copyInPath(t, addr, false /* allowEmpty */) if err != nil { @@ -1480,6 +1508,10 @@ func Readlinkat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sy return n, nil, err } +// LINT.ThenChange(vfs2/stat.go) + +// LINT.IfChange + func unlinkAt(t *kernel.Task, dirFD int32, addr usermem.Addr) error { path, dirPath, err := copyInPath(t, addr, false /* allowEmpty */) if err != nil { @@ -1491,7 +1523,7 @@ func unlinkAt(t *kernel.Task, dirFD int32, addr usermem.Addr) error { return syserror.ENOTDIR } - if err := fs.MayDelete(t, root, d, name); err != nil { + if err := d.MayDelete(t, root, name); err != nil { return err } @@ -1516,6 +1548,10 @@ func Unlinkat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysc return 0, nil, unlinkAt(t, dirFD, addr) } +// LINT.ThenChange(vfs2/filesystem.go) + +// LINT.IfChange + // Truncate implements linux syscall truncate(2). func Truncate(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { addr := args[0].Pointer() @@ -1614,6 +1650,8 @@ func Ftruncate(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sys return 0, nil, nil } +// LINT.ThenChange(vfs2/setstat.go) + // Umask implements linux syscall umask(2). func Umask(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { mask := args[0].ModeT() @@ -1621,6 +1659,8 @@ func Umask(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall return uintptr(mask), nil, nil } +// LINT.IfChange + // Change ownership of a file. // // uid and gid may be -1, in which case they will not be changed. @@ -1987,6 +2027,10 @@ func Futimesat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sys return 0, nil, utimes(t, dirFD, pathnameAddr, ts, true) } +// LINT.ThenChange(vfs2/setstat.go) + +// LINT.IfChange + func renameAt(t *kernel.Task, oldDirFD int32, oldAddr usermem.Addr, newDirFD int32, newAddr usermem.Addr) error { newPath, _, err := copyInPath(t, newAddr, false /* allowEmpty */) if err != nil { @@ -2042,6 +2086,8 @@ func Renameat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysc return 0, nil, renameAt(t, oldDirFD, oldPathAddr, newDirFD, newPathAddr) } +// LINT.ThenChange(vfs2/filesystem.go) + // Fallocate implements linux system call fallocate(2). func Fallocate(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { fd := args[0].Int() diff --git a/pkg/sentry/syscalls/linux/sys_getdents.go b/pkg/sentry/syscalls/linux/sys_getdents.go index f66f4ffde..b126fecc0 100644 --- a/pkg/sentry/syscalls/linux/sys_getdents.go +++ b/pkg/sentry/syscalls/linux/sys_getdents.go @@ -27,6 +27,8 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) +// LINT.IfChange + // Getdents implements linux syscall getdents(2) for 64bit systems. func Getdents(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { fd := args[0].Int() @@ -244,3 +246,5 @@ func (ds *direntSerializer) CopyOut(name string, attr fs.DentAttr) error { func (ds *direntSerializer) Written() int { return ds.written } + +// LINT.ThenChange(vfs2/getdents.go) diff --git a/pkg/sentry/syscalls/linux/sys_lseek.go b/pkg/sentry/syscalls/linux/sys_lseek.go index 297e920c4..3f7691eae 100644 --- a/pkg/sentry/syscalls/linux/sys_lseek.go +++ b/pkg/sentry/syscalls/linux/sys_lseek.go @@ -21,6 +21,8 @@ import ( "gvisor.dev/gvisor/pkg/syserror" ) +// LINT.IfChange + // Lseek implements linux syscall lseek(2). func Lseek(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { fd := args[0].Int() @@ -52,3 +54,5 @@ func Lseek(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall } return uintptr(offset), nil, err } + +// LINT.ThenChange(vfs2/read_write.go) diff --git a/pkg/sentry/syscalls/linux/sys_mmap.go b/pkg/sentry/syscalls/linux/sys_mmap.go index 9959f6e61..91694d374 100644 --- a/pkg/sentry/syscalls/linux/sys_mmap.go +++ b/pkg/sentry/syscalls/linux/sys_mmap.go @@ -35,6 +35,8 @@ func Brk(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallCo return uintptr(addr), nil, nil } +// LINT.IfChange + // Mmap implements linux syscall mmap(2). func Mmap(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { prot := args[2].Int() @@ -104,6 +106,8 @@ func Mmap(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallC return uintptr(rv), nil, err } +// LINT.ThenChange(vfs2/mmap.go) + // Munmap implements linux syscall munmap(2). func Munmap(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { return 0, nil, t.MemoryManager().MUnmap(t, args[0].Pointer(), args[1].Uint64()) diff --git a/pkg/sentry/syscalls/linux/sys_pipe.go b/pkg/sentry/syscalls/linux/sys_pipe.go index 798344042..43c510930 100644 --- a/pkg/sentry/syscalls/linux/sys_pipe.go +++ b/pkg/sentry/syscalls/linux/sys_pipe.go @@ -24,6 +24,8 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) +// LINT.IfChange + // pipe2 implements the actual system call with flags. func pipe2(t *kernel.Task, addr usermem.Addr, flags uint) (uintptr, error) { if flags&^(linux.O_NONBLOCK|linux.O_CLOEXEC) != 0 { @@ -45,10 +47,12 @@ func pipe2(t *kernel.Task, addr usermem.Addr, flags uint) (uintptr, error) { } if _, err := t.CopyOut(addr, fds); err != nil { - // The files are not closed in this case, the exact semantics - // of this error case are not well defined, but they could have - // already been observed by user space. - return 0, syserror.EFAULT + for _, fd := range fds { + if file, _ := t.FDTable().Remove(fd); file != nil { + file.DecRef() + } + } + return 0, err } return 0, nil } @@ -69,3 +73,5 @@ func Pipe2(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall n, err := pipe2(t, addr, flags) return n, nil, err } + +// LINT.ThenChange(vfs2/pipe.go) diff --git a/pkg/sentry/syscalls/linux/sys_prctl.go b/pkg/sentry/syscalls/linux/sys_prctl.go index 98db32d77..f92bf8096 100644 --- a/pkg/sentry/syscalls/linux/sys_prctl.go +++ b/pkg/sentry/syscalls/linux/sys_prctl.go @@ -20,6 +20,7 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/mm" @@ -135,7 +136,7 @@ func Prctl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall } // Set the underlying executable. - t.MemoryManager().SetExecutable(file.Dirent) + t.MemoryManager().SetExecutable(fsbridge.NewFSFile(file)) case linux.PR_SET_MM_AUXV, linux.PR_SET_MM_START_CODE, @@ -160,8 +161,8 @@ func Prctl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall if args[1].Int() != 1 || args[2].Int() != 0 || args[3].Int() != 0 || args[4].Int() != 0 { return 0, nil, syserror.EINVAL } - // no_new_privs is assumed to always be set. See - // kernel.Task.updateCredsForExec. + // PR_SET_NO_NEW_PRIVS is assumed to always be set. + // See kernel.Task.updateCredsForExecLocked. return 0, nil, nil case linux.PR_GET_NO_NEW_PRIVS: diff --git a/pkg/sentry/syscalls/linux/sys_read.go b/pkg/sentry/syscalls/linux/sys_read.go index 227692f06..071b4bacc 100644 --- a/pkg/sentry/syscalls/linux/sys_read.go +++ b/pkg/sentry/syscalls/linux/sys_read.go @@ -28,6 +28,8 @@ import ( "gvisor.dev/gvisor/pkg/waiter" ) +// LINT.IfChange + const ( // EventMaskRead contains events that can be triggered on reads. EventMaskRead = waiter.EventIn | waiter.EventHUp | waiter.EventErr @@ -94,8 +96,8 @@ func Readahead(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sys return 0, nil, syserror.EINVAL } - // Check that the offset is legitimate. - if offset < 0 { + // Check that the offset is legitimate and does not overflow. + if offset < 0 || offset+int64(size) < 0 { return 0, nil, syserror.EINVAL } @@ -118,8 +120,8 @@ func Pread64(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysca } defer file.DecRef() - // Check that the offset is legitimate. - if offset < 0 { + // Check that the offset is legitimate and does not overflow. + if offset < 0 || offset+int64(size) < 0 { return 0, nil, syserror.EINVAL } @@ -388,3 +390,5 @@ func preadv(t *kernel.Task, f *fs.File, dst usermem.IOSequence, offset int64) (i return total, err } + +// LINT.ThenChange(vfs2/read_write.go) diff --git a/pkg/sentry/syscalls/linux/sys_rlimit.go b/pkg/sentry/syscalls/linux/sys_rlimit.go index e08c333d6..d5d5b6959 100644 --- a/pkg/sentry/syscalls/linux/sys_rlimit.go +++ b/pkg/sentry/syscalls/linux/sys_rlimit.go @@ -197,7 +197,7 @@ func Prlimit64(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sys // saved set user IDs of the target process must match the real user ID of // the caller and the real, effective, and saved set group IDs of the // target process must match the real group ID of the caller." - if !t.HasCapabilityIn(linux.CAP_SYS_RESOURCE, t.PIDNamespace().UserNamespace()) { + if ot != t && !t.HasCapabilityIn(linux.CAP_SYS_RESOURCE, t.PIDNamespace().UserNamespace()) { cred, tcred := t.Credentials(), ot.Credentials() if cred.RealKUID != tcred.RealKUID || cred.RealKUID != tcred.EffectiveKUID || diff --git a/pkg/sentry/syscalls/linux/sys_socket.go b/pkg/sentry/syscalls/linux/sys_socket.go index 2919228d0..0760af77b 100644 --- a/pkg/sentry/syscalls/linux/sys_socket.go +++ b/pkg/sentry/syscalls/linux/sys_socket.go @@ -31,6 +31,8 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) +// LINT.IfChange + // minListenBacklog is the minimum reasonable backlog for listening sockets. const minListenBacklog = 8 @@ -244,7 +246,11 @@ func SocketPair(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sy // Copy the file descriptors out. if _, err := t.CopyOut(socks, fds); err != nil { - // Note that we don't close files here; see pipe(2) also. + for _, fd := range fds { + if file, _ := t.FDTable().Remove(fd); file != nil { + file.DecRef() + } + } return 0, nil, err } @@ -1128,3 +1134,5 @@ func SendTo(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal n, err := sendTo(t, fd, bufPtr, bufLen, flags, namePtr, nameLen) return n, nil, err } + +// LINT.ThenChange(./vfs2/socket.go) diff --git a/pkg/sentry/syscalls/linux/sys_splice.go b/pkg/sentry/syscalls/linux/sys_splice.go index fd642834b..df0d0f461 100644 --- a/pkg/sentry/syscalls/linux/sys_splice.go +++ b/pkg/sentry/syscalls/linux/sys_splice.go @@ -16,6 +16,7 @@ package linux import ( "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/log" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/kernel" @@ -25,10 +26,15 @@ import ( // doSplice implements a blocking splice operation. func doSplice(t *kernel.Task, outFile, inFile *fs.File, opts fs.SpliceOpts, nonBlocking bool) (int64, error) { - if opts.Length < 0 || opts.SrcStart < 0 || opts.DstStart < 0 { + log.Infof("NLAC: doSplice opts: %+v", opts) + if opts.Length < 0 || opts.SrcStart < 0 || opts.DstStart < 0 || (opts.SrcStart+opts.Length < 0) { return 0, syserror.EINVAL } + if opts.Length > int64(kernel.MAX_RW_COUNT) { + opts.Length = int64(kernel.MAX_RW_COUNT) + } + var ( total int64 n int64 diff --git a/pkg/sentry/syscalls/linux/sys_stat.go b/pkg/sentry/syscalls/linux/sys_stat.go index c841abccb..46ebf27a2 100644 --- a/pkg/sentry/syscalls/linux/sys_stat.go +++ b/pkg/sentry/syscalls/linux/sys_stat.go @@ -23,6 +23,8 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) +// LINT.IfChange + // Stat implements linux syscall stat(2). func Stat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { addr := args[0].Pointer() @@ -112,7 +114,9 @@ func stat(t *kernel.Task, d *fs.Dirent, dirPath bool, statAddr usermem.Addr) err if err != nil { return err } - return copyOutStat(t, statAddr, d.Inode.StableAttr, uattr) + s := statFromAttrs(t, d.Inode.StableAttr, uattr) + _, err = s.CopyOut(t, statAddr) + return err } // fstat implements fstat for the given *fs.File. @@ -121,7 +125,9 @@ func fstat(t *kernel.Task, f *fs.File, statAddr usermem.Addr) error { if err != nil { return err } - return copyOutStat(t, statAddr, f.Dirent.Inode.StableAttr, uattr) + s := statFromAttrs(t, f.Dirent.Inode.StableAttr, uattr) + _, err = s.CopyOut(t, statAddr) + return err } // Statx implements linux syscall statx(2). @@ -132,7 +138,10 @@ func Statx(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall mask := args[3].Uint() statxAddr := args[4].Pointer() - if mask&linux.STATX__RESERVED > 0 { + if mask&linux.STATX__RESERVED != 0 { + return 0, nil, syserror.EINVAL + } + if flags&^(linux.AT_SYMLINK_NOFOLLOW|linux.AT_EMPTY_PATH|linux.AT_STATX_SYNC_TYPE) != 0 { return 0, nil, syserror.EINVAL } if flags&linux.AT_STATX_SYNC_TYPE == linux.AT_STATX_SYNC_TYPE { @@ -277,3 +286,5 @@ func statfsImpl(t *kernel.Task, d *fs.Dirent, addr usermem.Addr) error { _, err = t.CopyOut(addr, &statfs) return err } + +// LINT.ThenChange(vfs2/stat.go) diff --git a/pkg/sentry/syscalls/linux/sys_stat_amd64.go b/pkg/sentry/syscalls/linux/sys_stat_amd64.go index 75a567bd4..0a04a6113 100644 --- a/pkg/sentry/syscalls/linux/sys_stat_amd64.go +++ b/pkg/sentry/syscalls/linux/sys_stat_amd64.go @@ -12,64 +12,34 @@ // See the License for the specific language governing permissions and // limitations under the License. -//+build amd64 +// +build amd64 package linux import ( "gvisor.dev/gvisor/pkg/abi/linux" - "gvisor.dev/gvisor/pkg/binary" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/kernel" - "gvisor.dev/gvisor/pkg/usermem" ) -// copyOutStat copies the attributes (sattr, uattr) to the struct stat at -// address dst in t's address space. It encodes the stat struct to bytes -// manually, as stat() is a very common syscall for many applications, and -// t.CopyObjectOut has noticeable performance impact due to its many slice -// allocations and use of reflection. -func copyOutStat(t *kernel.Task, dst usermem.Addr, sattr fs.StableAttr, uattr fs.UnstableAttr) error { - b := t.CopyScratchBuffer(int(linux.SizeOfStat))[:0] - - // Dev (uint64) - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(sattr.DeviceID)) - // Ino (uint64) - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(sattr.InodeID)) - // Nlink (uint64) - b = binary.AppendUint64(b, usermem.ByteOrder, uattr.Links) - // Mode (uint32) - b = binary.AppendUint32(b, usermem.ByteOrder, sattr.Type.LinuxType()|uint32(uattr.Perms.LinuxMode())) - // UID (uint32) - b = binary.AppendUint32(b, usermem.ByteOrder, uint32(uattr.Owner.UID.In(t.UserNamespace()).OrOverflow())) - // GID (uint32) - b = binary.AppendUint32(b, usermem.ByteOrder, uint32(uattr.Owner.GID.In(t.UserNamespace()).OrOverflow())) - // Padding (uint32) - b = binary.AppendUint32(b, usermem.ByteOrder, 0) - // Rdev (uint64) - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(linux.MakeDeviceID(sattr.DeviceFileMajor, sattr.DeviceFileMinor))) - // Size (uint64) - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(uattr.Size)) - // Blksize (uint64) - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(sattr.BlockSize)) - // Blocks (uint64) - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(uattr.Usage/512)) - - // ATime - atime := uattr.AccessTime.Timespec() - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(atime.Sec)) - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(atime.Nsec)) - - // MTime - mtime := uattr.ModificationTime.Timespec() - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(mtime.Sec)) - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(mtime.Nsec)) - - // CTime - ctime := uattr.StatusChangeTime.Timespec() - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(ctime.Sec)) - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(ctime.Nsec)) - - _, err := t.CopyOutBytes(dst, b) - return err +// LINT.IfChange + +func statFromAttrs(t *kernel.Task, sattr fs.StableAttr, uattr fs.UnstableAttr) linux.Stat { + return linux.Stat{ + Dev: sattr.DeviceID, + Ino: sattr.InodeID, + Nlink: uattr.Links, + Mode: sattr.Type.LinuxType() | uint32(uattr.Perms.LinuxMode()), + UID: uint32(uattr.Owner.UID.In(t.UserNamespace()).OrOverflow()), + GID: uint32(uattr.Owner.GID.In(t.UserNamespace()).OrOverflow()), + Rdev: uint64(linux.MakeDeviceID(sattr.DeviceFileMajor, sattr.DeviceFileMinor)), + Size: uattr.Size, + Blksize: sattr.BlockSize, + Blocks: uattr.Usage / 512, + ATime: uattr.AccessTime.Timespec(), + MTime: uattr.ModificationTime.Timespec(), + CTime: uattr.StatusChangeTime.Timespec(), + } } + +// LINT.ThenChange(vfs2/stat_amd64.go) diff --git a/pkg/sentry/syscalls/linux/sys_stat_arm64.go b/pkg/sentry/syscalls/linux/sys_stat_arm64.go index 80c98d05c..5a3b1bfad 100644 --- a/pkg/sentry/syscalls/linux/sys_stat_arm64.go +++ b/pkg/sentry/syscalls/linux/sys_stat_arm64.go @@ -12,66 +12,34 @@ // See the License for the specific language governing permissions and // limitations under the License. -//+build arm64 +// +build arm64 package linux import ( "gvisor.dev/gvisor/pkg/abi/linux" - "gvisor.dev/gvisor/pkg/binary" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/kernel" - "gvisor.dev/gvisor/pkg/usermem" ) -// copyOutStat copies the attributes (sattr, uattr) to the struct stat at -// address dst in t's address space. It encodes the stat struct to bytes -// manually, as stat() is a very common syscall for many applications, and -// t.CopyObjectOut has noticeable performance impact due to its many slice -// allocations and use of reflection. -func copyOutStat(t *kernel.Task, dst usermem.Addr, sattr fs.StableAttr, uattr fs.UnstableAttr) error { - b := t.CopyScratchBuffer(int(linux.SizeOfStat))[:0] - - // Dev (uint64) - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(sattr.DeviceID)) - // Ino (uint64) - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(sattr.InodeID)) - // Mode (uint32) - b = binary.AppendUint32(b, usermem.ByteOrder, sattr.Type.LinuxType()|uint32(uattr.Perms.LinuxMode())) - // Nlink (uint32) - b = binary.AppendUint32(b, usermem.ByteOrder, uint32(uattr.Links)) - // UID (uint32) - b = binary.AppendUint32(b, usermem.ByteOrder, uint32(uattr.Owner.UID.In(t.UserNamespace()).OrOverflow())) - // GID (uint32) - b = binary.AppendUint32(b, usermem.ByteOrder, uint32(uattr.Owner.GID.In(t.UserNamespace()).OrOverflow())) - // Rdev (uint64) - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(linux.MakeDeviceID(sattr.DeviceFileMajor, sattr.DeviceFileMinor))) - // Padding (uint64) - b = binary.AppendUint64(b, usermem.ByteOrder, 0) - // Size (uint64) - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(uattr.Size)) - // Blksize (uint32) - b = binary.AppendUint32(b, usermem.ByteOrder, uint32(sattr.BlockSize)) - // Padding (uint32) - b = binary.AppendUint32(b, usermem.ByteOrder, 0) - // Blocks (uint64) - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(uattr.Usage/512)) - - // ATime - atime := uattr.AccessTime.Timespec() - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(atime.Sec)) - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(atime.Nsec)) - - // MTime - mtime := uattr.ModificationTime.Timespec() - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(mtime.Sec)) - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(mtime.Nsec)) - - // CTime - ctime := uattr.StatusChangeTime.Timespec() - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(ctime.Sec)) - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(ctime.Nsec)) - - _, err := t.CopyOutBytes(dst, b) - return err +// LINT.IfChange + +func statFromAttrs(t *kernel.Task, sattr fs.StableAttr, uattr fs.UnstableAttr) linux.Stat { + return linux.Stat{ + Dev: sattr.DeviceID, + Ino: sattr.InodeID, + Nlink: uint32(uattr.Links), + Mode: sattr.Type.LinuxType() | uint32(uattr.Perms.LinuxMode()), + UID: uint32(uattr.Owner.UID.In(t.UserNamespace()).OrOverflow()), + GID: uint32(uattr.Owner.GID.In(t.UserNamespace()).OrOverflow()), + Rdev: uint64(linux.MakeDeviceID(sattr.DeviceFileMajor, sattr.DeviceFileMinor)), + Size: uattr.Size, + Blksize: int32(sattr.BlockSize), + Blocks: uattr.Usage / 512, + ATime: uattr.AccessTime.Timespec(), + MTime: uattr.ModificationTime.Timespec(), + CTime: uattr.StatusChangeTime.Timespec(), + } } + +// LINT.ThenChange(vfs2/stat_arm64.go) diff --git a/pkg/sentry/syscalls/linux/sys_sync.go b/pkg/sentry/syscalls/linux/sys_sync.go index 3e55235bd..5ad465ae3 100644 --- a/pkg/sentry/syscalls/linux/sys_sync.go +++ b/pkg/sentry/syscalls/linux/sys_sync.go @@ -22,6 +22,8 @@ import ( "gvisor.dev/gvisor/pkg/syserror" ) +// LINT.IfChange + // Sync implements linux system call sync(2). func Sync(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { t.MountNamespace().SyncAll(t) @@ -135,3 +137,5 @@ func SyncFileRange(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel return 0, nil, syserror.ConvertIntr(err, kernel.ERESTARTSYS) } + +// LINT.ThenChange(vfs2/sync.go) diff --git a/pkg/sentry/syscalls/linux/sys_thread.go b/pkg/sentry/syscalls/linux/sys_thread.go index 0c9e2255d..00915fdde 100644 --- a/pkg/sentry/syscalls/linux/sys_thread.go +++ b/pkg/sentry/syscalls/linux/sys_thread.go @@ -21,6 +21,7 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/kernel/sched" "gvisor.dev/gvisor/pkg/sentry/loader" @@ -119,7 +120,7 @@ func execveat(t *kernel.Task, dirFD int32, pathnameAddr, argvAddr, envvAddr user defer root.DecRef() var wd *fs.Dirent - var executable *fs.File + var executable fsbridge.File var closeOnExec bool if dirFD == linux.AT_FDCWD || path.IsAbs(pathname) { // Even if the pathname is absolute, we may still need the wd @@ -136,7 +137,15 @@ func execveat(t *kernel.Task, dirFD int32, pathnameAddr, argvAddr, envvAddr user closeOnExec = fdFlags.CloseOnExec if atEmptyPath && len(pathname) == 0 { - executable = f + // TODO(gvisor.dev/issue/160): Linux requires only execute permission, + // not read. However, our backing filesystems may prevent us from reading + // the file without read permission. Additionally, a task with a + // non-readable executable has additional constraints on access via + // ptrace and procfs. + if err := f.Dirent.Inode.CheckPermission(t, fs.PermMask{Read: true, Execute: true}); err != nil { + return 0, nil, err + } + executable = fsbridge.NewFSFile(f) } else { wd = f.Dirent wd.IncRef() @@ -152,9 +161,7 @@ func execveat(t *kernel.Task, dirFD int32, pathnameAddr, argvAddr, envvAddr user // Load the new TaskContext. remainingTraversals := uint(linux.MaxSymlinkTraversals) loadArgs := loader.LoadArgs{ - Mounts: t.MountNamespace(), - Root: root, - WorkingDirectory: wd, + Opener: fsbridge.NewFSLookup(t.MountNamespace(), root, wd), RemainingTraversals: &remainingTraversals, ResolveFinal: resolveFinal, Filename: pathname, diff --git a/pkg/sentry/syscalls/linux/sys_write.go b/pkg/sentry/syscalls/linux/sys_write.go index aba892939..6ec0de96e 100644 --- a/pkg/sentry/syscalls/linux/sys_write.go +++ b/pkg/sentry/syscalls/linux/sys_write.go @@ -28,6 +28,8 @@ import ( "gvisor.dev/gvisor/pkg/waiter" ) +// LINT.IfChange + const ( // EventMaskWrite contains events that can be triggered on writes. // @@ -85,8 +87,8 @@ func Pwrite64(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysc } defer file.DecRef() - // Check that the offset is legitimate. - if offset < 0 { + // Check that the offset is legitimate and does not overflow. + if offset < 0 || offset+int64(size) < 0 { return 0, nil, syserror.EINVAL } @@ -358,3 +360,5 @@ func pwritev(t *kernel.Task, f *fs.File, src usermem.IOSequence, offset int64) ( return total, err } + +// LINT.ThenChange(vfs2/read_write.go) diff --git a/pkg/sentry/syscalls/linux/sys_xattr.go b/pkg/sentry/syscalls/linux/sys_xattr.go index 9d8140b8a..2de5e3422 100644 --- a/pkg/sentry/syscalls/linux/sys_xattr.go +++ b/pkg/sentry/syscalls/linux/sys_xattr.go @@ -25,6 +25,8 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) +// LINT.IfChange + // GetXattr implements linux syscall getxattr(2). func GetXattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { return getXattrFromPath(t, args, true) @@ -418,3 +420,5 @@ func removeXattr(t *kernel.Task, d *fs.Dirent, nameAddr usermem.Addr) error { return d.Inode.RemoveXattr(t, d, name) } + +// LINT.ThenChange(vfs2/xattr.go) diff --git a/pkg/sentry/syscalls/linux/vfs2/BUILD b/pkg/sentry/syscalls/linux/vfs2/BUILD index 6b8a00b6e..6ff2d84d2 100644 --- a/pkg/sentry/syscalls/linux/vfs2/BUILD +++ b/pkg/sentry/syscalls/linux/vfs2/BUILD @@ -5,18 +5,57 @@ package(licenses = ["notice"]) go_library( name = "vfs2", srcs = [ + "epoll.go", + "epoll_unsafe.go", + "execve.go", + "fd.go", + "filesystem.go", + "fscontext.go", + "getdents.go", + "ioctl.go", "linux64.go", "linux64_override_amd64.go", "linux64_override_arm64.go", - "sys_read.go", + "mmap.go", + "path.go", + "pipe.go", + "poll.go", + "read_write.go", + "setstat.go", + "socket.go", + "stat.go", + "stat_amd64.go", + "stat_arm64.go", + "sync.go", + "sys_timerfd.go", + "xattr.go", ], + marshal = True, visibility = ["//:sandbox"], deps = [ + "//pkg/abi/linux", + "//pkg/binary", + "//pkg/bits", + "//pkg/fspath", + "//pkg/gohacks", "//pkg/sentry/arch", + "//pkg/sentry/fsbridge", + "//pkg/sentry/fsimpl/pipefs", "//pkg/sentry/kernel", + "//pkg/sentry/kernel/auth", + "//pkg/sentry/kernel/pipe", + "//pkg/sentry/kernel/time", + "//pkg/sentry/limits", + "//pkg/sentry/loader", + "//pkg/sentry/memmap", + "//pkg/sentry/socket", + "//pkg/sentry/socket/control", + "//pkg/sentry/socket/unix/transport", "//pkg/sentry/syscalls", "//pkg/sentry/syscalls/linux", "//pkg/sentry/vfs", + "//pkg/sync", + "//pkg/syserr", "//pkg/syserror", "//pkg/usermem", "//pkg/waiter", diff --git a/pkg/sentry/syscalls/linux/vfs2/epoll.go b/pkg/sentry/syscalls/linux/vfs2/epoll.go new file mode 100644 index 000000000..5a938cee2 --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/epoll.go @@ -0,0 +1,225 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "math" + "time" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" + "gvisor.dev/gvisor/pkg/waiter" +) + +// EpollCreate1 implements Linux syscall epoll_create1(2). +func EpollCreate1(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + flags := args[0].Int() + if flags&^linux.EPOLL_CLOEXEC != 0 { + return 0, nil, syserror.EINVAL + } + + file, err := t.Kernel().VFS().NewEpollInstanceFD() + if err != nil { + return 0, nil, err + } + defer file.DecRef() + + fd, err := t.NewFDFromVFS2(0, file, kernel.FDFlags{ + CloseOnExec: flags&linux.EPOLL_CLOEXEC != 0, + }) + if err != nil { + return 0, nil, err + } + return uintptr(fd), nil, nil +} + +// EpollCreate implements Linux syscall epoll_create(2). +func EpollCreate(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + size := args[0].Int() + + // "Since Linux 2.6.8, the size argument is ignored, but must be greater + // than zero" - epoll_create(2) + if size <= 0 { + return 0, nil, syserror.EINVAL + } + + file, err := t.Kernel().VFS().NewEpollInstanceFD() + if err != nil { + return 0, nil, err + } + defer file.DecRef() + + fd, err := t.NewFDFromVFS2(0, file, kernel.FDFlags{}) + if err != nil { + return 0, nil, err + } + return uintptr(fd), nil, nil +} + +// EpollCtl implements Linux syscall epoll_ctl(2). +func EpollCtl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + epfd := args[0].Int() + op := args[1].Int() + fd := args[2].Int() + eventAddr := args[3].Pointer() + + epfile := t.GetFileVFS2(epfd) + if epfile == nil { + return 0, nil, syserror.EBADF + } + defer epfile.DecRef() + ep, ok := epfile.Impl().(*vfs.EpollInstance) + if !ok { + return 0, nil, syserror.EINVAL + } + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + if epfile == file { + return 0, nil, syserror.EINVAL + } + + var event linux.EpollEvent + switch op { + case linux.EPOLL_CTL_ADD: + if _, err := event.CopyIn(t, eventAddr); err != nil { + return 0, nil, err + } + return 0, nil, ep.AddInterest(file, fd, event) + case linux.EPOLL_CTL_DEL: + return 0, nil, ep.DeleteInterest(file, fd) + case linux.EPOLL_CTL_MOD: + if _, err := event.CopyIn(t, eventAddr); err != nil { + return 0, nil, err + } + return 0, nil, ep.ModifyInterest(file, fd, event) + default: + return 0, nil, syserror.EINVAL + } +} + +// EpollWait implements Linux syscall epoll_wait(2). +func EpollWait(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + epfd := args[0].Int() + eventsAddr := args[1].Pointer() + maxEvents := int(args[2].Int()) + timeout := int(args[3].Int()) + + const _EP_MAX_EVENTS = math.MaxInt32 / sizeofEpollEvent // Linux: fs/eventpoll.c:EP_MAX_EVENTS + if maxEvents <= 0 || maxEvents > _EP_MAX_EVENTS { + return 0, nil, syserror.EINVAL + } + + epfile := t.GetFileVFS2(epfd) + if epfile == nil { + return 0, nil, syserror.EBADF + } + defer epfile.DecRef() + ep, ok := epfile.Impl().(*vfs.EpollInstance) + if !ok { + return 0, nil, syserror.EINVAL + } + + // Use a fixed-size buffer in a loop, instead of make([]linux.EpollEvent, + // maxEvents), so that the buffer can be allocated on the stack. + var ( + events [16]linux.EpollEvent + total int + ch chan struct{} + haveDeadline bool + deadline ktime.Time + ) + for { + batchEvents := len(events) + if batchEvents > maxEvents { + batchEvents = maxEvents + } + n := ep.ReadEvents(events[:batchEvents]) + maxEvents -= n + if n != 0 { + // Copy what we read out. + copiedEvents, err := copyOutEvents(t, eventsAddr, events[:n]) + eventsAddr += usermem.Addr(copiedEvents * sizeofEpollEvent) + total += copiedEvents + if err != nil { + if total != 0 { + return uintptr(total), nil, nil + } + return 0, nil, err + } + // If we've filled the application's event buffer, we're done. + if maxEvents == 0 { + return uintptr(total), nil, nil + } + // Loop if we read a full batch, under the expectation that there + // may be more events to read. + if n == batchEvents { + continue + } + } + // We get here if n != batchEvents. If we read any number of events + // (just now, or in a previous iteration of this loop), or if timeout + // is 0 (such that epoll_wait should be non-blocking), return the + // events we've read so far to the application. + if total != 0 || timeout == 0 { + return uintptr(total), nil, nil + } + // In the first iteration of this loop, register with the epoll + // instance for readability events, but then immediately continue the + // loop since we need to retry ReadEvents() before blocking. In all + // subsequent iterations, block until events are available, the timeout + // expires, or an interrupt arrives. + if ch == nil { + var w waiter.Entry + w, ch = waiter.NewChannelEntry(nil) + epfile.EventRegister(&w, waiter.EventIn) + defer epfile.EventUnregister(&w) + } else { + // Set up the timer if a timeout was specified. + if timeout > 0 && !haveDeadline { + timeoutDur := time.Duration(timeout) * time.Millisecond + deadline = t.Kernel().MonotonicClock().Now().Add(timeoutDur) + haveDeadline = true + } + if err := t.BlockWithDeadline(ch, haveDeadline, deadline); err != nil { + if err == syserror.ETIMEDOUT { + err = nil + } + // total must be 0 since otherwise we would have returned + // above. + return 0, nil, err + } + } + } +} + +// EpollPwait implements Linux syscall epoll_pwait(2). +func EpollPwait(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + maskAddr := args[4].Pointer() + maskSize := uint(args[5].Uint()) + + if err := setTempSignalSet(t, maskAddr, maskSize); err != nil { + return 0, nil, err + } + + return EpollWait(t, args) +} diff --git a/pkg/sentry/syscalls/linux/vfs2/epoll_unsafe.go b/pkg/sentry/syscalls/linux/vfs2/epoll_unsafe.go new file mode 100644 index 000000000..825f325bf --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/epoll_unsafe.go @@ -0,0 +1,44 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "reflect" + "runtime" + "unsafe" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/gohacks" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/usermem" +) + +const sizeofEpollEvent = int(unsafe.Sizeof(linux.EpollEvent{})) + +func copyOutEvents(t *kernel.Task, addr usermem.Addr, events []linux.EpollEvent) (int, error) { + if len(events) == 0 { + return 0, nil + } + // Cast events to a byte slice for copying. + var eventBytes []byte + eventBytesHdr := (*reflect.SliceHeader)(unsafe.Pointer(&eventBytes)) + eventBytesHdr.Data = uintptr(gohacks.Noescape(unsafe.Pointer(&events[0]))) + eventBytesHdr.Len = len(events) * sizeofEpollEvent + eventBytesHdr.Cap = len(events) * sizeofEpollEvent + copiedBytes, err := t.CopyOutBytes(addr, eventBytes) + runtime.KeepAlive(events) + copiedEvents := copiedBytes / sizeofEpollEvent // rounded down + return copiedEvents, err +} diff --git a/pkg/sentry/syscalls/linux/vfs2/execve.go b/pkg/sentry/syscalls/linux/vfs2/execve.go new file mode 100644 index 000000000..aef0078a8 --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/execve.go @@ -0,0 +1,137 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/loader" + slinux "gvisor.dev/gvisor/pkg/sentry/syscalls/linux" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +// Execve implements linux syscall execve(2). +func Execve(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + pathnameAddr := args[0].Pointer() + argvAddr := args[1].Pointer() + envvAddr := args[2].Pointer() + return execveat(t, linux.AT_FDCWD, pathnameAddr, argvAddr, envvAddr, 0 /* flags */) +} + +// Execveat implements linux syscall execveat(2). +func Execveat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + dirfd := args[0].Int() + pathnameAddr := args[1].Pointer() + argvAddr := args[2].Pointer() + envvAddr := args[3].Pointer() + flags := args[4].Int() + return execveat(t, dirfd, pathnameAddr, argvAddr, envvAddr, flags) +} + +func execveat(t *kernel.Task, dirfd int32, pathnameAddr, argvAddr, envvAddr usermem.Addr, flags int32) (uintptr, *kernel.SyscallControl, error) { + if flags&^(linux.AT_EMPTY_PATH|linux.AT_SYMLINK_NOFOLLOW) != 0 { + return 0, nil, syserror.EINVAL + } + + pathname, err := t.CopyInString(pathnameAddr, linux.PATH_MAX) + if err != nil { + return 0, nil, err + } + var argv, envv []string + if argvAddr != 0 { + var err error + argv, err = t.CopyInVector(argvAddr, slinux.ExecMaxElemSize, slinux.ExecMaxTotalSize) + if err != nil { + return 0, nil, err + } + } + if envvAddr != 0 { + var err error + envv, err = t.CopyInVector(envvAddr, slinux.ExecMaxElemSize, slinux.ExecMaxTotalSize) + if err != nil { + return 0, nil, err + } + } + + root := t.FSContext().RootDirectoryVFS2() + defer root.DecRef() + var executable fsbridge.File + closeOnExec := false + if path := fspath.Parse(pathname); dirfd != linux.AT_FDCWD && !path.Absolute { + // We must open the executable ourselves since dirfd is used as the + // starting point while resolving path, but the task working directory + // is used as the starting point while resolving interpreters (Linux: + // fs/binfmt_script.c:load_script() => fs/exec.c:open_exec() => + // do_open_execat(fd=AT_FDCWD)), and the loader package is currently + // incapable of handling this correctly. + if !path.HasComponents() && flags&linux.AT_EMPTY_PATH == 0 { + return 0, nil, syserror.ENOENT + } + dirfile, dirfileFlags := t.FDTable().GetVFS2(dirfd) + if dirfile == nil { + return 0, nil, syserror.EBADF + } + start := dirfile.VirtualDentry() + start.IncRef() + dirfile.DecRef() + closeOnExec = dirfileFlags.CloseOnExec + file, err := t.Kernel().VFS().OpenAt(t, t.Credentials(), &vfs.PathOperation{ + Root: root, + Start: start, + Path: path, + FollowFinalSymlink: flags&linux.AT_SYMLINK_NOFOLLOW == 0, + }, &vfs.OpenOptions{ + Flags: linux.O_RDONLY, + FileExec: true, + }) + start.DecRef() + if err != nil { + return 0, nil, err + } + defer file.DecRef() + executable = fsbridge.NewVFSFile(file) + } + + // Load the new TaskContext. + mntns := t.MountNamespaceVFS2() // FIXME(jamieliu): useless refcount change + defer mntns.DecRef() + wd := t.FSContext().WorkingDirectoryVFS2() + defer wd.DecRef() + remainingTraversals := uint(linux.MaxSymlinkTraversals) + loadArgs := loader.LoadArgs{ + Opener: fsbridge.NewVFSLookup(mntns, root, wd), + RemainingTraversals: &remainingTraversals, + ResolveFinal: flags&linux.AT_SYMLINK_NOFOLLOW == 0, + Filename: pathname, + File: executable, + CloseOnExec: closeOnExec, + Argv: argv, + Envv: envv, + Features: t.Arch().FeatureSet(), + } + + tc, se := t.Kernel().LoadTaskImage(t, loadArgs) + if se != nil { + return 0, nil, se.ToError() + } + + ctrl, err := t.Execve(tc) + return 0, ctrl, err +} diff --git a/pkg/sentry/syscalls/linux/vfs2/fd.go b/pkg/sentry/syscalls/linux/vfs2/fd.go new file mode 100644 index 000000000..8181d80f4 --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/fd.go @@ -0,0 +1,164 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/kernel/pipe" + slinux "gvisor.dev/gvisor/pkg/sentry/syscalls/linux" + "gvisor.dev/gvisor/pkg/syserror" +) + +// Close implements Linux syscall close(2). +func Close(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + + // Note that Remove provides a reference on the file that we may use to + // flush. It is still active until we drop the final reference below + // (and other reference-holding operations complete). + _, file := t.FDTable().Remove(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + err := file.OnClose(t) + return 0, nil, slinux.HandleIOErrorVFS2(t, false /* partial */, err, syserror.EINTR, "close", file) +} + +// Dup implements Linux syscall dup(2). +func Dup(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + newFD, err := t.NewFDFromVFS2(0, file, kernel.FDFlags{}) + if err != nil { + return 0, nil, syserror.EMFILE + } + return uintptr(newFD), nil, nil +} + +// Dup2 implements Linux syscall dup2(2). +func Dup2(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + oldfd := args[0].Int() + newfd := args[1].Int() + + if oldfd == newfd { + // As long as oldfd is valid, dup2() does nothing and returns newfd. + file := t.GetFileVFS2(oldfd) + if file == nil { + return 0, nil, syserror.EBADF + } + file.DecRef() + return uintptr(newfd), nil, nil + } + + return dup3(t, oldfd, newfd, 0) +} + +// Dup3 implements Linux syscall dup3(2). +func Dup3(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + oldfd := args[0].Int() + newfd := args[1].Int() + flags := args[2].Uint() + + if oldfd == newfd { + return 0, nil, syserror.EINVAL + } + + return dup3(t, oldfd, newfd, flags) +} + +func dup3(t *kernel.Task, oldfd, newfd int32, flags uint32) (uintptr, *kernel.SyscallControl, error) { + if flags&^linux.O_CLOEXEC != 0 { + return 0, nil, syserror.EINVAL + } + + file := t.GetFileVFS2(oldfd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + err := t.NewFDAtVFS2(newfd, file, kernel.FDFlags{ + CloseOnExec: flags&linux.O_CLOEXEC != 0, + }) + if err != nil { + return 0, nil, err + } + return uintptr(newfd), nil, nil +} + +// Fcntl implements linux syscall fcntl(2). +func Fcntl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + cmd := args[1].Int() + + file, flags := t.FDTable().GetVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + switch cmd { + case linux.F_DUPFD, linux.F_DUPFD_CLOEXEC: + minfd := args[2].Int() + fd, err := t.NewFDFromVFS2(minfd, file, kernel.FDFlags{ + CloseOnExec: cmd == linux.F_DUPFD_CLOEXEC, + }) + if err != nil { + return 0, nil, err + } + return uintptr(fd), nil, nil + case linux.F_GETFD: + return uintptr(flags.ToLinuxFDFlags()), nil, nil + case linux.F_SETFD: + flags := args[2].Uint() + t.FDTable().SetFlags(fd, kernel.FDFlags{ + CloseOnExec: flags&linux.FD_CLOEXEC != 0, + }) + return 0, nil, nil + case linux.F_GETFL: + return uintptr(file.StatusFlags()), nil, nil + case linux.F_SETFL: + return 0, nil, file.SetStatusFlags(t, t.Credentials(), args[2].Uint()) + case linux.F_SETPIPE_SZ: + pipefile, ok := file.Impl().(*pipe.VFSPipeFD) + if !ok { + return 0, nil, syserror.EBADF + } + n, err := pipefile.SetPipeSize(int64(args[2].Int())) + if err != nil { + return 0, nil, err + } + return uintptr(n), nil, nil + case linux.F_GETPIPE_SZ: + pipefile, ok := file.Impl().(*pipe.VFSPipeFD) + if !ok { + return 0, nil, syserror.EBADF + } + return uintptr(pipefile.PipeSize()), nil, nil + default: + // TODO(gvisor.dev/issue/1623): Everything else is not yet supported. + return 0, nil, syserror.EINVAL + } +} diff --git a/pkg/sentry/syscalls/linux/vfs2/filesystem.go b/pkg/sentry/syscalls/linux/vfs2/filesystem.go new file mode 100644 index 000000000..46d3e189c --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/filesystem.go @@ -0,0 +1,326 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +// Link implements Linux syscall link(2). +func Link(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + oldpathAddr := args[0].Pointer() + newpathAddr := args[1].Pointer() + return 0, nil, linkat(t, linux.AT_FDCWD, oldpathAddr, linux.AT_FDCWD, newpathAddr, 0 /* flags */) +} + +// Linkat implements Linux syscall linkat(2). +func Linkat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + olddirfd := args[0].Int() + oldpathAddr := args[1].Pointer() + newdirfd := args[2].Int() + newpathAddr := args[3].Pointer() + flags := args[4].Int() + return 0, nil, linkat(t, olddirfd, oldpathAddr, newdirfd, newpathAddr, flags) +} + +func linkat(t *kernel.Task, olddirfd int32, oldpathAddr usermem.Addr, newdirfd int32, newpathAddr usermem.Addr, flags int32) error { + if flags&^(linux.AT_EMPTY_PATH|linux.AT_SYMLINK_FOLLOW) != 0 { + return syserror.EINVAL + } + if flags&linux.AT_EMPTY_PATH != 0 && !t.HasCapability(linux.CAP_DAC_READ_SEARCH) { + return syserror.ENOENT + } + + oldpath, err := copyInPath(t, oldpathAddr) + if err != nil { + return err + } + oldtpop, err := getTaskPathOperation(t, olddirfd, oldpath, shouldAllowEmptyPath(flags&linux.AT_EMPTY_PATH != 0), shouldFollowFinalSymlink(flags&linux.AT_SYMLINK_FOLLOW != 0)) + if err != nil { + return err + } + defer oldtpop.Release() + + newpath, err := copyInPath(t, newpathAddr) + if err != nil { + return err + } + newtpop, err := getTaskPathOperation(t, newdirfd, newpath, disallowEmptyPath, nofollowFinalSymlink) + if err != nil { + return err + } + defer newtpop.Release() + + return t.Kernel().VFS().LinkAt(t, t.Credentials(), &oldtpop.pop, &newtpop.pop) +} + +// Mkdir implements Linux syscall mkdir(2). +func Mkdir(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + addr := args[0].Pointer() + mode := args[1].ModeT() + return 0, nil, mkdirat(t, linux.AT_FDCWD, addr, mode) +} + +// Mkdirat implements Linux syscall mkdirat(2). +func Mkdirat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + dirfd := args[0].Int() + addr := args[1].Pointer() + mode := args[2].ModeT() + return 0, nil, mkdirat(t, dirfd, addr, mode) +} + +func mkdirat(t *kernel.Task, dirfd int32, addr usermem.Addr, mode uint) error { + path, err := copyInPath(t, addr) + if err != nil { + return err + } + tpop, err := getTaskPathOperation(t, dirfd, path, disallowEmptyPath, nofollowFinalSymlink) + if err != nil { + return err + } + defer tpop.Release() + return t.Kernel().VFS().MkdirAt(t, t.Credentials(), &tpop.pop, &vfs.MkdirOptions{ + Mode: linux.FileMode(mode & (0777 | linux.S_ISVTX) &^ t.FSContext().Umask()), + }) +} + +// Mknod implements Linux syscall mknod(2). +func Mknod(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + addr := args[0].Pointer() + mode := args[1].ModeT() + dev := args[2].Uint() + return 0, nil, mknodat(t, linux.AT_FDCWD, addr, mode, dev) +} + +// Mknodat implements Linux syscall mknodat(2). +func Mknodat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + dirfd := args[0].Int() + addr := args[1].Pointer() + mode := args[2].ModeT() + dev := args[3].Uint() + return 0, nil, mknodat(t, dirfd, addr, mode, dev) +} + +func mknodat(t *kernel.Task, dirfd int32, addr usermem.Addr, mode uint, dev uint32) error { + path, err := copyInPath(t, addr) + if err != nil { + return err + } + tpop, err := getTaskPathOperation(t, dirfd, path, disallowEmptyPath, nofollowFinalSymlink) + if err != nil { + return err + } + defer tpop.Release() + major, minor := linux.DecodeDeviceID(dev) + return t.Kernel().VFS().MknodAt(t, t.Credentials(), &tpop.pop, &vfs.MknodOptions{ + Mode: linux.FileMode(mode &^ t.FSContext().Umask()), + DevMajor: uint32(major), + DevMinor: minor, + }) +} + +// Open implements Linux syscall open(2). +func Open(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + addr := args[0].Pointer() + flags := args[1].Uint() + mode := args[2].ModeT() + return openat(t, linux.AT_FDCWD, addr, flags, mode) +} + +// Openat implements Linux syscall openat(2). +func Openat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + dirfd := args[0].Int() + addr := args[1].Pointer() + flags := args[2].Uint() + mode := args[3].ModeT() + return openat(t, dirfd, addr, flags, mode) +} + +// Creat implements Linux syscall creat(2). +func Creat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + addr := args[0].Pointer() + mode := args[1].ModeT() + return openat(t, linux.AT_FDCWD, addr, linux.O_WRONLY|linux.O_CREAT|linux.O_TRUNC, mode) +} + +func openat(t *kernel.Task, dirfd int32, pathAddr usermem.Addr, flags uint32, mode uint) (uintptr, *kernel.SyscallControl, error) { + path, err := copyInPath(t, pathAddr) + if err != nil { + return 0, nil, err + } + tpop, err := getTaskPathOperation(t, dirfd, path, disallowEmptyPath, shouldFollowFinalSymlink(flags&linux.O_NOFOLLOW == 0)) + if err != nil { + return 0, nil, err + } + defer tpop.Release() + + file, err := t.Kernel().VFS().OpenAt(t, t.Credentials(), &tpop.pop, &vfs.OpenOptions{ + Flags: flags | linux.O_LARGEFILE, + Mode: linux.FileMode(mode & (0777 | linux.S_ISUID | linux.S_ISGID | linux.S_ISVTX) &^ t.FSContext().Umask()), + }) + if err != nil { + return 0, nil, err + } + defer file.DecRef() + + fd, err := t.NewFDFromVFS2(0, file, kernel.FDFlags{ + CloseOnExec: flags&linux.O_CLOEXEC != 0, + }) + return uintptr(fd), nil, err +} + +// Rename implements Linux syscall rename(2). +func Rename(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + oldpathAddr := args[0].Pointer() + newpathAddr := args[1].Pointer() + return 0, nil, renameat(t, linux.AT_FDCWD, oldpathAddr, linux.AT_FDCWD, newpathAddr, 0 /* flags */) +} + +// Renameat implements Linux syscall renameat(2). +func Renameat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + olddirfd := args[0].Int() + oldpathAddr := args[1].Pointer() + newdirfd := args[2].Int() + newpathAddr := args[3].Pointer() + return 0, nil, renameat(t, olddirfd, oldpathAddr, newdirfd, newpathAddr, 0 /* flags */) +} + +// Renameat2 implements Linux syscall renameat2(2). +func Renameat2(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + olddirfd := args[0].Int() + oldpathAddr := args[1].Pointer() + newdirfd := args[2].Int() + newpathAddr := args[3].Pointer() + flags := args[4].Uint() + return 0, nil, renameat(t, olddirfd, oldpathAddr, newdirfd, newpathAddr, flags) +} + +func renameat(t *kernel.Task, olddirfd int32, oldpathAddr usermem.Addr, newdirfd int32, newpathAddr usermem.Addr, flags uint32) error { + oldpath, err := copyInPath(t, oldpathAddr) + if err != nil { + return err + } + // "If oldpath refers to a symbolic link, the link is renamed" - rename(2) + oldtpop, err := getTaskPathOperation(t, olddirfd, oldpath, disallowEmptyPath, nofollowFinalSymlink) + if err != nil { + return err + } + defer oldtpop.Release() + + newpath, err := copyInPath(t, newpathAddr) + if err != nil { + return err + } + newtpop, err := getTaskPathOperation(t, newdirfd, newpath, disallowEmptyPath, nofollowFinalSymlink) + if err != nil { + return err + } + defer newtpop.Release() + + return t.Kernel().VFS().RenameAt(t, t.Credentials(), &oldtpop.pop, &newtpop.pop, &vfs.RenameOptions{ + Flags: flags, + }) +} + +// Rmdir implements Linux syscall rmdir(2). +func Rmdir(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + pathAddr := args[0].Pointer() + return 0, nil, rmdirat(t, linux.AT_FDCWD, pathAddr) +} + +func rmdirat(t *kernel.Task, dirfd int32, pathAddr usermem.Addr) error { + path, err := copyInPath(t, pathAddr) + if err != nil { + return err + } + tpop, err := getTaskPathOperation(t, dirfd, path, disallowEmptyPath, nofollowFinalSymlink) + if err != nil { + return err + } + defer tpop.Release() + return t.Kernel().VFS().RmdirAt(t, t.Credentials(), &tpop.pop) +} + +// Unlink implements Linux syscall unlink(2). +func Unlink(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + pathAddr := args[0].Pointer() + return 0, nil, unlinkat(t, linux.AT_FDCWD, pathAddr) +} + +func unlinkat(t *kernel.Task, dirfd int32, pathAddr usermem.Addr) error { + path, err := copyInPath(t, pathAddr) + if err != nil { + return err + } + tpop, err := getTaskPathOperation(t, dirfd, path, disallowEmptyPath, nofollowFinalSymlink) + if err != nil { + return err + } + defer tpop.Release() + return t.Kernel().VFS().UnlinkAt(t, t.Credentials(), &tpop.pop) +} + +// Unlinkat implements Linux syscall unlinkat(2). +func Unlinkat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + dirfd := args[0].Int() + pathAddr := args[1].Pointer() + flags := args[2].Int() + + if flags&^linux.AT_REMOVEDIR != 0 { + return 0, nil, syserror.EINVAL + } + + if flags&linux.AT_REMOVEDIR != 0 { + return 0, nil, rmdirat(t, dirfd, pathAddr) + } + return 0, nil, unlinkat(t, dirfd, pathAddr) +} + +// Symlink implements Linux syscall symlink(2). +func Symlink(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + targetAddr := args[0].Pointer() + linkpathAddr := args[1].Pointer() + return 0, nil, symlinkat(t, targetAddr, linux.AT_FDCWD, linkpathAddr) +} + +// Symlinkat implements Linux syscall symlinkat(2). +func Symlinkat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + targetAddr := args[0].Pointer() + newdirfd := args[1].Int() + linkpathAddr := args[2].Pointer() + return 0, nil, symlinkat(t, targetAddr, newdirfd, linkpathAddr) +} + +func symlinkat(t *kernel.Task, targetAddr usermem.Addr, newdirfd int32, linkpathAddr usermem.Addr) error { + target, err := t.CopyInString(targetAddr, linux.PATH_MAX) + if err != nil { + return err + } + linkpath, err := copyInPath(t, linkpathAddr) + if err != nil { + return err + } + tpop, err := getTaskPathOperation(t, newdirfd, linkpath, disallowEmptyPath, nofollowFinalSymlink) + if err != nil { + return err + } + defer tpop.Release() + return t.Kernel().VFS().SymlinkAt(t, t.Credentials(), &tpop.pop, target) +} diff --git a/pkg/sentry/syscalls/linux/vfs2/fscontext.go b/pkg/sentry/syscalls/linux/vfs2/fscontext.go new file mode 100644 index 000000000..317409a18 --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/fscontext.go @@ -0,0 +1,131 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" +) + +// Getcwd implements Linux syscall getcwd(2). +func Getcwd(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + addr := args[0].Pointer() + size := args[1].SizeT() + + root := t.FSContext().RootDirectoryVFS2() + wd := t.FSContext().WorkingDirectoryVFS2() + s, err := t.Kernel().VFS().PathnameForGetcwd(t, root, wd) + root.DecRef() + wd.DecRef() + if err != nil { + return 0, nil, err + } + + // Note this is >= because we need a terminator. + if uint(len(s)) >= size { + return 0, nil, syserror.ERANGE + } + + // Construct a byte slice containing a NUL terminator. + buf := t.CopyScratchBuffer(len(s) + 1) + copy(buf, s) + buf[len(buf)-1] = 0 + + // Write the pathname slice. + n, err := t.CopyOutBytes(addr, buf) + if err != nil { + return 0, nil, err + } + return uintptr(n), nil, nil +} + +// Chdir implements Linux syscall chdir(2). +func Chdir(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + addr := args[0].Pointer() + + path, err := copyInPath(t, addr) + if err != nil { + return 0, nil, err + } + tpop, err := getTaskPathOperation(t, linux.AT_FDCWD, path, disallowEmptyPath, followFinalSymlink) + if err != nil { + return 0, nil, err + } + defer tpop.Release() + + vd, err := t.Kernel().VFS().GetDentryAt(t, t.Credentials(), &tpop.pop, &vfs.GetDentryOptions{ + CheckSearchable: true, + }) + if err != nil { + return 0, nil, err + } + t.FSContext().SetWorkingDirectoryVFS2(vd) + vd.DecRef() + return 0, nil, nil +} + +// Fchdir implements Linux syscall fchdir(2). +func Fchdir(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + + tpop, err := getTaskPathOperation(t, fd, fspath.Path{}, allowEmptyPath, nofollowFinalSymlink) + if err != nil { + return 0, nil, err + } + defer tpop.Release() + + vd, err := t.Kernel().VFS().GetDentryAt(t, t.Credentials(), &tpop.pop, &vfs.GetDentryOptions{ + CheckSearchable: true, + }) + if err != nil { + return 0, nil, err + } + t.FSContext().SetWorkingDirectoryVFS2(vd) + vd.DecRef() + return 0, nil, nil +} + +// Chroot implements Linux syscall chroot(2). +func Chroot(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + addr := args[0].Pointer() + + if !t.HasCapability(linux.CAP_SYS_CHROOT) { + return 0, nil, syserror.EPERM + } + + path, err := copyInPath(t, addr) + if err != nil { + return 0, nil, err + } + tpop, err := getTaskPathOperation(t, linux.AT_FDCWD, path, disallowEmptyPath, followFinalSymlink) + if err != nil { + return 0, nil, err + } + defer tpop.Release() + + vd, err := t.Kernel().VFS().GetDentryAt(t, t.Credentials(), &tpop.pop, &vfs.GetDentryOptions{ + CheckSearchable: true, + }) + if err != nil { + return 0, nil, err + } + t.FSContext().SetRootDirectoryVFS2(vd) + vd.DecRef() + return 0, nil, nil +} diff --git a/pkg/sentry/syscalls/linux/vfs2/getdents.go b/pkg/sentry/syscalls/linux/vfs2/getdents.go new file mode 100644 index 000000000..62e98817d --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/getdents.go @@ -0,0 +1,161 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "fmt" + + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +// Getdents implements Linux syscall getdents(2). +func Getdents(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + return getdents(t, args, false /* isGetdents64 */) +} + +// Getdents64 implements Linux syscall getdents64(2). +func Getdents64(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + return getdents(t, args, true /* isGetdents64 */) +} + +func getdents(t *kernel.Task, args arch.SyscallArguments, isGetdents64 bool) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + addr := args[1].Pointer() + size := int(args[2].Uint()) + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + cb := getGetdentsCallback(t, addr, size, isGetdents64) + err := file.IterDirents(t, cb) + n := size - cb.remaining + putGetdentsCallback(cb) + if n == 0 { + return 0, nil, err + } + return uintptr(n), nil, nil +} + +type getdentsCallback struct { + t *kernel.Task + addr usermem.Addr + remaining int + isGetdents64 bool +} + +var getdentsCallbackPool = sync.Pool{ + New: func() interface{} { + return &getdentsCallback{} + }, +} + +func getGetdentsCallback(t *kernel.Task, addr usermem.Addr, size int, isGetdents64 bool) *getdentsCallback { + cb := getdentsCallbackPool.Get().(*getdentsCallback) + *cb = getdentsCallback{ + t: t, + addr: addr, + remaining: size, + isGetdents64: isGetdents64, + } + return cb +} + +func putGetdentsCallback(cb *getdentsCallback) { + cb.t = nil + getdentsCallbackPool.Put(cb) +} + +// Handle implements vfs.IterDirentsCallback.Handle. +func (cb *getdentsCallback) Handle(dirent vfs.Dirent) error { + var buf []byte + if cb.isGetdents64 { + // struct linux_dirent64 { + // ino64_t d_ino; /* 64-bit inode number */ + // off64_t d_off; /* 64-bit offset to next structure */ + // unsigned short d_reclen; /* Size of this dirent */ + // unsigned char d_type; /* File type */ + // char d_name[]; /* Filename (null-terminated) */ + // }; + size := 8 + 8 + 2 + 1 + 1 + len(dirent.Name) + size = (size + 7) &^ 7 // round up to multiple of 8 + if size > cb.remaining { + return syserror.EINVAL + } + buf = cb.t.CopyScratchBuffer(size) + usermem.ByteOrder.PutUint64(buf[0:8], dirent.Ino) + usermem.ByteOrder.PutUint64(buf[8:16], uint64(dirent.NextOff)) + usermem.ByteOrder.PutUint16(buf[16:18], uint16(size)) + buf[18] = dirent.Type + copy(buf[19:], dirent.Name) + // Zero out all remaining bytes in buf, including the NUL terminator + // after dirent.Name. + bufTail := buf[19+len(dirent.Name):] + for i := range bufTail { + bufTail[i] = 0 + } + } else { + // struct linux_dirent { + // unsigned long d_ino; /* Inode number */ + // unsigned long d_off; /* Offset to next linux_dirent */ + // unsigned short d_reclen; /* Length of this linux_dirent */ + // char d_name[]; /* Filename (null-terminated) */ + // /* length is actually (d_reclen - 2 - + // offsetof(struct linux_dirent, d_name)) */ + // /* + // char pad; // Zero padding byte + // char d_type; // File type (only since Linux + // // 2.6.4); offset is (d_reclen - 1) + // */ + // }; + if cb.t.Arch().Width() != 8 { + panic(fmt.Sprintf("unsupported sizeof(unsigned long): %d", cb.t.Arch().Width())) + } + size := 8 + 8 + 2 + 1 + 1 + 1 + len(dirent.Name) + size = (size + 7) &^ 7 // round up to multiple of sizeof(long) + if size > cb.remaining { + return syserror.EINVAL + } + buf = cb.t.CopyScratchBuffer(size) + usermem.ByteOrder.PutUint64(buf[0:8], dirent.Ino) + usermem.ByteOrder.PutUint64(buf[8:16], uint64(dirent.NextOff)) + usermem.ByteOrder.PutUint16(buf[16:18], uint16(size)) + copy(buf[18:], dirent.Name) + // Zero out all remaining bytes in buf, including the NUL terminator + // after dirent.Name and the zero padding byte between the name and + // dirent type. + bufTail := buf[18+len(dirent.Name):] + for i := range bufTail { + bufTail[i] = 0 + } + bufTail[2] = dirent.Type + } + n, err := cb.t.CopyOutBytes(cb.addr, buf) + if err != nil { + // Don't report partially-written dirents by advancing cb.addr or + // cb.remaining. + return err + } + cb.addr += usermem.Addr(n) + cb.remaining -= n + return nil +} diff --git a/pkg/sentry/syscalls/linux/vfs2/ioctl.go b/pkg/sentry/syscalls/linux/vfs2/ioctl.go new file mode 100644 index 000000000..5a2418da9 --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/ioctl.go @@ -0,0 +1,35 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/syserror" +) + +// Ioctl implements Linux syscall ioctl(2). +func Ioctl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + ret, err := file.Ioctl(t, t.MemoryManager(), args) + return ret, nil, err +} diff --git a/pkg/sentry/syscalls/linux/vfs2/linux64_override_amd64.go b/pkg/sentry/syscalls/linux/vfs2/linux64_override_amd64.go index c134714ee..21eb98444 100644 --- a/pkg/sentry/syscalls/linux/vfs2/linux64_override_amd64.go +++ b/pkg/sentry/syscalls/linux/vfs2/linux64_override_amd64.go @@ -12,6 +12,8 @@ // See the License for the specific language governing permissions and // limitations under the License. +// +build amd64 + package vfs2 import ( @@ -22,4 +24,146 @@ import ( // Override syscall table to add syscalls implementations from this package. func Override(table map[uintptr]kernel.Syscall) { table[0] = syscalls.Supported("read", Read) + table[1] = syscalls.Supported("write", Write) + table[2] = syscalls.Supported("open", Open) + table[3] = syscalls.Supported("close", Close) + table[4] = syscalls.Supported("stat", Stat) + table[5] = syscalls.Supported("fstat", Fstat) + table[6] = syscalls.Supported("lstat", Lstat) + table[7] = syscalls.Supported("poll", Poll) + table[8] = syscalls.Supported("lseek", Lseek) + table[9] = syscalls.Supported("mmap", Mmap) + table[16] = syscalls.Supported("ioctl", Ioctl) + table[17] = syscalls.Supported("pread64", Pread64) + table[18] = syscalls.Supported("pwrite64", Pwrite64) + table[19] = syscalls.Supported("readv", Readv) + table[20] = syscalls.Supported("writev", Writev) + table[21] = syscalls.Supported("access", Access) + table[22] = syscalls.Supported("pipe", Pipe) + table[23] = syscalls.Supported("select", Select) + table[32] = syscalls.Supported("dup", Dup) + table[33] = syscalls.Supported("dup2", Dup2) + delete(table, 40) // sendfile + // TODO(gvisor.dev/issue/1485): Port all socket variants to VFS2. + table[41] = syscalls.PartiallySupported("socket", Socket, "In process of porting socket syscalls to VFS2.", nil) + table[42] = syscalls.PartiallySupported("connect", Connect, "In process of porting socket syscalls to VFS2.", nil) + table[43] = syscalls.PartiallySupported("accept", Accept, "In process of porting socket syscalls to VFS2.", nil) + table[44] = syscalls.PartiallySupported("sendto", SendTo, "In process of porting socket syscalls to VFS2.", nil) + table[45] = syscalls.PartiallySupported("recvfrom", RecvFrom, "In process of porting socket syscalls to VFS2.", nil) + table[46] = syscalls.PartiallySupported("sendmsg", SendMsg, "In process of porting socket syscalls to VFS2.", nil) + table[47] = syscalls.PartiallySupported("recvmsg", RecvMsg, "In process of porting socket syscalls to VFS2.", nil) + table[48] = syscalls.PartiallySupported("shutdown", Shutdown, "In process of porting socket syscalls to VFS2.", nil) + table[49] = syscalls.PartiallySupported("bind", Bind, "In process of porting socket syscalls to VFS2.", nil) + table[50] = syscalls.PartiallySupported("listen", Listen, "In process of porting socket syscalls to VFS2.", nil) + table[51] = syscalls.PartiallySupported("getsockname", GetSockName, "In process of porting socket syscalls to VFS2.", nil) + table[52] = syscalls.PartiallySupported("getpeername", GetPeerName, "In process of porting socket syscalls to VFS2.", nil) + table[53] = syscalls.PartiallySupported("socketpair", SocketPair, "In process of porting socket syscalls to VFS2.", nil) + table[54] = syscalls.PartiallySupported("getsockopt", GetSockOpt, "In process of porting socket syscalls to VFS2.", nil) + table[55] = syscalls.PartiallySupported("setsockopt", SetSockOpt, "In process of porting socket syscalls to VFS2.", nil) + table[59] = syscalls.Supported("execve", Execve) + table[72] = syscalls.Supported("fcntl", Fcntl) + delete(table, 73) // flock + table[74] = syscalls.Supported("fsync", Fsync) + table[75] = syscalls.Supported("fdatasync", Fdatasync) + table[76] = syscalls.Supported("truncate", Truncate) + table[77] = syscalls.Supported("ftruncate", Ftruncate) + table[78] = syscalls.Supported("getdents", Getdents) + table[79] = syscalls.Supported("getcwd", Getcwd) + table[80] = syscalls.Supported("chdir", Chdir) + table[81] = syscalls.Supported("fchdir", Fchdir) + table[82] = syscalls.Supported("rename", Rename) + table[83] = syscalls.Supported("mkdir", Mkdir) + table[84] = syscalls.Supported("rmdir", Rmdir) + table[85] = syscalls.Supported("creat", Creat) + table[86] = syscalls.Supported("link", Link) + table[87] = syscalls.Supported("unlink", Unlink) + table[88] = syscalls.Supported("symlink", Symlink) + table[89] = syscalls.Supported("readlink", Readlink) + table[90] = syscalls.Supported("chmod", Chmod) + table[91] = syscalls.Supported("fchmod", Fchmod) + table[92] = syscalls.Supported("chown", Chown) + table[93] = syscalls.Supported("fchown", Fchown) + table[94] = syscalls.Supported("lchown", Lchown) + table[132] = syscalls.Supported("utime", Utime) + table[133] = syscalls.Supported("mknod", Mknod) + table[137] = syscalls.Supported("statfs", Statfs) + table[138] = syscalls.Supported("fstatfs", Fstatfs) + table[161] = syscalls.Supported("chroot", Chroot) + table[162] = syscalls.Supported("sync", Sync) + delete(table, 165) // mount + delete(table, 166) // umount2 + delete(table, 187) // readahead + table[188] = syscalls.Supported("setxattr", Setxattr) + table[189] = syscalls.Supported("lsetxattr", Lsetxattr) + table[190] = syscalls.Supported("fsetxattr", Fsetxattr) + table[191] = syscalls.Supported("getxattr", Getxattr) + table[192] = syscalls.Supported("lgetxattr", Lgetxattr) + table[193] = syscalls.Supported("fgetxattr", Fgetxattr) + table[194] = syscalls.Supported("listxattr", Listxattr) + table[195] = syscalls.Supported("llistxattr", Llistxattr) + table[196] = syscalls.Supported("flistxattr", Flistxattr) + table[197] = syscalls.Supported("removexattr", Removexattr) + table[198] = syscalls.Supported("lremovexattr", Lremovexattr) + table[199] = syscalls.Supported("fremovexattr", Fremovexattr) + delete(table, 206) // io_setup + delete(table, 207) // io_destroy + delete(table, 208) // io_getevents + delete(table, 209) // io_submit + delete(table, 210) // io_cancel + table[213] = syscalls.Supported("epoll_create", EpollCreate) + table[217] = syscalls.Supported("getdents64", Getdents64) + delete(table, 221) // fdavise64 + table[232] = syscalls.Supported("epoll_wait", EpollWait) + table[233] = syscalls.Supported("epoll_ctl", EpollCtl) + table[235] = syscalls.Supported("utimes", Utimes) + delete(table, 253) // inotify_init + delete(table, 254) // inotify_add_watch + delete(table, 255) // inotify_rm_watch + table[257] = syscalls.Supported("openat", Openat) + table[258] = syscalls.Supported("mkdirat", Mkdirat) + table[259] = syscalls.Supported("mknodat", Mknodat) + table[260] = syscalls.Supported("fchownat", Fchownat) + table[261] = syscalls.Supported("futimens", Futimens) + table[262] = syscalls.Supported("newfstatat", Newfstatat) + table[263] = syscalls.Supported("unlinkat", Unlinkat) + table[264] = syscalls.Supported("renameat", Renameat) + table[265] = syscalls.Supported("linkat", Linkat) + table[266] = syscalls.Supported("symlinkat", Symlinkat) + table[267] = syscalls.Supported("readlinkat", Readlinkat) + table[268] = syscalls.Supported("fchmodat", Fchmodat) + table[269] = syscalls.Supported("faccessat", Faccessat) + table[270] = syscalls.Supported("pselect", Pselect) + table[271] = syscalls.Supported("ppoll", Ppoll) + delete(table, 275) // splice + delete(table, 276) // tee + table[277] = syscalls.Supported("sync_file_range", SyncFileRange) + table[280] = syscalls.Supported("utimensat", Utimensat) + table[281] = syscalls.Supported("epoll_pwait", EpollPwait) + delete(table, 282) // signalfd + table[283] = syscalls.Supported("timerfd_create", TimerfdCreate) + delete(table, 284) // eventfd + delete(table, 285) // fallocate + table[286] = syscalls.Supported("timerfd_settime", TimerfdSettime) + table[287] = syscalls.Supported("timerfd_gettime", TimerfdGettime) + // TODO(gvisor.dev/issue/1485): Port all socket variants to VFS2. + table[288] = syscalls.PartiallySupported("accept4", Accept4, "In process of porting socket syscalls to VFS2.", nil) + delete(table, 289) // signalfd4 + delete(table, 290) // eventfd2 + table[291] = syscalls.Supported("epoll_create1", EpollCreate1) + table[292] = syscalls.Supported("dup3", Dup3) + table[293] = syscalls.Supported("pipe2", Pipe2) + delete(table, 294) // inotify_init1 + table[295] = syscalls.Supported("preadv", Preadv) + table[296] = syscalls.Supported("pwritev", Pwritev) + // TODO(gvisor.dev/issue/1485): Port all socket variants to VFS2. + table[299] = syscalls.PartiallySupported("recvmmsg", RecvMMsg, "In process of porting socket syscalls to VFS2.", nil) + table[306] = syscalls.Supported("syncfs", Syncfs) + // TODO(gvisor.dev/issue/1485): Port all socket variants to VFS2. + table[307] = syscalls.PartiallySupported("sendmmsg", SendMMsg, "In process of porting socket syscalls to VFS2.", nil) + table[316] = syscalls.Supported("renameat2", Renameat2) + delete(table, 319) // memfd_create + table[322] = syscalls.Supported("execveat", Execveat) + table[327] = syscalls.Supported("preadv2", Preadv2) + table[328] = syscalls.Supported("pwritev2", Pwritev2) + table[332] = syscalls.Supported("statx", Statx) } diff --git a/pkg/sentry/syscalls/linux/vfs2/linux64_override_arm64.go b/pkg/sentry/syscalls/linux/vfs2/linux64_override_arm64.go index 6af5c400f..a6b367468 100644 --- a/pkg/sentry/syscalls/linux/vfs2/linux64_override_arm64.go +++ b/pkg/sentry/syscalls/linux/vfs2/linux64_override_arm64.go @@ -12,6 +12,8 @@ // See the License for the specific language governing permissions and // limitations under the License. +// +build arm64 + package vfs2 import ( diff --git a/pkg/sentry/syscalls/linux/vfs2/mmap.go b/pkg/sentry/syscalls/linux/vfs2/mmap.go new file mode 100644 index 000000000..60a43f0a0 --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/mmap.go @@ -0,0 +1,92 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/memmap" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +// Mmap implements Linux syscall mmap(2). +func Mmap(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + prot := args[2].Int() + flags := args[3].Int() + fd := args[4].Int() + fixed := flags&linux.MAP_FIXED != 0 + private := flags&linux.MAP_PRIVATE != 0 + shared := flags&linux.MAP_SHARED != 0 + anon := flags&linux.MAP_ANONYMOUS != 0 + map32bit := flags&linux.MAP_32BIT != 0 + + // Require exactly one of MAP_PRIVATE and MAP_SHARED. + if private == shared { + return 0, nil, syserror.EINVAL + } + + opts := memmap.MMapOpts{ + Length: args[1].Uint64(), + Offset: args[5].Uint64(), + Addr: args[0].Pointer(), + Fixed: fixed, + Unmap: fixed, + Map32Bit: map32bit, + Private: private, + Perms: usermem.AccessType{ + Read: linux.PROT_READ&prot != 0, + Write: linux.PROT_WRITE&prot != 0, + Execute: linux.PROT_EXEC&prot != 0, + }, + MaxPerms: usermem.AnyAccess, + GrowsDown: linux.MAP_GROWSDOWN&flags != 0, + Precommit: linux.MAP_POPULATE&flags != 0, + } + if linux.MAP_LOCKED&flags != 0 { + opts.MLockMode = memmap.MLockEager + } + defer func() { + if opts.MappingIdentity != nil { + opts.MappingIdentity.DecRef() + } + }() + + if !anon { + // Convert the passed FD to a file reference. + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // mmap unconditionally requires that the FD is readable. + if !file.IsReadable() { + return 0, nil, syserror.EACCES + } + // MAP_SHARED requires that the FD be writable for PROT_WRITE. + if shared && !file.IsWritable() { + opts.MaxPerms.Write = false + } + + if err := file.ConfigureMMap(t, &opts); err != nil { + return 0, nil, err + } + } + + rv, err := t.MemoryManager().MMap(t, opts) + return uintptr(rv), nil, err +} diff --git a/pkg/sentry/syscalls/linux/vfs2/path.go b/pkg/sentry/syscalls/linux/vfs2/path.go new file mode 100644 index 000000000..97da6c647 --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/path.go @@ -0,0 +1,94 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +func copyInPath(t *kernel.Task, addr usermem.Addr) (fspath.Path, error) { + pathname, err := t.CopyInString(addr, linux.PATH_MAX) + if err != nil { + return fspath.Path{}, err + } + return fspath.Parse(pathname), nil +} + +type taskPathOperation struct { + pop vfs.PathOperation + haveStartRef bool +} + +func getTaskPathOperation(t *kernel.Task, dirfd int32, path fspath.Path, shouldAllowEmptyPath shouldAllowEmptyPath, shouldFollowFinalSymlink shouldFollowFinalSymlink) (taskPathOperation, error) { + root := t.FSContext().RootDirectoryVFS2() + start := root + haveStartRef := false + if !path.Absolute { + if !path.HasComponents() && !bool(shouldAllowEmptyPath) { + root.DecRef() + return taskPathOperation{}, syserror.ENOENT + } + if dirfd == linux.AT_FDCWD { + start = t.FSContext().WorkingDirectoryVFS2() + haveStartRef = true + } else { + dirfile := t.GetFileVFS2(dirfd) + if dirfile == nil { + root.DecRef() + return taskPathOperation{}, syserror.EBADF + } + start = dirfile.VirtualDentry() + start.IncRef() + haveStartRef = true + dirfile.DecRef() + } + } + return taskPathOperation{ + pop: vfs.PathOperation{ + Root: root, + Start: start, + Path: path, + FollowFinalSymlink: bool(shouldFollowFinalSymlink), + }, + haveStartRef: haveStartRef, + }, nil +} + +func (tpop *taskPathOperation) Release() { + tpop.pop.Root.DecRef() + if tpop.haveStartRef { + tpop.pop.Start.DecRef() + tpop.haveStartRef = false + } +} + +type shouldAllowEmptyPath bool + +const ( + disallowEmptyPath shouldAllowEmptyPath = false + allowEmptyPath shouldAllowEmptyPath = true +) + +type shouldFollowFinalSymlink bool + +const ( + nofollowFinalSymlink shouldFollowFinalSymlink = false + followFinalSymlink shouldFollowFinalSymlink = true +) diff --git a/pkg/sentry/syscalls/linux/vfs2/pipe.go b/pkg/sentry/syscalls/linux/vfs2/pipe.go new file mode 100644 index 000000000..4a01e4209 --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/pipe.go @@ -0,0 +1,63 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/pipefs" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +// Pipe implements Linux syscall pipe(2). +func Pipe(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + addr := args[0].Pointer() + return 0, nil, pipe2(t, addr, 0) +} + +// Pipe2 implements Linux syscall pipe2(2). +func Pipe2(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + addr := args[0].Pointer() + flags := args[1].Int() + return 0, nil, pipe2(t, addr, flags) +} + +func pipe2(t *kernel.Task, addr usermem.Addr, flags int32) error { + if flags&^(linux.O_NONBLOCK|linux.O_CLOEXEC) != 0 { + return syserror.EINVAL + } + r, w := pipefs.NewConnectedPipeFDs(t, t.Kernel().PipeMount(), uint32(flags&linux.O_NONBLOCK)) + defer r.DecRef() + defer w.DecRef() + + fds, err := t.NewFDsVFS2(0, []*vfs.FileDescription{r, w}, kernel.FDFlags{ + CloseOnExec: flags&linux.O_CLOEXEC != 0, + }) + if err != nil { + return err + } + if _, err := t.CopyOut(addr, fds); err != nil { + for _, fd := range fds { + if _, file := t.FDTable().Remove(fd); file != nil { + file.DecRef() + } + } + return err + } + return nil +} diff --git a/pkg/sentry/syscalls/linux/vfs2/poll.go b/pkg/sentry/syscalls/linux/vfs2/poll.go new file mode 100644 index 000000000..ff1b25d7b --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/poll.go @@ -0,0 +1,586 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "fmt" + "time" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" + "gvisor.dev/gvisor/pkg/sentry/limits" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" + "gvisor.dev/gvisor/pkg/waiter" +) + +// fileCap is the maximum allowable files for poll & select. This has no +// equivalent in Linux; it exists in gVisor since allocation failure in Go is +// unrecoverable. +const fileCap = 1024 * 1024 + +// Masks for "readable", "writable", and "exceptional" events as defined by +// select(2). +const ( + // selectReadEvents is analogous to the Linux kernel's + // fs/select.c:POLLIN_SET. + selectReadEvents = linux.POLLIN | linux.POLLHUP | linux.POLLERR + + // selectWriteEvents is analogous to the Linux kernel's + // fs/select.c:POLLOUT_SET. + selectWriteEvents = linux.POLLOUT | linux.POLLERR + + // selectExceptEvents is analogous to the Linux kernel's + // fs/select.c:POLLEX_SET. + selectExceptEvents = linux.POLLPRI +) + +// pollState tracks the associated file description and waiter of a PollFD. +type pollState struct { + file *vfs.FileDescription + waiter waiter.Entry +} + +// initReadiness gets the current ready mask for the file represented by the FD +// stored in pfd.FD. If a channel is passed in, the waiter entry in "state" is +// used to register with the file for event notifications, and a reference to +// the file is stored in "state". +func initReadiness(t *kernel.Task, pfd *linux.PollFD, state *pollState, ch chan struct{}) { + if pfd.FD < 0 { + pfd.REvents = 0 + return + } + + file := t.GetFileVFS2(pfd.FD) + if file == nil { + pfd.REvents = linux.POLLNVAL + return + } + + if ch == nil { + defer file.DecRef() + } else { + state.file = file + state.waiter, _ = waiter.NewChannelEntry(ch) + file.EventRegister(&state.waiter, waiter.EventMaskFromLinux(uint32(pfd.Events))) + } + + r := file.Readiness(waiter.EventMaskFromLinux(uint32(pfd.Events))) + pfd.REvents = int16(r.ToLinux()) & pfd.Events +} + +// releaseState releases all the pollState in "state". +func releaseState(state []pollState) { + for i := range state { + if state[i].file != nil { + state[i].file.EventUnregister(&state[i].waiter) + state[i].file.DecRef() + } + } +} + +// pollBlock polls the PollFDs in "pfd" with a bounded time specified in "timeout" +// when "timeout" is greater than zero. +// +// pollBlock returns the remaining timeout, which is always 0 on a timeout; and 0 or +// positive if interrupted by a signal. +func pollBlock(t *kernel.Task, pfd []linux.PollFD, timeout time.Duration) (time.Duration, uintptr, error) { + var ch chan struct{} + if timeout != 0 { + ch = make(chan struct{}, 1) + } + + // Register for event notification in the files involved if we may + // block (timeout not zero). Once we find a file that has a non-zero + // result, we stop registering for events but still go through all files + // to get their ready masks. + state := make([]pollState, len(pfd)) + defer releaseState(state) + n := uintptr(0) + for i := range pfd { + initReadiness(t, &pfd[i], &state[i], ch) + if pfd[i].REvents != 0 { + n++ + ch = nil + } + } + + if timeout == 0 { + return timeout, n, nil + } + + haveTimeout := timeout >= 0 + + for n == 0 { + var err error + // Wait for a notification. + timeout, err = t.BlockWithTimeout(ch, haveTimeout, timeout) + if err != nil { + if err == syserror.ETIMEDOUT { + err = nil + } + return timeout, 0, err + } + + // We got notified, count how many files are ready. If none, + // then this was a spurious notification, and we just go back + // to sleep with the remaining timeout. + for i := range state { + if state[i].file == nil { + continue + } + + r := state[i].file.Readiness(waiter.EventMaskFromLinux(uint32(pfd[i].Events))) + rl := int16(r.ToLinux()) & pfd[i].Events + if rl != 0 { + pfd[i].REvents = rl + n++ + } + } + } + + return timeout, n, nil +} + +// copyInPollFDs copies an array of struct pollfd unless nfds exceeds the max. +func copyInPollFDs(t *kernel.Task, addr usermem.Addr, nfds uint) ([]linux.PollFD, error) { + if uint64(nfds) > t.ThreadGroup().Limits().GetCapped(limits.NumberOfFiles, fileCap) { + return nil, syserror.EINVAL + } + + pfd := make([]linux.PollFD, nfds) + if nfds > 0 { + if _, err := t.CopyIn(addr, &pfd); err != nil { + return nil, err + } + } + + return pfd, nil +} + +func doPoll(t *kernel.Task, addr usermem.Addr, nfds uint, timeout time.Duration) (time.Duration, uintptr, error) { + pfd, err := copyInPollFDs(t, addr, nfds) + if err != nil { + return timeout, 0, err + } + + // Compatibility warning: Linux adds POLLHUP and POLLERR just before + // polling, in fs/select.c:do_pollfd(). Since pfd is copied out after + // polling, changing event masks here is an application-visible difference. + // (Linux also doesn't copy out event masks at all, only revents.) + for i := range pfd { + pfd[i].Events |= linux.POLLHUP | linux.POLLERR + } + remainingTimeout, n, err := pollBlock(t, pfd, timeout) + err = syserror.ConvertIntr(err, syserror.EINTR) + + // The poll entries are copied out regardless of whether + // any are set or not. This aligns with the Linux behavior. + if nfds > 0 && err == nil { + if _, err := t.CopyOut(addr, pfd); err != nil { + return remainingTimeout, 0, err + } + } + + return remainingTimeout, n, err +} + +// CopyInFDSet copies an fd set from select(2)/pselect(2). +func CopyInFDSet(t *kernel.Task, addr usermem.Addr, nBytes, nBitsInLastPartialByte int) ([]byte, error) { + set := make([]byte, nBytes) + + if addr != 0 { + if _, err := t.CopyIn(addr, &set); err != nil { + return nil, err + } + // If we only use part of the last byte, mask out the extraneous bits. + // + // N.B. This only works on little-endian architectures. + if nBitsInLastPartialByte != 0 { + set[nBytes-1] &^= byte(0xff) << nBitsInLastPartialByte + } + } + return set, nil +} + +func doSelect(t *kernel.Task, nfds int, readFDs, writeFDs, exceptFDs usermem.Addr, timeout time.Duration) (uintptr, error) { + if nfds < 0 || nfds > fileCap { + return 0, syserror.EINVAL + } + + // Calculate the size of the fd sets (one bit per fd). + nBytes := (nfds + 7) / 8 + nBitsInLastPartialByte := nfds % 8 + + // Capture all the provided input vectors. + r, err := CopyInFDSet(t, readFDs, nBytes, nBitsInLastPartialByte) + if err != nil { + return 0, err + } + w, err := CopyInFDSet(t, writeFDs, nBytes, nBitsInLastPartialByte) + if err != nil { + return 0, err + } + e, err := CopyInFDSet(t, exceptFDs, nBytes, nBitsInLastPartialByte) + if err != nil { + return 0, err + } + + // Count how many FDs are actually being requested so that we can build + // a PollFD array. + fdCount := 0 + for i := 0; i < nBytes; i++ { + v := r[i] | w[i] | e[i] + for v != 0 { + v &= (v - 1) + fdCount++ + } + } + + // Build the PollFD array. + pfd := make([]linux.PollFD, 0, fdCount) + var fd int32 + for i := 0; i < nBytes; i++ { + rV, wV, eV := r[i], w[i], e[i] + v := rV | wV | eV + m := byte(1) + for j := 0; j < 8; j++ { + if (v & m) != 0 { + // Make sure the fd is valid and decrement the reference + // immediately to ensure we don't leak. Note, another thread + // might be about to close fd. This is racy, but that's + // OK. Linux is racy in the same way. + file := t.GetFileVFS2(fd) + if file == nil { + return 0, syserror.EBADF + } + file.DecRef() + + var mask int16 + if (rV & m) != 0 { + mask |= selectReadEvents + } + + if (wV & m) != 0 { + mask |= selectWriteEvents + } + + if (eV & m) != 0 { + mask |= selectExceptEvents + } + + pfd = append(pfd, linux.PollFD{ + FD: fd, + Events: mask, + }) + } + + fd++ + m <<= 1 + } + } + + // Do the syscall, then count the number of bits set. + if _, _, err = pollBlock(t, pfd, timeout); err != nil { + return 0, syserror.ConvertIntr(err, syserror.EINTR) + } + + // r, w, and e are currently event mask bitsets; unset bits corresponding + // to events that *didn't* occur. + bitSetCount := uintptr(0) + for idx := range pfd { + events := pfd[idx].REvents + i, j := pfd[idx].FD/8, uint(pfd[idx].FD%8) + m := byte(1) << j + if r[i]&m != 0 { + if (events & selectReadEvents) != 0 { + bitSetCount++ + } else { + r[i] &^= m + } + } + if w[i]&m != 0 { + if (events & selectWriteEvents) != 0 { + bitSetCount++ + } else { + w[i] &^= m + } + } + if e[i]&m != 0 { + if (events & selectExceptEvents) != 0 { + bitSetCount++ + } else { + e[i] &^= m + } + } + } + + // Copy updated vectors back. + if readFDs != 0 { + if _, err := t.CopyOut(readFDs, r); err != nil { + return 0, err + } + } + + if writeFDs != 0 { + if _, err := t.CopyOut(writeFDs, w); err != nil { + return 0, err + } + } + + if exceptFDs != 0 { + if _, err := t.CopyOut(exceptFDs, e); err != nil { + return 0, err + } + } + + return bitSetCount, nil +} + +// timeoutRemaining returns the amount of time remaining for the specified +// timeout or 0 if it has elapsed. +// +// startNs must be from CLOCK_MONOTONIC. +func timeoutRemaining(t *kernel.Task, startNs ktime.Time, timeout time.Duration) time.Duration { + now := t.Kernel().MonotonicClock().Now() + remaining := timeout - now.Sub(startNs) + if remaining < 0 { + remaining = 0 + } + return remaining +} + +// copyOutTimespecRemaining copies the time remaining in timeout to timespecAddr. +// +// startNs must be from CLOCK_MONOTONIC. +func copyOutTimespecRemaining(t *kernel.Task, startNs ktime.Time, timeout time.Duration, timespecAddr usermem.Addr) error { + if timeout <= 0 { + return nil + } + remaining := timeoutRemaining(t, startNs, timeout) + tsRemaining := linux.NsecToTimespec(remaining.Nanoseconds()) + _, err := tsRemaining.CopyOut(t, timespecAddr) + return err +} + +// copyOutTimevalRemaining copies the time remaining in timeout to timevalAddr. +// +// startNs must be from CLOCK_MONOTONIC. +func copyOutTimevalRemaining(t *kernel.Task, startNs ktime.Time, timeout time.Duration, timevalAddr usermem.Addr) error { + if timeout <= 0 { + return nil + } + remaining := timeoutRemaining(t, startNs, timeout) + tvRemaining := linux.NsecToTimeval(remaining.Nanoseconds()) + _, err := tvRemaining.CopyOut(t, timevalAddr) + return err +} + +// pollRestartBlock encapsulates the state required to restart poll(2) via +// restart_syscall(2). +// +// +stateify savable +type pollRestartBlock struct { + pfdAddr usermem.Addr + nfds uint + timeout time.Duration +} + +// Restart implements kernel.SyscallRestartBlock.Restart. +func (p *pollRestartBlock) Restart(t *kernel.Task) (uintptr, error) { + return poll(t, p.pfdAddr, p.nfds, p.timeout) +} + +func poll(t *kernel.Task, pfdAddr usermem.Addr, nfds uint, timeout time.Duration) (uintptr, error) { + remainingTimeout, n, err := doPoll(t, pfdAddr, nfds, timeout) + // On an interrupt poll(2) is restarted with the remaining timeout. + if err == syserror.EINTR { + t.SetSyscallRestartBlock(&pollRestartBlock{ + pfdAddr: pfdAddr, + nfds: nfds, + timeout: remainingTimeout, + }) + return 0, kernel.ERESTART_RESTARTBLOCK + } + return n, err +} + +// Poll implements linux syscall poll(2). +func Poll(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + pfdAddr := args[0].Pointer() + nfds := uint(args[1].Uint()) // poll(2) uses unsigned long. + timeout := time.Duration(args[2].Int()) * time.Millisecond + n, err := poll(t, pfdAddr, nfds, timeout) + return n, nil, err +} + +// Ppoll implements linux syscall ppoll(2). +func Ppoll(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + pfdAddr := args[0].Pointer() + nfds := uint(args[1].Uint()) // poll(2) uses unsigned long. + timespecAddr := args[2].Pointer() + maskAddr := args[3].Pointer() + maskSize := uint(args[4].Uint()) + + timeout, err := copyTimespecInToDuration(t, timespecAddr) + if err != nil { + return 0, nil, err + } + + var startNs ktime.Time + if timeout > 0 { + startNs = t.Kernel().MonotonicClock().Now() + } + + if err := setTempSignalSet(t, maskAddr, maskSize); err != nil { + return 0, nil, err + } + + _, n, err := doPoll(t, pfdAddr, nfds, timeout) + copyErr := copyOutTimespecRemaining(t, startNs, timeout, timespecAddr) + // doPoll returns EINTR if interrupted, but ppoll is normally restartable + // if interrupted by something other than a signal handled by the + // application (i.e. returns ERESTARTNOHAND). However, if + // copyOutTimespecRemaining failed, then the restarted ppoll would use the + // wrong timeout, so the error should be left as EINTR. + // + // Note that this means that if err is nil but copyErr is not, copyErr is + // ignored. This is consistent with Linux. + if err == syserror.EINTR && copyErr == nil { + err = kernel.ERESTARTNOHAND + } + return n, nil, err +} + +// Select implements linux syscall select(2). +func Select(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + nfds := int(args[0].Int()) // select(2) uses an int. + readFDs := args[1].Pointer() + writeFDs := args[2].Pointer() + exceptFDs := args[3].Pointer() + timevalAddr := args[4].Pointer() + + // Use a negative Duration to indicate "no timeout". + timeout := time.Duration(-1) + if timevalAddr != 0 { + var timeval linux.Timeval + if _, err := timeval.CopyIn(t, timevalAddr); err != nil { + return 0, nil, err + } + if timeval.Sec < 0 || timeval.Usec < 0 { + return 0, nil, syserror.EINVAL + } + timeout = time.Duration(timeval.ToNsecCapped()) + } + startNs := t.Kernel().MonotonicClock().Now() + n, err := doSelect(t, nfds, readFDs, writeFDs, exceptFDs, timeout) + copyErr := copyOutTimevalRemaining(t, startNs, timeout, timevalAddr) + // See comment in Ppoll. + if err == syserror.EINTR && copyErr == nil { + err = kernel.ERESTARTNOHAND + } + return n, nil, err +} + +// Pselect implements linux syscall pselect(2). +func Pselect(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + nfds := int(args[0].Int()) // select(2) uses an int. + readFDs := args[1].Pointer() + writeFDs := args[2].Pointer() + exceptFDs := args[3].Pointer() + timespecAddr := args[4].Pointer() + maskWithSizeAddr := args[5].Pointer() + + timeout, err := copyTimespecInToDuration(t, timespecAddr) + if err != nil { + return 0, nil, err + } + + var startNs ktime.Time + if timeout > 0 { + startNs = t.Kernel().MonotonicClock().Now() + } + + if maskWithSizeAddr != 0 { + if t.Arch().Width() != 8 { + panic(fmt.Sprintf("unsupported sizeof(void*): %d", t.Arch().Width())) + } + var maskStruct sigSetWithSize + if _, err := maskStruct.CopyIn(t, maskWithSizeAddr); err != nil { + return 0, nil, err + } + if err := setTempSignalSet(t, usermem.Addr(maskStruct.sigsetAddr), uint(maskStruct.sizeofSigset)); err != nil { + return 0, nil, err + } + } + + n, err := doSelect(t, nfds, readFDs, writeFDs, exceptFDs, timeout) + copyErr := copyOutTimespecRemaining(t, startNs, timeout, timespecAddr) + // See comment in Ppoll. + if err == syserror.EINTR && copyErr == nil { + err = kernel.ERESTARTNOHAND + } + return n, nil, err +} + +// +marshal +type sigSetWithSize struct { + sigsetAddr uint64 + sizeofSigset uint64 +} + +// copyTimespecInToDuration copies a Timespec from the untrusted app range, +// validates it and converts it to a Duration. +// +// If the Timespec is larger than what can be represented in a Duration, the +// returned value is the maximum that Duration will allow. +// +// If timespecAddr is NULL, the returned value is negative. +func copyTimespecInToDuration(t *kernel.Task, timespecAddr usermem.Addr) (time.Duration, error) { + // Use a negative Duration to indicate "no timeout". + timeout := time.Duration(-1) + if timespecAddr != 0 { + var timespec linux.Timespec + if _, err := timespec.CopyIn(t, timespecAddr); err != nil { + return 0, err + } + if !timespec.Valid() { + return 0, syserror.EINVAL + } + timeout = time.Duration(timespec.ToNsecCapped()) + } + return timeout, nil +} + +func setTempSignalSet(t *kernel.Task, maskAddr usermem.Addr, maskSize uint) error { + if maskAddr == 0 { + return nil + } + if maskSize != linux.SignalSetSize { + return syserror.EINVAL + } + var mask linux.SignalSet + if _, err := mask.CopyIn(t, maskAddr); err != nil { + return err + } + mask &^= kernel.UnblockableSignals + oldmask := t.SignalMask() + t.SetSignalMask(mask) + t.SetSavedSignalMask(oldmask) + return nil +} diff --git a/pkg/sentry/syscalls/linux/vfs2/read_write.go b/pkg/sentry/syscalls/linux/vfs2/read_write.go new file mode 100644 index 000000000..6c6998f45 --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/read_write.go @@ -0,0 +1,511 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + slinux "gvisor.dev/gvisor/pkg/sentry/syscalls/linux" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" + "gvisor.dev/gvisor/pkg/waiter" +) + +const ( + eventMaskRead = waiter.EventIn | waiter.EventHUp | waiter.EventErr + eventMaskWrite = waiter.EventOut | waiter.EventHUp | waiter.EventErr +) + +// Read implements Linux syscall read(2). +func Read(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + addr := args[1].Pointer() + size := args[2].SizeT() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Check that the size is legitimate. + si := int(size) + if si < 0 { + return 0, nil, syserror.EINVAL + } + + // Get the destination of the read. + dst, err := t.SingleIOSequence(addr, si, usermem.IOOpts{ + AddressSpaceActive: true, + }) + if err != nil { + return 0, nil, err + } + + n, err := read(t, file, dst, vfs.ReadOptions{}) + t.IOUsage().AccountReadSyscall(n) + return uintptr(n), nil, slinux.HandleIOErrorVFS2(t, n != 0, err, kernel.ERESTARTSYS, "read", file) +} + +// Readv implements Linux syscall readv(2). +func Readv(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + addr := args[1].Pointer() + iovcnt := int(args[2].Int()) + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Get the destination of the read. + dst, err := t.IovecsIOSequence(addr, iovcnt, usermem.IOOpts{ + AddressSpaceActive: true, + }) + if err != nil { + return 0, nil, err + } + + n, err := read(t, file, dst, vfs.ReadOptions{}) + t.IOUsage().AccountReadSyscall(n) + return uintptr(n), nil, slinux.HandleIOErrorVFS2(t, n != 0, err, kernel.ERESTARTSYS, "readv", file) +} + +func read(t *kernel.Task, file *vfs.FileDescription, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { + n, err := file.Read(t, dst, opts) + if err != syserror.ErrWouldBlock || file.StatusFlags()&linux.O_NONBLOCK != 0 { + return n, err + } + + // Register for notifications. + w, ch := waiter.NewChannelEntry(nil) + file.EventRegister(&w, eventMaskRead) + + total := n + for { + // Shorten dst to reflect bytes previously read. + dst = dst.DropFirst(int(n)) + + // Issue the request and break out if it completes with anything other than + // "would block". + n, err = file.Read(t, dst, opts) + total += n + if err != syserror.ErrWouldBlock { + break + } + if err := t.Block(ch); err != nil { + break + } + } + file.EventUnregister(&w) + + return total, err +} + +// Pread64 implements Linux syscall pread64(2). +func Pread64(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + addr := args[1].Pointer() + size := args[2].SizeT() + offset := args[3].Int64() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Check that the offset is legitimate and does not overflow. + if offset < 0 || offset+int64(size) < 0 { + return 0, nil, syserror.EINVAL + } + + // Check that the size is legitimate. + si := int(size) + if si < 0 { + return 0, nil, syserror.EINVAL + } + + // Get the destination of the read. + dst, err := t.SingleIOSequence(addr, si, usermem.IOOpts{ + AddressSpaceActive: true, + }) + if err != nil { + return 0, nil, err + } + + n, err := pread(t, file, dst, offset, vfs.ReadOptions{}) + t.IOUsage().AccountReadSyscall(n) + return uintptr(n), nil, slinux.HandleIOErrorVFS2(t, n != 0, err, kernel.ERESTARTSYS, "pread64", file) +} + +// Preadv implements Linux syscall preadv(2). +func Preadv(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + addr := args[1].Pointer() + iovcnt := int(args[2].Int()) + offset := args[3].Int64() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Check that the offset is legitimate. + if offset < 0 { + return 0, nil, syserror.EINVAL + } + + // Get the destination of the read. + dst, err := t.IovecsIOSequence(addr, iovcnt, usermem.IOOpts{ + AddressSpaceActive: true, + }) + if err != nil { + return 0, nil, err + } + + n, err := pread(t, file, dst, offset, vfs.ReadOptions{}) + t.IOUsage().AccountReadSyscall(n) + return uintptr(n), nil, slinux.HandleIOErrorVFS2(t, n != 0, err, kernel.ERESTARTSYS, "preadv", file) +} + +// Preadv2 implements Linux syscall preadv2(2). +func Preadv2(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + // While the glibc signature is + // preadv2(int fd, struct iovec* iov, int iov_cnt, off_t offset, int flags) + // the actual syscall + // (https://elixir.bootlin.com/linux/v5.5/source/fs/read_write.c#L1142) + // splits the offset argument into a high/low value for compatibility with + // 32-bit architectures. The flags argument is the 6th argument (index 5). + fd := args[0].Int() + addr := args[1].Pointer() + iovcnt := int(args[2].Int()) + offset := args[3].Int64() + flags := args[5].Int() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Check that the offset is legitimate. + if offset < -1 { + return 0, nil, syserror.EINVAL + } + + // Get the destination of the read. + dst, err := t.IovecsIOSequence(addr, iovcnt, usermem.IOOpts{ + AddressSpaceActive: true, + }) + if err != nil { + return 0, nil, err + } + + opts := vfs.ReadOptions{ + Flags: uint32(flags), + } + var n int64 + if offset == -1 { + n, err = read(t, file, dst, opts) + } else { + n, err = pread(t, file, dst, offset, opts) + } + t.IOUsage().AccountReadSyscall(n) + return uintptr(n), nil, slinux.HandleIOErrorVFS2(t, n != 0, err, kernel.ERESTARTSYS, "preadv2", file) +} + +func pread(t *kernel.Task, file *vfs.FileDescription, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { + n, err := file.PRead(t, dst, offset, opts) + if err != syserror.ErrWouldBlock || file.StatusFlags()&linux.O_NONBLOCK != 0 { + return n, err + } + + // Register for notifications. + w, ch := waiter.NewChannelEntry(nil) + file.EventRegister(&w, eventMaskRead) + + total := n + for { + // Shorten dst to reflect bytes previously read. + dst = dst.DropFirst(int(n)) + + // Issue the request and break out if it completes with anything other than + // "would block". + n, err = file.PRead(t, dst, offset+total, opts) + total += n + if err != syserror.ErrWouldBlock { + break + } + if err := t.Block(ch); err != nil { + break + } + } + file.EventUnregister(&w) + + return total, err +} + +// Write implements Linux syscall write(2). +func Write(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + addr := args[1].Pointer() + size := args[2].SizeT() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Check that the size is legitimate. + si := int(size) + if si < 0 { + return 0, nil, syserror.EINVAL + } + + // Get the source of the write. + src, err := t.SingleIOSequence(addr, si, usermem.IOOpts{ + AddressSpaceActive: true, + }) + if err != nil { + return 0, nil, err + } + + n, err := write(t, file, src, vfs.WriteOptions{}) + t.IOUsage().AccountWriteSyscall(n) + return uintptr(n), nil, slinux.HandleIOErrorVFS2(t, n != 0, err, kernel.ERESTARTSYS, "write", file) +} + +// Writev implements Linux syscall writev(2). +func Writev(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + addr := args[1].Pointer() + iovcnt := int(args[2].Int()) + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Get the source of the write. + src, err := t.IovecsIOSequence(addr, iovcnt, usermem.IOOpts{ + AddressSpaceActive: true, + }) + if err != nil { + return 0, nil, err + } + + n, err := write(t, file, src, vfs.WriteOptions{}) + t.IOUsage().AccountWriteSyscall(n) + return uintptr(n), nil, slinux.HandleIOErrorVFS2(t, n != 0, err, kernel.ERESTARTSYS, "writev", file) +} + +func write(t *kernel.Task, file *vfs.FileDescription, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) { + n, err := file.Write(t, src, opts) + if err != syserror.ErrWouldBlock || file.StatusFlags()&linux.O_NONBLOCK != 0 { + return n, err + } + + // Register for notifications. + w, ch := waiter.NewChannelEntry(nil) + file.EventRegister(&w, eventMaskWrite) + + total := n + for { + // Shorten src to reflect bytes previously written. + src = src.DropFirst(int(n)) + + // Issue the request and break out if it completes with anything other than + // "would block". + n, err = file.Write(t, src, opts) + total += n + if err != syserror.ErrWouldBlock { + break + } + if err := t.Block(ch); err != nil { + break + } + } + file.EventUnregister(&w) + + return total, err +} + +// Pwrite64 implements Linux syscall pwrite64(2). +func Pwrite64(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + addr := args[1].Pointer() + size := args[2].SizeT() + offset := args[3].Int64() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Check that the offset is legitimate and does not overflow. + if offset < 0 || offset+int64(size) < 0 { + return 0, nil, syserror.EINVAL + } + + // Check that the size is legitimate. + si := int(size) + if si < 0 { + return 0, nil, syserror.EINVAL + } + + // Get the source of the write. + src, err := t.SingleIOSequence(addr, si, usermem.IOOpts{ + AddressSpaceActive: true, + }) + if err != nil { + return 0, nil, err + } + + n, err := pwrite(t, file, src, offset, vfs.WriteOptions{}) + t.IOUsage().AccountWriteSyscall(n) + return uintptr(n), nil, slinux.HandleIOErrorVFS2(t, n != 0, err, kernel.ERESTARTSYS, "pwrite64", file) +} + +// Pwritev implements Linux syscall pwritev(2). +func Pwritev(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + addr := args[1].Pointer() + iovcnt := int(args[2].Int()) + offset := args[3].Int64() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Check that the offset is legitimate. + if offset < 0 { + return 0, nil, syserror.EINVAL + } + + // Get the source of the write. + src, err := t.IovecsIOSequence(addr, iovcnt, usermem.IOOpts{ + AddressSpaceActive: true, + }) + if err != nil { + return 0, nil, err + } + + n, err := pwrite(t, file, src, offset, vfs.WriteOptions{}) + t.IOUsage().AccountReadSyscall(n) + return uintptr(n), nil, slinux.HandleIOErrorVFS2(t, n != 0, err, kernel.ERESTARTSYS, "pwritev", file) +} + +// Pwritev2 implements Linux syscall pwritev2(2). +func Pwritev2(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + // While the glibc signature is + // pwritev2(int fd, struct iovec* iov, int iov_cnt, off_t offset, int flags) + // the actual syscall + // (https://elixir.bootlin.com/linux/v5.5/source/fs/read_write.c#L1162) + // splits the offset argument into a high/low value for compatibility with + // 32-bit architectures. The flags argument is the 6th argument (index 5). + fd := args[0].Int() + addr := args[1].Pointer() + iovcnt := int(args[2].Int()) + offset := args[3].Int64() + flags := args[5].Int() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Check that the offset is legitimate. + if offset < -1 { + return 0, nil, syserror.EINVAL + } + + // Get the source of the write. + src, err := t.IovecsIOSequence(addr, iovcnt, usermem.IOOpts{ + AddressSpaceActive: true, + }) + if err != nil { + return 0, nil, err + } + + opts := vfs.WriteOptions{ + Flags: uint32(flags), + } + var n int64 + if offset == -1 { + n, err = write(t, file, src, opts) + } else { + n, err = pwrite(t, file, src, offset, opts) + } + t.IOUsage().AccountWriteSyscall(n) + return uintptr(n), nil, slinux.HandleIOErrorVFS2(t, n != 0, err, kernel.ERESTARTSYS, "pwritev2", file) +} + +func pwrite(t *kernel.Task, file *vfs.FileDescription, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) { + n, err := file.PWrite(t, src, offset, opts) + if err != syserror.ErrWouldBlock || file.StatusFlags()&linux.O_NONBLOCK != 0 { + return n, err + } + + // Register for notifications. + w, ch := waiter.NewChannelEntry(nil) + file.EventRegister(&w, eventMaskWrite) + + total := n + for { + // Shorten src to reflect bytes previously written. + src = src.DropFirst(int(n)) + + // Issue the request and break out if it completes with anything other than + // "would block". + n, err = file.PWrite(t, src, offset+total, opts) + total += n + if err != syserror.ErrWouldBlock { + break + } + if err := t.Block(ch); err != nil { + break + } + } + file.EventUnregister(&w) + + return total, err +} + +// Lseek implements Linux syscall lseek(2). +func Lseek(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + offset := args[1].Int64() + whence := args[2].Int() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + newoff, err := file.Seek(t, offset, whence) + return uintptr(newoff), nil, err +} diff --git a/pkg/sentry/syscalls/linux/vfs2/setstat.go b/pkg/sentry/syscalls/linux/vfs2/setstat.go new file mode 100644 index 000000000..4e61f1452 --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/setstat.go @@ -0,0 +1,391 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +const chmodMask = 0777 | linux.S_ISUID | linux.S_ISGID | linux.S_ISVTX + +// Chmod implements Linux syscall chmod(2). +func Chmod(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + pathAddr := args[0].Pointer() + mode := args[1].ModeT() + return 0, nil, fchmodat(t, linux.AT_FDCWD, pathAddr, mode) +} + +// Fchmodat implements Linux syscall fchmodat(2). +func Fchmodat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + dirfd := args[0].Int() + pathAddr := args[1].Pointer() + mode := args[2].ModeT() + return 0, nil, fchmodat(t, dirfd, pathAddr, mode) +} + +func fchmodat(t *kernel.Task, dirfd int32, pathAddr usermem.Addr, mode uint) error { + path, err := copyInPath(t, pathAddr) + if err != nil { + return err + } + + return setstatat(t, dirfd, path, disallowEmptyPath, followFinalSymlink, &vfs.SetStatOptions{ + Stat: linux.Statx{ + Mask: linux.STATX_MODE, + Mode: uint16(mode & chmodMask), + }, + }) +} + +// Fchmod implements Linux syscall fchmod(2). +func Fchmod(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + mode := args[1].ModeT() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + return 0, nil, file.SetStat(t, vfs.SetStatOptions{ + Stat: linux.Statx{ + Mask: linux.STATX_MODE, + Mode: uint16(mode & chmodMask), + }, + }) +} + +// Chown implements Linux syscall chown(2). +func Chown(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + pathAddr := args[0].Pointer() + owner := args[1].Int() + group := args[2].Int() + return 0, nil, fchownat(t, linux.AT_FDCWD, pathAddr, owner, group, 0 /* flags */) +} + +// Lchown implements Linux syscall lchown(2). +func Lchown(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + pathAddr := args[0].Pointer() + owner := args[1].Int() + group := args[2].Int() + return 0, nil, fchownat(t, linux.AT_FDCWD, pathAddr, owner, group, linux.AT_SYMLINK_NOFOLLOW) +} + +// Fchownat implements Linux syscall fchownat(2). +func Fchownat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + dirfd := args[0].Int() + pathAddr := args[1].Pointer() + owner := args[2].Int() + group := args[3].Int() + flags := args[4].Int() + return 0, nil, fchownat(t, dirfd, pathAddr, owner, group, flags) +} + +func fchownat(t *kernel.Task, dirfd int32, pathAddr usermem.Addr, owner, group, flags int32) error { + if flags&^(linux.AT_EMPTY_PATH|linux.AT_SYMLINK_NOFOLLOW) != 0 { + return syserror.EINVAL + } + + path, err := copyInPath(t, pathAddr) + if err != nil { + return err + } + + var opts vfs.SetStatOptions + if err := populateSetStatOptionsForChown(t, owner, group, &opts); err != nil { + return err + } + + return setstatat(t, dirfd, path, shouldAllowEmptyPath(flags&linux.AT_EMPTY_PATH != 0), shouldFollowFinalSymlink(flags&linux.AT_SYMLINK_NOFOLLOW == 0), &opts) +} + +func populateSetStatOptionsForChown(t *kernel.Task, owner, group int32, opts *vfs.SetStatOptions) error { + userns := t.UserNamespace() + if owner != -1 { + kuid := userns.MapToKUID(auth.UID(owner)) + if !kuid.Ok() { + return syserror.EINVAL + } + opts.Stat.Mask |= linux.STATX_UID + opts.Stat.UID = uint32(kuid) + } + if group != -1 { + kgid := userns.MapToKGID(auth.GID(group)) + if !kgid.Ok() { + return syserror.EINVAL + } + opts.Stat.Mask |= linux.STATX_GID + opts.Stat.GID = uint32(kgid) + } + return nil +} + +// Fchown implements Linux syscall fchown(2). +func Fchown(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + owner := args[1].Int() + group := args[2].Int() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + var opts vfs.SetStatOptions + if err := populateSetStatOptionsForChown(t, owner, group, &opts); err != nil { + return 0, nil, err + } + return 0, nil, file.SetStat(t, opts) +} + +// Truncate implements Linux syscall truncate(2). +func Truncate(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + addr := args[0].Pointer() + length := args[1].Int64() + + if length < 0 { + return 0, nil, syserror.EINVAL + } + + path, err := copyInPath(t, addr) + if err != nil { + return 0, nil, err + } + + err = setstatat(t, linux.AT_FDCWD, path, disallowEmptyPath, followFinalSymlink, &vfs.SetStatOptions{ + Stat: linux.Statx{ + Mask: linux.STATX_SIZE, + Size: uint64(length), + }, + }) + return 0, nil, handleSetSizeError(t, err) +} + +// Ftruncate implements Linux syscall ftruncate(2). +func Ftruncate(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + length := args[1].Int64() + + if length < 0 { + return 0, nil, syserror.EINVAL + } + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + err := file.SetStat(t, vfs.SetStatOptions{ + Stat: linux.Statx{ + Mask: linux.STATX_SIZE, + Size: uint64(length), + }, + }) + return 0, nil, handleSetSizeError(t, err) +} + +// Utime implements Linux syscall utime(2). +func Utime(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + pathAddr := args[0].Pointer() + timesAddr := args[1].Pointer() + + path, err := copyInPath(t, pathAddr) + if err != nil { + return 0, nil, err + } + + opts := vfs.SetStatOptions{ + Stat: linux.Statx{ + Mask: linux.STATX_ATIME | linux.STATX_MTIME, + }, + } + if timesAddr == 0 { + opts.Stat.Atime.Nsec = linux.UTIME_NOW + opts.Stat.Mtime.Nsec = linux.UTIME_NOW + } else { + var times linux.Utime + if _, err := times.CopyIn(t, timesAddr); err != nil { + return 0, nil, err + } + opts.Stat.Atime.Sec = times.Actime + opts.Stat.Mtime.Sec = times.Modtime + } + + return 0, nil, setstatat(t, linux.AT_FDCWD, path, disallowEmptyPath, followFinalSymlink, &opts) +} + +// Utimes implements Linux syscall utimes(2). +func Utimes(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + pathAddr := args[0].Pointer() + timesAddr := args[1].Pointer() + + path, err := copyInPath(t, pathAddr) + if err != nil { + return 0, nil, err + } + + opts := vfs.SetStatOptions{ + Stat: linux.Statx{ + Mask: linux.STATX_ATIME | linux.STATX_MTIME, + }, + } + if timesAddr == 0 { + opts.Stat.Atime.Nsec = linux.UTIME_NOW + opts.Stat.Mtime.Nsec = linux.UTIME_NOW + } else { + var times [2]linux.Timeval + if _, err := t.CopyIn(timesAddr, ×); err != nil { + return 0, nil, err + } + opts.Stat.Atime = linux.StatxTimestamp{ + Sec: times[0].Sec, + Nsec: uint32(times[0].Usec * 1000), + } + opts.Stat.Mtime = linux.StatxTimestamp{ + Sec: times[1].Sec, + Nsec: uint32(times[1].Usec * 1000), + } + } + + return 0, nil, setstatat(t, linux.AT_FDCWD, path, disallowEmptyPath, followFinalSymlink, &opts) +} + +// Utimensat implements Linux syscall utimensat(2). +func Utimensat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + dirfd := args[0].Int() + pathAddr := args[1].Pointer() + timesAddr := args[2].Pointer() + flags := args[3].Int() + + if flags&^linux.AT_SYMLINK_NOFOLLOW != 0 { + return 0, nil, syserror.EINVAL + } + + path, err := copyInPath(t, pathAddr) + if err != nil { + return 0, nil, err + } + + var opts vfs.SetStatOptions + if err := populateSetStatOptionsForUtimens(t, timesAddr, &opts); err != nil { + return 0, nil, err + } + + return 0, nil, setstatat(t, dirfd, path, disallowEmptyPath, followFinalSymlink, &opts) +} + +// Futimens implements Linux syscall futimens(2). +func Futimens(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + timesAddr := args[1].Pointer() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + var opts vfs.SetStatOptions + if err := populateSetStatOptionsForUtimens(t, timesAddr, &opts); err != nil { + return 0, nil, err + } + + return 0, nil, file.SetStat(t, opts) +} + +func populateSetStatOptionsForUtimens(t *kernel.Task, timesAddr usermem.Addr, opts *vfs.SetStatOptions) error { + if timesAddr == 0 { + opts.Stat.Mask = linux.STATX_ATIME | linux.STATX_MTIME + opts.Stat.Atime.Nsec = linux.UTIME_NOW + opts.Stat.Mtime.Nsec = linux.UTIME_NOW + return nil + } + var times [2]linux.Timespec + if _, err := t.CopyIn(timesAddr, ×); err != nil { + return err + } + if times[0].Nsec != linux.UTIME_OMIT { + opts.Stat.Mask |= linux.STATX_ATIME + opts.Stat.Atime = linux.StatxTimestamp{ + Sec: times[0].Sec, + Nsec: uint32(times[0].Nsec), + } + } + if times[1].Nsec != linux.UTIME_OMIT { + opts.Stat.Mask |= linux.STATX_MTIME + opts.Stat.Mtime = linux.StatxTimestamp{ + Sec: times[1].Sec, + Nsec: uint32(times[1].Nsec), + } + } + return nil +} + +func setstatat(t *kernel.Task, dirfd int32, path fspath.Path, shouldAllowEmptyPath shouldAllowEmptyPath, shouldFollowFinalSymlink shouldFollowFinalSymlink, opts *vfs.SetStatOptions) error { + root := t.FSContext().RootDirectoryVFS2() + defer root.DecRef() + start := root + if !path.Absolute { + if !path.HasComponents() && !bool(shouldAllowEmptyPath) { + return syserror.ENOENT + } + if dirfd == linux.AT_FDCWD { + start = t.FSContext().WorkingDirectoryVFS2() + defer start.DecRef() + } else { + dirfile := t.GetFileVFS2(dirfd) + if dirfile == nil { + return syserror.EBADF + } + if !path.HasComponents() { + // Use FileDescription.SetStat() instead of + // VirtualFilesystem.SetStatAt(), since the former may be able + // to use opened file state to expedite the SetStat. + err := dirfile.SetStat(t, *opts) + dirfile.DecRef() + return err + } + start = dirfile.VirtualDentry() + start.IncRef() + defer start.DecRef() + dirfile.DecRef() + } + } + return t.Kernel().VFS().SetStatAt(t, t.Credentials(), &vfs.PathOperation{ + Root: root, + Start: start, + Path: path, + FollowFinalSymlink: bool(shouldFollowFinalSymlink), + }, opts) +} + +func handleSetSizeError(t *kernel.Task, err error) error { + if err == syserror.ErrExceedsFileSizeLimit { + // Convert error to EFBIG and send a SIGXFSZ per setrlimit(2). + t.SendSignal(kernel.SignalInfoNoInfo(linux.SIGXFSZ, t, t)) + return syserror.EFBIG + } + return err +} diff --git a/pkg/sentry/syscalls/linux/vfs2/socket.go b/pkg/sentry/syscalls/linux/vfs2/socket.go new file mode 100644 index 000000000..b1ede32f0 --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/socket.go @@ -0,0 +1,1139 @@ +// Copyright 2018 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "time" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/binary" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" + "gvisor.dev/gvisor/pkg/sentry/socket" + "gvisor.dev/gvisor/pkg/sentry/socket/control" + "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" + slinux "gvisor.dev/gvisor/pkg/sentry/syscalls/linux" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserr" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +// minListenBacklog is the minimum reasonable backlog for listening sockets. +const minListenBacklog = 8 + +// maxListenBacklog is the maximum allowed backlog for listening sockets. +const maxListenBacklog = 1024 + +// maxAddrLen is the maximum socket address length we're willing to accept. +const maxAddrLen = 200 + +// maxOptLen is the maximum sockopt parameter length we're willing to accept. +const maxOptLen = 1024 * 8 + +// maxControlLen is the maximum length of the msghdr.msg_control buffer we're +// willing to accept. Note that this limit is smaller than Linux, which allows +// buffers upto INT_MAX. +const maxControlLen = 10 * 1024 * 1024 + +// nameLenOffset is the offset from the start of the MessageHeader64 struct to +// the NameLen field. +const nameLenOffset = 8 + +// controlLenOffset is the offset form the start of the MessageHeader64 struct +// to the ControlLen field. +const controlLenOffset = 40 + +// flagsOffset is the offset form the start of the MessageHeader64 struct +// to the Flags field. +const flagsOffset = 48 + +const sizeOfInt32 = 4 + +// messageHeader64Len is the length of a MessageHeader64 struct. +var messageHeader64Len = uint64(binary.Size(MessageHeader64{})) + +// multipleMessageHeader64Len is the length of a multipeMessageHeader64 struct. +var multipleMessageHeader64Len = uint64(binary.Size(multipleMessageHeader64{})) + +// baseRecvFlags are the flags that are accepted across recvmsg(2), +// recvmmsg(2), and recvfrom(2). +const baseRecvFlags = linux.MSG_OOB | linux.MSG_DONTROUTE | linux.MSG_DONTWAIT | linux.MSG_NOSIGNAL | linux.MSG_WAITALL | linux.MSG_TRUNC | linux.MSG_CTRUNC + +// MessageHeader64 is the 64-bit representation of the msghdr struct used in +// the recvmsg and sendmsg syscalls. +type MessageHeader64 struct { + // Name is the optional pointer to a network address buffer. + Name uint64 + + // NameLen is the length of the buffer pointed to by Name. + NameLen uint32 + _ uint32 + + // Iov is a pointer to an array of io vectors that describe the memory + // locations involved in the io operation. + Iov uint64 + + // IovLen is the length of the array pointed to by Iov. + IovLen uint64 + + // Control is the optional pointer to ancillary control data. + Control uint64 + + // ControlLen is the length of the data pointed to by Control. + ControlLen uint64 + + // Flags on the sent/received message. + Flags int32 + _ int32 +} + +// multipleMessageHeader64 is the 64-bit representation of the mmsghdr struct used in +// the recvmmsg and sendmmsg syscalls. +type multipleMessageHeader64 struct { + msgHdr MessageHeader64 + msgLen uint32 + _ int32 +} + +// CopyInMessageHeader64 copies a message header from user to kernel memory. +func CopyInMessageHeader64(t *kernel.Task, addr usermem.Addr, msg *MessageHeader64) error { + b := t.CopyScratchBuffer(52) + if _, err := t.CopyInBytes(addr, b); err != nil { + return err + } + + msg.Name = usermem.ByteOrder.Uint64(b[0:]) + msg.NameLen = usermem.ByteOrder.Uint32(b[8:]) + msg.Iov = usermem.ByteOrder.Uint64(b[16:]) + msg.IovLen = usermem.ByteOrder.Uint64(b[24:]) + msg.Control = usermem.ByteOrder.Uint64(b[32:]) + msg.ControlLen = usermem.ByteOrder.Uint64(b[40:]) + msg.Flags = int32(usermem.ByteOrder.Uint32(b[48:])) + + return nil +} + +// CaptureAddress allocates memory for and copies a socket address structure +// from the untrusted address space range. +func CaptureAddress(t *kernel.Task, addr usermem.Addr, addrlen uint32) ([]byte, error) { + if addrlen > maxAddrLen { + return nil, syserror.EINVAL + } + + addrBuf := make([]byte, addrlen) + if _, err := t.CopyInBytes(addr, addrBuf); err != nil { + return nil, err + } + + return addrBuf, nil +} + +// writeAddress writes a sockaddr structure and its length to an output buffer +// in the unstrusted address space range. If the address is bigger than the +// buffer, it is truncated. +func writeAddress(t *kernel.Task, addr interface{}, addrLen uint32, addrPtr usermem.Addr, addrLenPtr usermem.Addr) error { + // Get the buffer length. + var bufLen uint32 + if _, err := t.CopyIn(addrLenPtr, &bufLen); err != nil { + return err + } + + if int32(bufLen) < 0 { + return syserror.EINVAL + } + + // Write the length unconditionally. + if _, err := t.CopyOut(addrLenPtr, addrLen); err != nil { + return err + } + + if addr == nil { + return nil + } + + if bufLen > addrLen { + bufLen = addrLen + } + + // Copy as much of the address as will fit in the buffer. + encodedAddr := binary.Marshal(nil, usermem.ByteOrder, addr) + if bufLen > uint32(len(encodedAddr)) { + bufLen = uint32(len(encodedAddr)) + } + _, err := t.CopyOutBytes(addrPtr, encodedAddr[:int(bufLen)]) + return err +} + +// Socket implements the linux syscall socket(2). +func Socket(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + domain := int(args[0].Int()) + stype := args[1].Int() + protocol := int(args[2].Int()) + + // Check and initialize the flags. + if stype & ^(0xf|linux.SOCK_NONBLOCK|linux.SOCK_CLOEXEC) != 0 { + return 0, nil, syserror.EINVAL + } + + // Create the new socket. + s, e := socket.NewVFS2(t, domain, linux.SockType(stype&0xf), protocol) + if e != nil { + return 0, nil, e.ToError() + } + defer s.DecRef() + + if err := s.SetStatusFlags(t, t.Credentials(), uint32(stype&linux.SOCK_NONBLOCK)); err != nil { + return 0, nil, err + } + + fd, err := t.NewFDFromVFS2(0, s, kernel.FDFlags{ + CloseOnExec: stype&linux.SOCK_CLOEXEC != 0, + }) + if err != nil { + return 0, nil, err + } + + return uintptr(fd), nil, nil +} + +// SocketPair implements the linux syscall socketpair(2). +func SocketPair(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + domain := int(args[0].Int()) + stype := args[1].Int() + protocol := int(args[2].Int()) + addr := args[3].Pointer() + + // Check and initialize the flags. + if stype & ^(0xf|linux.SOCK_NONBLOCK|linux.SOCK_CLOEXEC) != 0 { + return 0, nil, syserror.EINVAL + } + + // Create the socket pair. + s1, s2, e := socket.PairVFS2(t, domain, linux.SockType(stype&0xf), protocol) + if e != nil { + return 0, nil, e.ToError() + } + // Adding to the FD table will cause an extra reference to be acquired. + defer s1.DecRef() + defer s2.DecRef() + + nonblocking := uint32(stype & linux.SOCK_NONBLOCK) + if err := s1.SetStatusFlags(t, t.Credentials(), nonblocking); err != nil { + return 0, nil, err + } + if err := s2.SetStatusFlags(t, t.Credentials(), nonblocking); err != nil { + return 0, nil, err + } + + // Create the FDs for the sockets. + flags := kernel.FDFlags{ + CloseOnExec: stype&linux.SOCK_CLOEXEC != 0, + } + fds, err := t.NewFDsVFS2(0, []*vfs.FileDescription{s1, s2}, flags) + if err != nil { + return 0, nil, err + } + + if _, err := t.CopyOut(addr, fds); err != nil { + for _, fd := range fds { + if _, file := t.FDTable().Remove(fd); file != nil { + file.DecRef() + } + } + return 0, nil, err + } + + return 0, nil, nil +} + +// Connect implements the linux syscall connect(2). +func Connect(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + addr := args[1].Pointer() + addrlen := args[2].Uint() + + // Get socket from the file descriptor. + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Extract the socket. + s, ok := file.Impl().(socket.SocketVFS2) + if !ok { + return 0, nil, syserror.ENOTSOCK + } + + // Capture address and call syscall implementation. + a, err := CaptureAddress(t, addr, addrlen) + if err != nil { + return 0, nil, err + } + + blocking := (file.StatusFlags() & linux.SOCK_NONBLOCK) == 0 + return 0, nil, syserror.ConvertIntr(s.Connect(t, a, blocking).ToError(), kernel.ERESTARTSYS) +} + +// accept is the implementation of the accept syscall. It is called by accept +// and accept4 syscall handlers. +func accept(t *kernel.Task, fd int32, addr usermem.Addr, addrLen usermem.Addr, flags int) (uintptr, error) { + // Check that no unsupported flags are passed in. + if flags & ^(linux.SOCK_NONBLOCK|linux.SOCK_CLOEXEC) != 0 { + return 0, syserror.EINVAL + } + + // Get socket from the file descriptor. + file := t.GetFileVFS2(fd) + if file == nil { + return 0, syserror.EBADF + } + defer file.DecRef() + + // Extract the socket. + s, ok := file.Impl().(socket.SocketVFS2) + if !ok { + return 0, syserror.ENOTSOCK + } + + // Call the syscall implementation for this socket, then copy the + // output address if one is specified. + blocking := (file.StatusFlags() & linux.SOCK_NONBLOCK) == 0 + + peerRequested := addrLen != 0 + nfd, peer, peerLen, e := s.Accept(t, peerRequested, flags, blocking) + if e != nil { + return 0, syserror.ConvertIntr(e.ToError(), kernel.ERESTARTSYS) + } + if peerRequested { + // NOTE(magi): Linux does not give you an error if it can't + // write the data back out so neither do we. + if err := writeAddress(t, peer, peerLen, addr, addrLen); err == syserror.EINVAL { + return 0, err + } + } + return uintptr(nfd), nil +} + +// Accept4 implements the linux syscall accept4(2). +func Accept4(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + addr := args[1].Pointer() + addrlen := args[2].Pointer() + flags := int(args[3].Int()) + + n, err := accept(t, fd, addr, addrlen, flags) + return n, nil, err +} + +// Accept implements the linux syscall accept(2). +func Accept(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + addr := args[1].Pointer() + addrlen := args[2].Pointer() + + n, err := accept(t, fd, addr, addrlen, 0) + return n, nil, err +} + +// Bind implements the linux syscall bind(2). +func Bind(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + addr := args[1].Pointer() + addrlen := args[2].Uint() + + // Get socket from the file descriptor. + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Extract the socket. + s, ok := file.Impl().(socket.SocketVFS2) + if !ok { + return 0, nil, syserror.ENOTSOCK + } + + // Capture address and call syscall implementation. + a, err := CaptureAddress(t, addr, addrlen) + if err != nil { + return 0, nil, err + } + + return 0, nil, s.Bind(t, a).ToError() +} + +// Listen implements the linux syscall listen(2). +func Listen(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + backlog := args[1].Int() + + // Get socket from the file descriptor. + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Extract the socket. + s, ok := file.Impl().(socket.SocketVFS2) + if !ok { + return 0, nil, syserror.ENOTSOCK + } + + // Per Linux, the backlog is silently capped to reasonable values. + if backlog <= 0 { + backlog = minListenBacklog + } + if backlog > maxListenBacklog { + backlog = maxListenBacklog + } + + return 0, nil, s.Listen(t, int(backlog)).ToError() +} + +// Shutdown implements the linux syscall shutdown(2). +func Shutdown(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + how := args[1].Int() + + // Get socket from the file descriptor. + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Extract the socket. + s, ok := file.Impl().(socket.SocketVFS2) + if !ok { + return 0, nil, syserror.ENOTSOCK + } + + // Validate how, then call syscall implementation. + switch how { + case linux.SHUT_RD, linux.SHUT_WR, linux.SHUT_RDWR: + default: + return 0, nil, syserror.EINVAL + } + + return 0, nil, s.Shutdown(t, int(how)).ToError() +} + +// GetSockOpt implements the linux syscall getsockopt(2). +func GetSockOpt(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + level := args[1].Int() + name := args[2].Int() + optValAddr := args[3].Pointer() + optLenAddr := args[4].Pointer() + + // Get socket from the file descriptor. + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Extract the socket. + s, ok := file.Impl().(socket.SocketVFS2) + if !ok { + return 0, nil, syserror.ENOTSOCK + } + + // Read the length. Reject negative values. + optLen := int32(0) + if _, err := t.CopyIn(optLenAddr, &optLen); err != nil { + return 0, nil, err + } + if optLen < 0 { + return 0, nil, syserror.EINVAL + } + + // Call syscall implementation then copy both value and value len out. + v, e := getSockOpt(t, s, int(level), int(name), optValAddr, int(optLen)) + if e != nil { + return 0, nil, e.ToError() + } + + vLen := int32(binary.Size(v)) + if _, err := t.CopyOut(optLenAddr, vLen); err != nil { + return 0, nil, err + } + + if v != nil { + if _, err := t.CopyOut(optValAddr, v); err != nil { + return 0, nil, err + } + } + + return 0, nil, nil +} + +// getSockOpt tries to handle common socket options, or dispatches to a specific +// socket implementation. +func getSockOpt(t *kernel.Task, s socket.SocketVFS2, level, name int, optValAddr usermem.Addr, len int) (interface{}, *syserr.Error) { + if level == linux.SOL_SOCKET { + switch name { + case linux.SO_TYPE, linux.SO_DOMAIN, linux.SO_PROTOCOL: + if len < sizeOfInt32 { + return nil, syserr.ErrInvalidArgument + } + } + + switch name { + case linux.SO_TYPE: + _, skType, _ := s.Type() + return int32(skType), nil + case linux.SO_DOMAIN: + family, _, _ := s.Type() + return int32(family), nil + case linux.SO_PROTOCOL: + _, _, protocol := s.Type() + return int32(protocol), nil + } + } + + return s.GetSockOpt(t, level, name, optValAddr, len) +} + +// SetSockOpt implements the linux syscall setsockopt(2). +// +// Note that unlike Linux, enabling SO_PASSCRED does not autobind the socket. +func SetSockOpt(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + level := args[1].Int() + name := args[2].Int() + optValAddr := args[3].Pointer() + optLen := args[4].Int() + + // Get socket from the file descriptor. + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Extract the socket. + s, ok := file.Impl().(socket.SocketVFS2) + if !ok { + return 0, nil, syserror.ENOTSOCK + } + + if optLen < 0 { + return 0, nil, syserror.EINVAL + } + if optLen > maxOptLen { + return 0, nil, syserror.EINVAL + } + buf := t.CopyScratchBuffer(int(optLen)) + if _, err := t.CopyIn(optValAddr, &buf); err != nil { + return 0, nil, err + } + + // Call syscall implementation. + if err := s.SetSockOpt(t, int(level), int(name), buf); err != nil { + return 0, nil, err.ToError() + } + + return 0, nil, nil +} + +// GetSockName implements the linux syscall getsockname(2). +func GetSockName(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + addr := args[1].Pointer() + addrlen := args[2].Pointer() + + // Get socket from the file descriptor. + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Extract the socket. + s, ok := file.Impl().(socket.SocketVFS2) + if !ok { + return 0, nil, syserror.ENOTSOCK + } + + // Get the socket name and copy it to the caller. + v, vl, err := s.GetSockName(t) + if err != nil { + return 0, nil, err.ToError() + } + + return 0, nil, writeAddress(t, v, vl, addr, addrlen) +} + +// GetPeerName implements the linux syscall getpeername(2). +func GetPeerName(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + addr := args[1].Pointer() + addrlen := args[2].Pointer() + + // Get socket from the file descriptor. + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Extract the socket. + s, ok := file.Impl().(socket.SocketVFS2) + if !ok { + return 0, nil, syserror.ENOTSOCK + } + + // Get the socket peer name and copy it to the caller. + v, vl, err := s.GetPeerName(t) + if err != nil { + return 0, nil, err.ToError() + } + + return 0, nil, writeAddress(t, v, vl, addr, addrlen) +} + +// RecvMsg implements the linux syscall recvmsg(2). +func RecvMsg(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + msgPtr := args[1].Pointer() + flags := args[2].Int() + + if t.Arch().Width() != 8 { + // We only handle 64-bit for now. + return 0, nil, syserror.EINVAL + } + + // Get socket from the file descriptor. + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Extract the socket. + s, ok := file.Impl().(socket.SocketVFS2) + if !ok { + return 0, nil, syserror.ENOTSOCK + } + + // Reject flags that we don't handle yet. + if flags & ^(baseRecvFlags|linux.MSG_PEEK|linux.MSG_CMSG_CLOEXEC|linux.MSG_ERRQUEUE) != 0 { + return 0, nil, syserror.EINVAL + } + + if (file.StatusFlags() & linux.SOCK_NONBLOCK) != 0 { + flags |= linux.MSG_DONTWAIT + } + + var haveDeadline bool + var deadline ktime.Time + if dl := s.RecvTimeout(); dl > 0 { + deadline = t.Kernel().MonotonicClock().Now().Add(time.Duration(dl) * time.Nanosecond) + haveDeadline = true + } else if dl < 0 { + flags |= linux.MSG_DONTWAIT + } + + n, err := recvSingleMsg(t, s, msgPtr, flags, haveDeadline, deadline) + return n, nil, err +} + +// RecvMMsg implements the linux syscall recvmmsg(2). +func RecvMMsg(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + msgPtr := args[1].Pointer() + vlen := args[2].Uint() + flags := args[3].Int() + toPtr := args[4].Pointer() + + if t.Arch().Width() != 8 { + // We only handle 64-bit for now. + return 0, nil, syserror.EINVAL + } + + // Reject flags that we don't handle yet. + if flags & ^(baseRecvFlags|linux.MSG_CMSG_CLOEXEC|linux.MSG_ERRQUEUE) != 0 { + return 0, nil, syserror.EINVAL + } + + // Get socket from the file descriptor. + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Extract the socket. + s, ok := file.Impl().(socket.SocketVFS2) + if !ok { + return 0, nil, syserror.ENOTSOCK + } + + if (file.StatusFlags() & linux.SOCK_NONBLOCK) != 0 { + flags |= linux.MSG_DONTWAIT + } + + var haveDeadline bool + var deadline ktime.Time + if toPtr != 0 { + var ts linux.Timespec + if _, err := ts.CopyIn(t, toPtr); err != nil { + return 0, nil, err + } + if !ts.Valid() { + return 0, nil, syserror.EINVAL + } + deadline = t.Kernel().MonotonicClock().Now().Add(ts.ToDuration()) + haveDeadline = true + } + + if !haveDeadline { + if dl := s.RecvTimeout(); dl > 0 { + deadline = t.Kernel().MonotonicClock().Now().Add(time.Duration(dl) * time.Nanosecond) + haveDeadline = true + } else if dl < 0 { + flags |= linux.MSG_DONTWAIT + } + } + + var count uint32 + var err error + for i := uint64(0); i < uint64(vlen); i++ { + mp, ok := msgPtr.AddLength(i * multipleMessageHeader64Len) + if !ok { + return 0, nil, syserror.EFAULT + } + var n uintptr + if n, err = recvSingleMsg(t, s, mp, flags, haveDeadline, deadline); err != nil { + break + } + + // Copy the received length to the caller. + lp, ok := mp.AddLength(messageHeader64Len) + if !ok { + return 0, nil, syserror.EFAULT + } + if _, err = t.CopyOut(lp, uint32(n)); err != nil { + break + } + count++ + } + + if count == 0 { + return 0, nil, err + } + return uintptr(count), nil, nil +} + +func recvSingleMsg(t *kernel.Task, s socket.SocketVFS2, msgPtr usermem.Addr, flags int32, haveDeadline bool, deadline ktime.Time) (uintptr, error) { + // Capture the message header and io vectors. + var msg MessageHeader64 + if err := CopyInMessageHeader64(t, msgPtr, &msg); err != nil { + return 0, err + } + + if msg.IovLen > linux.UIO_MAXIOV { + return 0, syserror.EMSGSIZE + } + dst, err := t.IovecsIOSequence(usermem.Addr(msg.Iov), int(msg.IovLen), usermem.IOOpts{ + AddressSpaceActive: true, + }) + if err != nil { + return 0, err + } + + // FIXME(b/63594852): Pretend we have an empty error queue. + if flags&linux.MSG_ERRQUEUE != 0 { + return 0, syserror.EAGAIN + } + + // Fast path when no control message nor name buffers are provided. + if msg.ControlLen == 0 && msg.NameLen == 0 { + n, mflags, _, _, cms, err := s.RecvMsg(t, dst, int(flags), haveDeadline, deadline, false, 0) + if err != nil { + return 0, syserror.ConvertIntr(err.ToError(), kernel.ERESTARTSYS) + } + if !cms.Unix.Empty() { + mflags |= linux.MSG_CTRUNC + cms.Release() + } + + if int(msg.Flags) != mflags { + // Copy out the flags to the caller. + if _, err := t.CopyOut(msgPtr+flagsOffset, int32(mflags)); err != nil { + return 0, err + } + } + + return uintptr(n), nil + } + + if msg.ControlLen > maxControlLen { + return 0, syserror.ENOBUFS + } + n, mflags, sender, senderLen, cms, e := s.RecvMsg(t, dst, int(flags), haveDeadline, deadline, msg.NameLen != 0, msg.ControlLen) + if e != nil { + return 0, syserror.ConvertIntr(e.ToError(), kernel.ERESTARTSYS) + } + defer cms.Release() + + controlData := make([]byte, 0, msg.ControlLen) + controlData = control.PackControlMessages(t, cms, controlData) + + if cr, ok := s.(transport.Credentialer); ok && cr.Passcred() { + creds, _ := cms.Unix.Credentials.(control.SCMCredentials) + controlData, mflags = control.PackCredentials(t, creds, controlData, mflags) + } + + if cms.Unix.Rights != nil { + controlData, mflags = control.PackRights(t, cms.Unix.Rights.(control.SCMRights), flags&linux.MSG_CMSG_CLOEXEC != 0, controlData, mflags) + } + + // Copy the address to the caller. + if msg.NameLen != 0 { + if err := writeAddress(t, sender, senderLen, usermem.Addr(msg.Name), usermem.Addr(msgPtr+nameLenOffset)); err != nil { + return 0, err + } + } + + // Copy the control data to the caller. + if _, err := t.CopyOut(msgPtr+controlLenOffset, uint64(len(controlData))); err != nil { + return 0, err + } + if len(controlData) > 0 { + if _, err := t.CopyOut(usermem.Addr(msg.Control), controlData); err != nil { + return 0, err + } + } + + // Copy out the flags to the caller. + if _, err := t.CopyOut(msgPtr+flagsOffset, int32(mflags)); err != nil { + return 0, err + } + + return uintptr(n), nil +} + +// recvFrom is the implementation of the recvfrom syscall. It is called by +// recvfrom and recv syscall handlers. +func recvFrom(t *kernel.Task, fd int32, bufPtr usermem.Addr, bufLen uint64, flags int32, namePtr usermem.Addr, nameLenPtr usermem.Addr) (uintptr, error) { + if int(bufLen) < 0 { + return 0, syserror.EINVAL + } + + // Reject flags that we don't handle yet. + if flags & ^(baseRecvFlags|linux.MSG_PEEK|linux.MSG_CONFIRM) != 0 { + return 0, syserror.EINVAL + } + + // Get socket from the file descriptor. + file := t.GetFileVFS2(fd) + if file == nil { + return 0, syserror.EBADF + } + defer file.DecRef() + + // Extract the socket. + s, ok := file.Impl().(socket.SocketVFS2) + if !ok { + return 0, syserror.ENOTSOCK + } + + if (file.StatusFlags() & linux.SOCK_NONBLOCK) != 0 { + flags |= linux.MSG_DONTWAIT + } + + dst, err := t.SingleIOSequence(bufPtr, int(bufLen), usermem.IOOpts{ + AddressSpaceActive: true, + }) + if err != nil { + return 0, err + } + + var haveDeadline bool + var deadline ktime.Time + if dl := s.RecvTimeout(); dl > 0 { + deadline = t.Kernel().MonotonicClock().Now().Add(time.Duration(dl) * time.Nanosecond) + haveDeadline = true + } else if dl < 0 { + flags |= linux.MSG_DONTWAIT + } + + n, _, sender, senderLen, cm, e := s.RecvMsg(t, dst, int(flags), haveDeadline, deadline, nameLenPtr != 0, 0) + cm.Release() + if e != nil { + return 0, syserror.ConvertIntr(e.ToError(), kernel.ERESTARTSYS) + } + + // Copy the address to the caller. + if nameLenPtr != 0 { + if err := writeAddress(t, sender, senderLen, namePtr, nameLenPtr); err != nil { + return 0, err + } + } + + return uintptr(n), nil +} + +// RecvFrom implements the linux syscall recvfrom(2). +func RecvFrom(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + bufPtr := args[1].Pointer() + bufLen := args[2].Uint64() + flags := args[3].Int() + namePtr := args[4].Pointer() + nameLenPtr := args[5].Pointer() + + n, err := recvFrom(t, fd, bufPtr, bufLen, flags, namePtr, nameLenPtr) + return n, nil, err +} + +// SendMsg implements the linux syscall sendmsg(2). +func SendMsg(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + msgPtr := args[1].Pointer() + flags := args[2].Int() + + if t.Arch().Width() != 8 { + // We only handle 64-bit for now. + return 0, nil, syserror.EINVAL + } + + // Get socket from the file descriptor. + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Extract the socket. + s, ok := file.Impl().(socket.SocketVFS2) + if !ok { + return 0, nil, syserror.ENOTSOCK + } + + // Reject flags that we don't handle yet. + if flags & ^(linux.MSG_DONTWAIT|linux.MSG_EOR|linux.MSG_MORE|linux.MSG_NOSIGNAL) != 0 { + return 0, nil, syserror.EINVAL + } + + if (file.StatusFlags() & linux.SOCK_NONBLOCK) != 0 { + flags |= linux.MSG_DONTWAIT + } + + n, err := sendSingleMsg(t, s, file, msgPtr, flags) + return n, nil, err +} + +// SendMMsg implements the linux syscall sendmmsg(2). +func SendMMsg(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + msgPtr := args[1].Pointer() + vlen := args[2].Uint() + flags := args[3].Int() + + if t.Arch().Width() != 8 { + // We only handle 64-bit for now. + return 0, nil, syserror.EINVAL + } + + // Get socket from the file descriptor. + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Extract the socket. + s, ok := file.Impl().(socket.SocketVFS2) + if !ok { + return 0, nil, syserror.ENOTSOCK + } + + // Reject flags that we don't handle yet. + if flags & ^(linux.MSG_DONTWAIT|linux.MSG_EOR|linux.MSG_MORE|linux.MSG_NOSIGNAL) != 0 { + return 0, nil, syserror.EINVAL + } + + if (file.StatusFlags() & linux.SOCK_NONBLOCK) != 0 { + flags |= linux.MSG_DONTWAIT + } + + var count uint32 + var err error + for i := uint64(0); i < uint64(vlen); i++ { + mp, ok := msgPtr.AddLength(i * multipleMessageHeader64Len) + if !ok { + return 0, nil, syserror.EFAULT + } + var n uintptr + if n, err = sendSingleMsg(t, s, file, mp, flags); err != nil { + break + } + + // Copy the received length to the caller. + lp, ok := mp.AddLength(messageHeader64Len) + if !ok { + return 0, nil, syserror.EFAULT + } + if _, err = t.CopyOut(lp, uint32(n)); err != nil { + break + } + count++ + } + + if count == 0 { + return 0, nil, err + } + return uintptr(count), nil, nil +} + +func sendSingleMsg(t *kernel.Task, s socket.SocketVFS2, file *vfs.FileDescription, msgPtr usermem.Addr, flags int32) (uintptr, error) { + // Capture the message header. + var msg MessageHeader64 + if err := CopyInMessageHeader64(t, msgPtr, &msg); err != nil { + return 0, err + } + + var controlData []byte + if msg.ControlLen > 0 { + // Put an upper bound to prevent large allocations. + if msg.ControlLen > maxControlLen { + return 0, syserror.ENOBUFS + } + controlData = make([]byte, msg.ControlLen) + if _, err := t.CopyIn(usermem.Addr(msg.Control), &controlData); err != nil { + return 0, err + } + } + + // Read the destination address if one is specified. + var to []byte + if msg.NameLen != 0 { + var err error + to, err = CaptureAddress(t, usermem.Addr(msg.Name), msg.NameLen) + if err != nil { + return 0, err + } + } + + // Read data then call the sendmsg implementation. + if msg.IovLen > linux.UIO_MAXIOV { + return 0, syserror.EMSGSIZE + } + src, err := t.IovecsIOSequence(usermem.Addr(msg.Iov), int(msg.IovLen), usermem.IOOpts{ + AddressSpaceActive: true, + }) + if err != nil { + return 0, err + } + + controlMessages, err := control.Parse(t, s, controlData) + if err != nil { + return 0, err + } + + var haveDeadline bool + var deadline ktime.Time + if dl := s.SendTimeout(); dl > 0 { + deadline = t.Kernel().MonotonicClock().Now().Add(time.Duration(dl) * time.Nanosecond) + haveDeadline = true + } else if dl < 0 { + flags |= linux.MSG_DONTWAIT + } + + // Call the syscall implementation. + n, e := s.SendMsg(t, src, to, int(flags), haveDeadline, deadline, controlMessages) + err = slinux.HandleIOErrorVFS2(t, n != 0, e.ToError(), kernel.ERESTARTSYS, "sendmsg", file) + if err != nil { + controlMessages.Release() + } + return uintptr(n), err +} + +// sendTo is the implementation of the sendto syscall. It is called by sendto +// and send syscall handlers. +func sendTo(t *kernel.Task, fd int32, bufPtr usermem.Addr, bufLen uint64, flags int32, namePtr usermem.Addr, nameLen uint32) (uintptr, error) { + bl := int(bufLen) + if bl < 0 { + return 0, syserror.EINVAL + } + + // Get socket from the file descriptor. + file := t.GetFileVFS2(fd) + if file == nil { + return 0, syserror.EBADF + } + defer file.DecRef() + + // Extract the socket. + s, ok := file.Impl().(socket.SocketVFS2) + if !ok { + return 0, syserror.ENOTSOCK + } + + if (file.StatusFlags() & linux.SOCK_NONBLOCK) != 0 { + flags |= linux.MSG_DONTWAIT + } + + // Read the destination address if one is specified. + var to []byte + var err error + if namePtr != 0 { + to, err = CaptureAddress(t, namePtr, nameLen) + if err != nil { + return 0, err + } + } + + src, err := t.SingleIOSequence(bufPtr, bl, usermem.IOOpts{ + AddressSpaceActive: true, + }) + if err != nil { + return 0, err + } + + var haveDeadline bool + var deadline ktime.Time + if dl := s.SendTimeout(); dl > 0 { + deadline = t.Kernel().MonotonicClock().Now().Add(time.Duration(dl) * time.Nanosecond) + haveDeadline = true + } else if dl < 0 { + flags |= linux.MSG_DONTWAIT + } + + // Call the syscall implementation. + n, e := s.SendMsg(t, src, to, int(flags), haveDeadline, deadline, socket.ControlMessages{Unix: control.New(t, s, nil)}) + return uintptr(n), slinux.HandleIOErrorVFS2(t, n != 0, e.ToError(), kernel.ERESTARTSYS, "sendto", file) +} + +// SendTo implements the linux syscall sendto(2). +func SendTo(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + bufPtr := args[1].Pointer() + bufLen := args[2].Uint64() + flags := args[3].Int() + namePtr := args[4].Pointer() + nameLen := args[5].Uint() + + n, err := sendTo(t, fd, bufPtr, bufLen, flags, namePtr, nameLen) + return n, nil, err +} diff --git a/pkg/sentry/syscalls/linux/vfs2/stat.go b/pkg/sentry/syscalls/linux/vfs2/stat.go new file mode 100644 index 000000000..bb1d5cac4 --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/stat.go @@ -0,0 +1,388 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/bits" + "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/gohacks" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +// Stat implements Linux syscall stat(2). +func Stat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + pathAddr := args[0].Pointer() + statAddr := args[1].Pointer() + return 0, nil, fstatat(t, linux.AT_FDCWD, pathAddr, statAddr, 0 /* flags */) +} + +// Lstat implements Linux syscall lstat(2). +func Lstat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + pathAddr := args[0].Pointer() + statAddr := args[1].Pointer() + return 0, nil, fstatat(t, linux.AT_FDCWD, pathAddr, statAddr, linux.AT_SYMLINK_NOFOLLOW) +} + +// Newfstatat implements Linux syscall newfstatat, which backs fstatat(2). +func Newfstatat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + dirfd := args[0].Int() + pathAddr := args[1].Pointer() + statAddr := args[2].Pointer() + flags := args[3].Int() + return 0, nil, fstatat(t, dirfd, pathAddr, statAddr, flags) +} + +func fstatat(t *kernel.Task, dirfd int32, pathAddr, statAddr usermem.Addr, flags int32) error { + if flags&^(linux.AT_EMPTY_PATH|linux.AT_SYMLINK_NOFOLLOW) != 0 { + return syserror.EINVAL + } + + opts := vfs.StatOptions{ + Mask: linux.STATX_BASIC_STATS, + } + + path, err := copyInPath(t, pathAddr) + if err != nil { + return err + } + + root := t.FSContext().RootDirectoryVFS2() + defer root.DecRef() + start := root + if !path.Absolute { + if !path.HasComponents() && flags&linux.AT_EMPTY_PATH == 0 { + return syserror.ENOENT + } + if dirfd == linux.AT_FDCWD { + start = t.FSContext().WorkingDirectoryVFS2() + defer start.DecRef() + } else { + dirfile := t.GetFileVFS2(dirfd) + if dirfile == nil { + return syserror.EBADF + } + if !path.HasComponents() { + // Use FileDescription.Stat() instead of + // VirtualFilesystem.StatAt() for fstatat(fd, ""), since the + // former may be able to use opened file state to expedite the + // Stat. + statx, err := dirfile.Stat(t, opts) + dirfile.DecRef() + if err != nil { + return err + } + var stat linux.Stat + convertStatxToUserStat(t, &statx, &stat) + _, err = stat.CopyOut(t, statAddr) + return err + } + start = dirfile.VirtualDentry() + start.IncRef() + defer start.DecRef() + dirfile.DecRef() + } + } + + statx, err := t.Kernel().VFS().StatAt(t, t.Credentials(), &vfs.PathOperation{ + Root: root, + Start: start, + Path: path, + FollowFinalSymlink: flags&linux.AT_SYMLINK_NOFOLLOW == 0, + }, &opts) + if err != nil { + return err + } + var stat linux.Stat + convertStatxToUserStat(t, &statx, &stat) + _, err = stat.CopyOut(t, statAddr) + return err +} + +func timespecFromStatxTimestamp(sxts linux.StatxTimestamp) linux.Timespec { + return linux.Timespec{ + Sec: sxts.Sec, + Nsec: int64(sxts.Nsec), + } +} + +// Fstat implements Linux syscall fstat(2). +func Fstat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + statAddr := args[1].Pointer() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + statx, err := file.Stat(t, vfs.StatOptions{ + Mask: linux.STATX_BASIC_STATS, + }) + if err != nil { + return 0, nil, err + } + var stat linux.Stat + convertStatxToUserStat(t, &statx, &stat) + _, err = stat.CopyOut(t, statAddr) + return 0, nil, err +} + +// Statx implements Linux syscall statx(2). +func Statx(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + dirfd := args[0].Int() + pathAddr := args[1].Pointer() + flags := args[2].Int() + mask := args[3].Uint() + statxAddr := args[4].Pointer() + + if flags&^(linux.AT_EMPTY_PATH|linux.AT_SYMLINK_NOFOLLOW|linux.AT_STATX_SYNC_TYPE) != 0 { + return 0, nil, syserror.EINVAL + } + // Make sure that only one sync type option is set. + syncType := uint32(flags & linux.AT_STATX_SYNC_TYPE) + if syncType != 0 && !bits.IsPowerOfTwo32(syncType) { + return 0, nil, syserror.EINVAL + } + if mask&linux.STATX__RESERVED != 0 { + return 0, nil, syserror.EINVAL + } + + opts := vfs.StatOptions{ + Mask: mask, + Sync: uint32(flags & linux.AT_STATX_SYNC_TYPE), + } + + path, err := copyInPath(t, pathAddr) + if err != nil { + return 0, nil, err + } + + root := t.FSContext().RootDirectoryVFS2() + defer root.DecRef() + start := root + if !path.Absolute { + if !path.HasComponents() && flags&linux.AT_EMPTY_PATH == 0 { + return 0, nil, syserror.ENOENT + } + if dirfd == linux.AT_FDCWD { + start = t.FSContext().WorkingDirectoryVFS2() + defer start.DecRef() + } else { + dirfile := t.GetFileVFS2(dirfd) + if dirfile == nil { + return 0, nil, syserror.EBADF + } + if !path.HasComponents() { + // Use FileDescription.Stat() instead of + // VirtualFilesystem.StatAt() for statx(fd, ""), since the + // former may be able to use opened file state to expedite the + // Stat. + statx, err := dirfile.Stat(t, opts) + dirfile.DecRef() + if err != nil { + return 0, nil, err + } + userifyStatx(t, &statx) + _, err = statx.CopyOut(t, statxAddr) + return 0, nil, err + } + start = dirfile.VirtualDentry() + start.IncRef() + defer start.DecRef() + dirfile.DecRef() + } + } + + statx, err := t.Kernel().VFS().StatAt(t, t.Credentials(), &vfs.PathOperation{ + Root: root, + Start: start, + Path: path, + FollowFinalSymlink: flags&linux.AT_SYMLINK_NOFOLLOW == 0, + }, &opts) + if err != nil { + return 0, nil, err + } + userifyStatx(t, &statx) + _, err = statx.CopyOut(t, statxAddr) + return 0, nil, err +} + +func userifyStatx(t *kernel.Task, statx *linux.Statx) { + userns := t.UserNamespace() + statx.UID = uint32(auth.KUID(statx.UID).In(userns).OrOverflow()) + statx.GID = uint32(auth.KGID(statx.GID).In(userns).OrOverflow()) +} + +// Readlink implements Linux syscall readlink(2). +func Readlink(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + pathAddr := args[0].Pointer() + bufAddr := args[1].Pointer() + size := args[2].SizeT() + return readlinkat(t, linux.AT_FDCWD, pathAddr, bufAddr, size) +} + +// Access implements Linux syscall access(2). +func Access(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + addr := args[0].Pointer() + mode := args[1].ModeT() + + return 0, nil, accessAt(t, linux.AT_FDCWD, addr, mode) +} + +// Faccessat implements Linux syscall faccessat(2). +// +// Note that the faccessat() system call does not take a flags argument: +// "The raw faccessat() system call takes only the first three arguments. The +// AT_EACCESS and AT_SYMLINK_NOFOLLOW flags are actually implemented within +// the glibc wrapper function for faccessat(). If either of these flags is +// specified, then the wrapper function employs fstatat(2) to determine access +// permissions." - faccessat(2) +func Faccessat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + dirfd := args[0].Int() + addr := args[1].Pointer() + mode := args[2].ModeT() + + return 0, nil, accessAt(t, dirfd, addr, mode) +} + +func accessAt(t *kernel.Task, dirfd int32, pathAddr usermem.Addr, mode uint) error { + const rOK = 4 + const wOK = 2 + const xOK = 1 + + // Sanity check the mode. + if mode&^(rOK|wOK|xOK) != 0 { + return syserror.EINVAL + } + + path, err := copyInPath(t, pathAddr) + if err != nil { + return err + } + tpop, err := getTaskPathOperation(t, dirfd, path, disallowEmptyPath, followFinalSymlink) + if err != nil { + return err + } + defer tpop.Release() + + // access(2) and faccessat(2) check permissions using real + // UID/GID, not effective UID/GID. + // + // "access() needs to use the real uid/gid, not the effective + // uid/gid. We do this by temporarily clearing all FS-related + // capabilities and switching the fsuid/fsgid around to the + // real ones." -fs/open.c:faccessat + creds := t.Credentials().Fork() + creds.EffectiveKUID = creds.RealKUID + creds.EffectiveKGID = creds.RealKGID + if creds.EffectiveKUID.In(creds.UserNamespace) == auth.RootUID { + creds.EffectiveCaps = creds.PermittedCaps + } else { + creds.EffectiveCaps = 0 + } + + return t.Kernel().VFS().AccessAt(t, creds, vfs.AccessTypes(mode), &tpop.pop) +} + +// Readlinkat implements Linux syscall mknodat(2). +func Readlinkat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + dirfd := args[0].Int() + pathAddr := args[1].Pointer() + bufAddr := args[2].Pointer() + size := args[3].SizeT() + return readlinkat(t, dirfd, pathAddr, bufAddr, size) +} + +func readlinkat(t *kernel.Task, dirfd int32, pathAddr, bufAddr usermem.Addr, size uint) (uintptr, *kernel.SyscallControl, error) { + if int(size) <= 0 { + return 0, nil, syserror.EINVAL + } + + path, err := copyInPath(t, pathAddr) + if err != nil { + return 0, nil, err + } + // "Since Linux 2.6.39, pathname can be an empty string, in which case the + // call operates on the symbolic link referred to by dirfd ..." - + // readlinkat(2) + tpop, err := getTaskPathOperation(t, dirfd, path, allowEmptyPath, nofollowFinalSymlink) + if err != nil { + return 0, nil, err + } + defer tpop.Release() + + target, err := t.Kernel().VFS().ReadlinkAt(t, t.Credentials(), &tpop.pop) + if err != nil { + return 0, nil, err + } + + if len(target) > int(size) { + target = target[:size] + } + n, err := t.CopyOutBytes(bufAddr, gohacks.ImmutableBytesFromString(target)) + if n == 0 { + return 0, nil, err + } + return uintptr(n), nil, nil +} + +// Statfs implements Linux syscall statfs(2). +func Statfs(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + pathAddr := args[0].Pointer() + bufAddr := args[1].Pointer() + + path, err := copyInPath(t, pathAddr) + if err != nil { + return 0, nil, err + } + tpop, err := getTaskPathOperation(t, linux.AT_FDCWD, path, disallowEmptyPath, followFinalSymlink) + if err != nil { + return 0, nil, err + } + defer tpop.Release() + + statfs, err := t.Kernel().VFS().StatFSAt(t, t.Credentials(), &tpop.pop) + if err != nil { + return 0, nil, err + } + _, err = statfs.CopyOut(t, bufAddr) + return 0, nil, err +} + +// Fstatfs implements Linux syscall fstatfs(2). +func Fstatfs(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + bufAddr := args[1].Pointer() + + tpop, err := getTaskPathOperation(t, fd, fspath.Path{}, allowEmptyPath, nofollowFinalSymlink) + if err != nil { + return 0, nil, err + } + defer tpop.Release() + + statfs, err := t.Kernel().VFS().StatFSAt(t, t.Credentials(), &tpop.pop) + if err != nil { + return 0, nil, err + } + _, err = statfs.CopyOut(t, bufAddr) + return 0, nil, err +} diff --git a/pkg/sentry/syscalls/linux/vfs2/stat_amd64.go b/pkg/sentry/syscalls/linux/vfs2/stat_amd64.go new file mode 100644 index 000000000..2da538fc6 --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/stat_amd64.go @@ -0,0 +1,46 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// +build amd64 + +package vfs2 + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" +) + +// This takes both input and output as pointer arguments to avoid copying large +// structs. +func convertStatxToUserStat(t *kernel.Task, statx *linux.Statx, stat *linux.Stat) { + // Linux just copies fields from struct kstat without regard to struct + // kstat::result_mask (fs/stat.c:cp_new_stat()), so we do too. + userns := t.UserNamespace() + *stat = linux.Stat{ + Dev: uint64(linux.MakeDeviceID(uint16(statx.DevMajor), statx.DevMinor)), + Ino: statx.Ino, + Nlink: uint64(statx.Nlink), + Mode: uint32(statx.Mode), + UID: uint32(auth.KUID(statx.UID).In(userns).OrOverflow()), + GID: uint32(auth.KGID(statx.GID).In(userns).OrOverflow()), + Rdev: uint64(linux.MakeDeviceID(uint16(statx.RdevMajor), statx.RdevMinor)), + Size: int64(statx.Size), + Blksize: int64(statx.Blksize), + Blocks: int64(statx.Blocks), + ATime: timespecFromStatxTimestamp(statx.Atime), + MTime: timespecFromStatxTimestamp(statx.Mtime), + CTime: timespecFromStatxTimestamp(statx.Ctime), + } +} diff --git a/pkg/sentry/syscalls/linux/vfs2/stat_arm64.go b/pkg/sentry/syscalls/linux/vfs2/stat_arm64.go new file mode 100644 index 000000000..88b9c7627 --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/stat_arm64.go @@ -0,0 +1,46 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// +build arm64 + +package vfs2 + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" +) + +// This takes both input and output as pointer arguments to avoid copying large +// structs. +func convertStatxToUserStat(t *kernel.Task, statx *linux.Statx, stat *linux.Stat) { + // Linux just copies fields from struct kstat without regard to struct + // kstat::result_mask (fs/stat.c:cp_new_stat()), so we do too. + userns := t.UserNamespace() + *stat = linux.Stat{ + Dev: uint64(linux.MakeDeviceID(uint16(statx.DevMajor), statx.DevMinor)), + Ino: statx.Ino, + Nlink: uint32(statx.Nlink), + Mode: uint32(statx.Mode), + UID: uint32(auth.KUID(statx.UID).In(userns).OrOverflow()), + GID: uint32(auth.KGID(statx.GID).In(userns).OrOverflow()), + Rdev: uint64(linux.MakeDeviceID(uint16(statx.RdevMajor), statx.RdevMinor)), + Size: int64(statx.Size), + Blksize: int32(statx.Blksize), + Blocks: int64(statx.Blocks), + ATime: timespecFromStatxTimestamp(statx.Atime), + MTime: timespecFromStatxTimestamp(statx.Mtime), + CTime: timespecFromStatxTimestamp(statx.Ctime), + } +} diff --git a/pkg/sentry/syscalls/linux/vfs2/sync.go b/pkg/sentry/syscalls/linux/vfs2/sync.go new file mode 100644 index 000000000..365250b0b --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/sync.go @@ -0,0 +1,87 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/syserror" +) + +// Sync implements Linux syscall sync(2). +func Sync(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + return 0, nil, t.Kernel().VFS().SyncAllFilesystems(t) +} + +// Syncfs implements Linux syscall syncfs(2). +func Syncfs(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + return 0, nil, file.SyncFS(t) +} + +// Fsync implements Linux syscall fsync(2). +func Fsync(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + return 0, nil, file.Sync(t) +} + +// Fdatasync implements Linux syscall fdatasync(2). +func Fdatasync(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + // TODO(gvisor.dev/issue/1897): Avoid writeback of unnecessary metadata. + return Fsync(t, args) +} + +// SyncFileRange implements Linux syscall sync_file_range(2). +func SyncFileRange(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + offset := args[1].Int64() + nbytes := args[2].Int64() + flags := args[3].Uint() + + if offset < 0 { + return 0, nil, syserror.EINVAL + } + if nbytes < 0 { + return 0, nil, syserror.EINVAL + } + if flags&^(linux.SYNC_FILE_RANGE_WAIT_BEFORE|linux.SYNC_FILE_RANGE_WRITE|linux.SYNC_FILE_RANGE_WAIT_AFTER) != 0 { + return 0, nil, syserror.EINVAL + } + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // TODO(gvisor.dev/issue/1897): Avoid writeback of data ranges outside of + // [offset, offset+nbytes). + return 0, nil, file.Sync(t) +} diff --git a/pkg/sentry/syscalls/linux/vfs2/sys_read.go b/pkg/sentry/syscalls/linux/vfs2/sys_read.go deleted file mode 100644 index 7667524c7..000000000 --- a/pkg/sentry/syscalls/linux/vfs2/sys_read.go +++ /dev/null @@ -1,95 +0,0 @@ -// Copyright 2020 The gVisor Authors. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -package vfs2 - -import ( - "gvisor.dev/gvisor/pkg/sentry/arch" - "gvisor.dev/gvisor/pkg/sentry/kernel" - "gvisor.dev/gvisor/pkg/sentry/syscalls/linux" - "gvisor.dev/gvisor/pkg/sentry/vfs" - "gvisor.dev/gvisor/pkg/syserror" - "gvisor.dev/gvisor/pkg/usermem" - "gvisor.dev/gvisor/pkg/waiter" -) - -const ( - // EventMaskRead contains events that can be triggered on reads. - EventMaskRead = waiter.EventIn | waiter.EventHUp | waiter.EventErr -) - -// Read implements linux syscall read(2). Note that we try to get a buffer that -// is exactly the size requested because some applications like qemu expect -// they can do large reads all at once. Bug for bug. Same for other read -// calls below. -func Read(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { - fd := args[0].Int() - addr := args[1].Pointer() - size := args[2].SizeT() - - file := t.GetFileVFS2(fd) - if file == nil { - return 0, nil, syserror.EBADF - } - defer file.DecRef() - - // Check that the size is legitimate. - si := int(size) - if si < 0 { - return 0, nil, syserror.EINVAL - } - - // Get the destination of the read. - dst, err := t.SingleIOSequence(addr, si, usermem.IOOpts{ - AddressSpaceActive: true, - }) - if err != nil { - return 0, nil, err - } - - n, err := read(t, file, dst, vfs.ReadOptions{}) - t.IOUsage().AccountReadSyscall(n) - return uintptr(n), nil, linux.HandleIOErrorVFS2(t, n != 0, err, kernel.ERESTARTSYS, "read", file) -} - -func read(t *kernel.Task, file *vfs.FileDescription, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { - n, err := file.Read(t, dst, opts) - if err != syserror.ErrWouldBlock { - return n, err - } - - // Register for notifications. - w, ch := waiter.NewChannelEntry(nil) - file.EventRegister(&w, EventMaskRead) - - total := n - for { - // Shorten dst to reflect bytes previously read. - dst = dst.DropFirst(int(n)) - - // Issue the request and break out if it completes with anything other than - // "would block". - n, err := file.Read(t, dst, opts) - total += n - if err != syserror.ErrWouldBlock { - break - } - if err := t.Block(ch); err != nil { - break - } - } - file.EventUnregister(&w) - - return total, err -} diff --git a/pkg/sentry/syscalls/linux/vfs2/sys_timerfd.go b/pkg/sentry/syscalls/linux/vfs2/sys_timerfd.go new file mode 100644 index 000000000..7938a5249 --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/sys_timerfd.go @@ -0,0 +1,123 @@ +// Copyright 2018 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" +) + +// TimerfdCreate implements Linux syscall timerfd_create(2). +func TimerfdCreate(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + clockID := args[0].Int() + flags := args[1].Int() + + if flags&^(linux.TFD_CLOEXEC|linux.TFD_NONBLOCK) != 0 { + return 0, nil, syserror.EINVAL + } + + var fileFlags uint32 + if flags&linux.TFD_NONBLOCK != 0 { + fileFlags = linux.O_NONBLOCK + } + + var clock ktime.Clock + switch clockID { + case linux.CLOCK_REALTIME: + clock = t.Kernel().RealtimeClock() + case linux.CLOCK_MONOTONIC, linux.CLOCK_BOOTTIME: + clock = t.Kernel().MonotonicClock() + default: + return 0, nil, syserror.EINVAL + } + file, err := t.Kernel().VFS().NewTimerFD(clock, fileFlags) + if err != nil { + return 0, nil, err + } + defer file.DecRef() + fd, err := t.NewFDFromVFS2(0, file, kernel.FDFlags{ + CloseOnExec: flags&linux.TFD_CLOEXEC != 0, + }) + if err != nil { + return 0, nil, err + } + return uintptr(fd), nil, nil +} + +// TimerfdSettime implements Linux syscall timerfd_settime(2). +func TimerfdSettime(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + flags := args[1].Int() + newValAddr := args[2].Pointer() + oldValAddr := args[3].Pointer() + + if flags&^(linux.TFD_TIMER_ABSTIME) != 0 { + return 0, nil, syserror.EINVAL + } + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + tfd, ok := file.Impl().(*vfs.TimerFileDescription) + if !ok { + return 0, nil, syserror.EINVAL + } + + var newVal linux.Itimerspec + if _, err := t.CopyIn(newValAddr, &newVal); err != nil { + return 0, nil, err + } + newS, err := ktime.SettingFromItimerspec(newVal, flags&linux.TFD_TIMER_ABSTIME != 0, tfd.Clock()) + if err != nil { + return 0, nil, err + } + tm, oldS := tfd.SetTime(newS) + if oldValAddr != 0 { + oldVal := ktime.ItimerspecFromSetting(tm, oldS) + if _, err := t.CopyOut(oldValAddr, &oldVal); err != nil { + return 0, nil, err + } + } + return 0, nil, nil +} + +// TimerfdGettime implements Linux syscall timerfd_gettime(2). +func TimerfdGettime(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + curValAddr := args[1].Pointer() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + tfd, ok := file.Impl().(*vfs.TimerFileDescription) + if !ok { + return 0, nil, syserror.EINVAL + } + + tm, s := tfd.GetTime() + curVal := ktime.ItimerspecFromSetting(tm, s) + _, err := t.CopyOut(curValAddr, &curVal) + return 0, nil, err +} diff --git a/pkg/sentry/syscalls/linux/vfs2/xattr.go b/pkg/sentry/syscalls/linux/vfs2/xattr.go new file mode 100644 index 000000000..af455d5c1 --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/xattr.go @@ -0,0 +1,356 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "bytes" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/gohacks" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +// Listxattr implements Linux syscall listxattr(2). +func Listxattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + return listxattr(t, args, followFinalSymlink) +} + +// Llistxattr implements Linux syscall llistxattr(2). +func Llistxattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + return listxattr(t, args, nofollowFinalSymlink) +} + +func listxattr(t *kernel.Task, args arch.SyscallArguments, shouldFollowFinalSymlink shouldFollowFinalSymlink) (uintptr, *kernel.SyscallControl, error) { + pathAddr := args[0].Pointer() + listAddr := args[1].Pointer() + size := args[2].SizeT() + + path, err := copyInPath(t, pathAddr) + if err != nil { + return 0, nil, err + } + tpop, err := getTaskPathOperation(t, linux.AT_FDCWD, path, disallowEmptyPath, shouldFollowFinalSymlink) + if err != nil { + return 0, nil, err + } + defer tpop.Release() + + names, err := t.Kernel().VFS().ListxattrAt(t, t.Credentials(), &tpop.pop, uint64(size)) + if err != nil { + return 0, nil, err + } + n, err := copyOutXattrNameList(t, listAddr, size, names) + if err != nil { + return 0, nil, err + } + return uintptr(n), nil, nil +} + +// Flistxattr implements Linux syscall flistxattr(2). +func Flistxattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + listAddr := args[1].Pointer() + size := args[2].SizeT() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + names, err := file.Listxattr(t, uint64(size)) + if err != nil { + return 0, nil, err + } + n, err := copyOutXattrNameList(t, listAddr, size, names) + if err != nil { + return 0, nil, err + } + return uintptr(n), nil, nil +} + +// Getxattr implements Linux syscall getxattr(2). +func Getxattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + return getxattr(t, args, followFinalSymlink) +} + +// Lgetxattr implements Linux syscall lgetxattr(2). +func Lgetxattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + return getxattr(t, args, nofollowFinalSymlink) +} + +func getxattr(t *kernel.Task, args arch.SyscallArguments, shouldFollowFinalSymlink shouldFollowFinalSymlink) (uintptr, *kernel.SyscallControl, error) { + pathAddr := args[0].Pointer() + nameAddr := args[1].Pointer() + valueAddr := args[2].Pointer() + size := args[3].SizeT() + + path, err := copyInPath(t, pathAddr) + if err != nil { + return 0, nil, err + } + tpop, err := getTaskPathOperation(t, linux.AT_FDCWD, path, disallowEmptyPath, shouldFollowFinalSymlink) + if err != nil { + return 0, nil, err + } + defer tpop.Release() + + name, err := copyInXattrName(t, nameAddr) + if err != nil { + return 0, nil, err + } + + value, err := t.Kernel().VFS().GetxattrAt(t, t.Credentials(), &tpop.pop, &vfs.GetxattrOptions{ + Name: name, + Size: uint64(size), + }) + if err != nil { + return 0, nil, err + } + n, err := copyOutXattrValue(t, valueAddr, size, value) + if err != nil { + return 0, nil, err + } + return uintptr(n), nil, nil +} + +// Fgetxattr implements Linux syscall fgetxattr(2). +func Fgetxattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + nameAddr := args[1].Pointer() + valueAddr := args[2].Pointer() + size := args[3].SizeT() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + name, err := copyInXattrName(t, nameAddr) + if err != nil { + return 0, nil, err + } + + value, err := file.Getxattr(t, &vfs.GetxattrOptions{Name: name, Size: uint64(size)}) + if err != nil { + return 0, nil, err + } + n, err := copyOutXattrValue(t, valueAddr, size, value) + if err != nil { + return 0, nil, err + } + return uintptr(n), nil, nil +} + +// Setxattr implements Linux syscall setxattr(2). +func Setxattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + return 0, nil, setxattr(t, args, followFinalSymlink) +} + +// Lsetxattr implements Linux syscall lsetxattr(2). +func Lsetxattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + return 0, nil, setxattr(t, args, nofollowFinalSymlink) +} + +func setxattr(t *kernel.Task, args arch.SyscallArguments, shouldFollowFinalSymlink shouldFollowFinalSymlink) error { + pathAddr := args[0].Pointer() + nameAddr := args[1].Pointer() + valueAddr := args[2].Pointer() + size := args[3].SizeT() + flags := args[4].Int() + + if flags&^(linux.XATTR_CREATE|linux.XATTR_REPLACE) != 0 { + return syserror.EINVAL + } + + path, err := copyInPath(t, pathAddr) + if err != nil { + return err + } + tpop, err := getTaskPathOperation(t, linux.AT_FDCWD, path, disallowEmptyPath, shouldFollowFinalSymlink) + if err != nil { + return err + } + defer tpop.Release() + + name, err := copyInXattrName(t, nameAddr) + if err != nil { + return err + } + value, err := copyInXattrValue(t, valueAddr, size) + if err != nil { + return err + } + + return t.Kernel().VFS().SetxattrAt(t, t.Credentials(), &tpop.pop, &vfs.SetxattrOptions{ + Name: name, + Value: value, + Flags: uint32(flags), + }) +} + +// Fsetxattr implements Linux syscall fsetxattr(2). +func Fsetxattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + nameAddr := args[1].Pointer() + valueAddr := args[2].Pointer() + size := args[3].SizeT() + flags := args[4].Int() + + if flags&^(linux.XATTR_CREATE|linux.XATTR_REPLACE) != 0 { + return 0, nil, syserror.EINVAL + } + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + name, err := copyInXattrName(t, nameAddr) + if err != nil { + return 0, nil, err + } + value, err := copyInXattrValue(t, valueAddr, size) + if err != nil { + return 0, nil, err + } + + return 0, nil, file.Setxattr(t, &vfs.SetxattrOptions{ + Name: name, + Value: value, + Flags: uint32(flags), + }) +} + +// Removexattr implements Linux syscall removexattr(2). +func Removexattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + return 0, nil, removexattr(t, args, followFinalSymlink) +} + +// Lremovexattr implements Linux syscall lremovexattr(2). +func Lremovexattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + return 0, nil, removexattr(t, args, nofollowFinalSymlink) +} + +func removexattr(t *kernel.Task, args arch.SyscallArguments, shouldFollowFinalSymlink shouldFollowFinalSymlink) error { + pathAddr := args[0].Pointer() + nameAddr := args[1].Pointer() + + path, err := copyInPath(t, pathAddr) + if err != nil { + return err + } + tpop, err := getTaskPathOperation(t, linux.AT_FDCWD, path, disallowEmptyPath, shouldFollowFinalSymlink) + if err != nil { + return err + } + defer tpop.Release() + + name, err := copyInXattrName(t, nameAddr) + if err != nil { + return err + } + + return t.Kernel().VFS().RemovexattrAt(t, t.Credentials(), &tpop.pop, name) +} + +// Fremovexattr implements Linux syscall fremovexattr(2). +func Fremovexattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + nameAddr := args[1].Pointer() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + name, err := copyInXattrName(t, nameAddr) + if err != nil { + return 0, nil, err + } + + return 0, nil, file.Removexattr(t, name) +} + +func copyInXattrName(t *kernel.Task, nameAddr usermem.Addr) (string, error) { + name, err := t.CopyInString(nameAddr, linux.XATTR_NAME_MAX+1) + if err != nil { + if err == syserror.ENAMETOOLONG { + return "", syserror.ERANGE + } + return "", err + } + if len(name) == 0 { + return "", syserror.ERANGE + } + return name, nil +} + +func copyOutXattrNameList(t *kernel.Task, listAddr usermem.Addr, size uint, names []string) (int, error) { + if size > linux.XATTR_LIST_MAX { + size = linux.XATTR_LIST_MAX + } + var buf bytes.Buffer + for _, name := range names { + buf.WriteString(name) + buf.WriteByte(0) + } + if size == 0 { + // Return the size that would be required to accomodate the list. + return buf.Len(), nil + } + if buf.Len() > int(size) { + if size >= linux.XATTR_LIST_MAX { + return 0, syserror.E2BIG + } + return 0, syserror.ERANGE + } + return t.CopyOutBytes(listAddr, buf.Bytes()) +} + +func copyInXattrValue(t *kernel.Task, valueAddr usermem.Addr, size uint) (string, error) { + if size > linux.XATTR_SIZE_MAX { + return "", syserror.E2BIG + } + buf := make([]byte, size) + if _, err := t.CopyInBytes(valueAddr, buf); err != nil { + return "", err + } + return gohacks.StringFromImmutableBytes(buf), nil +} + +func copyOutXattrValue(t *kernel.Task, valueAddr usermem.Addr, size uint, value string) (int, error) { + if size > linux.XATTR_SIZE_MAX { + size = linux.XATTR_SIZE_MAX + } + if size == 0 { + // Return the size that would be required to accomodate the value. + return len(value), nil + } + if len(value) > int(size) { + if size >= linux.XATTR_SIZE_MAX { + return 0, syserror.E2BIG + } + return 0, syserror.ERANGE + } + return t.CopyOutBytes(valueAddr, gohacks.ImmutableBytesFromString(value)) +} diff --git a/pkg/sentry/vfs/BUILD b/pkg/sentry/vfs/BUILD index 14b39eb9d..9aeb83fb0 100644 --- a/pkg/sentry/vfs/BUILD +++ b/pkg/sentry/vfs/BUILD @@ -36,21 +36,31 @@ go_library( "pathname.go", "permissions.go", "resolving_path.go", + "timerfd.go", "vfs.go", ], visibility = ["//pkg/sentry:internal"], deps = [ "//pkg/abi/linux", "//pkg/context", + "//pkg/fd", "//pkg/fspath", + "//pkg/gohacks", + "//pkg/log", + "//pkg/safemem", "//pkg/sentry/arch", + "//pkg/sentry/fs", "//pkg/sentry/fs/lock", "//pkg/sentry/kernel/auth", + "//pkg/sentry/kernel/time", + "//pkg/sentry/limits", "//pkg/sentry/memmap", + "//pkg/sentry/socket/unix/transport", "//pkg/sync", "//pkg/syserror", "//pkg/usermem", "//pkg/waiter", + "@org_golang_x_sys//unix:go_default_library", ], ) diff --git a/pkg/sentry/vfs/anonfs.go b/pkg/sentry/vfs/anonfs.go index 2db25be49..a64d86122 100644 --- a/pkg/sentry/vfs/anonfs.go +++ b/pkg/sentry/vfs/anonfs.go @@ -21,6 +21,7 @@ import ( "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/fspath" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" ) @@ -41,7 +42,27 @@ func (vfs *VirtualFilesystem) NewAnonVirtualDentry(name string) VirtualDentry { } } -const anonfsBlockSize = usermem.PageSize // via fs/libfs.c:pseudo_fs_fill_super() +const ( + anonfsBlockSize = usermem.PageSize // via fs/libfs.c:pseudo_fs_fill_super() + + // Mode, UID, and GID for a generic anonfs file. + anonFileMode = 0600 // no type is correct + anonFileUID = auth.RootKUID + anonFileGID = auth.RootKGID +) + +// anonFilesystemType implements FilesystemType. +type anonFilesystemType struct{} + +// GetFilesystem implements FilesystemType.GetFilesystem. +func (anonFilesystemType) GetFilesystem(context.Context, *VirtualFilesystem, *auth.Credentials, string, GetFilesystemOptions) (*Filesystem, *Dentry, error) { + panic("cannot instaniate an anon filesystem") +} + +// Name implemenents FilesystemType.Name. +func (anonFilesystemType) Name() string { + return "none" +} // anonFilesystem is the implementation of FilesystemImpl that backs // VirtualDentries returned by VirtualFilesystem.NewAnonVirtualDentry(). @@ -69,6 +90,16 @@ func (fs *anonFilesystem) Sync(ctx context.Context) error { return nil } +// AccessAt implements vfs.Filesystem.Impl.AccessAt. +// +// TODO(gvisor.dev/issue/1965): Implement access permissions. +func (fs *anonFilesystem) AccessAt(ctx context.Context, rp *ResolvingPath, creds *auth.Credentials, ats AccessTypes) error { + if !rp.Done() { + return syserror.ENOTDIR + } + return GenericCheckPermissions(creds, ats, anonFileMode, anonFileUID, anonFileGID) +} + // GetDentryAt implements FilesystemImpl.GetDentryAt. func (fs *anonFilesystem) GetDentryAt(ctx context.Context, rp *ResolvingPath, opts GetDentryOptions) (*Dentry, error) { if !rp.Done() { @@ -167,9 +198,9 @@ func (fs *anonFilesystem) StatAt(ctx context.Context, rp *ResolvingPath, opts St Mask: linux.STATX_TYPE | linux.STATX_MODE | linux.STATX_NLINK | linux.STATX_UID | linux.STATX_GID | linux.STATX_INO | linux.STATX_SIZE | linux.STATX_BLOCKS, Blksize: anonfsBlockSize, Nlink: 1, - UID: uint32(auth.RootKUID), - GID: uint32(auth.RootKGID), - Mode: 0600, // no type is correct + UID: uint32(anonFileUID), + GID: uint32(anonFileGID), + Mode: anonFileMode, Ino: 1, Size: 0, Blocks: 0, @@ -205,8 +236,16 @@ func (fs *anonFilesystem) UnlinkAt(ctx context.Context, rp *ResolvingPath) error return syserror.EPERM } +// BoundEndpointAt implements FilesystemImpl.BoundEndpointAt. +func (fs *anonFilesystem) BoundEndpointAt(ctx context.Context, rp *ResolvingPath) (transport.BoundEndpoint, error) { + if !rp.Final() { + return nil, syserror.ENOTDIR + } + return nil, syserror.ECONNREFUSED +} + // ListxattrAt implements FilesystemImpl.ListxattrAt. -func (fs *anonFilesystem) ListxattrAt(ctx context.Context, rp *ResolvingPath) ([]string, error) { +func (fs *anonFilesystem) ListxattrAt(ctx context.Context, rp *ResolvingPath, size uint64) ([]string, error) { if !rp.Done() { return nil, syserror.ENOTDIR } @@ -214,7 +253,7 @@ func (fs *anonFilesystem) ListxattrAt(ctx context.Context, rp *ResolvingPath) ([ } // GetxattrAt implements FilesystemImpl.GetxattrAt. -func (fs *anonFilesystem) GetxattrAt(ctx context.Context, rp *ResolvingPath, name string) (string, error) { +func (fs *anonFilesystem) GetxattrAt(ctx context.Context, rp *ResolvingPath, opts GetxattrOptions) (string, error) { if !rp.Done() { return "", syserror.ENOTDIR } diff --git a/pkg/sentry/vfs/context.go b/pkg/sentry/vfs/context.go index d97362b9a..82781e6d3 100644 --- a/pkg/sentry/vfs/context.go +++ b/pkg/sentry/vfs/context.go @@ -29,9 +29,10 @@ const ( CtxRoot ) -// MountNamespaceFromContext returns the MountNamespace used by ctx. It does -// not take a reference on the returned MountNamespace. If ctx is not -// associated with a MountNamespace, MountNamespaceFromContext returns nil. +// MountNamespaceFromContext returns the MountNamespace used by ctx. If ctx is +// not associated with a MountNamespace, MountNamespaceFromContext returns nil. +// +// A reference is taken on the returned MountNamespace. func MountNamespaceFromContext(ctx context.Context) *MountNamespace { if v := ctx.Value(CtxMountNamespace); v != nil { return v.(*MountNamespace) diff --git a/pkg/sentry/vfs/dentry.go b/pkg/sentry/vfs/dentry.go index 486a76475..8624dbd5d 100644 --- a/pkg/sentry/vfs/dentry.go +++ b/pkg/sentry/vfs/dentry.go @@ -15,34 +15,17 @@ package vfs import ( - "fmt" "sync/atomic" "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/syserror" ) -// Dentry represents a node in a Filesystem tree which may represent a file. +// Dentry represents a node in a Filesystem tree at which a file exists. // // Dentries are reference-counted. Unless otherwise specified, all Dentry // methods require that a reference is held. // -// A Dentry transitions through up to 3 different states through its lifetime: -// -// - Dentries are initially "independent". Independent Dentries have no parent, -// and consequently no name. -// -// - Dentry.InsertChild() causes an independent Dentry to become a "child" of -// another Dentry. A child node has a parent node, and a name in that parent, -// both of which are mutable by DentryMoveChild(). Each child Dentry's name is -// unique within its parent. -// -// - Dentry.RemoveChild() causes a child Dentry to become "disowned". A -// disowned Dentry can still refer to its former parent and its former name in -// said parent, but the disowned Dentry is no longer reachable from its parent, -// and a new Dentry with the same name may become a child of the parent. (This -// is analogous to a struct dentry being "unhashed" in Linux.) -// // Dentry is loosely analogous to Linux's struct dentry, but: // // - VFS does not associate Dentries with inodes. gVisor interacts primarily @@ -57,9 +40,6 @@ import ( // and/or FileDescriptionImpl methods in gVisor's VFS. Filesystems that do // support inodes may store appropriate state in implementations of DentryImpl. // -// - VFS does not provide synchronization for mutable Dentry fields, other than -// mount-related ones. -// // - VFS does not require that Dentries are instantiated for all paths accessed // through VFS, only those that are tracked beyond the scope of a single // Filesystem operation. This includes file descriptions, mount points, mount @@ -67,41 +47,34 @@ import ( // of Dentries for operations on mutable remote filesystems that can't actually // cache any state in the Dentry. // +// - VFS does not track filesystem structure (i.e. relationships between +// Dentries), since both the relevant state and synchronization are +// filesystem-specific. +// // - For the reasons above, VFS is not directly responsible for managing Dentry // lifetime. Dentry reference counts only indicate the extent to which VFS // requires Dentries to exist; Filesystems may elect to cache or discard // Dentries with zero references. +// +// +stateify savable type Dentry struct { - // parent is this Dentry's parent in this Filesystem. If this Dentry is - // independent, parent is nil. - parent *Dentry + // mu synchronizes deletion/invalidation and mounting over this Dentry. + mu sync.Mutex `state:"nosave"` - // name is this Dentry's name in parent. - name string - - flags uint32 + // dead is true if the file represented by this Dentry has been deleted (by + // CommitDeleteDentry or CommitRenameReplaceDentry) or invalidated (by + // InvalidateDentry). dead is protected by mu. + dead bool // mounts is the number of Mounts for which this Dentry is Mount.point. // mounts is accessed using atomic memory operations. mounts uint32 - // children are child Dentries. - children map[string]*Dentry - - // mu synchronizes disowning and mounting over this Dentry. - mu sync.Mutex - // impl is the DentryImpl associated with this Dentry. impl is immutable. // This should be the last field in Dentry. impl DentryImpl } -const ( - // dflagsDisownedMask is set in Dentry.flags if the Dentry has been - // disowned. - dflagsDisownedMask = 1 << iota -) - // Init must be called before first use of d. func (d *Dentry) Init(impl DentryImpl) { d.impl = impl @@ -132,20 +105,6 @@ type DentryImpl interface { DecRef() } -// IsDisowned returns true if d is disowned. -func (d *Dentry) IsDisowned() bool { - return atomic.LoadUint32(&d.flags)&dflagsDisownedMask != 0 -} - -// Preconditions: !d.IsDisowned(). -func (d *Dentry) setDisowned() { - atomic.AddUint32(&d.flags, dflagsDisownedMask) -} - -func (d *Dentry) isMounted() bool { - return atomic.LoadUint32(&d.mounts) != 0 -} - // IncRef increments d's reference count. func (d *Dentry) IncRef() { d.impl.IncRef() @@ -162,104 +121,26 @@ func (d *Dentry) DecRef() { d.impl.DecRef() } -// These functions are exported so that filesystem implementations can use -// them. The vfs package, and users of VFS, should not call these functions. -// Unless otherwise specified, these methods require that there are no -// concurrent mutators of d. - -// Name returns d's name in its parent in its owning Filesystem. If d is -// independent, Name returns an empty string. -func (d *Dentry) Name() string { - return d.name -} - -// Parent returns d's parent in its owning Filesystem. It does not take a -// reference on the returned Dentry. If d is independent, Parent returns nil. -func (d *Dentry) Parent() *Dentry { - return d.parent -} - -// ParentOrSelf is equivalent to Parent, but returns d if d is independent. -func (d *Dentry) ParentOrSelf() *Dentry { - if d.parent == nil { - return d - } - return d.parent -} - -// Child returns d's child with the given name in its owning Filesystem. It -// does not take a reference on the returned Dentry. If no such child exists, -// Child returns nil. -func (d *Dentry) Child(name string) *Dentry { - return d.children[name] -} - -// HasChildren returns true if d has any children. -func (d *Dentry) HasChildren() bool { - return len(d.children) != 0 -} - -// Children returns a map containing all of d's children. -func (d *Dentry) Children() map[string]*Dentry { - if !d.HasChildren() { - return nil - } - m := make(map[string]*Dentry) - for name, child := range d.children { - m[name] = child - } - return m +// IsDead returns true if d has been deleted or invalidated by its owning +// filesystem. +func (d *Dentry) IsDead() bool { + d.mu.Lock() + defer d.mu.Unlock() + return d.dead } -// InsertChild makes child a child of d with the given name. -// -// InsertChild is a mutator of d and child. -// -// Preconditions: child must be an independent Dentry. d and child must be from -// the same Filesystem. d must not already have a child with the given name. -func (d *Dentry) InsertChild(child *Dentry, name string) { - if checkInvariants { - if _, ok := d.children[name]; ok { - panic(fmt.Sprintf("parent already contains a child named %q", name)) - } - if child.parent != nil || child.name != "" { - panic(fmt.Sprintf("child is not independent: parent = %v, name = %q", child.parent, child.name)) - } - } - if d.children == nil { - d.children = make(map[string]*Dentry) - } - d.children[name] = child - child.parent = d - child.name = name +func (d *Dentry) isMounted() bool { + return atomic.LoadUint32(&d.mounts) != 0 } -// IsAncestorOf returns true if d is an ancestor of d2; that is, d is either -// d2's parent or an ancestor of d2's parent. -func (d *Dentry) IsAncestorOf(d2 *Dentry) bool { - for d2.parent != nil { - if d2.parent == d { - return true - } - d2 = d2.parent - } - return false -} +// The following functions are exported so that filesystem implementations can +// use them. The vfs package, and users of VFS, should not call these +// functions. // PrepareDeleteDentry must be called before attempting to delete the file // represented by d. If PrepareDeleteDentry succeeds, the caller must call // AbortDeleteDentry or CommitDeleteDentry depending on the deletion's outcome. -// -// Preconditions: d is a child Dentry. func (vfs *VirtualFilesystem) PrepareDeleteDentry(mntns *MountNamespace, d *Dentry) error { - if checkInvariants { - if d.parent == nil { - panic("d is independent") - } - if d.IsDisowned() { - panic("d is already disowned") - } - } vfs.mountMu.Lock() if mntns.mountpoints[d] != 0 { vfs.mountMu.Unlock() @@ -278,42 +159,27 @@ func (vfs *VirtualFilesystem) AbortDeleteDentry(d *Dentry) { d.mu.Unlock() } -// CommitDeleteDentry must be called after the file represented by d is -// deleted, and causes d to become disowned. -// -// CommitDeleteDentry is a mutator of d and d.Parent(). -// -// Preconditions: PrepareDeleteDentry was previously called on d. +// CommitDeleteDentry must be called after PrepareDeleteDentry if the deletion +// succeeds. func (vfs *VirtualFilesystem) CommitDeleteDentry(d *Dentry) { - if d.parent != nil { - delete(d.parent.children, d.name) - } - d.setDisowned() + d.dead = true d.mu.Unlock() if d.isMounted() { - vfs.forgetDisownedMountpoint(d) + vfs.forgetDeadMountpoint(d) } } -// ForceDeleteDentry causes d to become disowned. It should only be used in -// cases where VFS has no ability to stop the deletion (e.g. d represents the -// local state of a file on a remote filesystem on which the file has already -// been deleted). -// -// ForceDeleteDentry is a mutator of d and d.Parent(). -// -// Preconditions: d is a child Dentry. -func (vfs *VirtualFilesystem) ForceDeleteDentry(d *Dentry) { - if checkInvariants { - if d.parent == nil { - panic("d is independent") - } - if d.IsDisowned() { - panic("d is already disowned") - } - } +// InvalidateDentry is called when d ceases to represent the file it formerly +// did for reasons outside of VFS' control (e.g. d represents the local state +// of a file on a remote filesystem on which the file has already been +// deleted). +func (vfs *VirtualFilesystem) InvalidateDentry(d *Dentry) { d.mu.Lock() - vfs.CommitDeleteDentry(d) + d.dead = true + d.mu.Unlock() + if d.isMounted() { + vfs.forgetDeadMountpoint(d) + } } // PrepareRenameDentry must be called before attempting to rename the file @@ -322,25 +188,9 @@ func (vfs *VirtualFilesystem) ForceDeleteDentry(d *Dentry) { // caller must call AbortRenameDentry, CommitRenameReplaceDentry, or // CommitRenameExchangeDentry depending on the rename's outcome. // -// Preconditions: from is a child Dentry. If to is not nil, it must be a child -// Dentry from the same Filesystem. from != to. +// Preconditions: If to is not nil, it must be a child Dentry from the same +// Filesystem. from != to. func (vfs *VirtualFilesystem) PrepareRenameDentry(mntns *MountNamespace, from, to *Dentry) error { - if checkInvariants { - if from.parent == nil { - panic("from is independent") - } - if from.IsDisowned() { - panic("from is already disowned") - } - if to != nil { - if to.parent == nil { - panic("to is independent") - } - if to.IsDisowned() { - panic("to is already disowned") - } - } - } vfs.mountMu.Lock() if mntns.mountpoints[from] != 0 { vfs.mountMu.Unlock() @@ -374,24 +224,14 @@ func (vfs *VirtualFilesystem) AbortRenameDentry(from, to *Dentry) { // is renamed without RENAME_EXCHANGE. If to is not nil, it represents the file // that was replaced by from. // -// CommitRenameReplaceDentry is a mutator of from, to, from.Parent(), and -// to.Parent(). -// // Preconditions: PrepareRenameDentry was previously called on from and to. -// newParent.Child(newName) == to. -func (vfs *VirtualFilesystem) CommitRenameReplaceDentry(from, newParent *Dentry, newName string, to *Dentry) { - if newParent.children == nil { - newParent.children = make(map[string]*Dentry) - } - newParent.children[newName] = from - from.parent = newParent - from.name = newName +func (vfs *VirtualFilesystem) CommitRenameReplaceDentry(from, to *Dentry) { from.mu.Unlock() if to != nil { - to.setDisowned() + to.dead = true to.mu.Unlock() if to.isMounted() { - vfs.forgetDisownedMountpoint(to) + vfs.forgetDeadMountpoint(to) } } } @@ -399,25 +239,18 @@ func (vfs *VirtualFilesystem) CommitRenameReplaceDentry(from, newParent *Dentry, // CommitRenameExchangeDentry must be called after the files represented by // from and to are exchanged by rename(RENAME_EXCHANGE). // -// CommitRenameExchangeDentry is a mutator of from, to, from.Parent(), and -// to.Parent(). -// // Preconditions: PrepareRenameDentry was previously called on from and to. func (vfs *VirtualFilesystem) CommitRenameExchangeDentry(from, to *Dentry) { - from.parent, to.parent = to.parent, from.parent - from.name, to.name = to.name, from.name - from.parent.children[from.name] = from - to.parent.children[to.name] = to from.mu.Unlock() to.mu.Unlock() } -// forgetDisownedMountpoint is called when a mount point is deleted to umount -// all mounts using it in all other mount namespaces. +// forgetDeadMountpoint is called when a mount point is deleted or invalidated +// to umount all mounts using it in all other mount namespaces. // -// forgetDisownedMountpoint is analogous to Linux's +// forgetDeadMountpoint is analogous to Linux's // fs/namespace.c:__detach_mounts(). -func (vfs *VirtualFilesystem) forgetDisownedMountpoint(d *Dentry) { +func (vfs *VirtualFilesystem) forgetDeadMountpoint(d *Dentry) { var ( vdsToDecRef []VirtualDentry mountsToDecRef []*Mount diff --git a/pkg/sentry/vfs/device.go b/pkg/sentry/vfs/device.go index 3af2aa58d..bda5576fa 100644 --- a/pkg/sentry/vfs/device.go +++ b/pkg/sentry/vfs/device.go @@ -56,6 +56,7 @@ type Device interface { Open(ctx context.Context, mnt *Mount, d *Dentry, opts OpenOptions) (*FileDescription, error) } +// +stateify savable type registeredDevice struct { dev Device opts RegisterDeviceOptions @@ -63,6 +64,8 @@ type registeredDevice struct { // RegisterDeviceOptions contains options to // VirtualFilesystem.RegisterDevice(). +// +// +stateify savable type RegisterDeviceOptions struct { // GroupName is the name shown for this device registration in // /proc/devices. If GroupName is empty, this registration will not be diff --git a/pkg/sentry/vfs/epoll.go b/pkg/sentry/vfs/epoll.go index 7c83f9a5a..8e0b40841 100644 --- a/pkg/sentry/vfs/epoll.go +++ b/pkg/sentry/vfs/epoll.go @@ -85,8 +85,8 @@ type epollInterest struct { ready bool epollInterestEntry - // userData is the epoll_data_t associated with this epollInterest. - // userData is protected by epoll.mu. + // userData is the struct epoll_event::data associated with this + // epollInterest. userData is protected by epoll.mu. userData [2]int32 } @@ -99,6 +99,8 @@ func (vfs *VirtualFilesystem) NewEpollInstanceFD() (*FileDescription, error) { interest: make(map[epollInterestKey]*epollInterest), } if err := ep.vfsfd.Init(ep, linux.O_RDWR, vd.Mount(), vd.Dentry(), &FileDescriptionOptions{ + DenyPRead: true, + DenyPWrite: true, UseDentryMetadata: true, }); err != nil { return nil, err @@ -157,7 +159,7 @@ func (ep *EpollInstance) Seek(ctx context.Context, offset int64, whence int32) ( // AddInterest implements the semantics of EPOLL_CTL_ADD. // // Preconditions: A reference must be held on file. -func (ep *EpollInstance) AddInterest(file *FileDescription, num int32, mask uint32, userData [2]int32) error { +func (ep *EpollInstance) AddInterest(file *FileDescription, num int32, event linux.EpollEvent) error { // Check for cyclic polling if necessary. subep, _ := file.impl.(*EpollInstance) if subep != nil { @@ -183,12 +185,12 @@ func (ep *EpollInstance) AddInterest(file *FileDescription, num int32, mask uint } // Register interest in file. - mask |= linux.EPOLLERR | linux.EPOLLRDHUP + mask := event.Events | linux.EPOLLERR | linux.EPOLLRDHUP epi := &epollInterest{ epoll: ep, key: key, mask: mask, - userData: userData, + userData: event.Data, } ep.interest[key] = epi wmask := waiter.EventMaskFromLinux(mask) @@ -202,6 +204,9 @@ func (ep *EpollInstance) AddInterest(file *FileDescription, num int32, mask uint // Add epi to file.epolls so that it is removed when the last // FileDescription reference is dropped. file.epollMu.Lock() + if file.epolls == nil { + file.epolls = make(map[*epollInterest]struct{}) + } file.epolls[epi] = struct{}{} file.epollMu.Unlock() @@ -236,7 +241,7 @@ func (ep *EpollInstance) mightPollRecursive(ep2 *EpollInstance, remainingRecursi // ModifyInterest implements the semantics of EPOLL_CTL_MOD. // // Preconditions: A reference must be held on file. -func (ep *EpollInstance) ModifyInterest(file *FileDescription, num int32, mask uint32, userData [2]int32) error { +func (ep *EpollInstance) ModifyInterest(file *FileDescription, num int32, event linux.EpollEvent) error { ep.interestMu.Lock() defer ep.interestMu.Unlock() @@ -250,13 +255,13 @@ func (ep *EpollInstance) ModifyInterest(file *FileDescription, num int32, mask u } // Update epi for the next call to ep.ReadEvents(). + mask := event.Events | linux.EPOLLERR | linux.EPOLLRDHUP ep.mu.Lock() epi.mask = mask - epi.userData = userData + epi.userData = event.Data ep.mu.Unlock() // Re-register with the new mask. - mask |= linux.EPOLLERR | linux.EPOLLRDHUP file.EventUnregister(&epi.waiter) wmask := waiter.EventMaskFromLinux(mask) file.EventRegister(&epi.waiter, wmask) @@ -363,8 +368,7 @@ func (ep *EpollInstance) ReadEvents(events []linux.EpollEvent) int { // Report ievents. events[i] = linux.EpollEvent{ Events: ievents.ToLinux(), - Fd: epi.userData[0], - Data: epi.userData[1], + Data: epi.userData, } i++ if i == len(events) { diff --git a/pkg/sentry/vfs/file_description.go b/pkg/sentry/vfs/file_description.go index 5bac660c7..15cc091e2 100644 --- a/pkg/sentry/vfs/file_description.go +++ b/pkg/sentry/vfs/file_description.go @@ -84,6 +84,17 @@ type FileDescriptionOptions struct { // usually only the case if O_DIRECT would actually have an effect. AllowDirectIO bool + // If DenyPRead is true, calls to FileDescription.PRead() return ESPIPE. + DenyPRead bool + + // If DenyPWrite is true, calls to FileDescription.PWrite() return + // ESPIPE. + DenyPWrite bool + + // if InvalidWrite is true, calls to FileDescription.Write() return + // EINVAL. + InvalidWrite bool + // If UseDentryMetadata is true, calls to FileDescription methods that // interact with file and filesystem metadata (Stat, SetStat, StatFS, // Listxattr, Getxattr, Setxattr, Removexattr) are implemented by calling @@ -111,12 +122,13 @@ func (fd *FileDescription) Init(impl FileDescriptionImpl, statusFlags uint32, mn } fd.refs = 1 - fd.statusFlags = statusFlags | linux.O_LARGEFILE + fd.statusFlags = statusFlags fd.vd = VirtualDentry{ mount: mnt, dentry: d, } - fd.vd.IncRef() + mnt.IncRef() + d.IncRef() fd.opts = *opts fd.readable = MayReadFileWithOpenFlags(statusFlags) fd.writable = writable @@ -175,6 +187,12 @@ func (fd *FileDescription) DecRef() { } } +// Refs returns the current number of references. The returned count +// is inherently racy and is unsafe to use without external synchronization. +func (fd *FileDescription) Refs() int64 { + return atomic.LoadInt64(&fd.refs) +} + // Mount returns the mount on which fd was opened. It does not take a reference // on the returned Mount. func (fd *FileDescription) Mount() *Mount { @@ -286,7 +304,8 @@ type FileDescriptionImpl interface { Stat(ctx context.Context, opts StatOptions) (linux.Statx, error) // SetStat updates metadata for the file represented by the - // FileDescription. + // FileDescription. Implementations are responsible for checking if the + // operation can be performed (see vfs.CheckSetStat() for common checks). SetStat(ctx context.Context, opts SetStatOptions) error // StatFS returns metadata for the filesystem containing the file @@ -305,6 +324,7 @@ type FileDescriptionImpl interface { // - If opts.Flags specifies unsupported options, PRead returns EOPNOTSUPP. // // Preconditions: The FileDescription was opened for reading. + // FileDescriptionOptions.DenyPRead == false. PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts ReadOptions) (int64, error) // Read is similar to PRead, but does not specify an offset. @@ -336,6 +356,7 @@ type FileDescriptionImpl interface { // EOPNOTSUPP. // // Preconditions: The FileDescription was opened for writing. + // FileDescriptionOptions.DenyPWrite == false. PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts WriteOptions) (int64, error) // Write is similar to PWrite, but does not specify an offset, which is @@ -381,11 +402,11 @@ type FileDescriptionImpl interface { Ioctl(ctx context.Context, uio usermem.IO, args arch.SyscallArguments) (uintptr, error) // Listxattr returns all extended attribute names for the file. - Listxattr(ctx context.Context) ([]string, error) + Listxattr(ctx context.Context, size uint64) ([]string, error) // Getxattr returns the value associated with the given extended attribute // for the file. - Getxattr(ctx context.Context, name string) (string, error) + Getxattr(ctx context.Context, opts GetxattrOptions) (string, error) // Setxattr changes the value associated with the given extended attribute // for the file. @@ -435,11 +456,11 @@ type Dirent struct { // IterDirentsCallback receives Dirents from FileDescriptionImpl.IterDirents. type IterDirentsCallback interface { - // Handle handles the given iterated Dirent. It returns true if iteration - // should continue, and false if FileDescriptionImpl.IterDirents should - // terminate now and restart with the same Dirent the next time it is - // called. - Handle(dirent Dirent) bool + // Handle handles the given iterated Dirent. If Handle returns a non-nil + // error, FileDescriptionImpl.IterDirents must stop iteration and return + // the error; the next call to FileDescriptionImpl.IterDirents should + // restart with the same Dirent. + Handle(dirent Dirent) error } // OnClose is called when a file descriptor representing the FileDescription is @@ -514,6 +535,9 @@ func (fd *FileDescription) EventUnregister(e *waiter.Entry) { // offset, and returns the number of bytes read. PRead is permitted to return // partial reads with a nil error. func (fd *FileDescription) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts ReadOptions) (int64, error) { + if fd.opts.DenyPRead { + return 0, syserror.ESPIPE + } if !fd.readable { return 0, syserror.EBADF } @@ -532,6 +556,9 @@ func (fd *FileDescription) Read(ctx context.Context, dst usermem.IOSequence, opt // offset, and returns the number of bytes written. PWrite is permitted to // return partial writes with a nil error. func (fd *FileDescription) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts WriteOptions) (int64, error) { + if fd.opts.DenyPWrite { + return 0, syserror.ESPIPE + } if !fd.writable { return 0, syserror.EBADF } @@ -540,6 +567,9 @@ func (fd *FileDescription) PWrite(ctx context.Context, src usermem.IOSequence, o // Write is similar to PWrite, but does not specify an offset. func (fd *FileDescription) Write(ctx context.Context, src usermem.IOSequence, opts WriteOptions) (int64, error) { + if fd.opts.InvalidWrite { + return 0, syserror.EINVAL + } if !fd.writable { return 0, syserror.EBADF } @@ -576,18 +606,23 @@ func (fd *FileDescription) Ioctl(ctx context.Context, uio usermem.IO, args arch. // Listxattr returns all extended attribute names for the file represented by // fd. -func (fd *FileDescription) Listxattr(ctx context.Context) ([]string, error) { +// +// If the size of the list (including a NUL terminating byte after every entry) +// would exceed size, ERANGE may be returned. Note that implementations +// are free to ignore size entirely and return without error). In all cases, +// if size is 0, the list should be returned without error, regardless of size. +func (fd *FileDescription) Listxattr(ctx context.Context, size uint64) ([]string, error) { if fd.opts.UseDentryMetadata { vfsObj := fd.vd.mount.vfs rp := vfsObj.getResolvingPath(auth.CredentialsFromContext(ctx), &PathOperation{ Root: fd.vd, Start: fd.vd, }) - names, err := fd.vd.mount.fs.impl.ListxattrAt(ctx, rp) + names, err := fd.vd.mount.fs.impl.ListxattrAt(ctx, rp, size) vfsObj.putResolvingPath(rp) return names, err } - names, err := fd.impl.Listxattr(ctx) + names, err := fd.impl.Listxattr(ctx, size) if err == syserror.ENOTSUP { // Linux doesn't actually return ENOTSUP in this case; instead, // fs/xattr.c:vfs_listxattr() falls back to allowing the security @@ -600,34 +635,39 @@ func (fd *FileDescription) Listxattr(ctx context.Context) ([]string, error) { // Getxattr returns the value associated with the given extended attribute for // the file represented by fd. -func (fd *FileDescription) Getxattr(ctx context.Context, name string) (string, error) { +// +// If the size of the return value exceeds opts.Size, ERANGE may be returned +// (note that implementations are free to ignore opts.Size entirely and return +// without error). In all cases, if opts.Size is 0, the value should be +// returned without error, regardless of size. +func (fd *FileDescription) Getxattr(ctx context.Context, opts *GetxattrOptions) (string, error) { if fd.opts.UseDentryMetadata { vfsObj := fd.vd.mount.vfs rp := vfsObj.getResolvingPath(auth.CredentialsFromContext(ctx), &PathOperation{ Root: fd.vd, Start: fd.vd, }) - val, err := fd.vd.mount.fs.impl.GetxattrAt(ctx, rp, name) + val, err := fd.vd.mount.fs.impl.GetxattrAt(ctx, rp, *opts) vfsObj.putResolvingPath(rp) return val, err } - return fd.impl.Getxattr(ctx, name) + return fd.impl.Getxattr(ctx, *opts) } // Setxattr changes the value associated with the given extended attribute for // the file represented by fd. -func (fd *FileDescription) Setxattr(ctx context.Context, opts SetxattrOptions) error { +func (fd *FileDescription) Setxattr(ctx context.Context, opts *SetxattrOptions) error { if fd.opts.UseDentryMetadata { vfsObj := fd.vd.mount.vfs rp := vfsObj.getResolvingPath(auth.CredentialsFromContext(ctx), &PathOperation{ Root: fd.vd, Start: fd.vd, }) - err := fd.vd.mount.fs.impl.SetxattrAt(ctx, rp, opts) + err := fd.vd.mount.fs.impl.SetxattrAt(ctx, rp, *opts) vfsObj.putResolvingPath(rp) return err } - return fd.impl.Setxattr(ctx, opts) + return fd.impl.Setxattr(ctx, *opts) } // Removexattr removes the given extended attribute from the file represented diff --git a/pkg/sentry/vfs/file_description_impl_util.go b/pkg/sentry/vfs/file_description_impl_util.go index c2a52ec1b..f4c111926 100644 --- a/pkg/sentry/vfs/file_description_impl_util.go +++ b/pkg/sentry/vfs/file_description_impl_util.go @@ -33,8 +33,8 @@ import ( // implementations to adapt: // - Have a local fileDescription struct (containing FileDescription) which // embeds FileDescriptionDefaultImpl and overrides the default methods -// which are common to all fd implementations for that for that filesystem -// like StatusFlags, SetStatusFlags, Stat, SetStat, StatFS, etc. +// which are common to all fd implementations for that filesystem like +// StatusFlags, SetStatusFlags, Stat, SetStat, StatFS, etc. // - This should be embedded in all file description implementations as the // first field by value. // - Directory FDs would also embed DirectoryFileDescriptionDefaultImpl. @@ -130,14 +130,14 @@ func (FileDescriptionDefaultImpl) Ioctl(ctx context.Context, uio usermem.IO, arg // Listxattr implements FileDescriptionImpl.Listxattr analogously to // inode_operations::listxattr == NULL in Linux. -func (FileDescriptionDefaultImpl) Listxattr(ctx context.Context) ([]string, error) { +func (FileDescriptionDefaultImpl) Listxattr(ctx context.Context, size uint64) ([]string, error) { // This isn't exactly accurate; see FileDescription.Listxattr. return nil, syserror.ENOTSUP } // Getxattr implements FileDescriptionImpl.Getxattr analogously to // inode::i_opflags & IOP_XATTR == 0 in Linux. -func (FileDescriptionDefaultImpl) Getxattr(ctx context.Context, name string) (string, error) { +func (FileDescriptionDefaultImpl) Getxattr(ctx context.Context, opts GetxattrOptions) (string, error) { return "", syserror.ENOTSUP } @@ -339,6 +339,11 @@ func (fd *DynamicBytesFileDescriptionImpl) pwriteLocked(ctx context.Context, src if opts.Flags&^(linux.RWF_HIPRI|linux.RWF_DSYNC|linux.RWF_SYNC) != 0 { return 0, syserror.EOPNOTSUPP } + limit, err := CheckLimit(ctx, offset, src.NumBytes()) + if err != nil { + return 0, err + } + src = src.TakeFirst64(limit) writable, ok := fd.data.(WritableDynamicBytesSource) if !ok { diff --git a/pkg/sentry/vfs/file_description_impl_util_test.go b/pkg/sentry/vfs/file_description_impl_util_test.go index 8fa26418e..3a75d4d62 100644 --- a/pkg/sentry/vfs/file_description_impl_util_test.go +++ b/pkg/sentry/vfs/file_description_impl_util_test.go @@ -107,7 +107,10 @@ func (fd *testFD) SetStat(ctx context.Context, opts SetStatOptions) error { func TestGenCountFD(t *testing.T) { ctx := contexttest.Context(t) - vfsObj := New() // vfs.New() + vfsObj := &VirtualFilesystem{} + if err := vfsObj.Init(); err != nil { + t.Fatalf("VFS init: %v", err) + } fd := newTestFD(vfsObj, linux.O_RDWR, &genCount{}) defer fd.DecRef() @@ -162,7 +165,10 @@ func TestGenCountFD(t *testing.T) { func TestWritable(t *testing.T) { ctx := contexttest.Context(t) - vfsObj := New() // vfs.New() + vfsObj := &VirtualFilesystem{} + if err := vfsObj.Init(); err != nil { + t.Fatalf("VFS init: %v", err) + } fd := newTestFD(vfsObj, linux.O_RDWR, &storeData{data: "init"}) defer fd.DecRef() diff --git a/pkg/sentry/vfs/filesystem.go b/pkg/sentry/vfs/filesystem.go index a06a6caf3..74577bc2f 100644 --- a/pkg/sentry/vfs/filesystem.go +++ b/pkg/sentry/vfs/filesystem.go @@ -20,6 +20,8 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" ) // A Filesystem is a tree of nodes represented by Dentries, which forms part of @@ -29,6 +31,8 @@ import ( // Filesystem methods require that a reference is held. // // Filesystem is analogous to Linux's struct super_block. +// +// +stateify savable type Filesystem struct { // refs is the reference count. refs is accessed using atomic memory // operations. @@ -38,21 +42,30 @@ type Filesystem struct { // immutable. vfs *VirtualFilesystem + // fsType is the FilesystemType of this Filesystem. + fsType FilesystemType + // impl is the FilesystemImpl associated with this Filesystem. impl is // immutable. This should be the last field in Dentry. impl FilesystemImpl } // Init must be called before first use of fs. -func (fs *Filesystem) Init(vfsObj *VirtualFilesystem, impl FilesystemImpl) { +func (fs *Filesystem) Init(vfsObj *VirtualFilesystem, fsType FilesystemType, impl FilesystemImpl) { fs.refs = 1 fs.vfs = vfsObj + fs.fsType = fsType fs.impl = impl vfsObj.filesystemsMu.Lock() vfsObj.filesystems[fs] = struct{}{} vfsObj.filesystemsMu.Unlock() } +// FilesystemType returns the FilesystemType for this Filesystem. +func (fs *Filesystem) FilesystemType() FilesystemType { + return fs.fsType +} + // VirtualFilesystem returns the containing VirtualFilesystem. func (fs *Filesystem) VirtualFilesystem() *VirtualFilesystem { return fs.vfs @@ -142,6 +155,9 @@ type FilesystemImpl interface { // file data to be written to the underlying [filesystem]", as by syncfs(2). Sync(ctx context.Context) error + // AccessAt checks whether a user with creds can access the file at rp. + AccessAt(ctx context.Context, rp *ResolvingPath, creds *auth.Credentials, ats AccessTypes) error + // GetDentryAt returns a Dentry representing the file at rp. A reference is // taken on the returned Dentry. // @@ -330,7 +346,10 @@ type FilesystemImpl interface { // ENOTEMPTY. // // Preconditions: !rp.Done(). For the final path component in rp, - // !rp.ShouldFollowSymlink(). oldName is not "." or "..". + // !rp.ShouldFollowSymlink(). oldParentVD.Dentry() was obtained from a + // previous call to + // oldParentVD.Mount().Filesystem().Impl().GetParentDentryAt(). oldName is + // not "." or "..". // // Postconditions: If RenameAt returns an error returned by // ResolvingPath.Resolve*(), then !rp.Done(). @@ -360,7 +379,9 @@ type FilesystemImpl interface { // ResolvingPath.Resolve*(), then !rp.Done(). RmdirAt(ctx context.Context, rp *ResolvingPath) error - // SetStatAt updates metadata for the file at the given path. + // SetStatAt updates metadata for the file at the given path. Implementations + // are responsible for checking if the operation can be performed + // (see vfs.CheckSetStat() for common checks). // // Errors: // @@ -424,7 +445,13 @@ type FilesystemImpl interface { // - If extended attributes are not supported by the filesystem, // ListxattrAt returns nil. (See FileDescription.Listxattr for an // explanation.) - ListxattrAt(ctx context.Context, rp *ResolvingPath) ([]string, error) + // + // - If the size of the list (including a NUL terminating byte after every + // entry) would exceed size, ERANGE may be returned. Note that + // implementations are free to ignore size entirely and return without + // error). In all cases, if size is 0, the list should be returned without + // error, regardless of size. + ListxattrAt(ctx context.Context, rp *ResolvingPath, size uint64) ([]string, error) // GetxattrAt returns the value associated with the given extended // attribute for the file at rp. @@ -433,7 +460,15 @@ type FilesystemImpl interface { // // - If extended attributes are not supported by the filesystem, GetxattrAt // returns ENOTSUP. - GetxattrAt(ctx context.Context, rp *ResolvingPath, name string) (string, error) + // + // - If an extended attribute named opts.Name does not exist, ENODATA is + // returned. + // + // - If the size of the return value exceeds opts.Size, ERANGE may be + // returned (note that implementations are free to ignore opts.Size entirely + // and return without error). In all cases, if opts.Size is 0, the value + // should be returned without error, regardless of size. + GetxattrAt(ctx context.Context, rp *ResolvingPath, opts GetxattrOptions) (string, error) // SetxattrAt changes the value associated with the given extended // attribute for the file at rp. @@ -442,6 +477,10 @@ type FilesystemImpl interface { // // - If extended attributes are not supported by the filesystem, SetxattrAt // returns ENOTSUP. + // + // - If XATTR_CREATE is set in opts.Flag and opts.Name already exists, + // EEXIST is returned. If XATTR_REPLACE is set and opts.Name does not exist, + // ENODATA is returned. SetxattrAt(ctx context.Context, rp *ResolvingPath, opts SetxattrOptions) error // RemovexattrAt removes the given extended attribute from the file at rp. @@ -450,8 +489,15 @@ type FilesystemImpl interface { // // - If extended attributes are not supported by the filesystem, // RemovexattrAt returns ENOTSUP. + // + // - If name does not exist, ENODATA is returned. RemovexattrAt(ctx context.Context, rp *ResolvingPath, name string) error + // BoundEndpointAt returns the Unix socket endpoint bound at the path rp. + // + // - If a non-socket file exists at rp, then BoundEndpointAt returns ECONNREFUSED. + BoundEndpointAt(ctx context.Context, rp *ResolvingPath) (transport.BoundEndpoint, error) + // PrependPath prepends a path from vd to vd.Mount().Root() to b. // // If vfsroot.Ok(), it is the contextual VFS root; if it is encountered @@ -474,7 +520,7 @@ type FilesystemImpl interface { // Preconditions: vd.Mount().Filesystem().Impl() == this FilesystemImpl. PrependPath(ctx context.Context, vfsroot, vd VirtualDentry, b *fspath.Builder) error - // TODO: inotify_add_watch(); bind() + // TODO(gvisor.dev/issue/1479): inotify_add_watch() } // PrependPathAtVFSRootError is returned by implementations of diff --git a/pkg/sentry/vfs/filesystem_impl_util.go b/pkg/sentry/vfs/filesystem_impl_util.go index 7315a588e..465e610e0 100644 --- a/pkg/sentry/vfs/filesystem_impl_util.go +++ b/pkg/sentry/vfs/filesystem_impl_util.go @@ -16,8 +16,6 @@ package vfs import ( "strings" - - "gvisor.dev/gvisor/pkg/fspath" ) // GenericParseMountOptions parses a comma-separated list of options of the @@ -43,27 +41,3 @@ func GenericParseMountOptions(str string) map[string]string { } return m } - -// GenericPrependPath may be used by implementations of -// FilesystemImpl.PrependPath() for which a single statically-determined lock -// or set of locks is sufficient to ensure its preconditions (as opposed to -// e.g. per-Dentry locks). -// -// Preconditions: Dentry.Name() and Dentry.Parent() must be held constant for -// vd.Dentry() and all of its ancestors. -func GenericPrependPath(vfsroot, vd VirtualDentry, b *fspath.Builder) error { - mnt, d := vd.mount, vd.dentry - for { - if mnt == vfsroot.mount && d == vfsroot.dentry { - return PrependPathAtVFSRootError{} - } - if d == mnt.root { - return nil - } - if d.parent == nil { - return PrependPathAtNonMountRootError{} - } - b.PrependComponent(d.name) - d = d.parent - } -} diff --git a/pkg/sentry/vfs/filesystem_type.go b/pkg/sentry/vfs/filesystem_type.go index c58b70728..f2298f7f6 100644 --- a/pkg/sentry/vfs/filesystem_type.go +++ b/pkg/sentry/vfs/filesystem_type.go @@ -30,6 +30,9 @@ type FilesystemType interface { // along with its mount root. A reference is taken on the returned // Filesystem and Dentry. GetFilesystem(ctx context.Context, vfsObj *VirtualFilesystem, creds *auth.Credentials, source string, opts GetFilesystemOptions) (*Filesystem, *Dentry, error) + + // Name returns the name of this FilesystemType. + Name() string } // GetFilesystemOptions contains options to FilesystemType.GetFilesystem. @@ -44,6 +47,7 @@ type GetFilesystemOptions struct { InternalData interface{} } +// +stateify savable type registeredFilesystemType struct { fsType FilesystemType opts RegisterFilesystemTypeOptions diff --git a/pkg/sentry/vfs/genericfstree/BUILD b/pkg/sentry/vfs/genericfstree/BUILD new file mode 100644 index 000000000..d8fd92677 --- /dev/null +++ b/pkg/sentry/vfs/genericfstree/BUILD @@ -0,0 +1,16 @@ +load("//tools/go_generics:defs.bzl", "go_template") + +package( + default_visibility = ["//:sandbox"], + licenses = ["notice"], +) + +go_template( + name = "generic_fstree", + srcs = [ + "genericfstree.go", + ], + types = [ + "Dentry", + ], +) diff --git a/pkg/sentry/vfs/genericfstree/genericfstree.go b/pkg/sentry/vfs/genericfstree/genericfstree.go new file mode 100644 index 000000000..286510195 --- /dev/null +++ b/pkg/sentry/vfs/genericfstree/genericfstree.go @@ -0,0 +1,80 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// Package genericfstree provides tools for implementing vfs.FilesystemImpls +// where a single statically-determined lock or set of locks is sufficient to +// ensure that a Dentry's name and parent are contextually immutable. +// +// Clients using this package must use the go_template_instance rule in +// tools/go_generics/defs.bzl to create an instantiation of this template +// package, providing types to use in place of Dentry. +package genericfstree + +import ( + "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/sentry/vfs" +) + +// Dentry is a required type parameter that is a struct with the given fields. +type Dentry struct { + // vfsd is the embedded vfs.Dentry corresponding to this vfs.DentryImpl. + vfsd vfs.Dentry + + // parent is the parent of this Dentry in the filesystem's tree. If this + // Dentry is a filesystem root, parent is nil. + parent *Dentry + + // name is the name of this Dentry in its parent. If this Dentry is a + // filesystem root, name is unspecified. + name string +} + +// IsAncestorDentry returns true if d is an ancestor of d2; that is, d is +// either d2's parent or an ancestor of d2's parent. +func IsAncestorDentry(d, d2 *Dentry) bool { + for { + if d2.parent == d { + return true + } + if d2.parent == d2 { + return false + } + d2 = d2.parent + } +} + +// ParentOrSelf returns d.parent. If d.parent is nil, ParentOrSelf returns d. +func ParentOrSelf(d *Dentry) *Dentry { + if d.parent != nil { + return d.parent + } + return d +} + +// PrependPath is a generic implementation of FilesystemImpl.PrependPath(). +func PrependPath(vfsroot vfs.VirtualDentry, mnt *vfs.Mount, d *Dentry, b *fspath.Builder) error { + for { + if mnt == vfsroot.Mount() && &d.vfsd == vfsroot.Dentry() { + return vfs.PrependPathAtVFSRootError{} + } + if &d.vfsd == mnt.Root() { + return nil + } + if d.parent == nil { + return vfs.PrependPathAtNonMountRootError{} + } + b.PrependComponent(d.name) + d = d.parent + } +} diff --git a/pkg/sentry/vfs/memxattr/BUILD b/pkg/sentry/vfs/memxattr/BUILD new file mode 100644 index 000000000..d8c4d27b9 --- /dev/null +++ b/pkg/sentry/vfs/memxattr/BUILD @@ -0,0 +1,15 @@ +load("//tools:defs.bzl", "go_library") + +package(licenses = ["notice"]) + +go_library( + name = "memxattr", + srcs = ["xattr.go"], + visibility = ["//pkg/sentry:internal"], + deps = [ + "//pkg/abi/linux", + "//pkg/sentry/vfs", + "//pkg/sync", + "//pkg/syserror", + ], +) diff --git a/pkg/sentry/vfs/memxattr/xattr.go b/pkg/sentry/vfs/memxattr/xattr.go new file mode 100644 index 000000000..cc1e7d764 --- /dev/null +++ b/pkg/sentry/vfs/memxattr/xattr.go @@ -0,0 +1,102 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// Package memxattr provides a default, in-memory extended attribute +// implementation. +package memxattr + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/pkg/syserror" +) + +// SimpleExtendedAttributes implements extended attributes using a map of +// names to values. +// +// +stateify savable +type SimpleExtendedAttributes struct { + // mu protects the below fields. + mu sync.RWMutex `state:"nosave"` + xattrs map[string]string +} + +// Getxattr returns the value at 'name'. +func (x *SimpleExtendedAttributes) Getxattr(opts *vfs.GetxattrOptions) (string, error) { + x.mu.RLock() + value, ok := x.xattrs[opts.Name] + x.mu.RUnlock() + if !ok { + return "", syserror.ENODATA + } + // Check that the size of the buffer provided in getxattr(2) is large enough + // to contain the value. + if opts.Size != 0 && uint64(len(value)) > opts.Size { + return "", syserror.ERANGE + } + return value, nil +} + +// Setxattr sets 'value' at 'name'. +func (x *SimpleExtendedAttributes) Setxattr(opts *vfs.SetxattrOptions) error { + x.mu.Lock() + defer x.mu.Unlock() + if x.xattrs == nil { + if opts.Flags&linux.XATTR_REPLACE != 0 { + return syserror.ENODATA + } + x.xattrs = make(map[string]string) + } + + _, ok := x.xattrs[opts.Name] + if ok && opts.Flags&linux.XATTR_CREATE != 0 { + return syserror.EEXIST + } + if !ok && opts.Flags&linux.XATTR_REPLACE != 0 { + return syserror.ENODATA + } + + x.xattrs[opts.Name] = opts.Value + return nil +} + +// Listxattr returns all names in xattrs. +func (x *SimpleExtendedAttributes) Listxattr(size uint64) ([]string, error) { + // Keep track of the size of the buffer needed in listxattr(2) for the list. + listSize := 0 + x.mu.RLock() + names := make([]string, 0, len(x.xattrs)) + for n := range x.xattrs { + names = append(names, n) + // Add one byte per null terminator. + listSize += len(n) + 1 + } + x.mu.RUnlock() + if size != 0 && uint64(listSize) > size { + return nil, syserror.ERANGE + } + return names, nil +} + +// Removexattr removes the xattr at 'name'. +func (x *SimpleExtendedAttributes) Removexattr(name string) error { + x.mu.Lock() + defer x.mu.Unlock() + if _, ok := x.xattrs[name]; !ok { + return syserror.ENODATA + } + delete(x.xattrs, name) + return nil +} diff --git a/pkg/sentry/vfs/mount.go b/pkg/sentry/vfs/mount.go index 1fbb420f9..02850b65c 100644 --- a/pkg/sentry/vfs/mount.go +++ b/pkg/sentry/vfs/mount.go @@ -15,7 +15,11 @@ package vfs import ( + "bytes" + "fmt" "math" + "sort" + "strings" "sync/atomic" "gvisor.dev/gvisor/pkg/abi/linux" @@ -24,6 +28,9 @@ import ( "gvisor.dev/gvisor/pkg/syserror" ) +// lastMountID is used to allocate mount ids. Must be accessed atomically. +var lastMountID uint64 + // A Mount is a replacement of a Dentry (Mount.key.point) from one Filesystem // (Mount.key.parent.fs) with a Dentry (Mount.root) from another Filesystem // (Mount.fs), which applies to path resolution in the context of a particular @@ -38,14 +45,19 @@ import ( // // Mount is analogous to Linux's struct mount. (gVisor does not distinguish // between struct mount and struct vfsmount.) +// +// +stateify savable type Mount struct { - // vfs, fs, and root are immutable. References are held on fs and root. + // vfs, fs, root are immutable. References are held on fs and root. // // Invariant: root belongs to fs. vfs *VirtualFilesystem fs *Filesystem root *Dentry + // ID is the immutable mount ID. + ID uint64 + // key is protected by VirtualFilesystem.mountMu and // VirtualFilesystem.mounts.seq, and may be nil. References are held on // key.parent and key.point if they are not nil. @@ -72,6 +84,10 @@ type Mount struct { // umounted is true. umounted is protected by VirtualFilesystem.mountMu. umounted bool + // flags contains settings as specified for mount(2), e.g. MS_NOEXEC, except + // for MS_RDONLY which is tracked in "writers". + flags MountFlags + // The lower 63 bits of writers is the number of calls to // Mount.CheckBeginWrite() that have not yet been paired with a call to // Mount.EndWrite(). The MSB of writers is set if MS_RDONLY is in effect. @@ -79,12 +95,30 @@ type Mount struct { writers int64 } +func newMount(vfs *VirtualFilesystem, fs *Filesystem, root *Dentry, mntns *MountNamespace, opts *MountOptions) *Mount { + mnt := &Mount{ + ID: atomic.AddUint64(&lastMountID, 1), + vfs: vfs, + fs: fs, + root: root, + flags: opts.Flags, + ns: mntns, + refs: 1, + } + if opts.ReadOnly { + mnt.setReadOnlyLocked(true) + } + return mnt +} + // A MountNamespace is a collection of Mounts. // // MountNamespaces are reference-counted. Unless otherwise specified, all // MountNamespace methods require that a reference is held. // // MountNamespace is analogous to Linux's struct mnt_namespace. +// +// +stateify savable type MountNamespace struct { // root is the MountNamespace's root mount. root is immutable. root *Mount @@ -114,6 +148,7 @@ type MountNamespace struct { func (vfs *VirtualFilesystem) NewMountNamespace(ctx context.Context, creds *auth.Credentials, source, fsTypeName string, opts *GetFilesystemOptions) (*MountNamespace, error) { rft := vfs.getFilesystemType(fsTypeName) if rft == nil { + ctx.Warningf("Unknown filesystem: %s", fsTypeName) return nil, syserror.ENODEV } fs, root, err := rft.fsType.GetFilesystem(ctx, vfs, creds, source, *opts) @@ -124,16 +159,22 @@ func (vfs *VirtualFilesystem) NewMountNamespace(ctx context.Context, creds *auth refs: 1, mountpoints: make(map[*Dentry]uint32), } - mntns.root = &Mount{ - vfs: vfs, - fs: fs, - root: root, - ns: mntns, - refs: 1, - } + mntns.root = newMount(vfs, fs, root, mntns, &MountOptions{}) return mntns, nil } +// NewDisconnectedMount returns a Mount representing fs with the given root +// (which may be nil). The new Mount is not associated with any MountNamespace +// and is not connected to any other Mounts. References are taken on fs and +// root. +func (vfs *VirtualFilesystem) NewDisconnectedMount(fs *Filesystem, root *Dentry, opts *MountOptions) (*Mount, error) { + fs.IncRef() + if root != nil { + root.IncRef() + } + return newMount(vfs, fs, root, nil /* mntns */, opts), nil +} + // MountAt creates and mounts a Filesystem configured by the given arguments. func (vfs *VirtualFilesystem) MountAt(ctx context.Context, creds *auth.Credentials, source string, target *PathOperation, fsTypeName string, opts *MountOptions) error { rft := vfs.getFilesystemType(fsTypeName) @@ -147,6 +188,7 @@ func (vfs *VirtualFilesystem) MountAt(ctx context.Context, creds *auth.Credentia if err != nil { return err } + // We can't hold vfs.mountMu while calling FilesystemImpl methods due to // lock ordering. vd, err := vfs.GetDentryAt(ctx, creds, target, &GetDentryOptions{}) @@ -158,7 +200,7 @@ func (vfs *VirtualFilesystem) MountAt(ctx context.Context, creds *auth.Credentia vfs.mountMu.Lock() vd.dentry.mu.Lock() for { - if vd.dentry.IsDisowned() { + if vd.dentry.dead { vd.dentry.mu.Unlock() vfs.mountMu.Unlock() vd.DecRef() @@ -192,17 +234,11 @@ func (vfs *VirtualFilesystem) MountAt(ctx context.Context, creds *auth.Credentia } vd.dentry.mu.Lock() } - // TODO: Linux requires that either both the mount point and the mount root - // are directories, or neither are, and returns ENOTDIR if this is not the - // case. + // TODO(gvisor.dev/issue/1035): Linux requires that either both the mount + // point and the mount root are directories, or neither are, and returns + // ENOTDIR if this is not the case. mntns := vd.mount.ns - mnt := &Mount{ - vfs: vfs, - fs: fs, - root: root, - ns: mntns, - refs: 1, - } + mnt := newMount(vfs, fs, root, mntns, opts) vfs.mounts.seq.BeginWrite() vfs.connectLocked(mnt, vd, mntns) vfs.mounts.seq.EndWrite() @@ -231,14 +267,17 @@ func (vfs *VirtualFilesystem) UmountAt(ctx context.Context, creds *auth.Credenti return syserror.EINVAL } vfs.mountMu.Lock() - if mntns := MountNamespaceFromContext(ctx); mntns != nil && mntns != vd.mount.ns { - vfs.mountMu.Unlock() - return syserror.EINVAL + if mntns := MountNamespaceFromContext(ctx); mntns != nil { + defer mntns.DecRef() + if mntns != vd.mount.ns { + vfs.mountMu.Unlock() + return syserror.EINVAL + } } - // TODO(jamieliu): Linux special-cases umount of the caller's root, which - // we don't implement yet (we'll just fail it since the caller holds a - // reference on it). + // TODO(gvisor.dev/issue/1035): Linux special-cases umount of the caller's + // root, which we don't implement yet (we'll just fail it since the caller + // holds a reference on it). vfs.mounts.seq.BeginWrite() if opts.Flags&linux.MNT_DETACH == 0 { @@ -605,12 +644,34 @@ func (mnt *Mount) setReadOnlyLocked(ro bool) error { return nil } +func (mnt *Mount) readOnly() bool { + return atomic.LoadInt64(&mnt.writers) < 0 +} + // Filesystem returns the mounted Filesystem. It does not take a reference on // the returned Filesystem. func (mnt *Mount) Filesystem() *Filesystem { return mnt.fs } +// submountsLocked returns this Mount and all Mounts that are descendents of +// it. +// +// Precondition: mnt.vfs.mountMu must be held. +func (mnt *Mount) submountsLocked() []*Mount { + mounts := []*Mount{mnt} + for m := range mnt.children { + mounts = append(mounts, m.submountsLocked()...) + } + return mounts +} + +// Root returns the mount's root. It does not take a reference on the returned +// Dentry. +func (mnt *Mount) Root() *Dentry { + return mnt.root +} + // Root returns mntns' root. A reference is taken on the returned // VirtualDentry. func (mntns *MountNamespace) Root() VirtualDentry { @@ -621,3 +682,174 @@ func (mntns *MountNamespace) Root() VirtualDentry { vd.IncRef() return vd } + +// GenerateProcMounts emits the contents of /proc/[pid]/mounts for vfs to buf. +// +// Preconditions: taskRootDir.Ok(). +func (vfs *VirtualFilesystem) GenerateProcMounts(ctx context.Context, taskRootDir VirtualDentry, buf *bytes.Buffer) { + vfs.mountMu.Lock() + defer vfs.mountMu.Unlock() + rootMnt := taskRootDir.mount + mounts := rootMnt.submountsLocked() + sort.Slice(mounts, func(i, j int) bool { return mounts[i].ID < mounts[j].ID }) + for _, mnt := range mounts { + // Get the path to this mount relative to task root. + mntRootVD := VirtualDentry{ + mount: mnt, + dentry: mnt.root, + } + path, err := vfs.PathnameReachable(ctx, taskRootDir, mntRootVD) + if err != nil { + // For some reason we didn't get a path. Log a warning + // and run with empty path. + ctx.Warningf("Error getting pathname for mount root %+v: %v", mnt.root, err) + path = "" + } + if path == "" { + // Either an error occurred, or path is not reachable + // from root. + break + } + + opts := "rw" + if mnt.readOnly() { + opts = "ro" + } + if mnt.flags.NoExec { + opts += ",noexec" + } + + // Format: + // <special device or remote filesystem> <mount point> <filesystem type> <mount options> <needs dump> <fsck order> + // + // The "needs dump" and "fsck order" flags are always 0, which + // is allowed. + fmt.Fprintf(buf, "%s %s %s %s %d %d\n", "none", path, mnt.fs.FilesystemType().Name(), opts, 0, 0) + } +} + +// GenerateProcMountInfo emits the contents of /proc/[pid]/mountinfo for vfs to +// buf. +// +// Preconditions: taskRootDir.Ok(). +func (vfs *VirtualFilesystem) GenerateProcMountInfo(ctx context.Context, taskRootDir VirtualDentry, buf *bytes.Buffer) { + vfs.mountMu.Lock() + defer vfs.mountMu.Unlock() + rootMnt := taskRootDir.mount + mounts := rootMnt.submountsLocked() + sort.Slice(mounts, func(i, j int) bool { return mounts[i].ID < mounts[j].ID }) + for _, mnt := range mounts { + // Get the path to this mount relative to task root. + mntRootVD := VirtualDentry{ + mount: mnt, + dentry: mnt.root, + } + path, err := vfs.PathnameReachable(ctx, taskRootDir, mntRootVD) + if err != nil { + // For some reason we didn't get a path. Log a warning + // and run with empty path. + ctx.Warningf("Error getting pathname for mount root %+v: %v", mnt.root, err) + path = "" + } + if path == "" { + // Either an error occurred, or path is not reachable + // from root. + break + } + // Stat the mount root to get the major/minor device numbers. + pop := &PathOperation{ + Root: mntRootVD, + Start: mntRootVD, + } + statx, err := vfs.StatAt(ctx, auth.NewAnonymousCredentials(), pop, &StatOptions{}) + if err != nil { + // Well that's not good. Ignore this mount. + break + } + + // Format: + // 36 35 98:0 /mnt1 /mnt2 rw,noatime master:1 - ext3 /dev/root rw,errors=continue + // (1)(2)(3) (4) (5) (6) (7) (8) (9) (10) (11) + + // (1) Mount ID. + fmt.Fprintf(buf, "%d ", mnt.ID) + + // (2) Parent ID (or this ID if there is no parent). + pID := mnt.ID + if p := mnt.parent(); p != nil { + pID = p.ID + } + fmt.Fprintf(buf, "%d ", pID) + + // (3) Major:Minor device ID. We don't have a superblock, so we + // just use the root inode device number. + fmt.Fprintf(buf, "%d:%d ", statx.DevMajor, statx.DevMinor) + + // (4) Root: the pathname of the directory in the filesystem + // which forms the root of this mount. + // + // NOTE(b/78135857): This will always be "/" until we implement + // bind mounts. + fmt.Fprintf(buf, "/ ") + + // (5) Mount point (relative to process root). + fmt.Fprintf(buf, "%s ", manglePath(path)) + + // (6) Mount options. + opts := "rw" + if mnt.readOnly() { + opts = "ro" + } + if mnt.flags.NoExec { + opts += ",noexec" + } + // TODO(gvisor.dev/issue/1193): Add "noatime" if MS_NOATIME is + // set. + fmt.Fprintf(buf, "%s ", opts) + + // (7) Optional fields: zero or more fields of the form "tag[:value]". + // (8) Separator: the end of the optional fields is marked by a single hyphen. + fmt.Fprintf(buf, "- ") + + // (9) Filesystem type. + fmt.Fprintf(buf, "%s ", mnt.fs.FilesystemType().Name()) + + // (10) Mount source: filesystem-specific information or "none". + fmt.Fprintf(buf, "none ") + + // (11) Superblock options, and final newline. + fmt.Fprintf(buf, "%s\n", superBlockOpts(path, mnt)) + } +} + +// manglePath replaces ' ', '\t', '\n', and '\\' with their octal equivalents. +// See Linux fs/seq_file.c:mangle_path. +func manglePath(p string) string { + r := strings.NewReplacer(" ", "\\040", "\t", "\\011", "\n", "\\012", "\\", "\\134") + return r.Replace(p) +} + +// superBlockOpts returns the super block options string for the the mount at +// the given path. +func superBlockOpts(mountPath string, mnt *Mount) string { + // gVisor doesn't (yet) have a concept of super block options, so we + // use the ro/rw bit from the mount flag. + opts := "rw" + if mnt.readOnly() { + opts = "ro" + } + + // NOTE(b/147673608): If the mount is a cgroup, we also need to include + // the cgroup name in the options. For now we just read that from the + // path. + // + // TODO(gvisor.dev/issue/190): Once gVisor has full cgroup support, we + // should get this value from the cgroup itself, and not rely on the + // path. + if mnt.fs.FilesystemType().Name() == "cgroup" { + splitPath := strings.Split(mountPath, "/") + cgroupType := splitPath[len(splitPath)-1] + opts += "," + cgroupType + } + return opts +} diff --git a/pkg/sentry/vfs/mount_test.go b/pkg/sentry/vfs/mount_test.go index 3b933468d..3335e4057 100644 --- a/pkg/sentry/vfs/mount_test.go +++ b/pkg/sentry/vfs/mount_test.go @@ -55,7 +55,7 @@ func TestMountTableInsertLookup(t *testing.T) { } } -// TODO: concurrent lookup/insertion/removal +// TODO(gvisor.dev/issue/1035): concurrent lookup/insertion/removal. // must be powers of 2 var benchNumMounts = []int{1 << 2, 1 << 5, 1 << 8} diff --git a/pkg/sentry/vfs/mount_unsafe.go b/pkg/sentry/vfs/mount_unsafe.go index bd90d36c4..bc7581698 100644 --- a/pkg/sentry/vfs/mount_unsafe.go +++ b/pkg/sentry/vfs/mount_unsafe.go @@ -26,6 +26,7 @@ import ( "sync/atomic" "unsafe" + "gvisor.dev/gvisor/pkg/gohacks" "gvisor.dev/gvisor/pkg/sync" ) @@ -64,6 +65,8 @@ func (mnt *Mount) storeKey(vd VirtualDentry) { // (provided mutation is sufficiently uncommon). // // mountTable.Init() must be called on new mountTables before use. +// +// +stateify savable type mountTable struct { // mountTable is implemented as a seqcount-protected hash table that // resolves collisions with linear probing, featuring Robin Hood insertion @@ -75,8 +78,8 @@ type mountTable struct { // intrinsics and inline assembly, limiting the performance of this // approach.) - seq sync.SeqCount - seed uint32 // for hashing keys + seq sync.SeqCount `state:"nosave"` + seed uint32 // for hashing keys // size holds both length (number of elements) and capacity (number of // slots): capacity is stored as its base-2 log (referred to as order) in @@ -89,7 +92,7 @@ type mountTable struct { // length and cap in separate uint32s) for ~free. size uint64 - slots unsafe.Pointer // []mountSlot; never nil after Init + slots unsafe.Pointer `state:"nosave"` // []mountSlot; never nil after Init } type mountSlot struct { @@ -158,7 +161,7 @@ func newMountTableSlots(cap uintptr) unsafe.Pointer { // Lookup may be called even if there are concurrent mutators of mt. func (mt *mountTable) Lookup(parent *Mount, point *Dentry) *Mount { key := mountKey{parent: unsafe.Pointer(parent), point: unsafe.Pointer(point)} - hash := memhash(noescape(unsafe.Pointer(&key)), uintptr(mt.seed), mountKeyBytes) + hash := memhash(gohacks.Noescape(unsafe.Pointer(&key)), uintptr(mt.seed), mountKeyBytes) loop: for { @@ -359,12 +362,3 @@ func memhash(p unsafe.Pointer, seed, s uintptr) uintptr //go:linkname rand32 runtime.fastrand func rand32() uint32 - -// This is copy/pasted from runtime.noescape(), and is needed because arguments -// apparently escape from all functions defined by linkname. -// -//go:nosplit -func noescape(p unsafe.Pointer) unsafe.Pointer { - x := uintptr(p) - return unsafe.Pointer(x ^ 0) -} diff --git a/pkg/sentry/vfs/options.go b/pkg/sentry/vfs/options.go index b7774bf28..534528ce6 100644 --- a/pkg/sentry/vfs/options.go +++ b/pkg/sentry/vfs/options.go @@ -16,6 +16,7 @@ package vfs import ( "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" ) // GetDentryOptions contains options to VirtualFilesystem.GetDentryAt() and @@ -44,10 +45,27 @@ type MknodOptions struct { // DevMinor are the major and minor device numbers for the created device. DevMajor uint32 DevMinor uint32 + + // Endpoint is the endpoint to bind to the created file, if a socket file is + // being created for bind(2) on a Unix domain socket. + Endpoint transport.BoundEndpoint +} + +// MountFlags contains flags as specified for mount(2), e.g. MS_NOEXEC. +// MS_RDONLY is not part of MountFlags because it's tracked in Mount.writers. +type MountFlags struct { + // NoExec is equivalent to MS_NOEXEC. + NoExec bool } // MountOptions contains options to VirtualFilesystem.MountAt(). type MountOptions struct { + // Flags contains flags as specified for mount(2), e.g. MS_NOEXEC. + Flags MountFlags + + // ReadOnly is equivalent to MS_RDONLY. + ReadOnly bool + // GetFilesystemOptions contains options to FilesystemType.GetFilesystem(). GetFilesystemOptions GetFilesystemOptions @@ -61,7 +79,7 @@ type MountOptions struct { type OpenOptions struct { // Flags contains access mode and flags as specified for open(2). // - // FilesystemImpls is reponsible for implementing the following flags: + // FilesystemImpls are responsible for implementing the following flags: // O_RDONLY, O_WRONLY, O_RDWR, O_APPEND, O_CREAT, O_DIRECT, O_DSYNC, // O_EXCL, O_NOATIME, O_NOCTTY, O_NONBLOCK, O_PATH, O_SYNC, O_TMPFILE, and // O_TRUNC. VFS is responsible for handling O_DIRECTORY, O_LARGEFILE, and @@ -72,6 +90,12 @@ type OpenOptions struct { // If FilesystemImpl.OpenAt() creates a file, Mode is the file mode for the // created file. Mode linux.FileMode + + // FileExec is set when the file is being opened to be executed. + // VirtualFilesystem.OpenAt() checks that the caller has execute permissions + // on the file, that the file is a regular file, and that the mount doesn't + // have MS_NOEXEC set. + FileExec bool } // ReadOptions contains options to FileDescription.PRead(), @@ -108,6 +132,20 @@ type SetStatOptions struct { Stat linux.Statx } +// GetxattrOptions contains options to VirtualFilesystem.GetxattrAt(), +// FilesystemImpl.GetxattrAt(), FileDescription.Getxattr(), and +// FileDescriptionImpl.Getxattr(). +type GetxattrOptions struct { + // Name is the name of the extended attribute to retrieve. + Name string + + // Size is the maximum value size that the caller will tolerate. If the value + // is larger than size, getxattr methods may return ERANGE, but they are also + // free to ignore the hint entirely (i.e. the value returned may be larger + // than size). All size checking is done independently at the syscall layer. + Size uint64 +} + // SetxattrOptions contains options to VirtualFilesystem.SetxattrAt(), // FilesystemImpl.SetxattrAt(), FileDescription.Setxattr(), and // FileDescriptionImpl.Setxattr(). diff --git a/pkg/sentry/vfs/pathname.go b/pkg/sentry/vfs/pathname.go index b318c681a..cd78d66bc 100644 --- a/pkg/sentry/vfs/pathname.go +++ b/pkg/sentry/vfs/pathname.go @@ -58,7 +58,7 @@ loop: switch err.(type) { case nil: if vd.mount == vfsroot.mount && vd.mount.root == vfsroot.dentry { - // GenericPrependPath() will have returned + // genericfstree.PrependPath() will have returned // PrependPathAtVFSRootError in this case since it checks // against vfsroot before mnt.root, but other implementations // of FilesystemImpl.PrependPath() may return nil instead. @@ -84,16 +84,59 @@ loop: } } b.PrependByte('/') - if origD.IsDisowned() { + if origD.IsDead() { b.AppendString(" (deleted)") } return b.String(), nil } +// PathnameReachable returns an absolute pathname to vd, consistent with +// Linux's __d_path() (as used by seq_path_root()). If vfsroot.Ok() and vd is +// not reachable from vfsroot, such that seq_path_root() would return SEQ_SKIP +// (causing the entire containing entry to be skipped), PathnameReachable +// returns ("", nil). +func (vfs *VirtualFilesystem) PathnameReachable(ctx context.Context, vfsroot, vd VirtualDentry) (string, error) { + b := getFSPathBuilder() + defer putFSPathBuilder(b) + haveRef := false + defer func() { + if haveRef { + vd.DecRef() + } + }() +loop: + for { + err := vd.mount.fs.impl.PrependPath(ctx, vfsroot, vd, b) + switch err.(type) { + case nil: + if vd.mount == vfsroot.mount && vd.mount.root == vfsroot.dentry { + break loop + } + nextVD := vfs.getMountpointAt(vd.mount, vfsroot) + if !nextVD.Ok() { + return "", nil + } + if haveRef { + vd.DecRef() + } + vd = nextVD + haveRef = true + case PrependPathAtVFSRootError: + break loop + case PrependPathAtNonMountRootError, PrependPathSyntheticError: + return "", nil + default: + return "", err + } + } + b.PrependByte('/') + return b.String(), nil +} + // PathnameForGetcwd returns an absolute pathname to vd, consistent with // Linux's sys_getcwd(). func (vfs *VirtualFilesystem) PathnameForGetcwd(ctx context.Context, vfsroot, vd VirtualDentry) (string, error) { - if vd.dentry.IsDisowned() { + if vd.dentry.IsDead() { return "", syserror.ENOENT } diff --git a/pkg/sentry/vfs/permissions.go b/pkg/sentry/vfs/permissions.go index f664581f4..f9647f90e 100644 --- a/pkg/sentry/vfs/permissions.go +++ b/pkg/sentry/vfs/permissions.go @@ -15,8 +15,12 @@ package vfs import ( + "math" + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/limits" "gvisor.dev/gvisor/pkg/syserror" ) @@ -25,9 +29,9 @@ type AccessTypes uint16 // Bits in AccessTypes. const ( + MayExec AccessTypes = 1 + MayWrite AccessTypes = 2 MayRead AccessTypes = 4 - MayWrite = 2 - MayExec = 1 ) // OnlyRead returns true if access _only_ allows read. @@ -52,16 +56,17 @@ func (a AccessTypes) MayExec() bool { // GenericCheckPermissions checks that creds has the given access rights on a // file with the given permissions, UID, and GID, subject to the rules of -// fs/namei.c:generic_permission(). isDir is true if the file is a directory. -func GenericCheckPermissions(creds *auth.Credentials, ats AccessTypes, isDir bool, mode uint16, kuid auth.KUID, kgid auth.KGID) error { +// fs/namei.c:generic_permission(). +func GenericCheckPermissions(creds *auth.Credentials, ats AccessTypes, mode linux.FileMode, kuid auth.KUID, kgid auth.KGID) error { // Check permission bits. - perms := mode + perms := uint16(mode.Permissions()) if creds.EffectiveKUID == kuid { perms >>= 6 } else if creds.InGroup(kgid) { perms >>= 3 } if uint16(ats)&perms == uint16(ats) { + // All permission bits match, access granted. return nil } @@ -73,7 +78,7 @@ func GenericCheckPermissions(creds *auth.Credentials, ats AccessTypes, isDir boo } // CAP_DAC_READ_SEARCH allows the caller to read and search arbitrary // directories, and read arbitrary non-directory files. - if (isDir && !ats.MayWrite()) || ats.OnlyRead() { + if (mode.IsDir() && !ats.MayWrite()) || ats.OnlyRead() { if creds.HasCapability(linux.CAP_DAC_READ_SEARCH) { return nil } @@ -81,7 +86,7 @@ func GenericCheckPermissions(creds *auth.Credentials, ats AccessTypes, isDir boo // CAP_DAC_OVERRIDE allows arbitrary access to directories, read/write // access to non-directory files, and execute access to non-directory files // for which at least one execute bit is set. - if isDir || !ats.MayExec() || (mode&0111 != 0) { + if mode.IsDir() || !ats.MayExec() || (mode.Permissions()&0111 != 0) { if creds.HasCapability(linux.CAP_DAC_OVERRIDE) { return nil } @@ -103,17 +108,22 @@ func GenericCheckPermissions(creds *auth.Credentials, ats AccessTypes, isDir boo // AccessTypesForOpenFlags returns MayRead|MayWrite in this case. // // Use May{Read,Write}FileWithOpenFlags() for these checks instead. -func AccessTypesForOpenFlags(flags uint32) AccessTypes { - switch flags & linux.O_ACCMODE { +func AccessTypesForOpenFlags(opts *OpenOptions) AccessTypes { + ats := AccessTypes(0) + if opts.FileExec { + ats |= MayExec + } + + switch opts.Flags & linux.O_ACCMODE { case linux.O_RDONLY: - if flags&linux.O_TRUNC != 0 { - return MayRead | MayWrite + if opts.Flags&linux.O_TRUNC != 0 { + return ats | MayRead | MayWrite } - return MayRead + return ats | MayRead case linux.O_WRONLY: - return MayWrite + return ats | MayWrite default: - return MayRead | MayWrite + return ats | MayRead | MayWrite } } @@ -142,7 +152,16 @@ func MayWriteFileWithOpenFlags(flags uint32) bool { // CheckSetStat checks that creds has permission to change the metadata of a // file with the given permissions, UID, and GID as specified by stat, subject // to the rules of Linux's fs/attr.c:setattr_prepare(). -func CheckSetStat(creds *auth.Credentials, stat *linux.Statx, mode uint16, kuid auth.KUID, kgid auth.KGID) error { +func CheckSetStat(ctx context.Context, creds *auth.Credentials, stat *linux.Statx, mode linux.FileMode, kuid auth.KUID, kgid auth.KGID) error { + if stat.Mask&linux.STATX_SIZE != 0 { + limit, err := CheckLimit(ctx, 0, int64(stat.Size)) + if err != nil { + return err + } + if limit < int64(stat.Size) { + return syserror.ErrExceedsFileSizeLimit + } + } if stat.Mask&linux.STATX_MODE != 0 { if !CanActAsOwner(creds, kuid) { return syserror.EPERM @@ -172,11 +191,7 @@ func CheckSetStat(creds *auth.Credentials, stat *linux.Statx, mode uint16, kuid (stat.Mask&linux.STATX_CTIME != 0 && stat.Ctime.Nsec != linux.UTIME_NOW) { return syserror.EPERM } - // isDir is irrelevant in the following call to - // GenericCheckPermissions since ats == MayWrite means that - // CAP_DAC_READ_SEARCH does not apply, and CAP_DAC_OVERRIDE - // applies, regardless of isDir. - if err := GenericCheckPermissions(creds, MayWrite, false /* isDir */, mode, kuid, kgid); err != nil { + if err := GenericCheckPermissions(creds, MayWrite, mode, kuid, kgid); err != nil { return err } } @@ -200,3 +215,21 @@ func CanActAsOwner(creds *auth.Credentials, kuid auth.KUID) bool { func HasCapabilityOnFile(creds *auth.Credentials, cp linux.Capability, kuid auth.KUID, kgid auth.KGID) bool { return creds.HasCapability(cp) && creds.UserNamespace.MapFromKUID(kuid).Ok() && creds.UserNamespace.MapFromKGID(kgid).Ok() } + +// CheckLimit enforces file size rlimits. It returns error if the write +// operation must not proceed. Otherwise it returns the max length allowed to +// without violating the limit. +func CheckLimit(ctx context.Context, offset, size int64) (int64, error) { + fileSizeLimit := limits.FromContext(ctx).Get(limits.FileSize).Cur + if fileSizeLimit > math.MaxInt64 { + return size, nil + } + if offset >= int64(fileSizeLimit) { + return 0, syserror.ErrExceedsFileSizeLimit + } + remaining := int64(fileSizeLimit) - offset + if remaining < size { + return remaining, nil + } + return size, nil +} diff --git a/pkg/sentry/vfs/resolving_path.go b/pkg/sentry/vfs/resolving_path.go index 8a0b382f6..9d047ff88 100644 --- a/pkg/sentry/vfs/resolving_path.go +++ b/pkg/sentry/vfs/resolving_path.go @@ -29,7 +29,9 @@ import ( // // From the perspective of FilesystemImpl methods, a ResolvingPath represents a // starting Dentry on the associated Filesystem (on which a reference is -// already held) and a stream of path components relative to that Dentry. +// already held), a stream of path components relative to that Dentry, and +// elements of the invoking Context that are commonly required by +// FilesystemImpl methods. // // ResolvingPath is loosely analogous to Linux's struct nameidata. type ResolvingPath struct { @@ -228,7 +230,7 @@ func (rp *ResolvingPath) Advance() { rp.pit = next } else { // at end of path segment, continue with next one rp.curPart-- - rp.pit = rp.parts[rp.curPart-1] + rp.pit = rp.parts[rp.curPart] } } @@ -251,18 +253,17 @@ func (rp *ResolvingPath) relpathCommit() { rp.origParts[rp.curPart] = rp.pit } -// ResolveParent returns the VFS parent of d. It does not take a reference on -// the returned Dentry. -// -// Preconditions: There are no concurrent mutators of d. -// -// Postconditions: If the returned error is nil, then the returned Dentry is -// not nil. -func (rp *ResolvingPath) ResolveParent(d *Dentry) (*Dentry, error) { - var parent *Dentry +// CheckRoot is called before resolving the parent of the Dentry d. If the +// Dentry is contextually a VFS root, such that path resolution should treat +// d's parent as itself, CheckRoot returns (true, nil). If the Dentry is the +// root of a non-root mount, such that path resolution should switch to another +// Mount, CheckRoot returns (unspecified, non-nil error). Otherwise, path +// resolution should resolve d's parent normally, and CheckRoot returns (false, +// nil). +func (rp *ResolvingPath) CheckRoot(d *Dentry) (bool, error) { if d == rp.root.dentry && rp.mount == rp.root.mount { - // At contextual VFS root. - parent = d + // At contextual VFS root (due to e.g. chroot(2)). + return true, nil } else if d == rp.mount.root { // At mount root ... vd := rp.vfs.getMountpointAt(rp.mount, rp.root) @@ -270,69 +271,49 @@ func (rp *ResolvingPath) ResolveParent(d *Dentry) (*Dentry, error) { // ... of non-root mount. rp.nextMount = vd.mount rp.nextStart = vd.dentry - return nil, resolveMountRootOrJumpError{} + return false, resolveMountRootOrJumpError{} } // ... of root mount. - parent = d - } else if d.parent == nil { - // At filesystem root. - parent = d - } else { - parent = d.parent - } - if parent.isMounted() { - if mnt := rp.vfs.getMountAt(rp.mount, parent); mnt != nil { - rp.nextMount = mnt - return nil, resolveMountPointError{} - } + return true, nil } - return parent, nil + return false, nil } -// ResolveChild returns the VFS child of d with the given name. It does not -// take a reference on the returned Dentry. If no such child exists, -// ResolveChild returns (nil, nil). -// -// Preconditions: There are no concurrent mutators of d. -func (rp *ResolvingPath) ResolveChild(d *Dentry, name string) (*Dentry, error) { - child := d.children[name] - if child == nil { - return nil, nil +// CheckMount is called after resolving the parent or child of another Dentry +// to d. If d is a mount point, such that path resolution should switch to +// another Mount, CheckMount returns a non-nil error. Otherwise, CheckMount +// returns nil. +func (rp *ResolvingPath) CheckMount(d *Dentry) error { + if !d.isMounted() { + return nil } - if child.isMounted() { - if mnt := rp.vfs.getMountAt(rp.mount, child); mnt != nil { - rp.nextMount = mnt - return nil, resolveMountPointError{} - } - } - return child, nil -} - -// ResolveComponent returns the Dentry reached by starting at d and resolving -// the current path component in the stream represented by rp. It does not -// advance the stream. It does not take a reference on the returned Dentry. If -// no such Dentry exists, ResolveComponent returns (nil, nil). -// -// Preconditions: !rp.Done(). There are no concurrent mutators of d. -func (rp *ResolvingPath) ResolveComponent(d *Dentry) (*Dentry, error) { - switch pc := rp.Component(); pc { - case ".": - return d, nil - case "..": - return rp.ResolveParent(d) - default: - return rp.ResolveChild(d, pc) + if mnt := rp.vfs.getMountAt(rp.mount, d); mnt != nil { + rp.nextMount = mnt + return resolveMountPointError{} } + return nil } // ShouldFollowSymlink returns true if, supposing that the current path // component in pcs represents a symbolic link, the symbolic link should be // followed. // +// If path is terminated with '/', the '/' is considered the last element and +// any symlink before that is followed: +// - For most non-creating walks, the last path component is handled by +// fs/namei.c:lookup_last(), which sets LOOKUP_FOLLOW if the first byte +// after the path component is non-NULL (which is only possible if it's '/') +// and the path component is of type LAST_NORM. +// +// - For open/openat/openat2 without O_CREAT, the last path component is +// handled by fs/namei.c:do_last(), which does the same, though without the +// LAST_NORM check. +// // Preconditions: !rp.Done(). func (rp *ResolvingPath) ShouldFollowSymlink() bool { - // Non-final symlinks are always followed. - return rp.flags&rpflagsFollowFinalSymlink != 0 || !rp.Final() + // Non-final symlinks are always followed. Paths terminated with '/' are also + // always followed. + return rp.flags&rpflagsFollowFinalSymlink != 0 || !rp.Final() || rp.MustBeDir() } // HandleSymlink is called when the current path component is a symbolic link diff --git a/pkg/sentry/vfs/timerfd.go b/pkg/sentry/vfs/timerfd.go new file mode 100644 index 000000000..42b880656 --- /dev/null +++ b/pkg/sentry/vfs/timerfd.go @@ -0,0 +1,142 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs + +import ( + "sync/atomic" + + "gvisor.dev/gvisor/pkg/context" + ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" + "gvisor.dev/gvisor/pkg/waiter" +) + +// TimerFileDescription implements FileDescriptionImpl for timer fds. It also +// implements ktime.TimerListener. +type TimerFileDescription struct { + vfsfd FileDescription + FileDescriptionDefaultImpl + DentryMetadataFileDescriptionImpl + + events waiter.Queue + timer *ktime.Timer + + // val is the number of timer expirations since the last successful + // call to PRead, or SetTime. val must be accessed using atomic memory + // operations. + val uint64 +} + +var _ FileDescriptionImpl = (*TimerFileDescription)(nil) +var _ ktime.TimerListener = (*TimerFileDescription)(nil) + +// NewTimerFD returns a new timer fd. +func (vfs *VirtualFilesystem) NewTimerFD(clock ktime.Clock, flags uint32) (*FileDescription, error) { + vd := vfs.NewAnonVirtualDentry("[timerfd]") + defer vd.DecRef() + tfd := &TimerFileDescription{} + tfd.timer = ktime.NewTimer(clock, tfd) + if err := tfd.vfsfd.Init(tfd, flags, vd.Mount(), vd.Dentry(), &FileDescriptionOptions{ + UseDentryMetadata: true, + DenyPRead: true, + DenyPWrite: true, + InvalidWrite: true, + }); err != nil { + return nil, err + } + return &tfd.vfsfd, nil +} + +// Read implements FileDescriptionImpl.Read. +func (tfd *TimerFileDescription) Read(ctx context.Context, dst usermem.IOSequence, opts ReadOptions) (int64, error) { + const sizeofUint64 = 8 + if dst.NumBytes() < sizeofUint64 { + return 0, syserror.EINVAL + } + if val := atomic.SwapUint64(&tfd.val, 0); val != 0 { + var buf [sizeofUint64]byte + usermem.ByteOrder.PutUint64(buf[:], val) + if _, err := dst.CopyOut(ctx, buf[:]); err != nil { + // Linux does not undo consuming the number of + // expirations even if writing to userspace fails. + return 0, err + } + return sizeofUint64, nil + } + return 0, syserror.ErrWouldBlock +} + +// Clock returns the timer fd's Clock. +func (tfd *TimerFileDescription) Clock() ktime.Clock { + return tfd.timer.Clock() +} + +// GetTime returns the associated Timer's setting and the time at which it was +// observed. +func (tfd *TimerFileDescription) GetTime() (ktime.Time, ktime.Setting) { + return tfd.timer.Get() +} + +// SetTime atomically changes the associated Timer's setting, resets the number +// of expirations to 0, and returns the previous setting and the time at which +// it was observed. +func (tfd *TimerFileDescription) SetTime(s ktime.Setting) (ktime.Time, ktime.Setting) { + return tfd.timer.SwapAnd(s, func() { atomic.StoreUint64(&tfd.val, 0) }) +} + +// Readiness implements waiter.Waitable.Readiness. +func (tfd *TimerFileDescription) Readiness(mask waiter.EventMask) waiter.EventMask { + var ready waiter.EventMask + if atomic.LoadUint64(&tfd.val) != 0 { + ready |= waiter.EventIn + } + return ready +} + +// EventRegister implements waiter.Waitable.EventRegister. +func (tfd *TimerFileDescription) EventRegister(e *waiter.Entry, mask waiter.EventMask) { + tfd.events.EventRegister(e, mask) +} + +// EventUnregister implements waiter.Waitable.EventUnregister. +func (tfd *TimerFileDescription) EventUnregister(e *waiter.Entry) { + tfd.events.EventUnregister(e) +} + +// PauseTimer pauses the associated Timer. +func (tfd *TimerFileDescription) PauseTimer() { + tfd.timer.Pause() +} + +// ResumeTimer resumes the associated Timer. +func (tfd *TimerFileDescription) ResumeTimer() { + tfd.timer.Resume() +} + +// Release implements FileDescriptionImpl.Release() +func (tfd *TimerFileDescription) Release() { + tfd.timer.Destroy() +} + +// Notify implements ktime.TimerListener.Notify. +func (tfd *TimerFileDescription) Notify(exp uint64, setting ktime.Setting) (ktime.Setting, bool) { + atomic.AddUint64(&tfd.val, exp) + tfd.events.Notify(waiter.EventIn) + return ktime.Setting{}, false +} + +// Destroy implements ktime.TimerListener.Destroy. +func (tfd *TimerFileDescription) Destroy() {} diff --git a/pkg/sentry/vfs/vfs.go b/pkg/sentry/vfs/vfs.go index 908c69f91..cb5bbd781 100644 --- a/pkg/sentry/vfs/vfs.go +++ b/pkg/sentry/vfs/vfs.go @@ -38,6 +38,7 @@ import ( "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/fspath" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/syserror" ) @@ -46,11 +47,13 @@ import ( // // There is no analogue to the VirtualFilesystem type in Linux, as the // equivalent state in Linux is global. +// +// +stateify savable type VirtualFilesystem struct { // mountMu serializes mount mutations. // // mountMu is analogous to Linux's namespace_sem. - mountMu sync.Mutex + mountMu sync.Mutex `state:"nosave"` // mounts maps (mount parent, mount point) pairs to mounts. (Since mounts // are uniquely namespaced, including mount parent in the key correctly @@ -89,56 +92,59 @@ type VirtualFilesystem struct { // devices contains all registered Devices. devices is protected by // devicesMu. - devicesMu sync.RWMutex + devicesMu sync.RWMutex `state:"nosave"` devices map[devTuple]*registeredDevice // anonBlockDevMinor contains all allocated anonymous block device minor // numbers. anonBlockDevMinorNext is a lower bound for the smallest // unallocated anonymous block device number. anonBlockDevMinorNext and // anonBlockDevMinor are protected by anonBlockDevMinorMu. - anonBlockDevMinorMu sync.Mutex + anonBlockDevMinorMu sync.Mutex `state:"nosave"` anonBlockDevMinorNext uint32 anonBlockDevMinor map[uint32]struct{} // fsTypes contains all registered FilesystemTypes. fsTypes is protected by // fsTypesMu. - fsTypesMu sync.RWMutex + fsTypesMu sync.RWMutex `state:"nosave"` fsTypes map[string]*registeredFilesystemType // filesystems contains all Filesystems. filesystems is protected by // filesystemsMu. - filesystemsMu sync.Mutex + filesystemsMu sync.Mutex `state:"nosave"` filesystems map[*Filesystem]struct{} } -// New returns a new VirtualFilesystem with no mounts or FilesystemTypes. -func New() *VirtualFilesystem { - vfs := &VirtualFilesystem{ - mountpoints: make(map[*Dentry]map[*Mount]struct{}), - devices: make(map[devTuple]*registeredDevice), - anonBlockDevMinorNext: 1, - anonBlockDevMinor: make(map[uint32]struct{}), - fsTypes: make(map[string]*registeredFilesystemType), - filesystems: make(map[*Filesystem]struct{}), - } +// Init initializes a new VirtualFilesystem with no mounts or FilesystemTypes. +func (vfs *VirtualFilesystem) Init() error { + vfs.mountpoints = make(map[*Dentry]map[*Mount]struct{}) + vfs.devices = make(map[devTuple]*registeredDevice) + vfs.anonBlockDevMinorNext = 1 + vfs.anonBlockDevMinor = make(map[uint32]struct{}) + vfs.fsTypes = make(map[string]*registeredFilesystemType) + vfs.filesystems = make(map[*Filesystem]struct{}) vfs.mounts.Init() // Construct vfs.anonMount. anonfsDevMinor, err := vfs.GetAnonBlockDevMinor() if err != nil { - panic(fmt.Sprintf("VirtualFilesystem.GetAnonBlockDevMinor() failed during VirtualFilesystem construction: %v", err)) + // This shouldn't be possible since anonBlockDevMinorNext was + // initialized to 1 above (no device numbers have been allocated yet). + panic(fmt.Sprintf("VirtualFilesystem.Init: device number allocation for anonfs failed: %v", err)) } anonfs := anonFilesystem{ devMinor: anonfsDevMinor, } - anonfs.vfsfs.Init(vfs, &anonfs) - vfs.anonMount = &Mount{ - vfs: vfs, - fs: &anonfs.vfsfs, - refs: 1, + anonfs.vfsfs.Init(vfs, &anonFilesystemType{}, &anonfs) + defer anonfs.vfsfs.DecRef() + anonMount, err := vfs.NewDisconnectedMount(&anonfs.vfsfs, nil, &MountOptions{}) + if err != nil { + // We should not be passing any MountOptions that would cause + // construction of this mount to fail. + panic(fmt.Sprintf("VirtualFilesystem.Init: anonfs mount failed: %v", err)) } + vfs.anonMount = anonMount - return vfs + return nil } // PathOperation specifies the path operated on by a VFS method. @@ -169,6 +175,23 @@ type PathOperation struct { FollowFinalSymlink bool } +// AccessAt checks whether a user with creds has access to the file at +// the given path. +func (vfs *VirtualFilesystem) AccessAt(ctx context.Context, creds *auth.Credentials, ats AccessTypes, pop *PathOperation) error { + rp := vfs.getResolvingPath(creds, pop) + for { + err := rp.mount.fs.impl.AccessAt(ctx, rp, creds, ats) + if err == nil { + vfs.putResolvingPath(rp) + return nil + } + if !rp.handleError(err) { + vfs.putResolvingPath(rp) + return err + } + } +} + // GetDentryAt returns a VirtualDentry representing the given path, at which a // file must exist. A reference is taken on the returned VirtualDentry. func (vfs *VirtualFilesystem) GetDentryAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, opts *GetDentryOptions) (VirtualDentry, error) { @@ -208,7 +231,7 @@ func (vfs *VirtualFilesystem) getParentDirAndName(ctx context.Context, creds *au } if checkInvariants { if rp.canHandleError(err) && rp.Done() { - panic(fmt.Sprintf("%T.GetParentDentryAt() consumed all path components and returned %T", rp.mount.fs.impl, err)) + panic(fmt.Sprintf("%T.GetParentDentryAt() consumed all path components and returned %v", rp.mount.fs.impl, err)) } } if !rp.handleError(err) { @@ -249,7 +272,7 @@ func (vfs *VirtualFilesystem) LinkAt(ctx context.Context, creds *auth.Credential } if checkInvariants { if rp.canHandleError(err) && rp.Done() { - panic(fmt.Sprintf("%T.LinkAt() consumed all path components and returned %T", rp.mount.fs.impl, err)) + panic(fmt.Sprintf("%T.LinkAt() consumed all path components and returned %v", rp.mount.fs.impl, err)) } } if !rp.handleError(err) { @@ -285,7 +308,7 @@ func (vfs *VirtualFilesystem) MkdirAt(ctx context.Context, creds *auth.Credentia } if checkInvariants { if rp.canHandleError(err) && rp.Done() { - panic(fmt.Sprintf("%T.MkdirAt() consumed all path components and returned %T", rp.mount.fs.impl, err)) + panic(fmt.Sprintf("%T.MkdirAt() consumed all path components and returned %v", rp.mount.fs.impl, err)) } } if !rp.handleError(err) { @@ -312,13 +335,13 @@ func (vfs *VirtualFilesystem) MknodAt(ctx context.Context, creds *auth.Credentia rp := vfs.getResolvingPath(creds, pop) for { err := rp.mount.fs.impl.MknodAt(ctx, rp, *opts) - if err != nil { + if err == nil { vfs.putResolvingPath(rp) return nil } if checkInvariants { if rp.canHandleError(err) && rp.Done() { - panic(fmt.Sprintf("%T.MknodAt() consumed all path components and returned %T", rp.mount.fs.impl, err)) + panic(fmt.Sprintf("%T.MknodAt() consumed all path components and returned %v", rp.mount.fs.impl, err)) } } if !rp.handleError(err) { @@ -328,19 +351,43 @@ func (vfs *VirtualFilesystem) MknodAt(ctx context.Context, creds *auth.Credentia } } +// BoundEndpointAt gets the bound endpoint at the given path, if one exists. +func (vfs *VirtualFilesystem) BoundEndpointAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation) (transport.BoundEndpoint, error) { + if !pop.Path.Begin.Ok() { + if pop.Path.Absolute { + return nil, syserror.ECONNREFUSED + } + return nil, syserror.ENOENT + } + rp := vfs.getResolvingPath(creds, pop) + for { + bep, err := rp.mount.fs.impl.BoundEndpointAt(ctx, rp) + if err == nil { + vfs.putResolvingPath(rp) + return bep, nil + } + if checkInvariants { + if rp.canHandleError(err) && rp.Done() { + panic(fmt.Sprintf("%T.BoundEndpointAt() consumed all path components and returned %v", rp.mount.fs.impl, err)) + } + } + if !rp.handleError(err) { + vfs.putResolvingPath(rp) + return nil, err + } + } +} + // OpenAt returns a FileDescription providing access to the file at the given // path. A reference is taken on the returned FileDescription. func (vfs *VirtualFilesystem) OpenAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, opts *OpenOptions) (*FileDescription, error) { // Remove: // - // - O_LARGEFILE, which we always report in FileDescription status flags - // since only 64-bit architectures are supported at this time. - // // - O_CLOEXEC, which affects file descriptors and therefore must be // handled outside of VFS. // // - Unknown flags. - opts.Flags &= linux.O_ACCMODE | linux.O_CREAT | linux.O_EXCL | linux.O_NOCTTY | linux.O_TRUNC | linux.O_APPEND | linux.O_NONBLOCK | linux.O_DSYNC | linux.O_ASYNC | linux.O_DIRECT | linux.O_DIRECTORY | linux.O_NOFOLLOW | linux.O_NOATIME | linux.O_SYNC | linux.O_PATH | linux.O_TMPFILE + opts.Flags &= linux.O_ACCMODE | linux.O_CREAT | linux.O_EXCL | linux.O_NOCTTY | linux.O_TRUNC | linux.O_APPEND | linux.O_NONBLOCK | linux.O_DSYNC | linux.O_ASYNC | linux.O_DIRECT | linux.O_LARGEFILE | linux.O_DIRECTORY | linux.O_NOFOLLOW | linux.O_NOATIME | linux.O_SYNC | linux.O_PATH | linux.O_TMPFILE // Linux's __O_SYNC (which we call linux.O_SYNC) implies O_DSYNC. if opts.Flags&linux.O_SYNC != 0 { opts.Flags |= linux.O_DSYNC @@ -379,6 +426,25 @@ func (vfs *VirtualFilesystem) OpenAt(ctx context.Context, creds *auth.Credential fd, err := rp.mount.fs.impl.OpenAt(ctx, rp, *opts) if err == nil { vfs.putResolvingPath(rp) + + if opts.FileExec { + if fd.Mount().flags.NoExec { + fd.DecRef() + return nil, syserror.EACCES + } + + // Only a regular file can be executed. + stat, err := fd.Stat(ctx, StatOptions{Mask: linux.STATX_TYPE}) + if err != nil { + fd.DecRef() + return nil, err + } + if stat.Mask&linux.STATX_TYPE == 0 || stat.Mode&linux.S_IFMT != linux.S_IFREG { + fd.DecRef() + return nil, syserror.EACCES + } + } + return fd, nil } if !rp.handleError(err) { @@ -453,7 +519,7 @@ func (vfs *VirtualFilesystem) RenameAt(ctx context.Context, creds *auth.Credenti } if checkInvariants { if rp.canHandleError(err) && rp.Done() { - panic(fmt.Sprintf("%T.RenameAt() consumed all path components and returned %T", rp.mount.fs.impl, err)) + panic(fmt.Sprintf("%T.RenameAt() consumed all path components and returned %v", rp.mount.fs.impl, err)) } } if !rp.handleError(err) { @@ -486,7 +552,7 @@ func (vfs *VirtualFilesystem) RmdirAt(ctx context.Context, creds *auth.Credentia } if checkInvariants { if rp.canHandleError(err) && rp.Done() { - panic(fmt.Sprintf("%T.RmdirAt() consumed all path components and returned %T", rp.mount.fs.impl, err)) + panic(fmt.Sprintf("%T.RmdirAt() consumed all path components and returned %v", rp.mount.fs.impl, err)) } } if !rp.handleError(err) { @@ -567,7 +633,7 @@ func (vfs *VirtualFilesystem) SymlinkAt(ctx context.Context, creds *auth.Credent } if checkInvariants { if rp.canHandleError(err) && rp.Done() { - panic(fmt.Sprintf("%T.SymlinkAt() consumed all path components and returned %T", rp.mount.fs.impl, err)) + panic(fmt.Sprintf("%T.SymlinkAt() consumed all path components and returned %v", rp.mount.fs.impl, err)) } } if !rp.handleError(err) { @@ -599,7 +665,7 @@ func (vfs *VirtualFilesystem) UnlinkAt(ctx context.Context, creds *auth.Credenti } if checkInvariants { if rp.canHandleError(err) && rp.Done() { - panic(fmt.Sprintf("%T.UnlinkAt() consumed all path components and returned %T", rp.mount.fs.impl, err)) + panic(fmt.Sprintf("%T.UnlinkAt() consumed all path components and returned %v", rp.mount.fs.impl, err)) } } if !rp.handleError(err) { @@ -611,10 +677,10 @@ func (vfs *VirtualFilesystem) UnlinkAt(ctx context.Context, creds *auth.Credenti // ListxattrAt returns all extended attribute names for the file at the given // path. -func (vfs *VirtualFilesystem) ListxattrAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation) ([]string, error) { +func (vfs *VirtualFilesystem) ListxattrAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, size uint64) ([]string, error) { rp := vfs.getResolvingPath(creds, pop) for { - names, err := rp.mount.fs.impl.ListxattrAt(ctx, rp) + names, err := rp.mount.fs.impl.ListxattrAt(ctx, rp, size) if err == nil { vfs.putResolvingPath(rp) return names, nil @@ -636,10 +702,10 @@ func (vfs *VirtualFilesystem) ListxattrAt(ctx context.Context, creds *auth.Crede // GetxattrAt returns the value associated with the given extended attribute // for the file at the given path. -func (vfs *VirtualFilesystem) GetxattrAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, name string) (string, error) { +func (vfs *VirtualFilesystem) GetxattrAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, opts *GetxattrOptions) (string, error) { rp := vfs.getResolvingPath(creds, pop) for { - val, err := rp.mount.fs.impl.GetxattrAt(ctx, rp, name) + val, err := rp.mount.fs.impl.GetxattrAt(ctx, rp, *opts) if err == nil { vfs.putResolvingPath(rp) return val, nil @@ -724,6 +790,8 @@ func (vfs *VirtualFilesystem) SyncAllFilesystems(ctx context.Context) error { // VirtualDentry methods require that a reference is held on the VirtualDentry. // // VirtualDentry is analogous to Linux's struct path. +// +// +stateify savable type VirtualDentry struct { mount *Mount dentry *Dentry diff --git a/pkg/sentry/watchdog/watchdog.go b/pkg/sentry/watchdog/watchdog.go index bfb2fac26..fcc46420f 100644 --- a/pkg/sentry/watchdog/watchdog.go +++ b/pkg/sentry/watchdog/watchdog.go @@ -221,7 +221,7 @@ func (w *Watchdog) waitForStart() { return } var buf bytes.Buffer - buf.WriteString("Watchdog.Start() not called within %s:\n") + buf.WriteString(fmt.Sprintf("Watchdog.Start() not called within %s", w.StartupTimeout)) w.doAction(w.StartupTimeoutAction, false, &buf) } @@ -319,26 +319,25 @@ func (w *Watchdog) report(offenders map[*kernel.Task]*offender, newTaskFound boo // Dump stack only if a new task is detected or if it sometime has // passed since the last time a stack dump was generated. - skipStack := newTaskFound || time.Since(w.lastStackDump) >= stackDumpSameTaskPeriod - w.doAction(w.TaskTimeoutAction, skipStack, &buf) + showStack := newTaskFound || time.Since(w.lastStackDump) >= stackDumpSameTaskPeriod + w.doAction(w.TaskTimeoutAction, showStack, &buf) } func (w *Watchdog) reportStuckWatchdog() { var buf bytes.Buffer - buf.WriteString("Watchdog goroutine is stuck:\n") + buf.WriteString("Watchdog goroutine is stuck:") w.doAction(w.TaskTimeoutAction, false, &buf) } -// doAction will take the given action. If the action is LogWarnind and -// skipStack is true, then the stack printing will be skipped. -func (w *Watchdog) doAction(action Action, skipStack bool, msg *bytes.Buffer) { +// doAction will take the given action. If the action is LogWarning and +// showStack is false, then the stack printing will be skipped. +func (w *Watchdog) doAction(action Action, showStack bool, msg *bytes.Buffer) { switch action { case LogWarning: - if skipStack { + if !showStack { msg.WriteString("\n...[stack dump skipped]...") log.Warningf(msg.String()) return - } log.TracebackAll(msg.String()) w.lastStackDump = time.Now() @@ -359,7 +358,7 @@ func (w *Watchdog) doAction(action Action, skipStack bool, msg *bytes.Buffer) { case <-metricsEmitted: case <-time.After(1 * time.Second): } - panic(fmt.Sprintf("Stack for running G's are skipped while panicking.\n%s", msg.String())) + panic(fmt.Sprintf("%s\nStack for running G's are skipped while panicking.", msg.String())) default: panic(fmt.Sprintf("Unknown watchdog action %v", action)) diff --git a/pkg/sleep/commit_noasm.go b/pkg/sleep/commit_noasm.go index 3af447fb9..f59061f37 100644 --- a/pkg/sleep/commit_noasm.go +++ b/pkg/sleep/commit_noasm.go @@ -28,15 +28,6 @@ import "sync/atomic" // It is written in assembly because it is called from g0, so it doesn't have // a race context. func commitSleep(g uintptr, waitingG *uintptr) bool { - for { - // Check if the wait was aborted. - if atomic.LoadUintptr(waitingG) == 0 { - return false - } - - // Try to store the G so that wakers know who to wake. - if atomic.CompareAndSwapUintptr(waitingG, preparingG, g) { - return true - } - } + // Try to store the G so that wakers know who to wake. + return atomic.CompareAndSwapUintptr(waitingG, preparingG, g) } diff --git a/pkg/sleep/sleep_unsafe.go b/pkg/sleep/sleep_unsafe.go index acbf0229b..65bfcf778 100644 --- a/pkg/sleep/sleep_unsafe.go +++ b/pkg/sleep/sleep_unsafe.go @@ -299,20 +299,17 @@ func (s *Sleeper) enqueueAssertedWaker(w *Waker) { } } - for { - // Nothing to do if there isn't a G waiting. - g := atomic.LoadUintptr(&s.waitingG) - if g == 0 { - return - } + // Nothing to do if there isn't a G waiting. + if atomic.LoadUintptr(&s.waitingG) == 0 { + return + } - // Signal to the sleeper that a waker has been asserted. - if atomic.CompareAndSwapUintptr(&s.waitingG, g, 0) { - if g != preparingG { - // We managed to get a G. Wake it up. - goready(g, 0) - } - } + // Signal to the sleeper that a waker has been asserted. + switch g := atomic.SwapUintptr(&s.waitingG, 0); g { + case 0, preparingG: + default: + // We managed to get a G. Wake it up. + goready(g, 0) } } diff --git a/pkg/state/state.go b/pkg/state/state.go index dbe507ab4..03ae2dbb0 100644 --- a/pkg/state/state.go +++ b/pkg/state/state.go @@ -241,10 +241,7 @@ func Register(name string, instance interface{}, fns Fns) { // // This function is used by the stateify tool. func IsZeroValue(val interface{}) bool { - if val == nil { - return true - } - return reflect.DeepEqual(val, reflect.Zero(reflect.TypeOf(val)).Interface()) + return val == nil || reflect.ValueOf(val).Elem().IsZero() } // step captures one encoding / decoding step. On each step, there is up to one diff --git a/pkg/sync/BUILD b/pkg/sync/BUILD index 5340cf0d6..0e35d7d17 100644 --- a/pkg/sync/BUILD +++ b/pkg/sync/BUILD @@ -31,13 +31,13 @@ go_library( name = "sync", srcs = [ "aliases.go", - "downgradable_rwmutex_unsafe.go", "memmove_unsafe.go", + "mutex_unsafe.go", "norace_unsafe.go", "race_unsafe.go", + "rwmutex_unsafe.go", "seqcount.go", - "syncutil.go", - "tmutex_unsafe.go", + "sync.go", ], ) @@ -45,9 +45,9 @@ go_test( name = "sync_test", size = "small", srcs = [ - "downgradable_rwmutex_test.go", + "mutex_test.go", + "rwmutex_test.go", "seqcount_test.go", - "tmutex_test.go", ], library = ":sync", ) diff --git a/pkg/sync/aliases.go b/pkg/sync/aliases.go index d2d7132fa..0d4316254 100644 --- a/pkg/sync/aliases.go +++ b/pkg/sync/aliases.go @@ -29,3 +29,8 @@ type ( // Map is an alias of sync.Map. Map = sync.Map ) + +// NewCond is a wrapper around sync.NewCond. +func NewCond(l Locker) *Cond { + return sync.NewCond(l) +} diff --git a/pkg/sync/tmutex_test.go b/pkg/sync/mutex_test.go index 0838248b4..0838248b4 100644 --- a/pkg/sync/tmutex_test.go +++ b/pkg/sync/mutex_test.go diff --git a/pkg/sync/tmutex_unsafe.go b/pkg/sync/mutex_unsafe.go index 3dd15578b..3dd15578b 100644 --- a/pkg/sync/tmutex_unsafe.go +++ b/pkg/sync/mutex_unsafe.go diff --git a/pkg/sync/downgradable_rwmutex_test.go b/pkg/sync/rwmutex_test.go index ce667e825..ce667e825 100644 --- a/pkg/sync/downgradable_rwmutex_test.go +++ b/pkg/sync/rwmutex_test.go diff --git a/pkg/sync/downgradable_rwmutex_unsafe.go b/pkg/sync/rwmutex_unsafe.go index ea6cdc447..ea6cdc447 100644 --- a/pkg/sync/downgradable_rwmutex_unsafe.go +++ b/pkg/sync/rwmutex_unsafe.go diff --git a/pkg/sync/syncutil.go b/pkg/sync/sync.go index b16cf5333..b16cf5333 100644 --- a/pkg/sync/syncutil.go +++ b/pkg/sync/sync.go diff --git a/pkg/syncevent/BUILD b/pkg/syncevent/BUILD new file mode 100644 index 000000000..0500a22cf --- /dev/null +++ b/pkg/syncevent/BUILD @@ -0,0 +1,39 @@ +load("//tools:defs.bzl", "go_library", "go_test") + +licenses(["notice"]) + +go_library( + name = "syncevent", + srcs = [ + "broadcaster.go", + "receiver.go", + "source.go", + "syncevent.go", + "waiter_amd64.s", + "waiter_arm64.s", + "waiter_asm_unsafe.go", + "waiter_noasm_unsafe.go", + "waiter_unsafe.go", + ], + visibility = ["//:sandbox"], + deps = [ + "//pkg/atomicbitops", + "//pkg/sync", + ], +) + +go_test( + name = "syncevent_test", + size = "small", + srcs = [ + "broadcaster_test.go", + "syncevent_example_test.go", + "waiter_test.go", + ], + library = ":syncevent", + deps = [ + "//pkg/sleep", + "//pkg/sync", + "//pkg/waiter", + ], +) diff --git a/pkg/syncevent/broadcaster.go b/pkg/syncevent/broadcaster.go new file mode 100644 index 000000000..4bff59e7d --- /dev/null +++ b/pkg/syncevent/broadcaster.go @@ -0,0 +1,218 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package syncevent + +import ( + "gvisor.dev/gvisor/pkg/sync" +) + +// Broadcaster is an implementation of Source that supports any number of +// subscribed Receivers. +// +// The zero value of Broadcaster is valid and has no subscribed Receivers. +// Broadcaster is not copyable by value. +// +// All Broadcaster methods may be called concurrently from multiple goroutines. +type Broadcaster struct { + // Broadcaster is implemented as a hash table where keys are assigned by + // the Broadcaster and returned as SubscriptionIDs, making it safe to use + // the identity function for hashing. The hash table resolves collisions + // using linear probing and features Robin Hood insertion and backward + // shift deletion in order to support a relatively high load factor + // efficiently, which matters since the cost of Broadcast is linear in the + // size of the table. + + // mu protects the following fields. + mu sync.Mutex + + // Invariants: len(table) is 0 or a power of 2. + table []broadcasterSlot + + // load is the number of entries in table with receiver != nil. + load int + + lastID SubscriptionID +} + +type broadcasterSlot struct { + // Invariants: If receiver == nil, then filter == NoEvents and id == 0. + // Otherwise, id != 0. + receiver *Receiver + filter Set + id SubscriptionID +} + +const ( + broadcasterMinNonZeroTableSize = 2 // must be a power of 2 > 1 + + broadcasterMaxLoadNum = 13 + broadcasterMaxLoadDen = 16 +) + +// SubscribeEvents implements Source.SubscribeEvents. +func (b *Broadcaster) SubscribeEvents(r *Receiver, filter Set) SubscriptionID { + b.mu.Lock() + + // Assign an ID for this subscription. + b.lastID++ + id := b.lastID + + // Expand the table if over the maximum load factor: + // + // load / len(b.table) > broadcasterMaxLoadNum / broadcasterMaxLoadDen + // load * broadcasterMaxLoadDen > broadcasterMaxLoadNum * len(b.table) + b.load++ + if (b.load * broadcasterMaxLoadDen) > (broadcasterMaxLoadNum * len(b.table)) { + // Double the number of slots in the new table. + newlen := broadcasterMinNonZeroTableSize + if len(b.table) != 0 { + newlen = 2 * len(b.table) + } + if newlen <= cap(b.table) { + // Reuse excess capacity in the current table, moving entries not + // already in their first-probed positions to better ones. + newtable := b.table[:newlen] + newmask := uint64(newlen - 1) + for i := range b.table { + if b.table[i].receiver != nil && uint64(b.table[i].id)&newmask != uint64(i) { + entry := b.table[i] + b.table[i] = broadcasterSlot{} + broadcasterTableInsert(newtable, entry.id, entry.receiver, entry.filter) + } + } + b.table = newtable + } else { + newtable := make([]broadcasterSlot, newlen) + // Copy existing entries to the new table. + for i := range b.table { + if b.table[i].receiver != nil { + broadcasterTableInsert(newtable, b.table[i].id, b.table[i].receiver, b.table[i].filter) + } + } + // Switch to the new table. + b.table = newtable + } + } + + broadcasterTableInsert(b.table, id, r, filter) + b.mu.Unlock() + return id +} + +// Preconditions: table must not be full. len(table) is a power of 2. +func broadcasterTableInsert(table []broadcasterSlot, id SubscriptionID, r *Receiver, filter Set) { + entry := broadcasterSlot{ + receiver: r, + filter: filter, + id: id, + } + mask := uint64(len(table) - 1) + i := uint64(id) & mask + disp := uint64(0) + for { + if table[i].receiver == nil { + table[i] = entry + return + } + // If we've been displaced farther from our first-probed slot than the + // element stored in this one, swap elements and switch to inserting + // the replaced one. (This is Robin Hood insertion.) + slotDisp := (i - uint64(table[i].id)) & mask + if disp > slotDisp { + table[i], entry = entry, table[i] + disp = slotDisp + } + i = (i + 1) & mask + disp++ + } +} + +// UnsubscribeEvents implements Source.UnsubscribeEvents. +func (b *Broadcaster) UnsubscribeEvents(id SubscriptionID) { + b.mu.Lock() + + mask := uint64(len(b.table) - 1) + i := uint64(id) & mask + for { + if b.table[i].id == id { + // Found the element to remove. Move all subsequent elements + // backward until we either find an empty slot, or an element that + // is already in its first-probed slot. (This is backward shift + // deletion.) + for { + next := (i + 1) & mask + if b.table[next].receiver == nil { + break + } + if uint64(b.table[next].id)&mask == next { + break + } + b.table[i] = b.table[next] + i = next + } + b.table[i] = broadcasterSlot{} + break + } + i = (i + 1) & mask + } + + // If a table 1/4 of the current size would still be at or under the + // maximum load factor (i.e. the current table size is at least two + // expansions bigger than necessary), halve the size of the table to reduce + // the cost of Broadcast. Since we are concerned with iteration time and + // not memory usage, reuse the existing slice to reduce future allocations + // from table re-expansion. + b.load-- + if len(b.table) > broadcasterMinNonZeroTableSize && (b.load*(4*broadcasterMaxLoadDen)) <= (broadcasterMaxLoadNum*len(b.table)) { + newlen := len(b.table) / 2 + newtable := b.table[:newlen] + for i := newlen; i < len(b.table); i++ { + if b.table[i].receiver != nil { + broadcasterTableInsert(newtable, b.table[i].id, b.table[i].receiver, b.table[i].filter) + b.table[i] = broadcasterSlot{} + } + } + b.table = newtable + } + + b.mu.Unlock() +} + +// Broadcast notifies all Receivers subscribed to the Broadcaster of the subset +// of events to which they subscribed. The order in which Receivers are +// notified is unspecified. +func (b *Broadcaster) Broadcast(events Set) { + b.mu.Lock() + for i := range b.table { + if intersection := events & b.table[i].filter; intersection != 0 { + // We don't need to check if broadcasterSlot.receiver is nil, since + // if it is then broadcasterSlot.filter is 0. + b.table[i].receiver.Notify(intersection) + } + } + b.mu.Unlock() +} + +// FilteredEvents returns the set of events for which Broadcast will notify at +// least one Receiver, i.e. the union of filters for all subscribed Receivers. +func (b *Broadcaster) FilteredEvents() Set { + var es Set + b.mu.Lock() + for i := range b.table { + es |= b.table[i].filter + } + b.mu.Unlock() + return es +} diff --git a/pkg/syncevent/broadcaster_test.go b/pkg/syncevent/broadcaster_test.go new file mode 100644 index 000000000..e88779e23 --- /dev/null +++ b/pkg/syncevent/broadcaster_test.go @@ -0,0 +1,376 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package syncevent + +import ( + "fmt" + "math/rand" + "testing" + + "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/pkg/waiter" +) + +func TestBroadcasterFilter(t *testing.T) { + const numReceivers = 2 * MaxEvents + + var br Broadcaster + ws := make([]Waiter, numReceivers) + for i := range ws { + ws[i].Init() + br.SubscribeEvents(ws[i].Receiver(), 1<<(i%MaxEvents)) + } + for ev := 0; ev < MaxEvents; ev++ { + br.Broadcast(1 << ev) + for i := range ws { + want := NoEvents + if i%MaxEvents == ev { + want = 1 << ev + } + if got := ws[i].Receiver().PendingAndAckAll(); got != want { + t.Errorf("after Broadcast of event %d: waiter %d has pending event set %#x, wanted %#x", ev, i, got, want) + } + } + } +} + +// TestBroadcasterManySubscriptions tests that subscriptions are not lost by +// table expansion/compaction. +func TestBroadcasterManySubscriptions(t *testing.T) { + const numReceivers = 5000 // arbitrary + + var br Broadcaster + ws := make([]Waiter, numReceivers) + for i := range ws { + ws[i].Init() + } + + ids := make([]SubscriptionID, numReceivers) + for i := 0; i < numReceivers; i++ { + // Subscribe receiver i. + ids[i] = br.SubscribeEvents(ws[i].Receiver(), 1) + // Check that receivers [0, i] are subscribed. + br.Broadcast(1) + for j := 0; j <= i; j++ { + if ws[j].Pending() != 1 { + t.Errorf("receiver %d did not receive an event after subscription of receiver %d", j, i) + } + ws[j].Ack(1) + } + } + + // Generate a random order for unsubscriptions. + unsub := rand.Perm(numReceivers) + for i := 0; i < numReceivers; i++ { + // Unsubscribe receiver unsub[i]. + br.UnsubscribeEvents(ids[unsub[i]]) + // Check that receivers [unsub[0], unsub[i]] are not subscribed, and that + // receivers (unsub[i], unsub[numReceivers]) are still subscribed. + br.Broadcast(1) + for j := 0; j <= i; j++ { + if ws[unsub[j]].Pending() != 0 { + t.Errorf("unsub iteration %d: receiver %d received an event after unsubscription of receiver %d", i, unsub[j], unsub[i]) + } + } + for j := i + 1; j < numReceivers; j++ { + if ws[unsub[j]].Pending() != 1 { + t.Errorf("unsub iteration %d: receiver %d did not receive an event after unsubscription of receiver %d", i, unsub[j], unsub[i]) + } + ws[unsub[j]].Ack(1) + } + } +} + +var ( + receiverCountsNonZero = []int{1, 4, 16, 64} + receiverCountsIncludingZero = append([]int{0}, receiverCountsNonZero...) +) + +// BenchmarkBroadcasterX, BenchmarkMapX, and BenchmarkQueueX benchmark usage +// pattern X (described in terms of Broadcaster) with Broadcaster, a +// Mutex-protected map[*Receiver]Set, and waiter.Queue respectively. + +// BenchmarkXxxSubscribeUnsubscribe measures the cost of a Subscribe/Unsubscribe +// cycle. + +func BenchmarkBroadcasterSubscribeUnsubscribe(b *testing.B) { + var br Broadcaster + var w Waiter + w.Init() + + b.ResetTimer() + for i := 0; i < b.N; i++ { + id := br.SubscribeEvents(w.Receiver(), 1) + br.UnsubscribeEvents(id) + } +} + +func BenchmarkMapSubscribeUnsubscribe(b *testing.B) { + var mu sync.Mutex + m := make(map[*Receiver]Set) + var w Waiter + w.Init() + + b.ResetTimer() + for i := 0; i < b.N; i++ { + mu.Lock() + m[w.Receiver()] = Set(1) + mu.Unlock() + mu.Lock() + delete(m, w.Receiver()) + mu.Unlock() + } +} + +func BenchmarkQueueSubscribeUnsubscribe(b *testing.B) { + var q waiter.Queue + e, _ := waiter.NewChannelEntry(nil) + + b.ResetTimer() + for i := 0; i < b.N; i++ { + q.EventRegister(&e, 1) + q.EventUnregister(&e) + } +} + +// BenchmarkXxxSubscribeUnsubscribeBatch is similar to +// BenchmarkXxxSubscribeUnsubscribe, but subscribes and unsubscribes a large +// number of Receivers at a time in order to measure the amortized overhead of +// table expansion/compaction. (Since waiter.Queue is implemented using a +// linked list, BenchmarkQueueSubscribeUnsubscribe and +// BenchmarkQueueSubscribeUnsubscribeBatch should produce nearly the same +// result.) + +const numBatchReceivers = 1000 + +func BenchmarkBroadcasterSubscribeUnsubscribeBatch(b *testing.B) { + var br Broadcaster + ws := make([]Waiter, numBatchReceivers) + for i := range ws { + ws[i].Init() + } + ids := make([]SubscriptionID, numBatchReceivers) + + // Generate a random order for unsubscriptions. + unsub := rand.Perm(numBatchReceivers) + + b.ResetTimer() + for i := 0; i < b.N/numBatchReceivers; i++ { + for j := 0; j < numBatchReceivers; j++ { + ids[j] = br.SubscribeEvents(ws[j].Receiver(), 1) + } + for j := 0; j < numBatchReceivers; j++ { + br.UnsubscribeEvents(ids[unsub[j]]) + } + } +} + +func BenchmarkMapSubscribeUnsubscribeBatch(b *testing.B) { + var mu sync.Mutex + m := make(map[*Receiver]Set) + ws := make([]Waiter, numBatchReceivers) + for i := range ws { + ws[i].Init() + } + + // Generate a random order for unsubscriptions. + unsub := rand.Perm(numBatchReceivers) + + b.ResetTimer() + for i := 0; i < b.N/numBatchReceivers; i++ { + for j := 0; j < numBatchReceivers; j++ { + mu.Lock() + m[ws[j].Receiver()] = Set(1) + mu.Unlock() + } + for j := 0; j < numBatchReceivers; j++ { + mu.Lock() + delete(m, ws[unsub[j]].Receiver()) + mu.Unlock() + } + } +} + +func BenchmarkQueueSubscribeUnsubscribeBatch(b *testing.B) { + var q waiter.Queue + es := make([]waiter.Entry, numBatchReceivers) + for i := range es { + es[i], _ = waiter.NewChannelEntry(nil) + } + + // Generate a random order for unsubscriptions. + unsub := rand.Perm(numBatchReceivers) + + b.ResetTimer() + for i := 0; i < b.N/numBatchReceivers; i++ { + for j := 0; j < numBatchReceivers; j++ { + q.EventRegister(&es[j], 1) + } + for j := 0; j < numBatchReceivers; j++ { + q.EventUnregister(&es[unsub[j]]) + } + } +} + +// BenchmarkXxxBroadcastRedundant measures how long it takes to Broadcast +// already-pending events to multiple Receivers. + +func BenchmarkBroadcasterBroadcastRedundant(b *testing.B) { + for _, n := range receiverCountsIncludingZero { + b.Run(fmt.Sprintf("%d", n), func(b *testing.B) { + var br Broadcaster + ws := make([]Waiter, n) + for i := range ws { + ws[i].Init() + br.SubscribeEvents(ws[i].Receiver(), 1) + } + br.Broadcast(1) + + b.ResetTimer() + for i := 0; i < b.N; i++ { + br.Broadcast(1) + } + }) + } +} + +func BenchmarkMapBroadcastRedundant(b *testing.B) { + for _, n := range receiverCountsIncludingZero { + b.Run(fmt.Sprintf("%d", n), func(b *testing.B) { + var mu sync.Mutex + m := make(map[*Receiver]Set) + ws := make([]Waiter, n) + for i := range ws { + ws[i].Init() + m[ws[i].Receiver()] = Set(1) + } + mu.Lock() + for r := range m { + r.Notify(1) + } + mu.Unlock() + + b.ResetTimer() + for i := 0; i < b.N; i++ { + mu.Lock() + for r := range m { + r.Notify(1) + } + mu.Unlock() + } + }) + } +} + +func BenchmarkQueueBroadcastRedundant(b *testing.B) { + for _, n := range receiverCountsIncludingZero { + b.Run(fmt.Sprintf("%d", n), func(b *testing.B) { + var q waiter.Queue + for i := 0; i < n; i++ { + e, _ := waiter.NewChannelEntry(nil) + q.EventRegister(&e, 1) + } + q.Notify(1) + + b.ResetTimer() + for i := 0; i < b.N; i++ { + q.Notify(1) + } + }) + } +} + +// BenchmarkXxxBroadcastAck measures how long it takes to Broadcast events to +// multiple Receivers, check that all Receivers have received the event, and +// clear the event from all Receivers. + +func BenchmarkBroadcasterBroadcastAck(b *testing.B) { + for _, n := range receiverCountsNonZero { + b.Run(fmt.Sprintf("%d", n), func(b *testing.B) { + var br Broadcaster + ws := make([]Waiter, n) + for i := range ws { + ws[i].Init() + br.SubscribeEvents(ws[i].Receiver(), 1) + } + + b.ResetTimer() + for i := 0; i < b.N; i++ { + br.Broadcast(1) + for j := range ws { + if got, want := ws[j].Pending(), Set(1); got != want { + b.Fatalf("Receiver.Pending(): got %#x, wanted %#x", got, want) + } + ws[j].Ack(1) + } + } + }) + } +} + +func BenchmarkMapBroadcastAck(b *testing.B) { + for _, n := range receiverCountsNonZero { + b.Run(fmt.Sprintf("%d", n), func(b *testing.B) { + var mu sync.Mutex + m := make(map[*Receiver]Set) + ws := make([]Waiter, n) + for i := range ws { + ws[i].Init() + m[ws[i].Receiver()] = Set(1) + } + + b.ResetTimer() + for i := 0; i < b.N; i++ { + mu.Lock() + for r := range m { + r.Notify(1) + } + mu.Unlock() + for j := range ws { + if got, want := ws[j].Pending(), Set(1); got != want { + b.Fatalf("Receiver.Pending(): got %#x, wanted %#x", got, want) + } + ws[j].Ack(1) + } + } + }) + } +} + +func BenchmarkQueueBroadcastAck(b *testing.B) { + for _, n := range receiverCountsNonZero { + b.Run(fmt.Sprintf("%d", n), func(b *testing.B) { + var q waiter.Queue + chs := make([]chan struct{}, n) + for i := range chs { + e, ch := waiter.NewChannelEntry(nil) + q.EventRegister(&e, 1) + chs[i] = ch + } + + b.ResetTimer() + for i := 0; i < b.N; i++ { + q.Notify(1) + for _, ch := range chs { + select { + case <-ch: + default: + b.Fatalf("channel did not receive event") + } + } + } + }) + } +} diff --git a/pkg/syncevent/receiver.go b/pkg/syncevent/receiver.go new file mode 100644 index 000000000..5c86e5400 --- /dev/null +++ b/pkg/syncevent/receiver.go @@ -0,0 +1,103 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package syncevent + +import ( + "sync/atomic" + + "gvisor.dev/gvisor/pkg/atomicbitops" +) + +// Receiver is an event sink that holds pending events and invokes a callback +// whenever new events become pending. Receiver's methods may be called +// concurrently from multiple goroutines. +// +// Receiver.Init() must be called before first use. +type Receiver struct { + // pending is the set of pending events. pending is accessed using atomic + // memory operations. + pending uint64 + + // cb is notified when new events become pending. cb is immutable after + // Init(). + cb ReceiverCallback +} + +// ReceiverCallback receives callbacks from a Receiver. +type ReceiverCallback interface { + // NotifyPending is called when the corresponding Receiver has new pending + // events. + // + // NotifyPending is called synchronously from Receiver.Notify(), so + // implementations must not take locks that may be held by callers of + // Receiver.Notify(). NotifyPending may be called concurrently from + // multiple goroutines. + NotifyPending() +} + +// Init must be called before first use of r. +func (r *Receiver) Init(cb ReceiverCallback) { + r.cb = cb +} + +// Pending returns the set of pending events. +func (r *Receiver) Pending() Set { + return Set(atomic.LoadUint64(&r.pending)) +} + +// Notify sets the given events as pending. +func (r *Receiver) Notify(es Set) { + p := Set(atomic.LoadUint64(&r.pending)) + // Optimization: Skip the atomic CAS on r.pending if all events are + // already pending. + if p&es == es { + return + } + // When this is uncontended (the common case), CAS is faster than + // atomic-OR because the former is inlined and the latter (which we + // implement in assembly ourselves) is not. + if !atomic.CompareAndSwapUint64(&r.pending, uint64(p), uint64(p|es)) { + // If the CAS fails, fall back to atomic-OR. + atomicbitops.OrUint64(&r.pending, uint64(es)) + } + r.cb.NotifyPending() +} + +// Ack unsets the given events as pending. +func (r *Receiver) Ack(es Set) { + p := Set(atomic.LoadUint64(&r.pending)) + // Optimization: Skip the atomic CAS on r.pending if all events are + // already not pending. + if p&es == 0 { + return + } + // When this is uncontended (the common case), CAS is faster than + // atomic-AND because the former is inlined and the latter (which we + // implement in assembly ourselves) is not. + if !atomic.CompareAndSwapUint64(&r.pending, uint64(p), uint64(p&^es)) { + // If the CAS fails, fall back to atomic-AND. + atomicbitops.AndUint64(&r.pending, ^uint64(es)) + } +} + +// PendingAndAckAll unsets all events as pending and returns the set of +// previously-pending events. +// +// PendingAndAckAll should only be used in preference to a call to Pending +// followed by a conditional call to Ack when the caller expects events to be +// pending (e.g. after a call to ReceiverCallback.NotifyPending()). +func (r *Receiver) PendingAndAckAll() Set { + return Set(atomic.SwapUint64(&r.pending, 0)) +} diff --git a/pkg/syncevent/source.go b/pkg/syncevent/source.go new file mode 100644 index 000000000..ddffb171a --- /dev/null +++ b/pkg/syncevent/source.go @@ -0,0 +1,59 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package syncevent + +// Source represents an event source. +type Source interface { + // SubscribeEvents causes the Source to notify the given Receiver of the + // given subset of events. + // + // Preconditions: r != nil. The ReceiverCallback for r must not take locks + // that are ordered prior to the Source; for example, it cannot call any + // Source methods. + SubscribeEvents(r *Receiver, filter Set) SubscriptionID + + // UnsubscribeEvents causes the Source to stop notifying the Receiver + // subscribed by a previous call to SubscribeEvents that returned the given + // SubscriptionID. + // + // Preconditions: UnsubscribeEvents may be called at most once for any + // given SubscriptionID. + UnsubscribeEvents(id SubscriptionID) +} + +// SubscriptionID identifies a call to Source.SubscribeEvents. +type SubscriptionID uint64 + +// UnsubscribeAndAck is a convenience function that unsubscribes r from the +// given events from src and also clears them from r. +func UnsubscribeAndAck(src Source, r *Receiver, filter Set, id SubscriptionID) { + src.UnsubscribeEvents(id) + r.Ack(filter) +} + +// NoopSource implements Source by never sending events to subscribed +// Receivers. +type NoopSource struct{} + +// SubscribeEvents implements Source.SubscribeEvents. +func (NoopSource) SubscribeEvents(*Receiver, Set) SubscriptionID { + return 0 +} + +// UnsubscribeEvents implements Source.UnsubscribeEvents. +func (NoopSource) UnsubscribeEvents(SubscriptionID) { +} + +// See Broadcaster for a non-noop implementations of Source. diff --git a/pkg/syncevent/syncevent.go b/pkg/syncevent/syncevent.go new file mode 100644 index 000000000..9fb6a06de --- /dev/null +++ b/pkg/syncevent/syncevent.go @@ -0,0 +1,32 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// Package syncevent provides efficient primitives for goroutine +// synchronization based on event bitmasks. +package syncevent + +// Set is a bitmask where each bit represents a distinct user-defined event. +// The event package does not treat any bits in Set specially. +type Set uint64 + +const ( + // NoEvents is a Set containing no events. + NoEvents = Set(0) + + // AllEvents is a Set containing all possible events. + AllEvents = ^Set(0) + + // MaxEvents is the number of distinct events that can be represented by a Set. + MaxEvents = 64 +) diff --git a/pkg/syncevent/syncevent_example_test.go b/pkg/syncevent/syncevent_example_test.go new file mode 100644 index 000000000..bfb18e2ea --- /dev/null +++ b/pkg/syncevent/syncevent_example_test.go @@ -0,0 +1,108 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package syncevent + +import ( + "fmt" + "sync/atomic" + "time" +) + +func Example_ioReadinessInterrputible() { + const ( + evReady = Set(1 << iota) + evInterrupt + ) + errNotReady := fmt.Errorf("not ready for I/O") + + // State of some I/O object. + var ( + br Broadcaster + ready uint32 + ) + doIO := func() error { + if atomic.LoadUint32(&ready) == 0 { + return errNotReady + } + return nil + } + go func() { + // The I/O object eventually becomes ready for I/O. + time.Sleep(100 * time.Millisecond) + // When it does, it first ensures that future calls to isReady() return + // true, then broadcasts the readiness event to Receivers. + atomic.StoreUint32(&ready, 1) + br.Broadcast(evReady) + }() + + // Each user of the I/O object owns a Waiter. + var w Waiter + w.Init() + // The Waiter may be asynchronously interruptible, e.g. for signal + // handling in the sentry. + go func() { + time.Sleep(200 * time.Millisecond) + w.Receiver().Notify(evInterrupt) + }() + + // To use the I/O object: + // + // Optionally, if the I/O object is likely to be ready, attempt I/O first. + err := doIO() + if err == nil { + // Success, we're done. + return /* nil */ + } + if err != errNotReady { + // Failure, I/O failed for some reason other than readiness. + return /* err */ + } + // Subscribe for readiness events from the I/O object. + id := br.SubscribeEvents(w.Receiver(), evReady) + // When we are finished blocking, unsubscribe from readiness events and + // remove readiness events from the pending event set. + defer UnsubscribeAndAck(&br, w.Receiver(), evReady, id) + for { + // Attempt I/O again. This must be done after the call to SubscribeEvents, + // since the I/O object might have become ready between the previous call + // to doIO and the call to SubscribeEvents. + err = doIO() + if err == nil { + return /* nil */ + } + if err != errNotReady { + return /* err */ + } + // Block until either the I/O object indicates it is ready, or we are + // interrupted. + events := w.Wait() + if events&evInterrupt != 0 { + // In the specific case of sentry signal handling, signal delivery + // is handled by another system, so we aren't responsible for + // acknowledging evInterrupt. + return /* errInterrupted */ + } + // Note that, in a concurrent context, the I/O object might become + // ready and then not ready again. To handle this: + // + // - evReady must be acknowledged before calling doIO() again (rather + // than after), so that if the I/O object becomes ready *again* after + // the call to doIO(), the readiness event is not lost. + // + // - We must loop instead of just calling doIO() once after receiving + // evReady. + w.Ack(evReady) + } +} diff --git a/pkg/syncevent/waiter_amd64.s b/pkg/syncevent/waiter_amd64.s new file mode 100644 index 000000000..985b56ae5 --- /dev/null +++ b/pkg/syncevent/waiter_amd64.s @@ -0,0 +1,32 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "textflag.h" + +// See waiter_noasm_unsafe.go for a description of waiterUnlock. +// +// func waiterUnlock(g unsafe.Pointer, wg *unsafe.Pointer) bool +TEXT ·waiterUnlock(SB),NOSPLIT,$0-24 + MOVQ g+0(FP), DI + MOVQ wg+8(FP), SI + + MOVQ $·preparingG(SB), AX + LOCK + CMPXCHGQ DI, 0(SI) + + SETEQ AX + MOVB AX, ret+16(FP) + + RET + diff --git a/pkg/syncevent/waiter_arm64.s b/pkg/syncevent/waiter_arm64.s new file mode 100644 index 000000000..20d7ac23b --- /dev/null +++ b/pkg/syncevent/waiter_arm64.s @@ -0,0 +1,34 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "textflag.h" + +// See waiter_noasm_unsafe.go for a description of waiterUnlock. +// +// func waiterUnlock(g unsafe.Pointer, wg *unsafe.Pointer) bool +TEXT ·waiterUnlock(SB),NOSPLIT,$0-24 + MOVD wg+8(FP), R0 + MOVD $·preparingG(SB), R1 + MOVD g+0(FP), R2 +again: + LDAXR (R0), R3 + CMP R1, R3 + BNE ok + STLXR R2, (R0), R3 + CBNZ R3, again +ok: + CSET EQ, R0 + MOVB R0, ret+16(FP) + RET + diff --git a/pkg/syncevent/waiter_asm_unsafe.go b/pkg/syncevent/waiter_asm_unsafe.go new file mode 100644 index 000000000..0995e9053 --- /dev/null +++ b/pkg/syncevent/waiter_asm_unsafe.go @@ -0,0 +1,24 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// +build amd64 arm64 + +package syncevent + +import ( + "unsafe" +) + +// See waiter_noasm_unsafe.go for a description of waiterUnlock. +func waiterUnlock(g unsafe.Pointer, wg *unsafe.Pointer) bool diff --git a/pkg/syncevent/waiter_noasm_unsafe.go b/pkg/syncevent/waiter_noasm_unsafe.go new file mode 100644 index 000000000..1c4b0e39a --- /dev/null +++ b/pkg/syncevent/waiter_noasm_unsafe.go @@ -0,0 +1,39 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// waiterUnlock is called from g0, so when the race detector is enabled, +// waiterUnlock must be implemented in assembly since no race context is +// available. +// +// +build !race +// +build !amd64,!arm64 + +package syncevent + +import ( + "sync/atomic" + "unsafe" +) + +// waiterUnlock is the "unlock function" passed to runtime.gopark by +// Waiter.Wait*. wg is &Waiter.g, and g is a pointer to the calling runtime.g. +// waiterUnlock returns true if Waiter.Wait should sleep and false if sleeping +// should be aborted. +// +//go:nosplit +func waiterUnlock(g unsafe.Pointer, wg *unsafe.Pointer) bool { + // The only way this CAS can fail is if a call to Waiter.NotifyPending() + // has replaced *wg with nil, in which case we should not sleep. + return atomic.CompareAndSwapPointer(wg, (unsafe.Pointer)(&preparingG), g) +} diff --git a/pkg/syncevent/waiter_test.go b/pkg/syncevent/waiter_test.go new file mode 100644 index 000000000..3c8cbcdd8 --- /dev/null +++ b/pkg/syncevent/waiter_test.go @@ -0,0 +1,414 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package syncevent + +import ( + "sync/atomic" + "testing" + "time" + + "gvisor.dev/gvisor/pkg/sleep" + "gvisor.dev/gvisor/pkg/sync" +) + +func TestWaiterAlreadyPending(t *testing.T) { + var w Waiter + w.Init() + want := Set(1) + w.Notify(want) + if got := w.Wait(); got != want { + t.Errorf("Waiter.Wait: got %#x, wanted %#x", got, want) + } +} + +func TestWaiterAsyncNotify(t *testing.T) { + var w Waiter + w.Init() + want := Set(1) + go func() { + time.Sleep(100 * time.Millisecond) + w.Notify(want) + }() + if got := w.Wait(); got != want { + t.Errorf("Waiter.Wait: got %#x, wanted %#x", got, want) + } +} + +func TestWaiterWaitFor(t *testing.T) { + var w Waiter + w.Init() + evWaited := Set(1) + evOther := Set(2) + w.Notify(evOther) + notifiedEvent := uint32(0) + go func() { + time.Sleep(100 * time.Millisecond) + atomic.StoreUint32(¬ifiedEvent, 1) + w.Notify(evWaited) + }() + if got, want := w.WaitFor(evWaited), evWaited|evOther; got != want { + t.Errorf("Waiter.WaitFor: got %#x, wanted %#x", got, want) + } + if atomic.LoadUint32(¬ifiedEvent) == 0 { + t.Errorf("Waiter.WaitFor returned before goroutine notified waited-for event") + } +} + +func TestWaiterWaitAndAckAll(t *testing.T) { + var w Waiter + w.Init() + w.Notify(AllEvents) + if got := w.WaitAndAckAll(); got != AllEvents { + t.Errorf("Waiter.WaitAndAckAll: got %#x, wanted %#x", got, AllEvents) + } + if got := w.Pending(); got != NoEvents { + t.Errorf("Waiter.WaitAndAckAll did not ack all events: got %#x, wanted 0", got) + } +} + +// BenchmarkWaiterX, BenchmarkSleeperX, and BenchmarkChannelX benchmark usage +// pattern X (described in terms of Waiter) with Waiter, sleep.Sleeper, and +// buffered chan struct{} respectively. When the maximum number of event +// sources is relevant, we use 3 event sources because this is representative +// of the kernel.Task.block() use case: an interrupt source, a timeout source, +// and the actual event source being waited on. + +// Event set used by most benchmarks. +const evBench Set = 1 + +// BenchmarkXxxNotifyRedundant measures how long it takes to notify a Waiter of +// an event that is already pending. + +func BenchmarkWaiterNotifyRedundant(b *testing.B) { + var w Waiter + w.Init() + w.Notify(evBench) + + b.ResetTimer() + for i := 0; i < b.N; i++ { + w.Notify(evBench) + } +} + +func BenchmarkSleeperNotifyRedundant(b *testing.B) { + var s sleep.Sleeper + var w sleep.Waker + s.AddWaker(&w, 0) + w.Assert() + + b.ResetTimer() + for i := 0; i < b.N; i++ { + w.Assert() + } +} + +func BenchmarkChannelNotifyRedundant(b *testing.B) { + ch := make(chan struct{}, 1) + ch <- struct{}{} + + b.ResetTimer() + for i := 0; i < b.N; i++ { + select { + case ch <- struct{}{}: + default: + } + } +} + +// BenchmarkXxxNotifyWaitAck measures how long it takes to notify a Waiter an +// event, return that event using a blocking check, and then unset the event as +// pending. + +func BenchmarkWaiterNotifyWaitAck(b *testing.B) { + var w Waiter + w.Init() + + b.ResetTimer() + for i := 0; i < b.N; i++ { + w.Notify(evBench) + w.Wait() + w.Ack(evBench) + } +} + +func BenchmarkSleeperNotifyWaitAck(b *testing.B) { + var s sleep.Sleeper + var w sleep.Waker + s.AddWaker(&w, 0) + + b.ResetTimer() + for i := 0; i < b.N; i++ { + w.Assert() + s.Fetch(true) + } +} + +func BenchmarkChannelNotifyWaitAck(b *testing.B) { + ch := make(chan struct{}, 1) + + b.ResetTimer() + for i := 0; i < b.N; i++ { + // notify + select { + case ch <- struct{}{}: + default: + } + + // wait + ack + <-ch + } +} + +// BenchmarkSleeperMultiNotifyWaitAck is equivalent to +// BenchmarkSleeperNotifyWaitAck, but also includes allocation of a +// temporary sleep.Waker. This is necessary when multiple goroutines may wait +// for the same event, since each sleep.Waker can wake only a single +// sleep.Sleeper. +// +// The syncevent package does not require a distinct object for each +// waiter-waker relationship, so BenchmarkWaiterNotifyWaitAck and +// BenchmarkWaiterMultiNotifyWaitAck would be identical. The analogous state +// for channels, runtime.sudog, is inescapably runtime-allocated, so +// BenchmarkChannelNotifyWaitAck and BenchmarkChannelMultiNotifyWaitAck would +// also be identical. + +func BenchmarkSleeperMultiNotifyWaitAck(b *testing.B) { + var s sleep.Sleeper + // The sleep package doesn't provide sync.Pool allocation of Wakers; + // we do for a fairer comparison. + wakerPool := sync.Pool{ + New: func() interface{} { + return &sleep.Waker{} + }, + } + + b.ResetTimer() + for i := 0; i < b.N; i++ { + w := wakerPool.Get().(*sleep.Waker) + s.AddWaker(w, 0) + w.Assert() + s.Fetch(true) + s.Done() + wakerPool.Put(w) + } +} + +// BenchmarkXxxTempNotifyWaitAck is equivalent to NotifyWaitAck, but also +// includes allocation of a temporary Waiter. This models the case where a +// goroutine not already associated with a Waiter needs one in order to block. +// +// The analogous state for channels is built into runtime.g, so +// BenchmarkChannelNotifyWaitAck and BenchmarkChannelTempNotifyWaitAck would be +// identical. + +func BenchmarkWaiterTempNotifyWaitAck(b *testing.B) { + b.ResetTimer() + for i := 0; i < b.N; i++ { + w := GetWaiter() + w.Notify(evBench) + w.Wait() + w.Ack(evBench) + PutWaiter(w) + } +} + +func BenchmarkSleeperTempNotifyWaitAck(b *testing.B) { + // The sleep package doesn't provide sync.Pool allocation of Sleepers; + // we do for a fairer comparison. + sleeperPool := sync.Pool{ + New: func() interface{} { + return &sleep.Sleeper{} + }, + } + var w sleep.Waker + + b.ResetTimer() + for i := 0; i < b.N; i++ { + s := sleeperPool.Get().(*sleep.Sleeper) + s.AddWaker(&w, 0) + w.Assert() + s.Fetch(true) + s.Done() + sleeperPool.Put(s) + } +} + +// BenchmarkXxxNotifyWaitMultiAck is equivalent to NotifyWaitAck, but allows +// for multiple event sources. + +func BenchmarkWaiterNotifyWaitMultiAck(b *testing.B) { + var w Waiter + w.Init() + + b.ResetTimer() + for i := 0; i < b.N; i++ { + w.Notify(evBench) + if e := w.Wait(); e != evBench { + b.Fatalf("Wait: got %#x, wanted %#x", e, evBench) + } + w.Ack(evBench) + } +} + +func BenchmarkSleeperNotifyWaitMultiAck(b *testing.B) { + var s sleep.Sleeper + var ws [3]sleep.Waker + for i := range ws { + s.AddWaker(&ws[i], i) + } + + b.ResetTimer() + for i := 0; i < b.N; i++ { + ws[0].Assert() + if id, _ := s.Fetch(true); id != 0 { + b.Fatalf("Fetch: got %d, wanted 0", id) + } + } +} + +func BenchmarkChannelNotifyWaitMultiAck(b *testing.B) { + ch0 := make(chan struct{}, 1) + ch1 := make(chan struct{}, 1) + ch2 := make(chan struct{}, 1) + + b.ResetTimer() + for i := 0; i < b.N; i++ { + // notify + select { + case ch0 <- struct{}{}: + default: + } + + // wait + clear + select { + case <-ch0: + // ok + case <-ch1: + b.Fatalf("received from ch1") + case <-ch2: + b.Fatalf("received from ch2") + } + } +} + +// BenchmarkXxxNotifyAsyncWaitAck measures how long it takes to wait for an +// event while another goroutine signals the event. This assumes that a new +// goroutine doesn't run immediately (i.e. the creator of a new goroutine is +// allowed to go to sleep before the new goroutine has a chance to run). + +func BenchmarkWaiterNotifyAsyncWaitAck(b *testing.B) { + var w Waiter + w.Init() + + b.ResetTimer() + for i := 0; i < b.N; i++ { + go func() { + w.Notify(1) + }() + w.Wait() + w.Ack(evBench) + } +} + +func BenchmarkSleeperNotifyAsyncWaitAck(b *testing.B) { + var s sleep.Sleeper + var w sleep.Waker + s.AddWaker(&w, 0) + + b.ResetTimer() + for i := 0; i < b.N; i++ { + go func() { + w.Assert() + }() + s.Fetch(true) + } +} + +func BenchmarkChannelNotifyAsyncWaitAck(b *testing.B) { + ch := make(chan struct{}, 1) + + b.ResetTimer() + for i := 0; i < b.N; i++ { + go func() { + select { + case ch <- struct{}{}: + default: + } + }() + <-ch + } +} + +// BenchmarkXxxNotifyAsyncWaitMultiAck is equivalent to NotifyAsyncWaitAck, but +// allows for multiple event sources. + +func BenchmarkWaiterNotifyAsyncWaitMultiAck(b *testing.B) { + var w Waiter + w.Init() + + b.ResetTimer() + for i := 0; i < b.N; i++ { + go func() { + w.Notify(evBench) + }() + if e := w.Wait(); e != evBench { + b.Fatalf("Wait: got %#x, wanted %#x", e, evBench) + } + w.Ack(evBench) + } +} + +func BenchmarkSleeperNotifyAsyncWaitMultiAck(b *testing.B) { + var s sleep.Sleeper + var ws [3]sleep.Waker + for i := range ws { + s.AddWaker(&ws[i], i) + } + + b.ResetTimer() + for i := 0; i < b.N; i++ { + go func() { + ws[0].Assert() + }() + if id, _ := s.Fetch(true); id != 0 { + b.Fatalf("Fetch: got %d, expected 0", id) + } + } +} + +func BenchmarkChannelNotifyAsyncWaitMultiAck(b *testing.B) { + ch0 := make(chan struct{}, 1) + ch1 := make(chan struct{}, 1) + ch2 := make(chan struct{}, 1) + + b.ResetTimer() + for i := 0; i < b.N; i++ { + go func() { + select { + case ch0 <- struct{}{}: + default: + } + }() + + select { + case <-ch0: + // ok + case <-ch1: + b.Fatalf("received from ch1") + case <-ch2: + b.Fatalf("received from ch2") + } + } +} diff --git a/pkg/syncevent/waiter_unsafe.go b/pkg/syncevent/waiter_unsafe.go new file mode 100644 index 000000000..112e0e604 --- /dev/null +++ b/pkg/syncevent/waiter_unsafe.go @@ -0,0 +1,206 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// +build go1.11 +// +build !go1.15 + +// Check go:linkname function signatures when updating Go version. + +package syncevent + +import ( + "sync/atomic" + "unsafe" + + "gvisor.dev/gvisor/pkg/sync" +) + +//go:linkname gopark runtime.gopark +func gopark(unlockf func(unsafe.Pointer, *unsafe.Pointer) bool, wg *unsafe.Pointer, reason uint8, traceEv byte, traceskip int) + +//go:linkname goready runtime.goready +func goready(g unsafe.Pointer, traceskip int) + +const ( + waitReasonSelect = 9 // Go: src/runtime/runtime2.go + traceEvGoBlockSelect = 24 // Go: src/runtime/trace.go +) + +// Waiter allows a goroutine to block on pending events received by a Receiver. +// +// Waiter.Init() must be called before first use. +type Waiter struct { + r Receiver + + // g is one of: + // + // - nil: No goroutine is blocking in Wait. + // + // - &preparingG: A goroutine is in Wait preparing to sleep, but hasn't yet + // completed waiterUnlock(). Thus the wait can only be interrupted by + // replacing the value of g with nil (the G may not be in state Gwaiting + // yet, so we can't call goready.) + // + // - Otherwise: g is a pointer to the runtime.g in state Gwaiting for the + // goroutine blocked in Wait, which can only be woken by calling goready. + g unsafe.Pointer `state:"zerovalue"` +} + +// Sentinel object for Waiter.g. +var preparingG struct{} + +// Init must be called before first use of w. +func (w *Waiter) Init() { + w.r.Init(w) +} + +// Receiver returns the Receiver that receives events that unblock calls to +// w.Wait(). +func (w *Waiter) Receiver() *Receiver { + return &w.r +} + +// Pending returns the set of pending events. +func (w *Waiter) Pending() Set { + return w.r.Pending() +} + +// Wait blocks until at least one event is pending, then returns the set of +// pending events. It does not affect the set of pending events; callers must +// call w.Ack() to do so, or use w.WaitAndAck() instead. +// +// Precondition: Only one goroutine may call any Wait* method at a time. +func (w *Waiter) Wait() Set { + return w.WaitFor(AllEvents) +} + +// WaitFor blocks until at least one event in es is pending, then returns the +// set of pending events (including those not in es). It does not affect the +// set of pending events; callers must call w.Ack() to do so. +// +// Precondition: Only one goroutine may call any Wait* method at a time. +func (w *Waiter) WaitFor(es Set) Set { + for { + // Optimization: Skip the atomic store to w.g if an event is already + // pending. + if p := w.r.Pending(); p&es != NoEvents { + return p + } + + // Indicate that we're preparing to go to sleep. + atomic.StorePointer(&w.g, (unsafe.Pointer)(&preparingG)) + + // If an event is pending, abort the sleep. + if p := w.r.Pending(); p&es != NoEvents { + atomic.StorePointer(&w.g, nil) + return p + } + + // If w.g is still preparingG (i.e. w.NotifyPending() has not been + // called or has not reached atomic.SwapPointer()), go to sleep until + // w.NotifyPending() => goready(). + gopark(waiterUnlock, &w.g, waitReasonSelect, traceEvGoBlockSelect, 0) + } +} + +// Ack marks the given events as not pending. +func (w *Waiter) Ack(es Set) { + w.r.Ack(es) +} + +// WaitAndAckAll blocks until at least one event is pending, then marks all +// events as not pending and returns the set of previously-pending events. +// +// Precondition: Only one goroutine may call any Wait* method at a time. +func (w *Waiter) WaitAndAckAll() Set { + // Optimization: Skip the atomic store to w.g if an event is already + // pending. Call Pending() first since, in the common case that events are + // not yet pending, this skips an atomic swap on w.r.pending. + if w.r.Pending() != NoEvents { + if p := w.r.PendingAndAckAll(); p != NoEvents { + return p + } + } + + for { + // Indicate that we're preparing to go to sleep. + atomic.StorePointer(&w.g, (unsafe.Pointer)(&preparingG)) + + // If an event is pending, abort the sleep. + if w.r.Pending() != NoEvents { + if p := w.r.PendingAndAckAll(); p != NoEvents { + atomic.StorePointer(&w.g, nil) + return p + } + } + + // If w.g is still preparingG (i.e. w.NotifyPending() has not been + // called or has not reached atomic.SwapPointer()), go to sleep until + // w.NotifyPending() => goready(). + gopark(waiterUnlock, &w.g, waitReasonSelect, traceEvGoBlockSelect, 0) + + // Check for pending events. We call PendingAndAckAll() directly now since + // we only expect to be woken after events become pending. + if p := w.r.PendingAndAckAll(); p != NoEvents { + return p + } + } +} + +// Notify marks the given events as pending, possibly unblocking concurrent +// calls to w.Wait() or w.WaitFor(). +func (w *Waiter) Notify(es Set) { + w.r.Notify(es) +} + +// NotifyPending implements ReceiverCallback.NotifyPending. Users of Waiter +// should not call NotifyPending. +func (w *Waiter) NotifyPending() { + // Optimization: Skip the atomic swap on w.g if there is no sleeping + // goroutine. NotifyPending is called after w.r.Pending() is updated, so + // concurrent and future calls to w.Wait() will observe pending events and + // abort sleeping. + if atomic.LoadPointer(&w.g) == nil { + return + } + // Wake a sleeping G, or prevent a G that is preparing to sleep from doing + // so. Swap is needed here to ensure that only one call to NotifyPending + // calls goready. + if g := atomic.SwapPointer(&w.g, nil); g != nil && g != (unsafe.Pointer)(&preparingG) { + goready(g, 0) + } +} + +var waiterPool = sync.Pool{ + New: func() interface{} { + w := &Waiter{} + w.Init() + return w + }, +} + +// GetWaiter returns an unused Waiter. PutWaiter should be called to release +// the Waiter once it is no longer needed. +// +// Where possible, users should prefer to associate each goroutine that calls +// Waiter.Wait() with a distinct pre-allocated Waiter to avoid allocation of +// Waiters in hot paths. +func GetWaiter() *Waiter { + return waiterPool.Get().(*Waiter) +} + +// PutWaiter releases an unused Waiter previously returned by GetWaiter. +func PutWaiter(w *Waiter) { + waiterPool.Put(w) +} diff --git a/pkg/syserror/syserror.go b/pkg/syserror/syserror.go index 2269f6237..f86db0999 100644 --- a/pkg/syserror/syserror.go +++ b/pkg/syserror/syserror.go @@ -27,8 +27,10 @@ import ( var ( E2BIG = error(syscall.E2BIG) EACCES = error(syscall.EACCES) + EADDRINUSE = error(syscall.EADDRINUSE) EAGAIN = error(syscall.EAGAIN) EBADF = error(syscall.EBADF) + EBADFD = error(syscall.EBADFD) EBUSY = error(syscall.EBUSY) ECHILD = error(syscall.ECHILD) ECONNREFUSED = error(syscall.ECONNREFUSED) diff --git a/pkg/tcpip/BUILD b/pkg/tcpip/BUILD index 26f7ba86b..454e07662 100644 --- a/pkg/tcpip/BUILD +++ b/pkg/tcpip/BUILD @@ -5,8 +5,6 @@ package(licenses = ["notice"]) go_library( name = "tcpip", srcs = [ - "packet_buffer.go", - "packet_buffer_state.go", "tcpip.go", "time_unsafe.go", "timer.go", diff --git a/pkg/tcpip/adapters/gonet/BUILD b/pkg/tcpip/adapters/gonet/BUILD index a984f1712..e57d45f2a 100644 --- a/pkg/tcpip/adapters/gonet/BUILD +++ b/pkg/tcpip/adapters/gonet/BUILD @@ -22,6 +22,7 @@ go_test( size = "small", srcs = ["gonet_test.go"], library = ":gonet", + tags = ["flaky"], deps = [ "//pkg/tcpip", "//pkg/tcpip/header", diff --git a/pkg/tcpip/adapters/gonet/gonet_test.go b/pkg/tcpip/adapters/gonet/gonet_test.go index ea0a0409a..3c552988a 100644 --- a/pkg/tcpip/adapters/gonet/gonet_test.go +++ b/pkg/tcpip/adapters/gonet/gonet_test.go @@ -127,6 +127,10 @@ func TestCloseReader(t *testing.T) { if err != nil { t.Fatalf("newLoopbackStack() = %v", err) } + defer func() { + s.Close() + s.Wait() + }() addr := tcpip.FullAddress{NICID, tcpip.Address(net.IPv4(169, 254, 10, 1).To4()), 11211} @@ -175,6 +179,10 @@ func TestCloseReaderWithForwarder(t *testing.T) { if err != nil { t.Fatalf("newLoopbackStack() = %v", err) } + defer func() { + s.Close() + s.Wait() + }() addr := tcpip.FullAddress{NICID, tcpip.Address(net.IPv4(169, 254, 10, 1).To4()), 11211} s.AddAddress(NICID, ipv4.ProtocolNumber, addr.Addr) @@ -225,30 +233,21 @@ func TestCloseRead(t *testing.T) { if terr != nil { t.Fatalf("newLoopbackStack() = %v", terr) } + defer func() { + s.Close() + s.Wait() + }() addr := tcpip.FullAddress{NICID, tcpip.Address(net.IPv4(169, 254, 10, 1).To4()), 11211} s.AddAddress(NICID, ipv4.ProtocolNumber, addr.Addr) fwd := tcp.NewForwarder(s, 30000, 10, func(r *tcp.ForwarderRequest) { var wq waiter.Queue - ep, err := r.CreateEndpoint(&wq) + _, err := r.CreateEndpoint(&wq) if err != nil { t.Fatalf("r.CreateEndpoint() = %v", err) } - defer ep.Close() - r.Complete(false) - - c := NewTCPConn(&wq, ep) - - buf := make([]byte, 256) - n, e := c.Read(buf) - if e != nil || string(buf[:n]) != "abc123" { - t.Fatalf("c.Read() = (%d, %v), want (6, nil)", n, e) - } - - if n, e = c.Write([]byte("abc123")); e != nil { - t.Errorf("c.Write() = (%d, %v), want (6, nil)", n, e) - } + // Endpoint will be closed in deferred s.Close (above). }) s.SetTransportProtocolHandler(tcp.ProtocolNumber, fwd.HandlePacket) @@ -278,6 +277,10 @@ func TestCloseWrite(t *testing.T) { if terr != nil { t.Fatalf("newLoopbackStack() = %v", terr) } + defer func() { + s.Close() + s.Wait() + }() addr := tcpip.FullAddress{NICID, tcpip.Address(net.IPv4(169, 254, 10, 1).To4()), 11211} s.AddAddress(NICID, ipv4.ProtocolNumber, addr.Addr) @@ -334,6 +337,10 @@ func TestUDPForwarder(t *testing.T) { if terr != nil { t.Fatalf("newLoopbackStack() = %v", terr) } + defer func() { + s.Close() + s.Wait() + }() ip1 := tcpip.Address(net.IPv4(169, 254, 10, 1).To4()) addr1 := tcpip.FullAddress{NICID, ip1, 11211} @@ -391,6 +398,10 @@ func TestDeadlineChange(t *testing.T) { if err != nil { t.Fatalf("newLoopbackStack() = %v", err) } + defer func() { + s.Close() + s.Wait() + }() addr := tcpip.FullAddress{NICID, tcpip.Address(net.IPv4(169, 254, 10, 1).To4()), 11211} @@ -440,6 +451,10 @@ func TestPacketConnTransfer(t *testing.T) { if e != nil { t.Fatalf("newLoopbackStack() = %v", e) } + defer func() { + s.Close() + s.Wait() + }() ip1 := tcpip.Address(net.IPv4(169, 254, 10, 1).To4()) addr1 := tcpip.FullAddress{NICID, ip1, 11211} @@ -492,6 +507,10 @@ func TestConnectedPacketConnTransfer(t *testing.T) { if e != nil { t.Fatalf("newLoopbackStack() = %v", e) } + defer func() { + s.Close() + s.Wait() + }() ip := tcpip.Address(net.IPv4(169, 254, 10, 1).To4()) addr := tcpip.FullAddress{NICID, ip, 11211} @@ -562,6 +581,8 @@ func makePipe() (c1, c2 net.Conn, stop func(), err error) { stop = func() { c1.Close() c2.Close() + s.Close() + s.Wait() } if err := l.Close(); err != nil { @@ -624,6 +645,10 @@ func TestTCPDialError(t *testing.T) { if e != nil { t.Fatalf("newLoopbackStack() = %v", e) } + defer func() { + s.Close() + s.Wait() + }() ip := tcpip.Address(net.IPv4(169, 254, 10, 1).To4()) addr := tcpip.FullAddress{NICID, ip, 11211} @@ -641,6 +666,10 @@ func TestDialContextTCPCanceled(t *testing.T) { if err != nil { t.Fatalf("newLoopbackStack() = %v", err) } + defer func() { + s.Close() + s.Wait() + }() addr := tcpip.FullAddress{NICID, tcpip.Address(net.IPv4(169, 254, 10, 1).To4()), 11211} s.AddAddress(NICID, ipv4.ProtocolNumber, addr.Addr) @@ -659,6 +688,10 @@ func TestDialContextTCPTimeout(t *testing.T) { if err != nil { t.Fatalf("newLoopbackStack() = %v", err) } + defer func() { + s.Close() + s.Wait() + }() addr := tcpip.FullAddress{NICID, tcpip.Address(net.IPv4(169, 254, 10, 1).To4()), 11211} s.AddAddress(NICID, ipv4.ProtocolNumber, addr.Addr) diff --git a/pkg/tcpip/buffer/view.go b/pkg/tcpip/buffer/view.go index 150310c11..8ec5d5d5c 100644 --- a/pkg/tcpip/buffer/view.go +++ b/pkg/tcpip/buffer/view.go @@ -15,6 +15,11 @@ // Package buffer provides the implementation of a buffer view. package buffer +import ( + "bytes" + "io" +) + // View is a slice of a buffer, with convenience methods. type View []byte @@ -45,6 +50,13 @@ func (v *View) CapLength(length int) { *v = (*v)[:length:length] } +// Reader returns a bytes.Reader for v. +func (v *View) Reader() bytes.Reader { + var r bytes.Reader + r.Reset(*v) + return r +} + // ToVectorisedView returns a VectorisedView containing the receiver. func (v View) ToVectorisedView() VectorisedView { return NewVectorisedView(len(v), []View{v}) @@ -78,6 +90,47 @@ func (vv *VectorisedView) TrimFront(count int) { } } +// Read implements io.Reader. +func (vv *VectorisedView) Read(v View) (copied int, err error) { + count := len(v) + for count > 0 && len(vv.views) > 0 { + if count < len(vv.views[0]) { + vv.size -= count + copy(v[copied:], vv.views[0][:count]) + vv.views[0].TrimFront(count) + copied += count + return copied, nil + } + count -= len(vv.views[0]) + copy(v[copied:], vv.views[0]) + copied += len(vv.views[0]) + vv.RemoveFirst() + } + if copied == 0 { + return 0, io.EOF + } + return copied, nil +} + +// ReadToVV reads up to n bytes from vv to dstVV and removes them from vv. It +// returns the number of bytes copied. +func (vv *VectorisedView) ReadToVV(dstVV *VectorisedView, count int) (copied int) { + for count > 0 && len(vv.views) > 0 { + if count < len(vv.views[0]) { + vv.size -= count + dstVV.AppendView(vv.views[0][:count]) + vv.views[0].TrimFront(count) + copied += count + return + } + count -= len(vv.views[0]) + dstVV.AppendView(vv.views[0]) + copied += len(vv.views[0]) + vv.RemoveFirst() + } + return copied +} + // CapLength irreversibly reduces the length of the vectorised view. func (vv *VectorisedView) CapLength(length int) { if length < 0 { @@ -105,12 +158,12 @@ func (vv *VectorisedView) CapLength(length int) { // Clone returns a clone of this VectorisedView. // If the buffer argument is large enough to contain all the Views of this VectorisedView, // the method will avoid allocations and use the buffer to store the Views of the clone. -func (vv VectorisedView) Clone(buffer []View) VectorisedView { +func (vv *VectorisedView) Clone(buffer []View) VectorisedView { return VectorisedView{views: append(buffer[:0], vv.views...), size: vv.size} } // First returns the first view of the vectorised view. -func (vv VectorisedView) First() View { +func (vv *VectorisedView) First() View { if len(vv.views) == 0 { return nil } @@ -123,11 +176,12 @@ func (vv *VectorisedView) RemoveFirst() { return } vv.size -= len(vv.views[0]) + vv.views[0] = nil vv.views = vv.views[1:] } // Size returns the size in bytes of the entire content stored in the vectorised view. -func (vv VectorisedView) Size() int { +func (vv *VectorisedView) Size() int { return vv.size } @@ -135,7 +189,7 @@ func (vv VectorisedView) Size() int { // // If the vectorised view contains a single view, that view will be returned // directly. -func (vv VectorisedView) ToView() View { +func (vv *VectorisedView) ToView() View { if len(vv.views) == 1 { return vv.views[0] } @@ -147,7 +201,7 @@ func (vv VectorisedView) ToView() View { } // Views returns the slice containing the all views. -func (vv VectorisedView) Views() []View { +func (vv *VectorisedView) Views() []View { return vv.views } @@ -156,3 +210,18 @@ func (vv *VectorisedView) Append(vv2 VectorisedView) { vv.views = append(vv.views, vv2.views...) vv.size += vv2.size } + +// AppendView appends the given view into this vectorised view. +func (vv *VectorisedView) AppendView(v View) { + vv.views = append(vv.views, v) + vv.size += len(v) +} + +// Readers returns a bytes.Reader for each of vv's views. +func (vv *VectorisedView) Readers() []bytes.Reader { + readers := make([]bytes.Reader, 0, len(vv.views)) + for _, v := range vv.views { + readers = append(readers, v.Reader()) + } + return readers +} diff --git a/pkg/tcpip/buffer/view_test.go b/pkg/tcpip/buffer/view_test.go index ebc3a17b7..106e1994c 100644 --- a/pkg/tcpip/buffer/view_test.go +++ b/pkg/tcpip/buffer/view_test.go @@ -233,3 +233,140 @@ func TestToClone(t *testing.T) { }) } } + +func TestVVReadToVV(t *testing.T) { + testCases := []struct { + comment string + vv VectorisedView + bytesToRead int + wantBytes string + leftVV VectorisedView + }{ + { + comment: "large VV, short read", + vv: vv(30, "012345678901234567890123456789"), + bytesToRead: 10, + wantBytes: "0123456789", + leftVV: vv(20, "01234567890123456789"), + }, + { + comment: "largeVV, multiple views, short read", + vv: vv(13, "123", "345", "567", "8910"), + bytesToRead: 6, + wantBytes: "123345", + leftVV: vv(7, "567", "8910"), + }, + { + comment: "smallVV (multiple views), large read", + vv: vv(3, "1", "2", "3"), + bytesToRead: 10, + wantBytes: "123", + leftVV: vv(0, ""), + }, + { + comment: "smallVV (single view), large read", + vv: vv(1, "1"), + bytesToRead: 10, + wantBytes: "1", + leftVV: vv(0, ""), + }, + { + comment: "emptyVV, large read", + vv: vv(0, ""), + bytesToRead: 10, + wantBytes: "", + leftVV: vv(0, ""), + }, + } + + for _, tc := range testCases { + t.Run(tc.comment, func(t *testing.T) { + var readTo VectorisedView + inSize := tc.vv.Size() + copied := tc.vv.ReadToVV(&readTo, tc.bytesToRead) + if got, want := copied, len(tc.wantBytes); got != want { + t.Errorf("incorrect number of bytes copied returned in ReadToVV got: %d, want: %d, tc: %+v", got, want, tc) + } + if got, want := string(readTo.ToView()), tc.wantBytes; got != want { + t.Errorf("unexpected content in readTo got: %s, want: %s", got, want) + } + if got, want := tc.vv.Size(), inSize-copied; got != want { + t.Errorf("test VV has incorrect size after reading got: %d, want: %d, tc.vv: %+v", got, want, tc.vv) + } + if got, want := string(tc.vv.ToView()), string(tc.leftVV.ToView()); got != want { + t.Errorf("unexpected data left in vv after read got: %+v, want: %+v", got, want) + } + }) + } +} + +func TestVVRead(t *testing.T) { + testCases := []struct { + comment string + vv VectorisedView + bytesToRead int + readBytes string + leftBytes string + wantError bool + }{ + { + comment: "large VV, short read", + vv: vv(30, "012345678901234567890123456789"), + bytesToRead: 10, + readBytes: "0123456789", + leftBytes: "01234567890123456789", + }, + { + comment: "largeVV, multiple buffers, short read", + vv: vv(13, "123", "345", "567", "8910"), + bytesToRead: 6, + readBytes: "123345", + leftBytes: "5678910", + }, + { + comment: "smallVV, large read", + vv: vv(3, "1", "2", "3"), + bytesToRead: 10, + readBytes: "123", + leftBytes: "", + }, + { + comment: "smallVV, large read", + vv: vv(1, "1"), + bytesToRead: 10, + readBytes: "1", + leftBytes: "", + }, + { + comment: "emptyVV, large read", + vv: vv(0, ""), + bytesToRead: 10, + readBytes: "", + wantError: true, + }, + } + + for _, tc := range testCases { + t.Run(tc.comment, func(t *testing.T) { + readTo := NewView(tc.bytesToRead) + inSize := tc.vv.Size() + copied, err := tc.vv.Read(readTo) + if !tc.wantError && err != nil { + t.Fatalf("unexpected error in tc.vv.Read(..) = %s", err) + } + readTo = readTo[:copied] + if got, want := copied, len(tc.readBytes); got != want { + t.Errorf("incorrect number of bytes copied returned in ReadToVV got: %d, want: %d, tc.vv: %+v", got, want, tc.vv) + } + if got, want := string(readTo), tc.readBytes; got != want { + t.Errorf("unexpected data in readTo got: %s, want: %s", got, want) + } + if got, want := tc.vv.Size(), inSize-copied; got != want { + t.Errorf("test VV has incorrect size after reading got: %d, want: %d, tc.vv: %+v", got, want, tc.vv) + } + if got, want := string(tc.vv.ToView()), tc.leftBytes; got != want { + t.Errorf("vv has incorrect data after Read got: %s, want: %s", got, want) + } + }) + } +} diff --git a/pkg/tcpip/checker/checker.go b/pkg/tcpip/checker/checker.go index 4d6ae0871..c1745ba6a 100644 --- a/pkg/tcpip/checker/checker.go +++ b/pkg/tcpip/checker/checker.go @@ -107,6 +107,8 @@ func DstAddr(addr tcpip.Address) NetworkChecker { // TTL creates a checker that checks the TTL (ipv4) or HopLimit (ipv6). func TTL(ttl uint8) NetworkChecker { return func(t *testing.T, h []header.Network) { + t.Helper() + var v uint8 switch ip := h[0].(type) { case header.IPv4: @@ -161,6 +163,20 @@ func FragmentFlags(flags uint8) NetworkChecker { } } +// ReceiveTClass creates a checker that checks the TCLASS field in +// ControlMessages. +func ReceiveTClass(want uint32) ControlMessagesChecker { + return func(t *testing.T, cm tcpip.ControlMessages) { + t.Helper() + if !cm.HasTClass { + t.Fatalf("got cm.HasTClass = %t, want cm.TClass = %d", cm.HasTClass, want) + } + if got := cm.TClass; got != want { + t.Fatalf("got cm.TClass = %d, want %d", got, want) + } + } +} + // ReceiveTOS creates a checker that checks the TOS field in ControlMessages. func ReceiveTOS(want uint8) ControlMessagesChecker { return func(t *testing.T, cm tcpip.ControlMessages) { @@ -296,6 +312,8 @@ func SrcPort(port uint16) TransportChecker { // DstPort creates a checker that checks the destination port. func DstPort(port uint16) TransportChecker { return func(t *testing.T, h header.Transport) { + t.Helper() + if p := h.DestinationPort(); p != port { t.Errorf("Bad destination port, got %v, want %v", p, port) } @@ -322,6 +340,7 @@ func SeqNum(seq uint32) TransportChecker { func AckNum(seq uint32) TransportChecker { return func(t *testing.T, h header.Transport) { t.Helper() + tcp, ok := h.(header.TCP) if !ok { return @@ -336,6 +355,8 @@ func AckNum(seq uint32) TransportChecker { // Window creates a checker that checks the tcp window. func Window(window uint16) TransportChecker { return func(t *testing.T, h header.Transport) { + t.Helper() + tcp, ok := h.(header.TCP) if !ok { return @@ -367,6 +388,8 @@ func TCPFlags(flags uint8) TransportChecker { // given mask, match the supplied flags. func TCPFlagsMatch(flags, mask uint8) TransportChecker { return func(t *testing.T, h header.Transport) { + t.Helper() + tcp, ok := h.(header.TCP) if !ok { return @@ -384,6 +407,8 @@ func TCPFlagsMatch(flags, mask uint8) TransportChecker { // If wndscale is negative, the window scale option must not be present. func TCPSynOptions(wantOpts header.TCPSynOptions) TransportChecker { return func(t *testing.T, h header.Transport) { + t.Helper() + tcp, ok := h.(header.TCP) if !ok { return @@ -480,6 +505,8 @@ func TCPSynOptions(wantOpts header.TCPSynOptions) TransportChecker { // skipped. func TCPTimestampChecker(wantTS bool, wantTSVal uint32, wantTSEcr uint32) TransportChecker { return func(t *testing.T, h header.Transport) { + t.Helper() + tcp, ok := h.(header.TCP) if !ok { return @@ -598,6 +625,8 @@ func TCPSACKBlockChecker(sackBlocks []header.SACKBlock) TransportChecker { // Payload creates a checker that checks the payload. func Payload(want []byte) TransportChecker { return func(t *testing.T, h header.Transport) { + t.Helper() + if got := h.Payload(); !reflect.DeepEqual(got, want) { t.Errorf("Wrong payload, got %v, want %v", got, want) } @@ -630,6 +659,7 @@ func ICMPv4(checkers ...TransportChecker) NetworkChecker { func ICMPv4Type(want header.ICMPv4Type) TransportChecker { return func(t *testing.T, h header.Transport) { t.Helper() + icmpv4, ok := h.(header.ICMPv4) if !ok { t.Fatalf("unexpected transport header passed to checker got: %+v, want: header.ICMPv4", h) @@ -644,6 +674,7 @@ func ICMPv4Type(want header.ICMPv4Type) TransportChecker { func ICMPv4Code(want byte) TransportChecker { return func(t *testing.T, h header.Transport) { t.Helper() + icmpv4, ok := h.(header.ICMPv4) if !ok { t.Fatalf("unexpected transport header passed to checker got: %+v, want: header.ICMPv4", h) @@ -686,6 +717,7 @@ func ICMPv6(checkers ...TransportChecker) NetworkChecker { func ICMPv6Type(want header.ICMPv6Type) TransportChecker { return func(t *testing.T, h header.Transport) { t.Helper() + icmpv6, ok := h.(header.ICMPv6) if !ok { t.Fatalf("unexpected transport header passed to checker got: %+v, want: header.ICMPv6", h) @@ -700,6 +732,7 @@ func ICMPv6Type(want header.ICMPv6Type) TransportChecker { func ICMPv6Code(want byte) TransportChecker { return func(t *testing.T, h header.Transport) { t.Helper() + icmpv6, ok := h.(header.ICMPv6) if !ok { t.Fatalf("unexpected transport header passed to checker got: %+v, want: header.ICMPv6", h) @@ -714,7 +747,7 @@ func ICMPv6Code(want byte) TransportChecker { // message for type of ty, with potentially additional checks specified by // checkers. // -// checkers may assume that a valid ICMPv6 is passed to it containing a valid +// Checkers may assume that a valid ICMPv6 is passed to it containing a valid // NDP message as far as the size of the message (minSize) is concerned. The // values within the message are up to checkers to validate. func NDP(msgType header.ICMPv6Type, minSize int, checkers ...TransportChecker) NetworkChecker { @@ -746,9 +779,9 @@ func NDP(msgType header.ICMPv6Type, minSize int, checkers ...TransportChecker) N // Neighbor Solicitation message (as per the raw wire format), with potentially // additional checks specified by checkers. // -// checkers may assume that a valid ICMPv6 is passed to it containing a valid -// NDPNS message as far as the size of the messages concerned. The values within -// the message are up to checkers to validate. +// Checkers may assume that a valid ICMPv6 is passed to it containing a valid +// NDPNS message as far as the size of the message is concerned. The values +// within the message are up to checkers to validate. func NDPNS(checkers ...TransportChecker) NetworkChecker { return NDP(header.ICMPv6NeighborSolicit, header.NDPNSMinimumSize, checkers...) } @@ -766,63 +799,162 @@ func NDPNSTargetAddress(want tcpip.Address) TransportChecker { ns := header.NDPNeighborSolicit(icmp.NDPPayload()) if got := ns.TargetAddress(); got != want { - t.Fatalf("got %T.TargetAddress = %s, want = %s", ns, got, want) + t.Errorf("got %T.TargetAddress() = %s, want = %s", ns, got, want) } } } -// NDPNSOptions creates a checker that checks that the packet contains the -// provided NDP options within an NDP Neighbor Solicitation message. +// NDPNA creates a checker that checks that the packet contains a valid NDP +// Neighbor Advertisement message (as per the raw wire format), with potentially +// additional checks specified by checkers. +// +// Checkers may assume that a valid ICMPv6 is passed to it containing a valid +// NDPNA message as far as the size of the message is concerned. The values +// within the message are up to checkers to validate. +func NDPNA(checkers ...TransportChecker) NetworkChecker { + return NDP(header.ICMPv6NeighborAdvert, header.NDPNAMinimumSize, checkers...) +} + +// NDPNATargetAddress creates a checker that checks the Target Address field of +// a header.NDPNeighborAdvert. // // The returned TransportChecker assumes that a valid ICMPv6 is passed to it -// containing a valid NDPNS message as far as the size is concerned. -func NDPNSOptions(opts []header.NDPOption) TransportChecker { +// containing a valid NDPNA message as far as the size is concerned. +func NDPNATargetAddress(want tcpip.Address) TransportChecker { return func(t *testing.T, h header.Transport) { t.Helper() icmp := h.(header.ICMPv6) - ns := header.NDPNeighborSolicit(icmp.NDPPayload()) - it, err := ns.Options().Iter(true) - if err != nil { - t.Errorf("opts.Iter(true): %s", err) - return + na := header.NDPNeighborAdvert(icmp.NDPPayload()) + + if got := na.TargetAddress(); got != want { + t.Errorf("got %T.TargetAddress() = %s, want = %s", na, got, want) } + } +} - i := 0 - for { - opt, done, _ := it.Next() - if done { - break - } +// NDPNASolicitedFlag creates a checker that checks the Solicited field of +// a header.NDPNeighborAdvert. +// +// The returned TransportChecker assumes that a valid ICMPv6 is passed to it +// containing a valid NDPNA message as far as the size is concerned. +func NDPNASolicitedFlag(want bool) TransportChecker { + return func(t *testing.T, h header.Transport) { + t.Helper() - if i >= len(opts) { - t.Errorf("got unexpected option: %s", opt) - continue - } + icmp := h.(header.ICMPv6) + na := header.NDPNeighborAdvert(icmp.NDPPayload()) - switch wantOpt := opts[i].(type) { - case header.NDPSourceLinkLayerAddressOption: - gotOpt, ok := opt.(header.NDPSourceLinkLayerAddressOption) - if !ok { - t.Errorf("got type = %T at index = %d; want = %T", opt, i, wantOpt) - } else if got, want := gotOpt.EthernetAddress(), wantOpt.EthernetAddress(); got != want { - t.Errorf("got EthernetAddress() = %s at index %d, want = %s", got, i, want) - } - default: - panic("not implemented") - } + if got := na.SolicitedFlag(); got != want { + t.Errorf("got %T.SolicitedFlag = %t, want = %t", na, got, want) + } + } +} + +// ndpOptions checks that optsBuf only contains opts. +func ndpOptions(t *testing.T, optsBuf header.NDPOptions, opts []header.NDPOption) { + t.Helper() + + it, err := optsBuf.Iter(true) + if err != nil { + t.Errorf("optsBuf.Iter(true): %s", err) + return + } - i++ + i := 0 + for { + opt, done, err := it.Next() + if err != nil { + // This should never happen as Iter(true) above did not return an error. + t.Fatalf("unexpected error when iterating over NDP options: %s", err) + } + if done { + break + } + + if i >= len(opts) { + t.Errorf("got unexpected option: %s", opt) + continue } - if missing := opts[i:]; len(missing) > 0 { - t.Errorf("missing options: %s", missing) + switch wantOpt := opts[i].(type) { + case header.NDPSourceLinkLayerAddressOption: + gotOpt, ok := opt.(header.NDPSourceLinkLayerAddressOption) + if !ok { + t.Errorf("got type = %T at index = %d; want = %T", opt, i, wantOpt) + } else if got, want := gotOpt.EthernetAddress(), wantOpt.EthernetAddress(); got != want { + t.Errorf("got EthernetAddress() = %s at index %d, want = %s", got, i, want) + } + case header.NDPTargetLinkLayerAddressOption: + gotOpt, ok := opt.(header.NDPTargetLinkLayerAddressOption) + if !ok { + t.Errorf("got type = %T at index = %d; want = %T", opt, i, wantOpt) + } else if got, want := gotOpt.EthernetAddress(), wantOpt.EthernetAddress(); got != want { + t.Errorf("got EthernetAddress() = %s at index %d, want = %s", got, i, want) + } + default: + t.Fatalf("checker not implemented for expected NDP option: %T", wantOpt) } + + i++ + } + + if missing := opts[i:]; len(missing) > 0 { + t.Errorf("missing options: %s", missing) + } +} + +// NDPNAOptions creates a checker that checks that the packet contains the +// provided NDP options within an NDP Neighbor Solicitation message. +// +// The returned TransportChecker assumes that a valid ICMPv6 is passed to it +// containing a valid NDPNA message as far as the size is concerned. +func NDPNAOptions(opts []header.NDPOption) TransportChecker { + return func(t *testing.T, h header.Transport) { + t.Helper() + + icmp := h.(header.ICMPv6) + na := header.NDPNeighborAdvert(icmp.NDPPayload()) + ndpOptions(t, na.Options(), opts) + } +} + +// NDPNSOptions creates a checker that checks that the packet contains the +// provided NDP options within an NDP Neighbor Solicitation message. +// +// The returned TransportChecker assumes that a valid ICMPv6 is passed to it +// containing a valid NDPNS message as far as the size is concerned. +func NDPNSOptions(opts []header.NDPOption) TransportChecker { + return func(t *testing.T, h header.Transport) { + t.Helper() + + icmp := h.(header.ICMPv6) + ns := header.NDPNeighborSolicit(icmp.NDPPayload()) + ndpOptions(t, ns.Options(), opts) } } // NDPRS creates a checker that checks that the packet contains a valid NDP // Router Solicitation message (as per the raw wire format). -func NDPRS() NetworkChecker { - return NDP(header.ICMPv6RouterSolicit, header.NDPRSMinimumSize) +// +// Checkers may assume that a valid ICMPv6 is passed to it containing a valid +// NDPRS as far as the size of the message is concerned. The values within the +// message are up to checkers to validate. +func NDPRS(checkers ...TransportChecker) NetworkChecker { + return NDP(header.ICMPv6RouterSolicit, header.NDPRSMinimumSize, checkers...) +} + +// NDPRSOptions creates a checker that checks that the packet contains the +// provided NDP options within an NDP Router Solicitation message. +// +// The returned TransportChecker assumes that a valid ICMPv6 is passed to it +// containing a valid NDPRS message as far as the size is concerned. +func NDPRSOptions(opts []header.NDPOption) TransportChecker { + return func(t *testing.T, h header.Transport) { + t.Helper() + + icmp := h.(header.ICMPv6) + rs := header.NDPRouterSolicit(icmp.NDPPayload()) + ndpOptions(t, rs.Options(), opts) + } } diff --git a/pkg/tcpip/header/BUILD b/pkg/tcpip/header/BUILD index 9da0d71f8..0cde694dc 100644 --- a/pkg/tcpip/header/BUILD +++ b/pkg/tcpip/header/BUILD @@ -14,12 +14,14 @@ go_library( "interfaces.go", "ipv4.go", "ipv6.go", + "ipv6_extension_headers.go", "ipv6_fragment.go", "ndp_neighbor_advert.go", "ndp_neighbor_solicit.go", "ndp_options.go", "ndp_router_advert.go", "ndp_router_solicit.go", + "ndpoptionidentifier_string.go", "tcp.go", "udp.go", ], @@ -55,11 +57,13 @@ go_test( size = "small", srcs = [ "eth_test.go", + "ipv6_extension_headers_test.go", "ndp_test.go", ], library = ":header", deps = [ "//pkg/tcpip", + "//pkg/tcpip/buffer", "@com_github_google_go-cmp//cmp:go_default_library", ], ) diff --git a/pkg/tcpip/header/eth_test.go b/pkg/tcpip/header/eth_test.go index 7a0014ad9..14413f2ce 100644 --- a/pkg/tcpip/header/eth_test.go +++ b/pkg/tcpip/header/eth_test.go @@ -88,7 +88,7 @@ func TestEthernetAddressFromMulticastIPv4Address(t *testing.T) { for _, test := range tests { t.Run(test.name, func(t *testing.T) { if got := EthernetAddressFromMulticastIPv4Address(test.addr); got != test.expectedLinkAddr { - t.Fatalf("got EthernetAddressFromMulticastIPv4Address(%s) = %s, want = %s", got, test.expectedLinkAddr) + t.Fatalf("got EthernetAddressFromMulticastIPv4Address(%s) = %s, want = %s", test.addr, got, test.expectedLinkAddr) } }) } diff --git a/pkg/tcpip/header/ipv4.go b/pkg/tcpip/header/ipv4.go index e5360e7c1..76839eb92 100644 --- a/pkg/tcpip/header/ipv4.go +++ b/pkg/tcpip/header/ipv4.go @@ -38,7 +38,8 @@ const ( // IPv4Fields contains the fields of an IPv4 packet. It is used to describe the // fields of a packet that needs to be encoded. type IPv4Fields struct { - // IHL is the "internet header length" field of an IPv4 packet. + // IHL is the "internet header length" field of an IPv4 packet. The value + // is in bytes. IHL uint8 // TOS is the "type of service" field of an IPv4 packet. @@ -138,7 +139,7 @@ func IPVersion(b []byte) int { } // HeaderLength returns the value of the "header length" field of the ipv4 -// header. +// header. The length returned is in bytes. func (b IPv4) HeaderLength() uint8 { return (b[versIHL] & 0xf) * 4 } diff --git a/pkg/tcpip/header/ipv6.go b/pkg/tcpip/header/ipv6.go index 70e6ce095..76e88e9b3 100644 --- a/pkg/tcpip/header/ipv6.go +++ b/pkg/tcpip/header/ipv6.go @@ -115,6 +115,19 @@ const ( // for the secret key used to generate an opaque interface identifier as // outlined by RFC 7217. OpaqueIIDSecretKeyMinBytes = 16 + + // ipv6MulticastAddressScopeByteIdx is the byte where the scope (scop) field + // is located within a multicast IPv6 address, as per RFC 4291 section 2.7. + ipv6MulticastAddressScopeByteIdx = 1 + + // ipv6MulticastAddressScopeMask is the mask for the scope (scop) field, + // within the byte holding the field, as per RFC 4291 section 2.7. + ipv6MulticastAddressScopeMask = 0xF + + // ipv6LinkLocalMulticastScope is the value of the scope (scop) field within + // a multicast IPv6 address that indicates the address has link-local scope, + // as per RFC 4291 section 2.7. + ipv6LinkLocalMulticastScope = 2 ) // IPv6EmptySubnet is the empty IPv6 subnet. It may also be known as the @@ -340,6 +353,12 @@ func IsV6LinkLocalAddress(addr tcpip.Address) bool { return addr[0] == 0xfe && (addr[1]&0xc0) == 0x80 } +// IsV6LinkLocalMulticastAddress determines if the provided address is an IPv6 +// link-local multicast address. +func IsV6LinkLocalMulticastAddress(addr tcpip.Address) bool { + return IsV6MulticastAddress(addr) && addr[ipv6MulticastAddressScopeByteIdx]&ipv6MulticastAddressScopeMask == ipv6LinkLocalMulticastScope +} + // IsV6UniqueLocalAddress determines if the provided address is an IPv6 // unique-local address (within the prefix FC00::/7). func IsV6UniqueLocalAddress(addr tcpip.Address) bool { @@ -411,6 +430,9 @@ func ScopeForIPv6Address(addr tcpip.Address) (IPv6AddressScope, *tcpip.Error) { } switch { + case IsV6LinkLocalMulticastAddress(addr): + return LinkLocalScope, nil + case IsV6LinkLocalAddress(addr): return LinkLocalScope, nil diff --git a/pkg/tcpip/header/ipv6_extension_headers.go b/pkg/tcpip/header/ipv6_extension_headers.go new file mode 100644 index 000000000..2c4591409 --- /dev/null +++ b/pkg/tcpip/header/ipv6_extension_headers.go @@ -0,0 +1,544 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package header + +import ( + "bufio" + "bytes" + "encoding/binary" + "fmt" + "io" + + "gvisor.dev/gvisor/pkg/tcpip/buffer" +) + +// IPv6ExtensionHeaderIdentifier is an IPv6 extension header identifier. +type IPv6ExtensionHeaderIdentifier uint8 + +const ( + // IPv6HopByHopOptionsExtHdrIdentifier is the header identifier of a Hop by + // Hop Options extension header, as per RFC 8200 section 4.3. + IPv6HopByHopOptionsExtHdrIdentifier IPv6ExtensionHeaderIdentifier = 0 + + // IPv6RoutingExtHdrIdentifier is the header identifier of a Routing extension + // header, as per RFC 8200 section 4.4. + IPv6RoutingExtHdrIdentifier IPv6ExtensionHeaderIdentifier = 43 + + // IPv6FragmentExtHdrIdentifier is the header identifier of a Fragment + // extension header, as per RFC 8200 section 4.5. + IPv6FragmentExtHdrIdentifier IPv6ExtensionHeaderIdentifier = 44 + + // IPv6DestinationOptionsExtHdrIdentifier is the header identifier of a + // Destination Options extension header, as per RFC 8200 section 4.6. + IPv6DestinationOptionsExtHdrIdentifier IPv6ExtensionHeaderIdentifier = 60 + + // IPv6NoNextHeaderIdentifier is the header identifier used to signify the end + // of an IPv6 payload, as per RFC 8200 section 4.7. + IPv6NoNextHeaderIdentifier IPv6ExtensionHeaderIdentifier = 59 +) + +const ( + // ipv6UnknownExtHdrOptionActionMask is the mask of the action to take when + // a node encounters an unrecognized option. + ipv6UnknownExtHdrOptionActionMask = 192 + + // ipv6UnknownExtHdrOptionActionShift is the least significant bits to discard + // from the action value for an unrecognized option identifier. + ipv6UnknownExtHdrOptionActionShift = 6 + + // ipv6RoutingExtHdrSegmentsLeftIdx is the index to the Segments Left field + // within an IPv6RoutingExtHdr. + ipv6RoutingExtHdrSegmentsLeftIdx = 1 + + // IPv6FragmentExtHdrLength is the length of an IPv6 extension header, in + // bytes. + IPv6FragmentExtHdrLength = 8 + + // ipv6FragmentExtHdrFragmentOffsetOffset is the offset to the start of the + // Fragment Offset field within an IPv6FragmentExtHdr. + ipv6FragmentExtHdrFragmentOffsetOffset = 0 + + // ipv6FragmentExtHdrFragmentOffsetShift is the least significant bits to + // discard from the Fragment Offset. + ipv6FragmentExtHdrFragmentOffsetShift = 3 + + // ipv6FragmentExtHdrFlagsIdx is the index to the flags field within an + // IPv6FragmentExtHdr. + ipv6FragmentExtHdrFlagsIdx = 1 + + // ipv6FragmentExtHdrMFlagMask is the mask of the More (M) flag within the + // flags field of an IPv6FragmentExtHdr. + ipv6FragmentExtHdrMFlagMask = 1 + + // ipv6FragmentExtHdrIdentificationOffset is the offset to the Identification + // field within an IPv6FragmentExtHdr. + ipv6FragmentExtHdrIdentificationOffset = 2 + + // ipv6ExtHdrLenBytesPerUnit is the unit size of an extension header's length + // field. That is, given a Length field of 2, the extension header expects + // 16 bytes following the first 8 bytes (see ipv6ExtHdrLenBytesExcluded for + // details about the first 8 bytes' exclusion from the Length field). + ipv6ExtHdrLenBytesPerUnit = 8 + + // ipv6ExtHdrLenBytesExcluded is the number of bytes excluded from an + // extension header's Length field following the Length field. + // + // The Length field excludes the first 8 bytes, but the Next Header and Length + // field take up the first 2 of the 8 bytes so we expect (at minimum) 6 bytes + // after the Length field. + // + // This ensures that every extension header is at least 8 bytes. + ipv6ExtHdrLenBytesExcluded = 6 + + // IPv6FragmentExtHdrFragmentOffsetBytesPerUnit is the unit size of a Fragment + // extension header's Fragment Offset field. That is, given a Fragment Offset + // of 2, the extension header is indiciating that the fragment's payload + // starts at the 16th byte in the reassembled packet. + IPv6FragmentExtHdrFragmentOffsetBytesPerUnit = 8 +) + +// IPv6PayloadHeader is implemented by the various headers that can be found +// in an IPv6 payload. +// +// These headers include IPv6 extension headers or upper layer data. +type IPv6PayloadHeader interface { + isIPv6PayloadHeader() +} + +// IPv6RawPayloadHeader the remainder of an IPv6 payload after an iterator +// encounters a Next Header field it does not recognize as an IPv6 extension +// header. +type IPv6RawPayloadHeader struct { + Identifier IPv6ExtensionHeaderIdentifier + Buf buffer.VectorisedView +} + +// isIPv6PayloadHeader implements IPv6PayloadHeader.isIPv6PayloadHeader. +func (IPv6RawPayloadHeader) isIPv6PayloadHeader() {} + +// ipv6OptionsExtHdr is an IPv6 extension header that holds options. +type ipv6OptionsExtHdr []byte + +// Iter returns an iterator over the IPv6 extension header options held in b. +func (b ipv6OptionsExtHdr) Iter() IPv6OptionsExtHdrOptionsIterator { + it := IPv6OptionsExtHdrOptionsIterator{} + it.reader.Reset(b) + return it +} + +// IPv6OptionsExtHdrOptionsIterator is an iterator over IPv6 extension header +// options. +// +// Note, between when an IPv6OptionsExtHdrOptionsIterator is obtained and last +// used, no changes to the underlying buffer may happen. Doing so may cause +// undefined and unexpected behaviour. It is fine to obtain an +// IPv6OptionsExtHdrOptionsIterator, iterate over the first few options then +// modify the backing payload so long as the IPv6OptionsExtHdrOptionsIterator +// obtained before modification is no longer used. +type IPv6OptionsExtHdrOptionsIterator struct { + reader bytes.Reader +} + +// IPv6OptionUnknownAction is the action that must be taken if the processing +// IPv6 node does not recognize the option, as outlined in RFC 8200 section 4.2. +type IPv6OptionUnknownAction int + +const ( + // IPv6OptionUnknownActionSkip indicates that the unrecognized option must + // be skipped and the node should continue processing the header. + IPv6OptionUnknownActionSkip IPv6OptionUnknownAction = 0 + + // IPv6OptionUnknownActionDiscard indicates that the packet must be silently + // discarded. + IPv6OptionUnknownActionDiscard IPv6OptionUnknownAction = 1 + + // IPv6OptionUnknownActionDiscardSendICMP indicates that the packet must be + // discarded and the node must send an ICMP Parameter Problem, Code 2, message + // to the packet's source, regardless of whether or not the packet's + // Destination was a multicast address. + IPv6OptionUnknownActionDiscardSendICMP IPv6OptionUnknownAction = 2 + + // IPv6OptionUnknownActionDiscardSendICMPNoMulticastDest indicates that the + // packet must be discarded and the node must send an ICMP Parameter Problem, + // Code 2, message to the packet's source only if the packet's Destination was + // not a multicast address. + IPv6OptionUnknownActionDiscardSendICMPNoMulticastDest IPv6OptionUnknownAction = 3 +) + +// IPv6ExtHdrOption is implemented by the various IPv6 extension header options. +type IPv6ExtHdrOption interface { + // UnknownAction returns the action to take in response to an unrecognized + // option. + UnknownAction() IPv6OptionUnknownAction + + // isIPv6ExtHdrOption is used to "lock" this interface so it is not + // implemented by other packages. + isIPv6ExtHdrOption() +} + +// IPv6ExtHdrOptionIndentifier is an IPv6 extension header option identifier. +type IPv6ExtHdrOptionIndentifier uint8 + +const ( + // ipv6Pad1ExtHdrOptionIdentifier is the identifier for a padding option that + // provides 1 byte padding, as outlined in RFC 8200 section 4.2. + ipv6Pad1ExtHdrOptionIdentifier IPv6ExtHdrOptionIndentifier = 0 + + // ipv6PadBExtHdrOptionIdentifier is the identifier for a padding option that + // provides variable length byte padding, as outlined in RFC 8200 section 4.2. + ipv6PadNExtHdrOptionIdentifier IPv6ExtHdrOptionIndentifier = 1 +) + +// IPv6UnknownExtHdrOption holds the identifier and data for an IPv6 extension +// header option that is unknown by the parsing utilities. +type IPv6UnknownExtHdrOption struct { + Identifier IPv6ExtHdrOptionIndentifier + Data []byte +} + +// UnknownAction implements IPv6OptionUnknownAction.UnknownAction. +func (o *IPv6UnknownExtHdrOption) UnknownAction() IPv6OptionUnknownAction { + return IPv6OptionUnknownAction((o.Identifier & ipv6UnknownExtHdrOptionActionMask) >> ipv6UnknownExtHdrOptionActionShift) +} + +// isIPv6ExtHdrOption implements IPv6ExtHdrOption.isIPv6ExtHdrOption. +func (*IPv6UnknownExtHdrOption) isIPv6ExtHdrOption() {} + +// Next returns the next option in the options data. +// +// If the next item is not a known extension header option, +// IPv6UnknownExtHdrOption will be returned with the option identifier and data. +// +// The return is of the format (option, done, error). done will be true when +// Next is unable to return anything because the iterator has reached the end of +// the options data, or an error occured. +func (i *IPv6OptionsExtHdrOptionsIterator) Next() (IPv6ExtHdrOption, bool, error) { + for { + temp, err := i.reader.ReadByte() + if err != nil { + // If we can't read the first byte of a new option, then we know the + // options buffer has been exhausted and we are done iterating. + return nil, true, nil + } + id := IPv6ExtHdrOptionIndentifier(temp) + + // If the option identifier indicates the option is a Pad1 option, then we + // know the option does not have Length and Data fields. End processing of + // the Pad1 option and continue processing the buffer as a new option. + if id == ipv6Pad1ExtHdrOptionIdentifier { + continue + } + + length, err := i.reader.ReadByte() + if err != nil { + if err != io.EOF { + // ReadByte should only ever return nil or io.EOF. + panic(fmt.Sprintf("unexpected error when reading the option's Length field for option with id = %d: %s", id, err)) + } + + // We use io.ErrUnexpectedEOF as exhausting the buffer is unexpected once + // we start parsing an option; we expect the reader to contain enough + // bytes for the whole option. + return nil, true, fmt.Errorf("error when reading the option's Length field for option with id = %d: %w", id, io.ErrUnexpectedEOF) + } + + // Special-case the variable length padding option to avoid a copy. + if id == ipv6PadNExtHdrOptionIdentifier { + // Do we have enough bytes in the reader for the PadN option? + if n := i.reader.Len(); n < int(length) { + // Reset the reader to effectively consume the remaining buffer. + i.reader.Reset(nil) + + // We return the same error as if we failed to read a non-padding option + // so consumers of this iterator don't need to differentiate between + // padding and non-padding options. + return nil, true, fmt.Errorf("read %d out of %d option data bytes for option with id = %d: %w", n, length, id, io.ErrUnexpectedEOF) + } + + if _, err := i.reader.Seek(int64(length), io.SeekCurrent); err != nil { + panic(fmt.Sprintf("error when skipping PadN (N = %d) option's data bytes: %s", length, err)) + } + + // End processing of the PadN option and continue processing the buffer as + // a new option. + continue + } + + bytes := make([]byte, length) + if n, err := io.ReadFull(&i.reader, bytes); err != nil { + // io.ReadFull may return io.EOF if i.reader has been exhausted. We use + // io.ErrUnexpectedEOF instead as the io.EOF is unexpected given the + // Length field found in the option. + if err == io.EOF { + err = io.ErrUnexpectedEOF + } + + return nil, true, fmt.Errorf("read %d out of %d option data bytes for option with id = %d: %w", n, length, id, err) + } + + return &IPv6UnknownExtHdrOption{Identifier: id, Data: bytes}, false, nil + } +} + +// IPv6HopByHopOptionsExtHdr is a buffer holding the Hop By Hop Options +// extension header. +type IPv6HopByHopOptionsExtHdr struct { + ipv6OptionsExtHdr +} + +// isIPv6PayloadHeader implements IPv6PayloadHeader.isIPv6PayloadHeader. +func (IPv6HopByHopOptionsExtHdr) isIPv6PayloadHeader() {} + +// IPv6DestinationOptionsExtHdr is a buffer holding the Destination Options +// extension header. +type IPv6DestinationOptionsExtHdr struct { + ipv6OptionsExtHdr +} + +// isIPv6PayloadHeader implements IPv6PayloadHeader.isIPv6PayloadHeader. +func (IPv6DestinationOptionsExtHdr) isIPv6PayloadHeader() {} + +// IPv6RoutingExtHdr is a buffer holding the Routing extension header specific +// data as outlined in RFC 8200 section 4.4. +type IPv6RoutingExtHdr []byte + +// isIPv6PayloadHeader implements IPv6PayloadHeader.isIPv6PayloadHeader. +func (IPv6RoutingExtHdr) isIPv6PayloadHeader() {} + +// SegmentsLeft returns the Segments Left field. +func (b IPv6RoutingExtHdr) SegmentsLeft() uint8 { + return b[ipv6RoutingExtHdrSegmentsLeftIdx] +} + +// IPv6FragmentExtHdr is a buffer holding the Fragment extension header specific +// data as outlined in RFC 8200 section 4.5. +// +// Note, the buffer does not include the Next Header and Reserved fields. +type IPv6FragmentExtHdr [6]byte + +// isIPv6PayloadHeader implements IPv6PayloadHeader.isIPv6PayloadHeader. +func (IPv6FragmentExtHdr) isIPv6PayloadHeader() {} + +// FragmentOffset returns the Fragment Offset field. +// +// This value indicates where the buffer following the Fragment extension header +// starts in the target (reassembled) packet. +func (b IPv6FragmentExtHdr) FragmentOffset() uint16 { + return binary.BigEndian.Uint16(b[ipv6FragmentExtHdrFragmentOffsetOffset:]) >> ipv6FragmentExtHdrFragmentOffsetShift +} + +// More returns the More (M) flag. +// +// This indicates whether any fragments are expected to succeed b. +func (b IPv6FragmentExtHdr) More() bool { + return b[ipv6FragmentExtHdrFlagsIdx]&ipv6FragmentExtHdrMFlagMask != 0 +} + +// ID returns the Identification field. +// +// This value is used to uniquely identify the packet, between a +// souce and destination. +func (b IPv6FragmentExtHdr) ID() uint32 { + return binary.BigEndian.Uint32(b[ipv6FragmentExtHdrIdentificationOffset:]) +} + +// IPv6PayloadIterator is an iterator over the contents of an IPv6 payload. +// +// The IPv6 payload may contain IPv6 extension headers before any upper layer +// data. +// +// Note, between when an IPv6PayloadIterator is obtained and last used, no +// changes to the payload may happen. Doing so may cause undefined and +// unexpected behaviour. It is fine to obtain an IPv6PayloadIterator, iterate +// over the first few headers then modify the backing payload so long as the +// IPv6PayloadIterator obtained before modification is no longer used. +type IPv6PayloadIterator struct { + // The identifier of the next header to parse. + nextHdrIdentifier IPv6ExtensionHeaderIdentifier + + // reader is an io.Reader over payload. + reader bufio.Reader + payload buffer.VectorisedView + + // Indicates to the iterator that it should return the remaining payload as a + // raw payload on the next call to Next. + forceRaw bool +} + +// MakeIPv6PayloadIterator returns an iterator over the IPv6 payload containing +// extension headers, or a raw payload if the payload cannot be parsed. +func MakeIPv6PayloadIterator(nextHdrIdentifier IPv6ExtensionHeaderIdentifier, payload buffer.VectorisedView) IPv6PayloadIterator { + readers := payload.Readers() + readerPs := make([]io.Reader, 0, len(readers)) + for i := range readers { + readerPs = append(readerPs, &readers[i]) + } + + return IPv6PayloadIterator{ + nextHdrIdentifier: nextHdrIdentifier, + payload: payload.Clone(nil), + // We need a buffer of size 1 for calls to bufio.Reader.ReadByte. + reader: *bufio.NewReaderSize(io.MultiReader(readerPs...), 1), + } +} + +// AsRawHeader returns the remaining payload of i as a raw header and +// optionally consumes the iterator. +// +// If consume is true, calls to Next after calling AsRawHeader on i will +// indicate that the iterator is done. +func (i *IPv6PayloadIterator) AsRawHeader(consume bool) IPv6RawPayloadHeader { + identifier := i.nextHdrIdentifier + + var buf buffer.VectorisedView + if consume { + // Since we consume the iterator, we return the payload as is. + buf = i.payload + + // Mark i as done. + *i = IPv6PayloadIterator{ + nextHdrIdentifier: IPv6NoNextHeaderIdentifier, + } + } else { + buf = i.payload.Clone(nil) + } + + return IPv6RawPayloadHeader{Identifier: identifier, Buf: buf} +} + +// Next returns the next item in the payload. +// +// If the next item is not a known IPv6 extension header, IPv6RawPayloadHeader +// will be returned with the remaining bytes and next header identifier. +// +// The return is of the format (header, done, error). done will be true when +// Next is unable to return anything because the iterator has reached the end of +// the payload, or an error occured. +func (i *IPv6PayloadIterator) Next() (IPv6PayloadHeader, bool, error) { + // We could be forced to return i as a raw header when the previous header was + // a fragment extension header as the data following the fragment extension + // header may not be complete. + if i.forceRaw { + return i.AsRawHeader(true /* consume */), false, nil + } + + // Is the header we are parsing a known extension header? + switch i.nextHdrIdentifier { + case IPv6HopByHopOptionsExtHdrIdentifier: + nextHdrIdentifier, bytes, err := i.nextHeaderData(false /* fragmentHdr */, nil) + if err != nil { + return nil, true, err + } + + i.nextHdrIdentifier = nextHdrIdentifier + return IPv6HopByHopOptionsExtHdr{ipv6OptionsExtHdr: bytes}, false, nil + case IPv6RoutingExtHdrIdentifier: + nextHdrIdentifier, bytes, err := i.nextHeaderData(false /* fragmentHdr */, nil) + if err != nil { + return nil, true, err + } + + i.nextHdrIdentifier = nextHdrIdentifier + return IPv6RoutingExtHdr(bytes), false, nil + case IPv6FragmentExtHdrIdentifier: + var data [6]byte + // We ignore the returned bytes becauase we know the fragment extension + // header specific data will fit in data. + nextHdrIdentifier, _, err := i.nextHeaderData(true /* fragmentHdr */, data[:]) + if err != nil { + return nil, true, err + } + + fragmentExtHdr := IPv6FragmentExtHdr(data) + + // If the packet is not the first fragment, do not attempt to parse anything + // after the fragment extension header as the payload following the fragment + // extension header should not contain any headers; the first fragment must + // hold all the headers up to and including any upper layer headers, as per + // RFC 8200 section 4.5. + if fragmentExtHdr.FragmentOffset() != 0 { + i.forceRaw = true + } + + i.nextHdrIdentifier = nextHdrIdentifier + return fragmentExtHdr, false, nil + case IPv6DestinationOptionsExtHdrIdentifier: + nextHdrIdentifier, bytes, err := i.nextHeaderData(false /* fragmentHdr */, nil) + if err != nil { + return nil, true, err + } + + i.nextHdrIdentifier = nextHdrIdentifier + return IPv6DestinationOptionsExtHdr{ipv6OptionsExtHdr: bytes}, false, nil + case IPv6NoNextHeaderIdentifier: + // This indicates the end of the IPv6 payload. + return nil, true, nil + + default: + // The header we are parsing is not a known extension header. Return the + // raw payload. + return i.AsRawHeader(true /* consume */), false, nil + } +} + +// nextHeaderData returns the extension header's Next Header field and raw data. +// +// fragmentHdr indicates that the extension header being parsed is the Fragment +// extension header so the Length field should be ignored as it is Reserved +// for the Fragment extension header. +// +// If bytes is not nil, extension header specific data will be read into bytes +// if it has enough capacity. If bytes is provided but does not have enough +// capacity for the data, nextHeaderData will panic. +func (i *IPv6PayloadIterator) nextHeaderData(fragmentHdr bool, bytes []byte) (IPv6ExtensionHeaderIdentifier, []byte, error) { + // We ignore the number of bytes read because we know we will only ever read + // at max 1 bytes since rune has a length of 1. If we read 0 bytes, the Read + // would return io.EOF to indicate that io.Reader has reached the end of the + // payload. + nextHdrIdentifier, err := i.reader.ReadByte() + i.payload.TrimFront(1) + if err != nil { + return 0, nil, fmt.Errorf("error when reading the Next Header field for extension header with id = %d: %w", i.nextHdrIdentifier, err) + } + + var length uint8 + length, err = i.reader.ReadByte() + i.payload.TrimFront(1) + if err != nil { + if fragmentHdr { + return 0, nil, fmt.Errorf("error when reading the Length field for extension header with id = %d: %w", i.nextHdrIdentifier, err) + } + + return 0, nil, fmt.Errorf("error when reading the Reserved field for extension header with id = %d: %w", i.nextHdrIdentifier, err) + } + if fragmentHdr { + length = 0 + } + + bytesLen := int(length)*ipv6ExtHdrLenBytesPerUnit + ipv6ExtHdrLenBytesExcluded + if bytes == nil { + bytes = make([]byte, bytesLen) + } else if n := len(bytes); n < bytesLen { + panic(fmt.Sprintf("bytes only has space for %d bytes but need space for %d bytes (length = %d) for extension header with id = %d", n, bytesLen, length, i.nextHdrIdentifier)) + } + + n, err := io.ReadFull(&i.reader, bytes) + i.payload.TrimFront(n) + if err != nil { + return 0, nil, fmt.Errorf("read %d out of %d extension header data bytes (length = %d) for header with id = %d: %w", n, bytesLen, length, i.nextHdrIdentifier, err) + } + + return IPv6ExtensionHeaderIdentifier(nextHdrIdentifier), bytes, nil +} diff --git a/pkg/tcpip/header/ipv6_extension_headers_test.go b/pkg/tcpip/header/ipv6_extension_headers_test.go new file mode 100644 index 000000000..ab20c5f37 --- /dev/null +++ b/pkg/tcpip/header/ipv6_extension_headers_test.go @@ -0,0 +1,992 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package header + +import ( + "bytes" + "errors" + "io" + "testing" + + "github.com/google/go-cmp/cmp" + "gvisor.dev/gvisor/pkg/tcpip/buffer" +) + +// Equal returns true of a and b are equivalent. +// +// Note, Equal will return true if a and b hold the same Identifier value and +// contain the same bytes in Buf, even if the bytes are split across views +// differently. +// +// Needed to use cmp.Equal on IPv6RawPayloadHeader as it contains unexported +// fields. +func (a IPv6RawPayloadHeader) Equal(b IPv6RawPayloadHeader) bool { + return a.Identifier == b.Identifier && bytes.Equal(a.Buf.ToView(), b.Buf.ToView()) +} + +// Equal returns true of a and b are equivalent. +// +// Note, Equal will return true if a and b hold equivalent ipv6OptionsExtHdrs. +// +// Needed to use cmp.Equal on IPv6RawPayloadHeader as it contains unexported +// fields. +func (a IPv6HopByHopOptionsExtHdr) Equal(b IPv6HopByHopOptionsExtHdr) bool { + return bytes.Equal(a.ipv6OptionsExtHdr, b.ipv6OptionsExtHdr) +} + +// Equal returns true of a and b are equivalent. +// +// Note, Equal will return true if a and b hold equivalent ipv6OptionsExtHdrs. +// +// Needed to use cmp.Equal on IPv6RawPayloadHeader as it contains unexported +// fields. +func (a IPv6DestinationOptionsExtHdr) Equal(b IPv6DestinationOptionsExtHdr) bool { + return bytes.Equal(a.ipv6OptionsExtHdr, b.ipv6OptionsExtHdr) +} + +func TestIPv6UnknownExtHdrOption(t *testing.T) { + tests := []struct { + name string + identifier IPv6ExtHdrOptionIndentifier + expectedUnknownAction IPv6OptionUnknownAction + }{ + { + name: "Skip with zero LSBs", + identifier: 0, + expectedUnknownAction: IPv6OptionUnknownActionSkip, + }, + { + name: "Discard with zero LSBs", + identifier: 64, + expectedUnknownAction: IPv6OptionUnknownActionDiscard, + }, + { + name: "Discard and ICMP with zero LSBs", + identifier: 128, + expectedUnknownAction: IPv6OptionUnknownActionDiscardSendICMP, + }, + { + name: "Discard and ICMP for non multicast destination with zero LSBs", + identifier: 192, + expectedUnknownAction: IPv6OptionUnknownActionDiscardSendICMPNoMulticastDest, + }, + { + name: "Skip with non-zero LSBs", + identifier: 63, + expectedUnknownAction: IPv6OptionUnknownActionSkip, + }, + { + name: "Discard with non-zero LSBs", + identifier: 127, + expectedUnknownAction: IPv6OptionUnknownActionDiscard, + }, + { + name: "Discard and ICMP with non-zero LSBs", + identifier: 191, + expectedUnknownAction: IPv6OptionUnknownActionDiscardSendICMP, + }, + { + name: "Discard and ICMP for non multicast destination with non-zero LSBs", + identifier: 255, + expectedUnknownAction: IPv6OptionUnknownActionDiscardSendICMPNoMulticastDest, + }, + } + + for _, test := range tests { + t.Run(test.name, func(t *testing.T) { + opt := &IPv6UnknownExtHdrOption{Identifier: test.identifier, Data: []byte{1, 2, 3, 4}} + if a := opt.UnknownAction(); a != test.expectedUnknownAction { + t.Fatalf("got UnknownAction() = %d, want = %d", a, test.expectedUnknownAction) + } + }) + } + +} + +func TestIPv6OptionsExtHdrIterErr(t *testing.T) { + tests := []struct { + name string + bytes []byte + err error + }{ + { + name: "Single unknown with zero length", + bytes: []byte{255, 0}, + }, + { + name: "Single unknown with non-zero length", + bytes: []byte{255, 3, 1, 2, 3}, + }, + { + name: "Two options", + bytes: []byte{ + 255, 0, + 254, 1, 1, + }, + }, + { + name: "Three options", + bytes: []byte{ + 255, 0, + 254, 1, 1, + 253, 4, 2, 3, 4, 5, + }, + }, + { + name: "Single unknown only identifier", + bytes: []byte{255}, + err: io.ErrUnexpectedEOF, + }, + { + name: "Single unknown too small with length = 1", + bytes: []byte{255, 1}, + err: io.ErrUnexpectedEOF, + }, + { + name: "Single unknown too small with length = 2", + bytes: []byte{255, 2, 1}, + err: io.ErrUnexpectedEOF, + }, + { + name: "Valid first with second unknown only identifier", + bytes: []byte{ + 255, 0, + 254, + }, + err: io.ErrUnexpectedEOF, + }, + { + name: "Valid first with second unknown missing data", + bytes: []byte{ + 255, 0, + 254, 1, + }, + err: io.ErrUnexpectedEOF, + }, + { + name: "Valid first with second unknown too small", + bytes: []byte{ + 255, 0, + 254, 2, 1, + }, + err: io.ErrUnexpectedEOF, + }, + { + name: "One Pad1", + bytes: []byte{0}, + }, + { + name: "Multiple Pad1", + bytes: []byte{0, 0, 0}, + }, + { + name: "Multiple PadN", + bytes: []byte{ + // Pad3 + 1, 1, 1, + + // Pad5 + 1, 3, 1, 2, 3, + }, + }, + { + name: "Pad5 too small middle of data buffer", + bytes: []byte{1, 3, 1, 2}, + err: io.ErrUnexpectedEOF, + }, + { + name: "Pad5 no data", + bytes: []byte{1, 3}, + err: io.ErrUnexpectedEOF, + }, + } + + check := func(t *testing.T, it IPv6OptionsExtHdrOptionsIterator, expectedErr error) { + for i := 0; ; i++ { + _, done, err := it.Next() + if err != nil { + // If we encountered a non-nil error while iterating, make sure it is + // is the same error as expectedErr. + if !errors.Is(err, expectedErr) { + t.Fatalf("got %d-th Next() = %v, want = %v", i, err, expectedErr) + } + + return + } + if done { + // If we are done (without an error), make sure that we did not expect + // an error. + if expectedErr != nil { + t.Fatalf("expected error when iterating; want = %s", expectedErr) + } + + return + } + } + } + + for _, test := range tests { + t.Run(test.name, func(t *testing.T) { + t.Run("Hop By Hop", func(t *testing.T) { + extHdr := IPv6HopByHopOptionsExtHdr{ipv6OptionsExtHdr: test.bytes} + check(t, extHdr.Iter(), test.err) + }) + + t.Run("Destination", func(t *testing.T) { + extHdr := IPv6DestinationOptionsExtHdr{ipv6OptionsExtHdr: test.bytes} + check(t, extHdr.Iter(), test.err) + }) + }) + } +} + +func TestIPv6OptionsExtHdrIter(t *testing.T) { + tests := []struct { + name string + bytes []byte + expected []IPv6ExtHdrOption + }{ + { + name: "Single unknown with zero length", + bytes: []byte{255, 0}, + expected: []IPv6ExtHdrOption{ + &IPv6UnknownExtHdrOption{Identifier: 255, Data: []byte{}}, + }, + }, + { + name: "Single unknown with non-zero length", + bytes: []byte{255, 3, 1, 2, 3}, + expected: []IPv6ExtHdrOption{ + &IPv6UnknownExtHdrOption{Identifier: 255, Data: []byte{1, 2, 3}}, + }, + }, + { + name: "Single Pad1", + bytes: []byte{0}, + }, + { + name: "Two Pad1", + bytes: []byte{0, 0}, + }, + { + name: "Single Pad3", + bytes: []byte{1, 1, 1}, + }, + { + name: "Single Pad5", + bytes: []byte{1, 3, 1, 2, 3}, + }, + { + name: "Multiple Pad", + bytes: []byte{ + // Pad1 + 0, + + // Pad2 + 1, 0, + + // Pad3 + 1, 1, 1, + + // Pad4 + 1, 2, 1, 2, + + // Pad5 + 1, 3, 1, 2, 3, + }, + }, + { + name: "Multiple options", + bytes: []byte{ + // Pad1 + 0, + + // Unknown + 255, 0, + + // Pad2 + 1, 0, + + // Unknown + 254, 1, 1, + + // Pad3 + 1, 1, 1, + + // Unknown + 253, 4, 2, 3, 4, 5, + + // Pad4 + 1, 2, 1, 2, + }, + expected: []IPv6ExtHdrOption{ + &IPv6UnknownExtHdrOption{Identifier: 255, Data: []byte{}}, + &IPv6UnknownExtHdrOption{Identifier: 254, Data: []byte{1}}, + &IPv6UnknownExtHdrOption{Identifier: 253, Data: []byte{2, 3, 4, 5}}, + }, + }, + } + + checkIter := func(t *testing.T, it IPv6OptionsExtHdrOptionsIterator, expected []IPv6ExtHdrOption) { + for i, e := range expected { + opt, done, err := it.Next() + if err != nil { + t.Errorf("(i=%d) Next(): %s", i, err) + } + if done { + t.Errorf("(i=%d) unexpectedly done iterating", i) + } + if diff := cmp.Diff(e, opt); diff != "" { + t.Errorf("(i=%d) got option mismatch (-want +got):\n%s", i, diff) + } + + if t.Failed() { + t.FailNow() + } + } + + opt, done, err := it.Next() + if err != nil { + t.Errorf("(last) Next(): %s", err) + } + if !done { + t.Errorf("(last) iterator unexpectedly not done") + } + if opt != nil { + t.Errorf("(last) got Next() = %T, want = nil", opt) + } + } + + for _, test := range tests { + t.Run(test.name, func(t *testing.T) { + t.Run("Hop By Hop", func(t *testing.T) { + extHdr := IPv6HopByHopOptionsExtHdr{ipv6OptionsExtHdr: test.bytes} + checkIter(t, extHdr.Iter(), test.expected) + }) + + t.Run("Destination", func(t *testing.T) { + extHdr := IPv6DestinationOptionsExtHdr{ipv6OptionsExtHdr: test.bytes} + checkIter(t, extHdr.Iter(), test.expected) + }) + }) + } +} + +func TestIPv6RoutingExtHdr(t *testing.T) { + tests := []struct { + name string + bytes []byte + segmentsLeft uint8 + }{ + { + name: "Zeroes", + bytes: []byte{0, 0, 0, 0, 0, 0}, + segmentsLeft: 0, + }, + { + name: "Ones", + bytes: []byte{1, 1, 1, 1, 1, 1}, + segmentsLeft: 1, + }, + { + name: "Mixed", + bytes: []byte{1, 2, 3, 4, 5, 6}, + segmentsLeft: 2, + }, + } + + for _, test := range tests { + t.Run(test.name, func(t *testing.T) { + extHdr := IPv6RoutingExtHdr(test.bytes) + if got := extHdr.SegmentsLeft(); got != test.segmentsLeft { + t.Errorf("got SegmentsLeft() = %d, want = %d", got, test.segmentsLeft) + } + }) + } +} + +func TestIPv6FragmentExtHdr(t *testing.T) { + tests := []struct { + name string + bytes [6]byte + fragmentOffset uint16 + more bool + id uint32 + }{ + { + name: "Zeroes", + bytes: [6]byte{0, 0, 0, 0, 0, 0}, + fragmentOffset: 0, + more: false, + id: 0, + }, + { + name: "Ones", + bytes: [6]byte{0, 9, 0, 0, 0, 1}, + fragmentOffset: 1, + more: true, + id: 1, + }, + { + name: "Mixed", + bytes: [6]byte{68, 9, 128, 4, 2, 1}, + fragmentOffset: 2177, + more: true, + id: 2147746305, + }, + } + + for _, test := range tests { + t.Run(test.name, func(t *testing.T) { + extHdr := IPv6FragmentExtHdr(test.bytes) + if got := extHdr.FragmentOffset(); got != test.fragmentOffset { + t.Errorf("got FragmentOffset() = %d, want = %d", got, test.fragmentOffset) + } + if got := extHdr.More(); got != test.more { + t.Errorf("got More() = %t, want = %t", got, test.more) + } + if got := extHdr.ID(); got != test.id { + t.Errorf("got ID() = %d, want = %d", got, test.id) + } + }) + } +} + +func makeVectorisedViewFromByteBuffers(bs ...[]byte) buffer.VectorisedView { + size := 0 + var vs []buffer.View + + for _, b := range bs { + vs = append(vs, buffer.View(b)) + size += len(b) + } + + return buffer.NewVectorisedView(size, vs) +} + +func TestIPv6ExtHdrIterErr(t *testing.T) { + tests := []struct { + name string + firstNextHdr IPv6ExtensionHeaderIdentifier + payload buffer.VectorisedView + err error + }{ + { + name: "Upper layer only without data", + firstNextHdr: 255, + }, + { + name: "Upper layer only with data", + firstNextHdr: 255, + payload: makeVectorisedViewFromByteBuffers([]byte{1, 2, 3, 4}), + }, + { + name: "No next header", + firstNextHdr: IPv6NoNextHeaderIdentifier, + }, + { + name: "No next header with data", + firstNextHdr: IPv6NoNextHeaderIdentifier, + payload: makeVectorisedViewFromByteBuffers([]byte{1, 2, 3, 4}), + }, + { + name: "Valid single hop by hop", + firstNextHdr: IPv6HopByHopOptionsExtHdrIdentifier, + payload: makeVectorisedViewFromByteBuffers([]byte{255, 0, 1, 4, 1, 2, 3, 4}), + }, + { + name: "Hop by hop too small", + firstNextHdr: IPv6HopByHopOptionsExtHdrIdentifier, + payload: makeVectorisedViewFromByteBuffers([]byte{255, 0, 1, 4, 1, 2, 3}), + err: io.ErrUnexpectedEOF, + }, + { + name: "Valid single fragment", + firstNextHdr: IPv6FragmentExtHdrIdentifier, + payload: makeVectorisedViewFromByteBuffers([]byte{255, 0, 68, 9, 128, 4, 2, 1}), + }, + { + name: "Fragment too small", + firstNextHdr: IPv6FragmentExtHdrIdentifier, + payload: makeVectorisedViewFromByteBuffers([]byte{255, 0, 68, 9, 128, 4, 2}), + err: io.ErrUnexpectedEOF, + }, + { + name: "Valid single destination", + firstNextHdr: IPv6DestinationOptionsExtHdrIdentifier, + payload: makeVectorisedViewFromByteBuffers([]byte{255, 0, 1, 4, 1, 2, 3, 4}), + }, + { + name: "Destination too small", + firstNextHdr: IPv6DestinationOptionsExtHdrIdentifier, + payload: makeVectorisedViewFromByteBuffers([]byte{255, 0, 1, 4, 1, 2, 3}), + err: io.ErrUnexpectedEOF, + }, + { + name: "Valid single routing", + firstNextHdr: IPv6RoutingExtHdrIdentifier, + payload: makeVectorisedViewFromByteBuffers([]byte{255, 0, 1, 2, 3, 4, 5, 6}), + }, + { + name: "Valid single routing across views", + firstNextHdr: IPv6RoutingExtHdrIdentifier, + payload: makeVectorisedViewFromByteBuffers([]byte{255, 0, 1, 2}, []byte{3, 4, 5, 6}), + }, + { + name: "Routing too small with zero length field", + firstNextHdr: IPv6RoutingExtHdrIdentifier, + payload: makeVectorisedViewFromByteBuffers([]byte{255, 0, 1, 2, 3, 4, 5}), + err: io.ErrUnexpectedEOF, + }, + { + name: "Valid routing with non-zero length field", + firstNextHdr: IPv6RoutingExtHdrIdentifier, + payload: makeVectorisedViewFromByteBuffers([]byte{255, 1, 1, 2, 3, 4, 5, 6, 1, 2, 3, 4, 5, 6, 7, 8}), + }, + { + name: "Valid routing with non-zero length field across views", + firstNextHdr: IPv6RoutingExtHdrIdentifier, + payload: makeVectorisedViewFromByteBuffers([]byte{255, 1, 1, 2, 3, 4, 5, 6}, []byte{1, 2, 3, 4, 5, 6, 7, 8}), + }, + { + name: "Routing too small with non-zero length field", + firstNextHdr: IPv6RoutingExtHdrIdentifier, + payload: makeVectorisedViewFromByteBuffers([]byte{255, 1, 1, 2, 3, 4, 5, 6, 1, 2, 3, 4, 5, 6, 7}), + err: io.ErrUnexpectedEOF, + }, + { + name: "Routing too small with non-zero length field across views", + firstNextHdr: IPv6RoutingExtHdrIdentifier, + payload: makeVectorisedViewFromByteBuffers([]byte{255, 1, 1, 2, 3, 4, 5, 6}, []byte{1, 2, 3, 4, 5, 6, 7}), + err: io.ErrUnexpectedEOF, + }, + { + name: "Mixed", + firstNextHdr: IPv6HopByHopOptionsExtHdrIdentifier, + payload: makeVectorisedViewFromByteBuffers([]byte{ + // Hop By Hop Options extension header. + uint8(IPv6FragmentExtHdrIdentifier), 0, 1, 4, 1, 2, 3, 4, + + // (Atomic) Fragment extension header. + // + // Reserved bits are 1 which should not affect anything. + uint8(IPv6RoutingExtHdrIdentifier), 255, 0, 6, 128, 4, 2, 1, + + // Routing extension header. + uint8(IPv6DestinationOptionsExtHdrIdentifier), 0, 1, 2, 3, 4, 5, 6, + + // Destination Options extension header. + 255, 0, 255, 4, 1, 2, 3, 4, + + // Upper layer data. + 1, 2, 3, 4, + }), + }, + { + name: "Mixed without upper layer data", + firstNextHdr: IPv6HopByHopOptionsExtHdrIdentifier, + payload: makeVectorisedViewFromByteBuffers([]byte{ + // Hop By Hop Options extension header. + uint8(IPv6FragmentExtHdrIdentifier), 0, 1, 4, 1, 2, 3, 4, + + // (Atomic) Fragment extension header. + // + // Reserved bits are 1 which should not affect anything. + uint8(IPv6RoutingExtHdrIdentifier), 255, 0, 6, 128, 4, 2, 1, + + // Routing extension header. + uint8(IPv6DestinationOptionsExtHdrIdentifier), 0, 1, 2, 3, 4, 5, 6, + + // Destination Options extension header. + 255, 0, 255, 4, 1, 2, 3, 4, + }), + }, + { + name: "Mixed without upper layer data but last ext hdr too small", + firstNextHdr: IPv6HopByHopOptionsExtHdrIdentifier, + payload: makeVectorisedViewFromByteBuffers([]byte{ + // Hop By Hop Options extension header. + uint8(IPv6FragmentExtHdrIdentifier), 0, 1, 4, 1, 2, 3, 4, + + // (Atomic) Fragment extension header. + // + // Reserved bits are 1 which should not affect anything. + uint8(IPv6RoutingExtHdrIdentifier), 255, 0, 6, 128, 4, 2, 1, + + // Routing extension header. + uint8(IPv6DestinationOptionsExtHdrIdentifier), 0, 1, 2, 3, 4, 5, 6, + + // Destination Options extension header. + 255, 0, 255, 4, 1, 2, 3, + }), + err: io.ErrUnexpectedEOF, + }, + } + + for _, test := range tests { + t.Run(test.name, func(t *testing.T) { + it := MakeIPv6PayloadIterator(test.firstNextHdr, test.payload) + + for i := 0; ; i++ { + _, done, err := it.Next() + if err != nil { + // If we encountered a non-nil error while iterating, make sure it is + // is the same error as test.err. + if !errors.Is(err, test.err) { + t.Fatalf("got %d-th Next() = %v, want = %v", i, err, test.err) + } + + return + } + if done { + // If we are done (without an error), make sure that we did not expect + // an error. + if test.err != nil { + t.Fatalf("expected error when iterating; want = %s", test.err) + } + + return + } + } + }) + } +} + +func TestIPv6ExtHdrIter(t *testing.T) { + routingExtHdrWithUpperLayerData := buffer.View([]byte{255, 0, 1, 2, 3, 4, 5, 6, 1, 2, 3, 4}) + upperLayerData := buffer.View([]byte{1, 2, 3, 4}) + tests := []struct { + name string + firstNextHdr IPv6ExtensionHeaderIdentifier + payload buffer.VectorisedView + expected []IPv6PayloadHeader + }{ + // With a non-atomic fragment that is not the first fragment, the payload + // after the fragment will not be parsed because the payload is expected to + // only hold upper layer data. + { + name: "hopbyhop - fragment (not first) - routing - upper", + firstNextHdr: IPv6HopByHopOptionsExtHdrIdentifier, + payload: makeVectorisedViewFromByteBuffers([]byte{ + // Hop By Hop extension header. + uint8(IPv6FragmentExtHdrIdentifier), 0, 1, 4, 1, 2, 3, 4, + + // Fragment extension header. + // + // More = 1, Fragment Offset = 2117, ID = 2147746305 + uint8(IPv6RoutingExtHdrIdentifier), 0, 68, 9, 128, 4, 2, 1, + + // Routing extension header. + // + // Even though we have a routing ext header here, it should be + // be interpretted as raw bytes as only the first fragment is expected + // to hold headers. + 255, 0, 1, 2, 3, 4, 5, 6, + + // Upper layer data. + 1, 2, 3, 4, + }), + expected: []IPv6PayloadHeader{ + IPv6HopByHopOptionsExtHdr{ipv6OptionsExtHdr: []byte{1, 4, 1, 2, 3, 4}}, + IPv6FragmentExtHdr([6]byte{68, 9, 128, 4, 2, 1}), + IPv6RawPayloadHeader{ + Identifier: IPv6RoutingExtHdrIdentifier, + Buf: routingExtHdrWithUpperLayerData.ToVectorisedView(), + }, + }, + }, + { + name: "hopbyhop - fragment (first) - routing - upper", + firstNextHdr: IPv6HopByHopOptionsExtHdrIdentifier, + payload: makeVectorisedViewFromByteBuffers([]byte{ + // Hop By Hop extension header. + uint8(IPv6FragmentExtHdrIdentifier), 0, 1, 4, 1, 2, 3, 4, + + // Fragment extension header. + // + // More = 1, Fragment Offset = 0, ID = 2147746305 + uint8(IPv6RoutingExtHdrIdentifier), 0, 0, 1, 128, 4, 2, 1, + + // Routing extension header. + 255, 0, 1, 2, 3, 4, 5, 6, + + // Upper layer data. + 1, 2, 3, 4, + }), + expected: []IPv6PayloadHeader{ + IPv6HopByHopOptionsExtHdr{ipv6OptionsExtHdr: []byte{1, 4, 1, 2, 3, 4}}, + IPv6FragmentExtHdr([6]byte{0, 1, 128, 4, 2, 1}), + IPv6RoutingExtHdr([]byte{1, 2, 3, 4, 5, 6}), + IPv6RawPayloadHeader{ + Identifier: 255, + Buf: upperLayerData.ToVectorisedView(), + }, + }, + }, + { + name: "fragment - routing - upper (across views)", + firstNextHdr: IPv6FragmentExtHdrIdentifier, + payload: makeVectorisedViewFromByteBuffers([]byte{ + // Fragment extension header. + uint8(IPv6RoutingExtHdrIdentifier), 0, 68, 9, 128, 4, 2, 1, + + // Routing extension header. + 255, 0, 1, 2}, []byte{3, 4, 5, 6, + + // Upper layer data. + 1, 2, 3, 4, + }), + expected: []IPv6PayloadHeader{ + IPv6FragmentExtHdr([6]byte{68, 9, 128, 4, 2, 1}), + IPv6RawPayloadHeader{ + Identifier: IPv6RoutingExtHdrIdentifier, + Buf: routingExtHdrWithUpperLayerData.ToVectorisedView(), + }, + }, + }, + + // If we have an atomic fragment, the payload following the fragment + // extension header should be parsed normally. + { + name: "atomic fragment - routing - destination - upper", + firstNextHdr: IPv6FragmentExtHdrIdentifier, + payload: makeVectorisedViewFromByteBuffers([]byte{ + // Fragment extension header. + // + // Reserved bits are 1 which should not affect anything. + uint8(IPv6RoutingExtHdrIdentifier), 255, 0, 6, 128, 4, 2, 1, + + // Routing extension header. + uint8(IPv6DestinationOptionsExtHdrIdentifier), 0, 1, 2, 3, 4, 5, 6, + + // Destination Options extension header. + 255, 0, 1, 4, 1, 2, 3, 4, + + // Upper layer data. + 1, 2, 3, 4, + }), + expected: []IPv6PayloadHeader{ + IPv6FragmentExtHdr([6]byte{0, 6, 128, 4, 2, 1}), + IPv6RoutingExtHdr([]byte{1, 2, 3, 4, 5, 6}), + IPv6DestinationOptionsExtHdr{ipv6OptionsExtHdr: []byte{1, 4, 1, 2, 3, 4}}, + IPv6RawPayloadHeader{ + Identifier: 255, + Buf: upperLayerData.ToVectorisedView(), + }, + }, + }, + { + name: "atomic fragment - routing - upper (across views)", + firstNextHdr: IPv6FragmentExtHdrIdentifier, + payload: makeVectorisedViewFromByteBuffers([]byte{ + // Fragment extension header. + // + // Reserved bits are 1 which should not affect anything. + uint8(IPv6RoutingExtHdrIdentifier), 255, 0, 6}, []byte{128, 4, 2, 1, + + // Routing extension header. + 255, 0, 1, 2}, []byte{3, 4, 5, 6, + + // Upper layer data. + 1, 2}, []byte{3, 4}), + expected: []IPv6PayloadHeader{ + IPv6FragmentExtHdr([6]byte{0, 6, 128, 4, 2, 1}), + IPv6RoutingExtHdr([]byte{1, 2, 3, 4, 5, 6}), + IPv6RawPayloadHeader{ + Identifier: 255, + Buf: makeVectorisedViewFromByteBuffers(upperLayerData[:2], upperLayerData[2:]), + }, + }, + }, + { + name: "atomic fragment - destination - no next header", + firstNextHdr: IPv6FragmentExtHdrIdentifier, + payload: makeVectorisedViewFromByteBuffers([]byte{ + // Fragment extension header. + // + // Res (Reserved) bits are 1 which should not affect anything. + uint8(IPv6DestinationOptionsExtHdrIdentifier), 0, 0, 6, 128, 4, 2, 1, + + // Destination Options extension header. + uint8(IPv6NoNextHeaderIdentifier), 0, 1, 4, 1, 2, 3, 4, + + // Random data. + 1, 2, 3, 4, + }), + expected: []IPv6PayloadHeader{ + IPv6FragmentExtHdr([6]byte{0, 6, 128, 4, 2, 1}), + IPv6DestinationOptionsExtHdr{ipv6OptionsExtHdr: []byte{1, 4, 1, 2, 3, 4}}, + }, + }, + { + name: "routing - atomic fragment - no next header", + firstNextHdr: IPv6RoutingExtHdrIdentifier, + payload: makeVectorisedViewFromByteBuffers([]byte{ + // Routing extension header. + uint8(IPv6FragmentExtHdrIdentifier), 0, 1, 2, 3, 4, 5, 6, + + // Fragment extension header. + // + // Reserved bits are 1 which should not affect anything. + uint8(IPv6NoNextHeaderIdentifier), 0, 0, 6, 128, 4, 2, 1, + + // Random data. + 1, 2, 3, 4, + }), + expected: []IPv6PayloadHeader{ + IPv6RoutingExtHdr([]byte{1, 2, 3, 4, 5, 6}), + IPv6FragmentExtHdr([6]byte{0, 6, 128, 4, 2, 1}), + }, + }, + { + name: "routing - atomic fragment - no next header (across views)", + firstNextHdr: IPv6RoutingExtHdrIdentifier, + payload: makeVectorisedViewFromByteBuffers([]byte{ + // Routing extension header. + uint8(IPv6FragmentExtHdrIdentifier), 0, 1, 2, 3, 4, 5, 6, + + // Fragment extension header. + // + // Reserved bits are 1 which should not affect anything. + uint8(IPv6NoNextHeaderIdentifier), 255, 0, 6}, []byte{128, 4, 2, 1, + + // Random data. + 1, 2, 3, 4, + }), + expected: []IPv6PayloadHeader{ + IPv6RoutingExtHdr([]byte{1, 2, 3, 4, 5, 6}), + IPv6FragmentExtHdr([6]byte{0, 6, 128, 4, 2, 1}), + }, + }, + { + name: "hopbyhop - routing - fragment - no next header", + firstNextHdr: IPv6HopByHopOptionsExtHdrIdentifier, + payload: makeVectorisedViewFromByteBuffers([]byte{ + // Hop By Hop Options extension header. + uint8(IPv6RoutingExtHdrIdentifier), 0, 1, 4, 1, 2, 3, 4, + + // Routing extension header. + uint8(IPv6FragmentExtHdrIdentifier), 0, 1, 2, 3, 4, 5, 6, + + // Fragment extension header. + // + // Fragment Offset = 32; Res = 6. + uint8(IPv6NoNextHeaderIdentifier), 0, 1, 6, 128, 4, 2, 1, + + // Random data. + 1, 2, 3, 4, + }), + expected: []IPv6PayloadHeader{ + IPv6HopByHopOptionsExtHdr{ipv6OptionsExtHdr: []byte{1, 4, 1, 2, 3, 4}}, + IPv6RoutingExtHdr([]byte{1, 2, 3, 4, 5, 6}), + IPv6FragmentExtHdr([6]byte{1, 6, 128, 4, 2, 1}), + IPv6RawPayloadHeader{ + Identifier: IPv6NoNextHeaderIdentifier, + Buf: upperLayerData.ToVectorisedView(), + }, + }, + }, + + // Test the raw payload for common transport layer protocol numbers. + { + name: "TCP raw payload", + firstNextHdr: IPv6ExtensionHeaderIdentifier(TCPProtocolNumber), + payload: makeVectorisedViewFromByteBuffers(upperLayerData), + expected: []IPv6PayloadHeader{IPv6RawPayloadHeader{ + Identifier: IPv6ExtensionHeaderIdentifier(TCPProtocolNumber), + Buf: upperLayerData.ToVectorisedView(), + }}, + }, + { + name: "UDP raw payload", + firstNextHdr: IPv6ExtensionHeaderIdentifier(UDPProtocolNumber), + payload: makeVectorisedViewFromByteBuffers(upperLayerData), + expected: []IPv6PayloadHeader{IPv6RawPayloadHeader{ + Identifier: IPv6ExtensionHeaderIdentifier(UDPProtocolNumber), + Buf: upperLayerData.ToVectorisedView(), + }}, + }, + { + name: "ICMPv4 raw payload", + firstNextHdr: IPv6ExtensionHeaderIdentifier(ICMPv4ProtocolNumber), + payload: makeVectorisedViewFromByteBuffers(upperLayerData), + expected: []IPv6PayloadHeader{IPv6RawPayloadHeader{ + Identifier: IPv6ExtensionHeaderIdentifier(ICMPv4ProtocolNumber), + Buf: upperLayerData.ToVectorisedView(), + }}, + }, + { + name: "ICMPv6 raw payload", + firstNextHdr: IPv6ExtensionHeaderIdentifier(ICMPv6ProtocolNumber), + payload: makeVectorisedViewFromByteBuffers(upperLayerData), + expected: []IPv6PayloadHeader{IPv6RawPayloadHeader{ + Identifier: IPv6ExtensionHeaderIdentifier(ICMPv6ProtocolNumber), + Buf: upperLayerData.ToVectorisedView(), + }}, + }, + { + name: "Unknwon next header raw payload", + firstNextHdr: 255, + payload: makeVectorisedViewFromByteBuffers(upperLayerData), + expected: []IPv6PayloadHeader{IPv6RawPayloadHeader{ + Identifier: 255, + Buf: upperLayerData.ToVectorisedView(), + }}, + }, + { + name: "Unknwon next header raw payload (across views)", + firstNextHdr: 255, + payload: makeVectorisedViewFromByteBuffers(upperLayerData[:2], upperLayerData[2:]), + expected: []IPv6PayloadHeader{IPv6RawPayloadHeader{ + Identifier: 255, + Buf: makeVectorisedViewFromByteBuffers(upperLayerData[:2], upperLayerData[2:]), + }}, + }, + } + + for _, test := range tests { + t.Run(test.name, func(t *testing.T) { + it := MakeIPv6PayloadIterator(test.firstNextHdr, test.payload) + + for i, e := range test.expected { + extHdr, done, err := it.Next() + if err != nil { + t.Errorf("(i=%d) Next(): %s", i, err) + } + if done { + t.Errorf("(i=%d) unexpectedly done iterating", i) + } + if diff := cmp.Diff(e, extHdr); diff != "" { + t.Errorf("(i=%d) got ext hdr mismatch (-want +got):\n%s", i, diff) + } + + if t.Failed() { + t.FailNow() + } + } + + extHdr, done, err := it.Next() + if err != nil { + t.Errorf("(last) Next(): %s", err) + } + if !done { + t.Errorf("(last) iterator unexpectedly not done") + } + if extHdr != nil { + t.Errorf("(last) got Next() = %T, want = nil", extHdr) + } + }) + } +} diff --git a/pkg/tcpip/header/ipv6_test.go b/pkg/tcpip/header/ipv6_test.go index c3ad503aa..426a873b1 100644 --- a/pkg/tcpip/header/ipv6_test.go +++ b/pkg/tcpip/header/ipv6_test.go @@ -27,11 +27,12 @@ import ( ) const ( - linkAddr = tcpip.LinkAddress("\x02\x02\x03\x04\x05\x06") - linkLocalAddr = tcpip.Address("\xfe\x80\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x01") - uniqueLocalAddr1 = tcpip.Address("\xfc\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x01") - uniqueLocalAddr2 = tcpip.Address("\xfd\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x02") - globalAddr = tcpip.Address("\xa0\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x01") + linkAddr = tcpip.LinkAddress("\x02\x02\x03\x04\x05\x06") + linkLocalAddr = tcpip.Address("\xfe\x80\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x01") + linkLocalMulticastAddr = tcpip.Address("\xff\x02\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x01") + uniqueLocalAddr1 = tcpip.Address("\xfc\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x01") + uniqueLocalAddr2 = tcpip.Address("\xfd\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x02") + globalAddr = tcpip.Address("\xa0\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x01") ) func TestEthernetAdddressToModifiedEUI64(t *testing.T) { @@ -256,6 +257,85 @@ func TestIsV6UniqueLocalAddress(t *testing.T) { } } +func TestIsV6LinkLocalMulticastAddress(t *testing.T) { + tests := []struct { + name string + addr tcpip.Address + expected bool + }{ + { + name: "Valid Link Local Multicast", + addr: linkLocalMulticastAddr, + expected: true, + }, + { + name: "Valid Link Local Multicast with flags", + addr: "\xff\xf2\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x01", + expected: true, + }, + { + name: "Link Local Unicast", + addr: linkLocalAddr, + expected: false, + }, + { + name: "IPv4 Multicast", + addr: "\xe0\x00\x00\x01", + expected: false, + }, + } + + for _, test := range tests { + t.Run(test.name, func(t *testing.T) { + if got := header.IsV6LinkLocalMulticastAddress(test.addr); got != test.expected { + t.Errorf("got header.IsV6LinkLocalMulticastAddress(%s) = %t, want = %t", test.addr, got, test.expected) + } + }) + } +} + +func TestIsV6LinkLocalAddress(t *testing.T) { + tests := []struct { + name string + addr tcpip.Address + expected bool + }{ + { + name: "Valid Link Local Unicast", + addr: linkLocalAddr, + expected: true, + }, + { + name: "Link Local Multicast", + addr: linkLocalMulticastAddr, + expected: false, + }, + { + name: "Unique Local", + addr: uniqueLocalAddr1, + expected: false, + }, + { + name: "Global", + addr: globalAddr, + expected: false, + }, + { + name: "IPv4 Link Local", + addr: "\xa9\xfe\x00\x01", + expected: false, + }, + } + + for _, test := range tests { + t.Run(test.name, func(t *testing.T) { + if got := header.IsV6LinkLocalAddress(test.addr); got != test.expected { + t.Errorf("got header.IsV6LinkLocalAddress(%s) = %t, want = %t", test.addr, got, test.expected) + } + }) + } +} + func TestScopeForIPv6Address(t *testing.T) { tests := []struct { name string @@ -270,12 +350,18 @@ func TestScopeForIPv6Address(t *testing.T) { err: nil, }, { - name: "Link Local", + name: "Link Local Unicast", addr: linkLocalAddr, scope: header.LinkLocalScope, err: nil, }, { + name: "Link Local Multicast", + addr: linkLocalMulticastAddr, + scope: header.LinkLocalScope, + err: nil, + }, + { name: "Global", addr: globalAddr, scope: header.GlobalScope, diff --git a/pkg/tcpip/header/ndp_options.go b/pkg/tcpip/header/ndp_options.go index e6a6ad39b..5d3975c56 100644 --- a/pkg/tcpip/header/ndp_options.go +++ b/pkg/tcpip/header/ndp_options.go @@ -15,32 +15,47 @@ package header import ( + "bytes" "encoding/binary" "errors" "fmt" + "io" "math" "time" "gvisor.dev/gvisor/pkg/tcpip" ) +// NDPOptionIdentifier is an NDP option type identifier. +type NDPOptionIdentifier uint8 + const ( // NDPSourceLinkLayerAddressOptionType is the type of the Source Link Layer // Address option, as per RFC 4861 section 4.6.1. - NDPSourceLinkLayerAddressOptionType = 1 + NDPSourceLinkLayerAddressOptionType NDPOptionIdentifier = 1 // NDPTargetLinkLayerAddressOptionType is the type of the Target Link Layer // Address option, as per RFC 4861 section 4.6.1. - NDPTargetLinkLayerAddressOptionType = 2 + NDPTargetLinkLayerAddressOptionType NDPOptionIdentifier = 2 + + // NDPPrefixInformationType is the type of the Prefix Information + // option, as per RFC 4861 section 4.6.2. + NDPPrefixInformationType NDPOptionIdentifier = 3 + + // NDPRecursiveDNSServerOptionType is the type of the Recursive DNS + // Server option, as per RFC 8106 section 5.1. + NDPRecursiveDNSServerOptionType NDPOptionIdentifier = 25 + // NDPDNSSearchListOptionType is the type of the DNS Search List option, + // as per RFC 8106 section 5.2. + NDPDNSSearchListOptionType = 31 +) + +const ( // NDPLinkLayerAddressSize is the size of a Source or Target Link Layer // Address option for an Ethernet address. NDPLinkLayerAddressSize = 8 - // NDPPrefixInformationType is the type of the Prefix Information - // option, as per RFC 4861 section 4.6.2. - NDPPrefixInformationType = 3 - // ndpPrefixInformationLength is the expected length, in bytes, of the // body of an NDP Prefix Information option, as per RFC 4861 section // 4.6.2 which specifies that the Length field is 4. Given this, the @@ -91,10 +106,6 @@ const ( // within an NDPPrefixInformation. ndpPrefixInformationPrefixOffset = 14 - // NDPRecursiveDNSServerOptionType is the type of the Recursive DNS - // Server option, as per RFC 8106 section 5.1. - NDPRecursiveDNSServerOptionType = 25 - // ndpRecursiveDNSServerLifetimeOffset is the start of the 4-byte // Lifetime field within an NDPRecursiveDNSServer. ndpRecursiveDNSServerLifetimeOffset = 2 @@ -103,10 +114,31 @@ const ( // for IPv6 Recursive DNS Servers within an NDPRecursiveDNSServer. ndpRecursiveDNSServerAddressesOffset = 6 - // minNDPRecursiveDNSServerLength is the minimum NDP Recursive DNS - // Server option's length field value when it contains at least one - // IPv6 address. - minNDPRecursiveDNSServerLength = 3 + // minNDPRecursiveDNSServerLength is the minimum NDP Recursive DNS Server + // option's body size when it contains at least one IPv6 address, as per + // RFC 8106 section 5.3.1. + minNDPRecursiveDNSServerBodySize = 22 + + // ndpDNSSearchListLifetimeOffset is the start of the 4-byte + // Lifetime field within an NDPDNSSearchList. + ndpDNSSearchListLifetimeOffset = 2 + + // ndpDNSSearchListDomainNamesOffset is the start of the DNS search list + // domain names within an NDPDNSSearchList. + ndpDNSSearchListDomainNamesOffset = 6 + + // minNDPDNSSearchListBodySize is the minimum NDP DNS Search List option's + // body size when it contains at least one domain name, as per RFC 8106 + // section 5.3.1. + minNDPDNSSearchListBodySize = 14 + + // maxDomainNameLabelLength is the maximum length of a domain name + // label, as per RFC 1035 section 3.1. + maxDomainNameLabelLength = 63 + + // maxDomainNameLength is the maximum length of a domain name, including + // label AND label length octet, as per RFC 1035 section 3.1. + maxDomainNameLength = 255 // lengthByteUnits is the multiplier factor for the Length field of an // NDP option. That is, the length field for NDP options is in units of @@ -132,16 +164,13 @@ var ( // few NDPOption then modify the backing NDPOptions so long as the // NDPOptionIterator obtained before modification is no longer used. type NDPOptionIterator struct { - // The NDPOptions this NDPOptionIterator is iterating over. - opts NDPOptions + opts *bytes.Buffer } // Potential errors when iterating over an NDPOptions. var ( - ErrNDPOptBufExhausted = errors.New("Buffer unexpectedly exhausted") - ErrNDPOptZeroLength = errors.New("NDP option has zero-valued Length field") - ErrNDPOptMalformedBody = errors.New("NDP option has a malformed body") - ErrNDPInvalidLength = errors.New("NDP option's Length value is invalid as per relevant RFC") + ErrNDPOptMalformedBody = errors.New("NDP option has a malformed body") + ErrNDPOptMalformedHeader = errors.New("NDP option has a malformed header") ) // Next returns the next element in the backing NDPOptions, or true if we are @@ -152,48 +181,50 @@ var ( func (i *NDPOptionIterator) Next() (NDPOption, bool, error) { for { // Do we still have elements to look at? - if len(i.opts) == 0 { + if i.opts.Len() == 0 { return nil, true, nil } - // Do we have enough bytes for an NDP option that has a Length - // field of at least 1? Note, 0 in the Length field is invalid. - if len(i.opts) < lengthByteUnits { - return nil, true, ErrNDPOptBufExhausted - } - // Get the Type field. - t := i.opts[0] - - // Get the Length field. - l := i.opts[1] + temp, err := i.opts.ReadByte() + if err != nil { + if err != io.EOF { + // ReadByte should only ever return nil or io.EOF. + panic(fmt.Sprintf("unexpected error when reading the option's Type field: %s", err)) + } - // This would indicate an erroneous NDP option as the Length - // field should never be 0. - if l == 0 { - return nil, true, ErrNDPOptZeroLength + // We use io.ErrUnexpectedEOF as exhausting the buffer is unexpected once + // we start parsing an option; we expect the buffer to contain enough + // bytes for the whole option. + return nil, true, fmt.Errorf("unexpectedly exhausted buffer when reading the option's Type field: %w", io.ErrUnexpectedEOF) } + kind := NDPOptionIdentifier(temp) - // How many bytes are in the option body? - numBytes := int(l) * lengthByteUnits - numBodyBytes := numBytes - 2 - - potentialBody := i.opts[2:] + // Get the Length field. + length, err := i.opts.ReadByte() + if err != nil { + if err != io.EOF { + panic(fmt.Sprintf("unexpected error when reading the option's Length field for %s: %s", kind, err)) + } - // This would indicate an erroenous NDPOptions buffer as we ran - // out of the buffer in the middle of an NDP option. - if left := len(potentialBody); left < numBodyBytes { - return nil, true, ErrNDPOptBufExhausted + return nil, true, fmt.Errorf("unexpectedly exhausted buffer when reading the option's Length field for %s: %w", kind, io.ErrUnexpectedEOF) } - // Get only the options body, leaving the rest of the options - // buffer alone. - body := potentialBody[:numBodyBytes] + // This would indicate an erroneous NDP option as the Length field should + // never be 0. + if length == 0 { + return nil, true, fmt.Errorf("zero valued Length field for %s: %w", kind, ErrNDPOptMalformedHeader) + } - // Update opts with the remaining options body. - i.opts = i.opts[numBytes:] + // Get the body. + numBytes := int(length) * lengthByteUnits + numBodyBytes := numBytes - 2 + body := i.opts.Next(numBodyBytes) + if len(body) < numBodyBytes { + return nil, true, fmt.Errorf("unexpectedly exhausted buffer when reading the option's Body for %s: %w", kind, io.ErrUnexpectedEOF) + } - switch t { + switch kind { case NDPSourceLinkLayerAddressOptionType: return NDPSourceLinkLayerAddressOption(body), false, nil @@ -205,22 +236,23 @@ func (i *NDPOptionIterator) Next() (NDPOption, bool, error) { // body is ndpPrefixInformationLength, as per RFC 4861 // section 4.6.2. if numBodyBytes != ndpPrefixInformationLength { - return nil, true, ErrNDPOptMalformedBody + return nil, true, fmt.Errorf("got %d bytes for NDP Prefix Information option's body, expected %d bytes: %w", numBodyBytes, ndpPrefixInformationLength, ErrNDPOptMalformedBody) } return NDPPrefixInformation(body), false, nil case NDPRecursiveDNSServerOptionType: - // RFC 8106 section 5.3.1 outlines that the RDNSS option - // must have a minimum length of 3 so it contains at - // least one IPv6 address. - if l < minNDPRecursiveDNSServerLength { - return nil, true, ErrNDPInvalidLength + opt := NDPRecursiveDNSServer(body) + if err := opt.checkAddresses(); err != nil { + return nil, true, err } - opt := NDPRecursiveDNSServer(body) - if len(opt.Addresses()) == 0 { - return nil, true, ErrNDPOptMalformedBody + return opt, false, nil + + case NDPDNSSearchListOptionType: + opt := NDPDNSSearchList(body) + if err := opt.checkDomainNames(); err != nil { + return nil, true, err } return opt, false, nil @@ -247,10 +279,16 @@ type NDPOptions []byte // // See NDPOptionIterator for more information. func (b NDPOptions) Iter(check bool) (NDPOptionIterator, error) { - it := NDPOptionIterator{opts: b} + it := NDPOptionIterator{ + opts: bytes.NewBuffer(b), + } if check { - for it2 := it; true; { + it2 := NDPOptionIterator{ + opts: bytes.NewBuffer(b), + } + + for { if _, done, err := it2.Next(); err != nil || done { return it, err } @@ -278,7 +316,7 @@ func (b NDPOptions) Serialize(s NDPOptionsSerializer) int { continue } - b[0] = o.Type() + b[0] = byte(o.Type()) // We know this safe because paddedLength would have returned // 0 if o had an invalid length (> 255 * lengthByteUnits). @@ -304,7 +342,7 @@ type NDPOption interface { fmt.Stringer // Type returns the type of the receiver. - Type() uint8 + Type() NDPOptionIdentifier // Length returns the length of the body of the receiver, in bytes. Length() int @@ -386,7 +424,7 @@ func (b NDPOptionsSerializer) Length() int { type NDPSourceLinkLayerAddressOption tcpip.LinkAddress // Type implements NDPOption.Type. -func (o NDPSourceLinkLayerAddressOption) Type() uint8 { +func (o NDPSourceLinkLayerAddressOption) Type() NDPOptionIdentifier { return NDPSourceLinkLayerAddressOptionType } @@ -426,7 +464,7 @@ func (o NDPSourceLinkLayerAddressOption) EthernetAddress() tcpip.LinkAddress { type NDPTargetLinkLayerAddressOption tcpip.LinkAddress // Type implements NDPOption.Type. -func (o NDPTargetLinkLayerAddressOption) Type() uint8 { +func (o NDPTargetLinkLayerAddressOption) Type() NDPOptionIdentifier { return NDPTargetLinkLayerAddressOptionType } @@ -466,7 +504,7 @@ func (o NDPTargetLinkLayerAddressOption) EthernetAddress() tcpip.LinkAddress { type NDPPrefixInformation []byte // Type implements NDPOption.Type. -func (o NDPPrefixInformation) Type() uint8 { +func (o NDPPrefixInformation) Type() NDPOptionIdentifier { return NDPPrefixInformationType } @@ -590,7 +628,7 @@ type NDPRecursiveDNSServer []byte // Type returns the type of an NDP Recursive DNS Server option. // // Type implements NDPOption.Type. -func (NDPRecursiveDNSServer) Type() uint8 { +func (NDPRecursiveDNSServer) Type() NDPOptionIdentifier { return NDPRecursiveDNSServerOptionType } @@ -613,7 +651,12 @@ func (o NDPRecursiveDNSServer) serializeInto(b []byte) int { // String implements fmt.Stringer.String. func (o NDPRecursiveDNSServer) String() string { - return fmt.Sprintf("%T(%s valid for %s)", o, o.Addresses(), o.Lifetime()) + lt := o.Lifetime() + addrs, err := o.Addresses() + if err != nil { + return fmt.Sprintf("%T([] valid for %s; err = %s)", o, lt, err) + } + return fmt.Sprintf("%T(%s valid for %s)", o, addrs, lt) } // Lifetime returns the length of time that the DNS server addresses @@ -632,29 +675,225 @@ func (o NDPRecursiveDNSServer) Lifetime() time.Duration { // Addresses returns the recursive DNS server IPv6 addresses that may be // used for name resolution. // -// Note, some of the addresses returned MAY be link-local addresses. +// Note, the addresses MAY be link-local addresses. +func (o NDPRecursiveDNSServer) Addresses() ([]tcpip.Address, error) { + var addrs []tcpip.Address + return addrs, o.iterAddresses(func(addr tcpip.Address) { addrs = append(addrs, addr) }) +} + +// checkAddresses iterates over the addresses in an NDP Recursive DNS Server +// option and returns any error it encounters. +func (o NDPRecursiveDNSServer) checkAddresses() error { + return o.iterAddresses(nil) +} + +// iterAddresses iterates over the addresses in an NDP Recursive DNS Server +// option and calls a function with each valid unicast IPv6 address. // -// Addresses may panic if o does not hold valid IPv6 addresses. -func (o NDPRecursiveDNSServer) Addresses() []tcpip.Address { - l := len(o) - if l < ndpRecursiveDNSServerAddressesOffset { - return nil +// Note, the addresses MAY be link-local addresses. +func (o NDPRecursiveDNSServer) iterAddresses(fn func(tcpip.Address)) error { + if l := len(o); l < minNDPRecursiveDNSServerBodySize { + return fmt.Errorf("got %d bytes for NDP Recursive DNS Server option's body, expected at least %d bytes: %w", l, minNDPRecursiveDNSServerBodySize, io.ErrUnexpectedEOF) } - l -= ndpRecursiveDNSServerAddressesOffset + o = o[ndpRecursiveDNSServerAddressesOffset:] + l := len(o) if l%IPv6AddressSize != 0 { - return nil + return fmt.Errorf("NDP Recursive DNS Server option's body ends in the middle of an IPv6 address (addresses body size = %d bytes): %w", l, ErrNDPOptMalformedBody) } - buf := o[ndpRecursiveDNSServerAddressesOffset:] - var addrs []tcpip.Address - for len(buf) > 0 { - addr := tcpip.Address(buf[:IPv6AddressSize]) + for i := 0; len(o) != 0; i++ { + addr := tcpip.Address(o[:IPv6AddressSize]) if !IsV6UnicastAddress(addr) { - return nil + return fmt.Errorf("%d-th address (%s) in NDP Recursive DNS Server option is not a valid unicast IPv6 address: %w", i, addr, ErrNDPOptMalformedBody) + } + + if fn != nil { + fn(addr) } - addrs = append(addrs, addr) - buf = buf[IPv6AddressSize:] + + o = o[IPv6AddressSize:] } - return addrs + + return nil +} + +// NDPDNSSearchList is the NDP DNS Search List option, as defined by +// RFC 8106 section 5.2. +type NDPDNSSearchList []byte + +// Type implements NDPOption.Type. +func (o NDPDNSSearchList) Type() NDPOptionIdentifier { + return NDPDNSSearchListOptionType +} + +// Length implements NDPOption.Length. +func (o NDPDNSSearchList) Length() int { + return len(o) +} + +// serializeInto implements NDPOption.serializeInto. +func (o NDPDNSSearchList) serializeInto(b []byte) int { + used := copy(b, o) + + // Zero out the reserved bytes that are before the Lifetime field. + for i := 0; i < ndpDNSSearchListLifetimeOffset; i++ { + b[i] = 0 + } + + return used +} + +// String implements fmt.Stringer.String. +func (o NDPDNSSearchList) String() string { + lt := o.Lifetime() + domainNames, err := o.DomainNames() + if err != nil { + return fmt.Sprintf("%T([] valid for %s; err = %s)", o, lt, err) + } + return fmt.Sprintf("%T(%s valid for %s)", o, domainNames, lt) +} + +// Lifetime returns the length of time that the DNS search list of domain names +// in this option may be used for name resolution. +// +// Note, a value of 0 implies the domain names should no longer be used, +// and a value of infinity/forever is represented by NDPInfiniteLifetime. +func (o NDPDNSSearchList) Lifetime() time.Duration { + // The field is the time in seconds, as per RFC 8106 section 5.1. + return time.Second * time.Duration(binary.BigEndian.Uint32(o[ndpDNSSearchListLifetimeOffset:])) +} + +// DomainNames returns a DNS search list of domain names. +// +// DomainNames will parse the backing buffer as outlined by RFC 1035 section +// 3.1 and return a list of strings, with all domain names in lower case. +func (o NDPDNSSearchList) DomainNames() ([]string, error) { + var domainNames []string + return domainNames, o.iterDomainNames(func(domainName string) { domainNames = append(domainNames, domainName) }) +} + +// checkDomainNames iterates over the domain names in an NDP DNS Search List +// option and returns any error it encounters. +func (o NDPDNSSearchList) checkDomainNames() error { + return o.iterDomainNames(nil) +} + +// iterDomainNames iterates over the domain names in an NDP DNS Search List +// option and calls a function with each valid domain name. +func (o NDPDNSSearchList) iterDomainNames(fn func(string)) error { + if l := len(o); l < minNDPDNSSearchListBodySize { + return fmt.Errorf("got %d bytes for NDP DNS Search List option's body, expected at least %d bytes: %w", l, minNDPDNSSearchListBodySize, io.ErrUnexpectedEOF) + } + + var searchList bytes.Reader + searchList.Reset(o[ndpDNSSearchListDomainNamesOffset:]) + + var scratch [maxDomainNameLength]byte + domainName := bytes.NewBuffer(scratch[:]) + + // Parse the domain names, as per RFC 1035 section 3.1. + for searchList.Len() != 0 { + domainName.Reset() + + // Parse a label within a domain name, as per RFC 1035 section 3.1. + for { + // The first byte is the label length. + labelLenByte, err := searchList.ReadByte() + if err != nil { + if err != io.EOF { + // ReadByte should only ever return nil or io.EOF. + panic(fmt.Sprintf("unexpected error when reading a label's length: %s", err)) + } + + // We use io.ErrUnexpectedEOF as exhausting the buffer is unexpected + // once we start parsing a domain name; we expect the buffer to contain + // enough bytes for the whole domain name. + return fmt.Errorf("unexpected exhausted buffer while parsing a new label for a domain from NDP Search List option: %w", io.ErrUnexpectedEOF) + } + labelLen := int(labelLenByte) + + // A zero-length label implies the end of a domain name. + if labelLen == 0 { + // If the domain name is empty or we have no callback function, do + // nothing further with the current domain name. + if domainName.Len() == 0 || fn == nil { + break + } + + // Ignore the trailing period in the parsed domain name. + domainName.Truncate(domainName.Len() - 1) + fn(domainName.String()) + break + } + + // The label's length must not exceed the maximum length for a label. + if labelLen > maxDomainNameLabelLength { + return fmt.Errorf("label length of %d bytes is greater than the max label length of %d bytes for an NDP Search List option: %w", labelLen, maxDomainNameLabelLength, ErrNDPOptMalformedBody) + } + + // The label (and trailing period) must not make the domain name too long. + if labelLen+1 > domainName.Cap()-domainName.Len() { + return fmt.Errorf("label would make an NDP Search List option's domain name longer than the max domain name length of %d bytes: %w", maxDomainNameLength, ErrNDPOptMalformedBody) + } + + // Copy the label and add a trailing period. + for i := 0; i < labelLen; i++ { + b, err := searchList.ReadByte() + if err != nil { + if err != io.EOF { + panic(fmt.Sprintf("unexpected error when reading domain name's label: %s", err)) + } + + return fmt.Errorf("read %d out of %d bytes for a domain name's label from NDP Search List option: %w", i, labelLen, io.ErrUnexpectedEOF) + } + + // As per RFC 1035 section 2.3.1: + // 1) the label must only contain ASCII include letters, digits and + // hyphens + // 2) the first character in a label must be a letter + // 3) the last letter in a label must be a letter or digit + + if !isLetter(b) { + if i == 0 { + return fmt.Errorf("first character of a domain name's label in an NDP Search List option must be a letter, got character code = %d: %w", b, ErrNDPOptMalformedBody) + } + + if b == '-' { + if i == labelLen-1 { + return fmt.Errorf("last character of a domain name's label in an NDP Search List option must not be a hyphen (-): %w", ErrNDPOptMalformedBody) + } + } else if !isDigit(b) { + return fmt.Errorf("domain name's label in an NDP Search List option may only contain letters, digits and hyphens, got character code = %d: %w", b, ErrNDPOptMalformedBody) + } + } + + // If b is an upper case character, make it lower case. + if isUpperLetter(b) { + b = b - 'A' + 'a' + } + + if err := domainName.WriteByte(b); err != nil { + panic(fmt.Sprintf("unexpected error writing label to domain name buffer: %s", err)) + } + } + if err := domainName.WriteByte('.'); err != nil { + panic(fmt.Sprintf("unexpected error writing trailing period to domain name buffer: %s", err)) + } + } + } + + return nil +} + +func isLetter(b byte) bool { + return b >= 'a' && b <= 'z' || isUpperLetter(b) +} + +func isUpperLetter(b byte) bool { + return b >= 'A' && b <= 'Z' +} + +func isDigit(b byte) bool { + return b >= '0' && b <= '9' } diff --git a/pkg/tcpip/header/ndp_test.go b/pkg/tcpip/header/ndp_test.go index 1cb9f5dc8..dc4591253 100644 --- a/pkg/tcpip/header/ndp_test.go +++ b/pkg/tcpip/header/ndp_test.go @@ -16,6 +16,10 @@ package header import ( "bytes" + "errors" + "fmt" + "io" + "regexp" "testing" "time" @@ -115,7 +119,7 @@ func TestNDPNeighborAdvert(t *testing.T) { // Make sure flags got updated in the backing buffer. if got := b[ndpNAFlagsOffset]; got != 64 { - t.Errorf("got flags byte = %d, want = 64") + t.Errorf("got flags byte = %d, want = 64", got) } } @@ -543,8 +547,12 @@ func TestNDPRecursiveDNSServerOptionSerialize(t *testing.T) { want := []tcpip.Address{ "\x00\x01\x02\x03\x04\x05\x06\x07\x08\x09\x0a\x0b\x0c\x0d\x0e\x0f", } - if got := opt.Addresses(); !cmp.Equal(got, want) { - t.Errorf("got Addresses = %v, want = %v", got, want) + addrs, err := opt.Addresses() + if err != nil { + t.Errorf("opt.Addresses() = %s", err) + } + if diff := cmp.Diff(addrs, want); diff != "" { + t.Errorf("mismatched addresses (-want +got):\n%s", diff) } // Iterator should not return anything else. @@ -638,8 +646,12 @@ func TestNDPRecursiveDNSServerOption(t *testing.T) { if got := opt.Lifetime(); got != test.lifetime { t.Errorf("got Lifetime = %d, want = %d", got, test.lifetime) } - if got := opt.Addresses(); !cmp.Equal(got, test.addrs) { - t.Errorf("got Addresses = %v, want = %v", got, test.addrs) + addrs, err := opt.Addresses() + if err != nil { + t.Errorf("opt.Addresses() = %s", err) + } + if diff := cmp.Diff(addrs, test.addrs); diff != "" { + t.Errorf("mismatched addresses (-want +got):\n%s", diff) } // Iterator should not return anything else. @@ -657,42 +669,513 @@ func TestNDPRecursiveDNSServerOption(t *testing.T) { } } +// TestNDPDNSSearchListOption tests the getters of NDPDNSSearchList. +func TestNDPDNSSearchListOption(t *testing.T) { + tests := []struct { + name string + buf []byte + lifetime time.Duration + domainNames []string + err error + }{ + { + name: "Valid1Label", + buf: []byte{ + 0, 0, + 0, 0, 0, 1, + 3, 'a', 'b', 'c', + 0, + 0, 0, 0, + }, + lifetime: time.Second, + domainNames: []string{ + "abc", + }, + err: nil, + }, + { + name: "Valid2Label", + buf: []byte{ + 0, 0, + 0, 0, 0, 5, + 3, 'a', 'b', 'c', + 4, 'a', 'b', 'c', 'd', + 0, + 0, 0, 0, 0, 0, 0, + }, + lifetime: 5 * time.Second, + domainNames: []string{ + "abc.abcd", + }, + err: nil, + }, + { + name: "Valid3Label", + buf: []byte{ + 0, 0, + 1, 0, 0, 0, + 3, 'a', 'b', 'c', + 4, 'a', 'b', 'c', 'd', + 1, 'e', + 0, + 0, 0, 0, 0, + }, + lifetime: 16777216 * time.Second, + domainNames: []string{ + "abc.abcd.e", + }, + err: nil, + }, + { + name: "Valid2Domains", + buf: []byte{ + 0, 0, + 1, 2, 3, 4, + 3, 'a', 'b', 'c', + 0, + 2, 'd', 'e', + 3, 'x', 'y', 'z', + 0, + 0, 0, 0, + }, + lifetime: 16909060 * time.Second, + domainNames: []string{ + "abc", + "de.xyz", + }, + err: nil, + }, + { + name: "Valid3DomainsMixedCase", + buf: []byte{ + 0, 0, + 0, 0, 0, 0, + 3, 'a', 'B', 'c', + 0, + 2, 'd', 'E', + 3, 'X', 'y', 'z', + 0, + 1, 'J', + 0, + }, + lifetime: 0, + domainNames: []string{ + "abc", + "de.xyz", + "j", + }, + err: nil, + }, + { + name: "ValidDomainAfterNULL", + buf: []byte{ + 0, 0, + 0, 0, 0, 0, + 3, 'a', 'B', 'c', + 0, 0, 0, 0, + 2, 'd', 'E', + 3, 'X', 'y', 'z', + 0, + }, + lifetime: 0, + domainNames: []string{ + "abc", + "de.xyz", + }, + err: nil, + }, + { + name: "Valid0Domains", + buf: []byte{ + 0, 0, + 0, 0, 0, 0, + 0, + 0, 0, 0, 0, 0, 0, 0, + }, + lifetime: 0, + domainNames: nil, + err: nil, + }, + { + name: "NoTrailingNull", + buf: []byte{ + 0, 0, + 0, 0, 0, 0, + 7, 'a', 'b', 'c', 'd', 'e', 'f', 'g', + }, + lifetime: 0, + domainNames: nil, + err: io.ErrUnexpectedEOF, + }, + { + name: "IncorrectLength", + buf: []byte{ + 0, 0, + 0, 0, 0, 0, + 8, 'a', 'b', 'c', 'd', 'e', 'f', 'g', + }, + lifetime: 0, + domainNames: nil, + err: io.ErrUnexpectedEOF, + }, + { + name: "IncorrectLengthWithNULL", + buf: []byte{ + 0, 0, + 0, 0, 0, 0, + 7, 'a', 'b', 'c', 'd', 'e', 'f', + 0, + }, + lifetime: 0, + domainNames: nil, + err: ErrNDPOptMalformedBody, + }, + { + name: "LabelOfLength63", + buf: []byte{ + 0, 0, + 0, 0, 0, 0, + 63, 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', + 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', + 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', + 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', + 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', + 0, + }, + lifetime: 0, + domainNames: []string{ + "abcdefghijklmnopqrstuvwxyzabcdefghijklmnopqrstuvwxyzabcdefghijk", + }, + err: nil, + }, + { + name: "LabelOfLength64", + buf: []byte{ + 0, 0, + 0, 0, 0, 0, + 64, 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', + 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', + 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', + 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', + 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', 'l', + 0, + }, + lifetime: 0, + domainNames: nil, + err: ErrNDPOptMalformedBody, + }, + { + name: "DomainNameOfLength255", + buf: []byte{ + 0, 0, + 0, 0, 0, 0, + 63, 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', + 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', + 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', + 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', + 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', + 63, 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', + 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', + 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', + 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', + 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', + 63, 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', + 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', + 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', + 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', + 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', + 62, 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', + 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', + 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', + 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', + 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', + 0, + }, + lifetime: 0, + domainNames: []string{ + "abcdefghijklmnopqrstuvwxyzabcdefghijklmnopqrstuvwxyzabcdefghijk.abcdefghijklmnopqrstuvwxyzabcdefghijklmnopqrstuvwxyzabcdefghijk.abcdefghijklmnopqrstuvwxyzabcdefghijklmnopqrstuvwxyzabcdefghijk.abcdefghijklmnopqrstuvwxyzabcdefghijklmnopqrstuvwxyzabcdefghij", + }, + err: nil, + }, + { + name: "DomainNameOfLength256", + buf: []byte{ + 0, 0, + 0, 0, 0, 0, + 63, 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', + 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', + 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', + 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', + 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', + 63, 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', + 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', + 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', + 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', + 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', + 63, 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', + 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', + 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', + 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', + 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', + 63, 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', + 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', + 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', + 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', + 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', + 0, + }, + lifetime: 0, + domainNames: nil, + err: ErrNDPOptMalformedBody, + }, + { + name: "StartingDigitForLabel", + buf: []byte{ + 0, 0, + 0, 0, 0, 1, + 3, '9', 'b', 'c', + 0, + 0, 0, 0, + }, + lifetime: time.Second, + domainNames: nil, + err: ErrNDPOptMalformedBody, + }, + { + name: "StartingHyphenForLabel", + buf: []byte{ + 0, 0, + 0, 0, 0, 1, + 3, '-', 'b', 'c', + 0, + 0, 0, 0, + }, + lifetime: time.Second, + domainNames: nil, + err: ErrNDPOptMalformedBody, + }, + { + name: "EndingHyphenForLabel", + buf: []byte{ + 0, 0, + 0, 0, 0, 1, + 3, 'a', 'b', '-', + 0, + 0, 0, 0, + }, + lifetime: time.Second, + domainNames: nil, + err: ErrNDPOptMalformedBody, + }, + { + name: "EndingDigitForLabel", + buf: []byte{ + 0, 0, + 0, 0, 0, 1, + 3, 'a', 'b', '9', + 0, + 0, 0, 0, + }, + lifetime: time.Second, + domainNames: []string{ + "ab9", + }, + err: nil, + }, + } + + for _, test := range tests { + t.Run(test.name, func(t *testing.T) { + opt := NDPDNSSearchList(test.buf) + + if got := opt.Lifetime(); got != test.lifetime { + t.Errorf("got Lifetime = %d, want = %d", got, test.lifetime) + } + domainNames, err := opt.DomainNames() + if !errors.Is(err, test.err) { + t.Errorf("opt.DomainNames() = %s", err) + } + if diff := cmp.Diff(domainNames, test.domainNames); diff != "" { + t.Errorf("mismatched domain names (-want +got):\n%s", diff) + } + }) + } +} + +func TestNDPSearchListOptionDomainNameLabelInvalidSymbols(t *testing.T) { + for r := rune(0); r <= 255; r++ { + t.Run(fmt.Sprintf("RuneVal=%d", r), func(t *testing.T) { + buf := []byte{ + 0, 0, + 0, 0, 0, 0, + 3, 'a', 0 /* will be replaced */, 'c', + 0, + 0, 0, 0, + } + buf[8] = uint8(r) + opt := NDPDNSSearchList(buf) + + // As per RFC 1035 section 2.3.1, the label must only include ASCII + // letters, digits and hyphens (a-z, A-Z, 0-9, -). + var expectedErr error + re := regexp.MustCompile(`[a-zA-Z0-9-]`) + if !re.Match([]byte{byte(r)}) { + expectedErr = ErrNDPOptMalformedBody + } + + if domainNames, err := opt.DomainNames(); !errors.Is(err, expectedErr) { + t.Errorf("got opt.DomainNames() = (%s, %v), want = (_, %v)", domainNames, err, ErrNDPOptMalformedBody) + } + }) + } +} + +func TestNDPDNSSearchListOptionSerialize(t *testing.T) { + b := []byte{ + 9, 8, + 1, 0, 0, 0, + 3, 'a', 'b', 'c', + 4, 'a', 'b', 'c', 'd', + 1, 'e', + 0, + } + targetBuf := []byte{1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1} + expected := []byte{ + 31, 3, 0, 0, + 1, 0, 0, 0, + 3, 'a', 'b', 'c', + 4, 'a', 'b', 'c', 'd', + 1, 'e', + 0, + 0, 0, 0, 0, + } + opts := NDPOptions(targetBuf) + serializer := NDPOptionsSerializer{ + NDPDNSSearchList(b), + } + if got, want := opts.Serialize(serializer), len(expected); got != want { + t.Errorf("got Serialize = %d, want = %d", got, want) + } + if !bytes.Equal(targetBuf, expected) { + t.Fatalf("got targetBuf = %x, want = %x", targetBuf, expected) + } + + it, err := opts.Iter(true) + if err != nil { + t.Fatalf("got Iter = (_, %s), want = (_, nil)", err) + } + + next, done, err := it.Next() + if err != nil { + t.Fatalf("got Next = (_, _, %s), want = (_, _, nil)", err) + } + if done { + t.Fatal("got Next = (_, true, _), want = (_, false, _)") + } + if got := next.Type(); got != NDPDNSSearchListOptionType { + t.Errorf("got Type = %d, want = %d", got, NDPDNSSearchListOptionType) + } + + opt, ok := next.(NDPDNSSearchList) + if !ok { + t.Fatalf("next (type = %T) cannot be casted to an NDPDNSSearchList", next) + } + if got := opt.Type(); got != 31 { + t.Errorf("got Type = %d, want = 31", got) + } + if got := opt.Length(); got != 22 { + t.Errorf("got Length = %d, want = 22", got) + } + if got, want := opt.Lifetime(), 16777216*time.Second; got != want { + t.Errorf("got Lifetime = %s, want = %s", got, want) + } + domainNames, err := opt.DomainNames() + if err != nil { + t.Errorf("opt.DomainNames() = %s", err) + } + if diff := cmp.Diff(domainNames, []string{"abc.abcd.e"}); diff != "" { + t.Errorf("domain names mismatch (-want +got):\n%s", diff) + } + + // Iterator should not return anything else. + next, done, err = it.Next() + if err != nil { + t.Errorf("got Next = (_, _, %s), want = (_, _, nil)", err) + } + if !done { + t.Error("got Next = (_, false, _), want = (_, true, _)") + } + if next != nil { + t.Errorf("got Next = (%x, _, _), want = (nil, _, _)", next) + } +} + // TestNDPOptionsIterCheck tests that Iter will return false if the NDPOptions // the iterator was returned for is malformed. func TestNDPOptionsIterCheck(t *testing.T) { tests := []struct { - name string - buf []byte - expected error + name string + buf []byte + expectedErr error }{ { - "ZeroLengthField", - []byte{0, 0, 0, 0, 0, 0, 0, 0}, - ErrNDPOptZeroLength, + name: "ZeroLengthField", + buf: []byte{0, 0, 0, 0, 0, 0, 0, 0}, + expectedErr: ErrNDPOptMalformedHeader, }, { - "ValidSourceLinkLayerAddressOption", - []byte{1, 1, 1, 2, 3, 4, 5, 6}, - nil, + name: "ValidSourceLinkLayerAddressOption", + buf: []byte{1, 1, 1, 2, 3, 4, 5, 6}, + expectedErr: nil, }, { - "TooSmallSourceLinkLayerAddressOption", - []byte{1, 1, 1, 2, 3, 4, 5}, - ErrNDPOptBufExhausted, + name: "TooSmallSourceLinkLayerAddressOption", + buf: []byte{1, 1, 1, 2, 3, 4, 5}, + expectedErr: io.ErrUnexpectedEOF, }, { - "ValidTargetLinkLayerAddressOption", - []byte{2, 1, 1, 2, 3, 4, 5, 6}, - nil, + name: "ValidTargetLinkLayerAddressOption", + buf: []byte{2, 1, 1, 2, 3, 4, 5, 6}, + expectedErr: nil, }, { - "TooSmallTargetLinkLayerAddressOption", - []byte{2, 1, 1, 2, 3, 4, 5}, - ErrNDPOptBufExhausted, + name: "TooSmallTargetLinkLayerAddressOption", + buf: []byte{2, 1, 1, 2, 3, 4, 5}, + expectedErr: io.ErrUnexpectedEOF, }, { - "ValidPrefixInformation", - []byte{ + name: "ValidPrefixInformation", + buf: []byte{ 3, 4, 43, 64, 1, 2, 3, 4, 5, 6, 7, 8, @@ -702,11 +1185,11 @@ func TestNDPOptionsIterCheck(t *testing.T) { 17, 18, 19, 20, 21, 22, 23, 24, }, - nil, + expectedErr: nil, }, { - "TooSmallPrefixInformation", - []byte{ + name: "TooSmallPrefixInformation", + buf: []byte{ 3, 4, 43, 64, 1, 2, 3, 4, 5, 6, 7, 8, @@ -716,11 +1199,11 @@ func TestNDPOptionsIterCheck(t *testing.T) { 17, 18, 19, 20, 21, 22, 23, }, - ErrNDPOptBufExhausted, + expectedErr: io.ErrUnexpectedEOF, }, { - "InvalidPrefixInformationLength", - []byte{ + name: "InvalidPrefixInformationLength", + buf: []byte{ 3, 3, 43, 64, 1, 2, 3, 4, 5, 6, 7, 8, @@ -728,11 +1211,11 @@ func TestNDPOptionsIterCheck(t *testing.T) { 9, 10, 11, 12, 13, 14, 15, 16, }, - ErrNDPOptMalformedBody, + expectedErr: ErrNDPOptMalformedBody, }, { - "ValidSourceAndTargetLinkLayerAddressWithPrefixInformation", - []byte{ + name: "ValidSourceAndTargetLinkLayerAddressWithPrefixInformation", + buf: []byte{ // Source Link-Layer Address. 1, 1, 1, 2, 3, 4, 5, 6, @@ -749,11 +1232,11 @@ func TestNDPOptionsIterCheck(t *testing.T) { 17, 18, 19, 20, 21, 22, 23, 24, }, - nil, + expectedErr: nil, }, { - "ValidSourceAndTargetLinkLayerAddressWithPrefixInformationWithUnrecognized", - []byte{ + name: "ValidSourceAndTargetLinkLayerAddressWithPrefixInformationWithUnrecognized", + buf: []byte{ // Source Link-Layer Address. 1, 1, 1, 2, 3, 4, 5, 6, @@ -775,52 +1258,153 @@ func TestNDPOptionsIterCheck(t *testing.T) { 17, 18, 19, 20, 21, 22, 23, 24, }, - nil, + expectedErr: nil, }, { - "InvalidRecursiveDNSServerCutsOffAddress", - []byte{ + name: "InvalidRecursiveDNSServerCutsOffAddress", + buf: []byte{ 25, 4, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 1, 2, 3, 4, 5, 6, 7, }, - ErrNDPOptMalformedBody, + expectedErr: ErrNDPOptMalformedBody, }, { - "InvalidRecursiveDNSServerInvalidLengthField", - []byte{ + name: "InvalidRecursiveDNSServerInvalidLengthField", + buf: []byte{ 25, 2, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8, }, - ErrNDPInvalidLength, + expectedErr: io.ErrUnexpectedEOF, }, { - "RecursiveDNSServerTooSmall", - []byte{ + name: "RecursiveDNSServerTooSmall", + buf: []byte{ 25, 1, 0, 0, 0, 0, 0, }, - ErrNDPOptBufExhausted, + expectedErr: io.ErrUnexpectedEOF, }, { - "RecursiveDNSServerMulticast", - []byte{ + name: "RecursiveDNSServerMulticast", + buf: []byte{ 25, 3, 0, 0, 0, 0, 0, 0, 255, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, }, - ErrNDPOptMalformedBody, + expectedErr: ErrNDPOptMalformedBody, }, { - "RecursiveDNSServerUnspecified", - []byte{ + name: "RecursiveDNSServerUnspecified", + buf: []byte{ 25, 3, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, }, - ErrNDPOptMalformedBody, + expectedErr: ErrNDPOptMalformedBody, + }, + { + name: "DNSSearchListLargeCompliantRFC1035", + buf: []byte{ + 31, 33, 0, 0, + 0, 0, 0, 0, + 63, 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', + 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', + 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', + 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', + 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', + 63, 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', + 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', + 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', + 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', + 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', + 63, 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', + 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', + 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', + 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', + 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', + 62, 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', + 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', + 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', + 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', + 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', + 0, + }, + expectedErr: nil, + }, + { + name: "DNSSearchListNonCompliantRFC1035", + buf: []byte{ + 31, 33, 0, 0, + 0, 0, 0, 0, + 63, 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', + 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', + 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', + 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', + 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', + 63, 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', + 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', + 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', + 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', + 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', + 63, 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', + 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', + 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', + 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', + 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', + 63, 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', + 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', + 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', + 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', + 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', + 'i', 'j', 'k', + 0, + 0, 0, 0, 0, 0, 0, 0, 0, + }, + expectedErr: ErrNDPOptMalformedBody, + }, + { + name: "DNSSearchListValidSmall", + buf: []byte{ + 31, 2, 0, 0, + 0, 0, 0, 0, + 6, 'a', 'b', 'c', 'd', 'e', 'f', + 0, + }, + expectedErr: nil, + }, + { + name: "DNSSearchListTooSmall", + buf: []byte{ + 31, 1, 0, 0, + 0, 0, 0, + }, + expectedErr: io.ErrUnexpectedEOF, }, } @@ -828,8 +1412,8 @@ func TestNDPOptionsIterCheck(t *testing.T) { t.Run(test.name, func(t *testing.T) { opts := NDPOptions(test.buf) - if _, err := opts.Iter(true); err != test.expected { - t.Fatalf("got Iter(true) = (_, %v), want = (_, %v)", err, test.expected) + if _, err := opts.Iter(true); !errors.Is(err, test.expectedErr) { + t.Fatalf("got Iter(true) = (_, %v), want = (_, %v)", err, test.expectedErr) } // test.buf may be malformed but we chose not to check diff --git a/pkg/tcpip/header/ndpoptionidentifier_string.go b/pkg/tcpip/header/ndpoptionidentifier_string.go new file mode 100644 index 000000000..6fe9a336b --- /dev/null +++ b/pkg/tcpip/header/ndpoptionidentifier_string.go @@ -0,0 +1,50 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// Code generated by "stringer -type NDPOptionIdentifier ."; DO NOT EDIT. + +package header + +import "strconv" + +func _() { + // An "invalid array index" compiler error signifies that the constant values have changed. + // Re-run the stringer command to generate them again. + var x [1]struct{} + _ = x[NDPSourceLinkLayerAddressOptionType-1] + _ = x[NDPTargetLinkLayerAddressOptionType-2] + _ = x[NDPPrefixInformationType-3] + _ = x[NDPRecursiveDNSServerOptionType-25] +} + +const ( + _NDPOptionIdentifier_name_0 = "NDPSourceLinkLayerAddressOptionTypeNDPTargetLinkLayerAddressOptionTypeNDPPrefixInformationType" + _NDPOptionIdentifier_name_1 = "NDPRecursiveDNSServerOptionType" +) + +var ( + _NDPOptionIdentifier_index_0 = [...]uint8{0, 35, 70, 94} +) + +func (i NDPOptionIdentifier) String() string { + switch { + case 1 <= i && i <= 3: + i -= 1 + return _NDPOptionIdentifier_name_0[_NDPOptionIdentifier_index_0[i]:_NDPOptionIdentifier_index_0[i+1]] + case i == 25: + return _NDPOptionIdentifier_name_1 + default: + return "NDPOptionIdentifier(" + strconv.FormatInt(int64(i), 10) + ")" + } +} diff --git a/pkg/tcpip/header/tcp.go b/pkg/tcpip/header/tcp.go index 82cfe785c..13480687d 100644 --- a/pkg/tcpip/header/tcp.go +++ b/pkg/tcpip/header/tcp.go @@ -81,7 +81,8 @@ type TCPFields struct { // AckNum is the "acknowledgement number" field of a TCP packet. AckNum uint32 - // DataOffset is the "data offset" field of a TCP packet. + // DataOffset is the "data offset" field of a TCP packet. It is the length of + // the TCP header in bytes. DataOffset uint8 // Flags is the "flags" field of a TCP packet. @@ -213,7 +214,8 @@ func (b TCP) AckNumber() uint32 { return binary.BigEndian.Uint32(b[TCPAckNumOffset:]) } -// DataOffset returns the "data offset" field of the tcp header. +// DataOffset returns the "data offset" field of the tcp header. The return +// value is the length of the TCP header in bytes. func (b TCP) DataOffset() uint8 { return (b[TCPDataOffset] >> 4) * 4 } @@ -238,6 +240,11 @@ func (b TCP) Checksum() uint16 { return binary.BigEndian.Uint16(b[TCPChecksumOffset:]) } +// UrgentPointer returns the "urgent pointer" field of the tcp header. +func (b TCP) UrgentPointer() uint16 { + return binary.BigEndian.Uint16(b[TCPUrgentPtrOffset:]) +} + // SetSourcePort sets the "source port" field of the tcp header. func (b TCP) SetSourcePort(port uint16) { binary.BigEndian.PutUint16(b[TCPSrcPortOffset:], port) @@ -253,6 +260,37 @@ func (b TCP) SetChecksum(checksum uint16) { binary.BigEndian.PutUint16(b[TCPChecksumOffset:], checksum) } +// SetDataOffset sets the data offset field of the tcp header. headerLen should +// be the length of the TCP header in bytes. +func (b TCP) SetDataOffset(headerLen uint8) { + b[TCPDataOffset] = (headerLen / 4) << 4 +} + +// SetSequenceNumber sets the sequence number field of the tcp header. +func (b TCP) SetSequenceNumber(seqNum uint32) { + binary.BigEndian.PutUint32(b[TCPSeqNumOffset:], seqNum) +} + +// SetAckNumber sets the ack number field of the tcp header. +func (b TCP) SetAckNumber(ackNum uint32) { + binary.BigEndian.PutUint32(b[TCPAckNumOffset:], ackNum) +} + +// SetFlags sets the flags field of the tcp header. +func (b TCP) SetFlags(flags uint8) { + b[TCPFlagsOffset] = flags +} + +// SetWindowSize sets the window size field of the tcp header. +func (b TCP) SetWindowSize(rcvwnd uint16) { + binary.BigEndian.PutUint16(b[TCPWinSizeOffset:], rcvwnd) +} + +// SetUrgentPoiner sets the window size field of the tcp header. +func (b TCP) SetUrgentPoiner(urgentPointer uint16) { + binary.BigEndian.PutUint16(b[TCPUrgentPtrOffset:], urgentPointer) +} + // CalculateChecksum calculates the checksum of the tcp segment. // partialChecksum is the checksum of the network-layer pseudo-header // and the checksum of the segment data. diff --git a/pkg/tcpip/header/udp.go b/pkg/tcpip/header/udp.go index 74412c894..9339d637f 100644 --- a/pkg/tcpip/header/udp.go +++ b/pkg/tcpip/header/udp.go @@ -99,6 +99,11 @@ func (b UDP) SetChecksum(checksum uint16) { binary.BigEndian.PutUint16(b[udpChecksum:], checksum) } +// SetLength sets the "length" field of the udp header. +func (b UDP) SetLength(length uint16) { + binary.BigEndian.PutUint16(b[udpLength:], length) +} + // CalculateChecksum calculates the checksum of the udp packet, given the // checksum of the network-layer pseudo-header and the checksum of the payload. func (b UDP) CalculateChecksum(partialChecksum uint16) uint16 { diff --git a/pkg/tcpip/iptables/BUILD b/pkg/tcpip/iptables/BUILD deleted file mode 100644 index d1b73cfdf..000000000 --- a/pkg/tcpip/iptables/BUILD +++ /dev/null @@ -1,18 +0,0 @@ -load("//tools:defs.bzl", "go_library") - -package(licenses = ["notice"]) - -go_library( - name = "iptables", - srcs = [ - "iptables.go", - "targets.go", - "types.go", - ], - visibility = ["//visibility:public"], - deps = [ - "//pkg/log", - "//pkg/tcpip", - "//pkg/tcpip/header", - ], -) diff --git a/pkg/tcpip/iptables/iptables.go b/pkg/tcpip/iptables/iptables.go deleted file mode 100644 index 1b9485bbd..000000000 --- a/pkg/tcpip/iptables/iptables.go +++ /dev/null @@ -1,216 +0,0 @@ -// Copyright 2019 The gVisor Authors. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -// Package iptables supports packet filtering and manipulation via the iptables -// tool. -package iptables - -import ( - "fmt" - - "gvisor.dev/gvisor/pkg/tcpip" - "gvisor.dev/gvisor/pkg/tcpip/header" -) - -// Table names. -const ( - TablenameNat = "nat" - TablenameMangle = "mangle" - TablenameFilter = "filter" -) - -// Chain names as defined by net/ipv4/netfilter/ip_tables.c. -const ( - ChainNamePrerouting = "PREROUTING" - ChainNameInput = "INPUT" - ChainNameForward = "FORWARD" - ChainNameOutput = "OUTPUT" - ChainNamePostrouting = "POSTROUTING" -) - -// HookUnset indicates that there is no hook set for an entrypoint or -// underflow. -const HookUnset = -1 - -// DefaultTables returns a default set of tables. Each chain is set to accept -// all packets. -func DefaultTables() IPTables { - // TODO(gvisor.dev/issue/170): We may be able to swap out some strings for - // iotas. - return IPTables{ - Tables: map[string]Table{ - TablenameNat: Table{ - Rules: []Rule{ - Rule{Target: UnconditionalAcceptTarget{}}, - Rule{Target: UnconditionalAcceptTarget{}}, - Rule{Target: UnconditionalAcceptTarget{}}, - Rule{Target: UnconditionalAcceptTarget{}}, - Rule{Target: ErrorTarget{}}, - }, - BuiltinChains: map[Hook]int{ - Prerouting: 0, - Input: 1, - Output: 2, - Postrouting: 3, - }, - Underflows: map[Hook]int{ - Prerouting: 0, - Input: 1, - Output: 2, - Postrouting: 3, - }, - UserChains: map[string]int{}, - }, - TablenameMangle: Table{ - Rules: []Rule{ - Rule{Target: UnconditionalAcceptTarget{}}, - Rule{Target: UnconditionalAcceptTarget{}}, - Rule{Target: ErrorTarget{}}, - }, - BuiltinChains: map[Hook]int{ - Prerouting: 0, - Output: 1, - }, - Underflows: map[Hook]int{ - Prerouting: 0, - Output: 1, - }, - UserChains: map[string]int{}, - }, - TablenameFilter: Table{ - Rules: []Rule{ - Rule{Target: UnconditionalAcceptTarget{}}, - Rule{Target: UnconditionalAcceptTarget{}}, - Rule{Target: UnconditionalAcceptTarget{}}, - Rule{Target: ErrorTarget{}}, - }, - BuiltinChains: map[Hook]int{ - Input: 0, - Forward: 1, - Output: 2, - }, - Underflows: map[Hook]int{ - Input: 0, - Forward: 1, - Output: 2, - }, - UserChains: map[string]int{}, - }, - }, - Priorities: map[Hook][]string{ - Input: []string{TablenameNat, TablenameFilter}, - Prerouting: []string{TablenameMangle, TablenameNat}, - Output: []string{TablenameMangle, TablenameNat, TablenameFilter}, - }, - } -} - -// EmptyFilterTable returns a Table with no rules and the filter table chains -// mapped to HookUnset. -func EmptyFilterTable() Table { - return Table{ - Rules: []Rule{}, - BuiltinChains: map[Hook]int{ - Input: HookUnset, - Forward: HookUnset, - Output: HookUnset, - }, - Underflows: map[Hook]int{ - Input: HookUnset, - Forward: HookUnset, - Output: HookUnset, - }, - UserChains: map[string]int{}, - } -} - -// Check runs pkt through the rules for hook. It returns true when the packet -// should continue traversing the network stack and false when it should be -// dropped. -// -// Precondition: pkt.NetworkHeader is set. -func (it *IPTables) Check(hook Hook, pkt tcpip.PacketBuffer) bool { - // TODO(gvisor.dev/issue/170): A lot of this is uncomplicated because - // we're missing features. Jumps, the call stack, etc. aren't checked - // for yet because we're yet to support them. - - // Go through each table containing the hook. - for _, tablename := range it.Priorities[hook] { - switch verdict := it.checkTable(hook, pkt, tablename); verdict { - // If the table returns Accept, move on to the next table. - case Accept: - continue - // The Drop verdict is final. - case Drop: - return false - case Stolen, Queue, Repeat, None, Jump, Return, Continue: - panic(fmt.Sprintf("Unimplemented verdict %v.", verdict)) - default: - panic(fmt.Sprintf("Unknown verdict %v.", verdict)) - } - } - - // Every table returned Accept. - return true -} - -// Precondition: pkt.NetworkHeader is set. -func (it *IPTables) checkTable(hook Hook, pkt tcpip.PacketBuffer, tablename string) Verdict { - // Start from ruleIdx and walk the list of rules until a rule gives us - // a verdict. - table := it.Tables[tablename] - for ruleIdx := table.BuiltinChains[hook]; ruleIdx < len(table.Rules); ruleIdx++ { - switch verdict := it.checkRule(hook, pkt, table, ruleIdx); verdict { - // In either of these cases, this table is done with the packet. - case Accept, Drop: - return verdict - // Continue traversing the rules of the table. - case Continue: - continue - case Stolen, Queue, Repeat, None, Jump, Return: - panic(fmt.Sprintf("Unimplemented verdict %v.", verdict)) - default: - panic(fmt.Sprintf("Unknown verdict %v.", verdict)) - } - } - - panic(fmt.Sprintf("Traversed past the entire list of iptables rules in table %q.", tablename)) -} - -// Precondition: pk.NetworkHeader is set. -func (it *IPTables) checkRule(hook Hook, pkt tcpip.PacketBuffer, table Table, ruleIdx int) Verdict { - rule := table.Rules[ruleIdx] - - // First check whether the packet matches the IP header filter. - // TODO(gvisor.dev/issue/170): Support other fields of the filter. - if rule.Filter.Protocol != 0 && rule.Filter.Protocol != header.IPv4(pkt.NetworkHeader).TransportProtocol() { - return Continue - } - - // Go through each rule matcher. If they all match, run - // the rule target. - for _, matcher := range rule.Matchers { - matches, hotdrop := matcher.Match(hook, pkt, "") - if hotdrop { - return Drop - } - if !matches { - return Continue - } - } - - // All the matchers matched, so run the target. - verdict, _ := rule.Target.Action(pkt) - return verdict -} diff --git a/pkg/tcpip/iptables/targets.go b/pkg/tcpip/iptables/targets.go deleted file mode 100644 index 4dd281371..000000000 --- a/pkg/tcpip/iptables/targets.go +++ /dev/null @@ -1,48 +0,0 @@ -// Copyright 2019 The gVisor Authors. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -// This file contains various Targets. - -package iptables - -import ( - "gvisor.dev/gvisor/pkg/log" - "gvisor.dev/gvisor/pkg/tcpip" -) - -// UnconditionalAcceptTarget accepts all packets. -type UnconditionalAcceptTarget struct{} - -// Action implements Target.Action. -func (UnconditionalAcceptTarget) Action(packet tcpip.PacketBuffer) (Verdict, string) { - return Accept, "" -} - -// UnconditionalDropTarget denies all packets. -type UnconditionalDropTarget struct{} - -// Action implements Target.Action. -func (UnconditionalDropTarget) Action(packet tcpip.PacketBuffer) (Verdict, string) { - return Drop, "" -} - -// ErrorTarget logs an error and drops the packet. It represents a target that -// should be unreachable. -type ErrorTarget struct{} - -// Action implements Target.Action. -func (ErrorTarget) Action(packet tcpip.PacketBuffer) (Verdict, string) { - log.Warningf("ErrorTarget triggered.") - return Drop, "" -} diff --git a/pkg/tcpip/link/channel/BUILD b/pkg/tcpip/link/channel/BUILD index 3974c464e..b8b93e78e 100644 --- a/pkg/tcpip/link/channel/BUILD +++ b/pkg/tcpip/link/channel/BUILD @@ -7,6 +7,7 @@ go_library( srcs = ["channel.go"], visibility = ["//visibility:public"], deps = [ + "//pkg/sync", "//pkg/tcpip", "//pkg/tcpip/buffer", "//pkg/tcpip/stack", diff --git a/pkg/tcpip/link/channel/channel.go b/pkg/tcpip/link/channel/channel.go index 78d447acd..9bf67686d 100644 --- a/pkg/tcpip/link/channel/channel.go +++ b/pkg/tcpip/link/channel/channel.go @@ -20,6 +20,7 @@ package channel import ( "context" + "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/tcpip" "gvisor.dev/gvisor/pkg/tcpip/buffer" "gvisor.dev/gvisor/pkg/tcpip/stack" @@ -27,12 +28,102 @@ import ( // PacketInfo holds all the information about an outbound packet. type PacketInfo struct { - Pkt tcpip.PacketBuffer + Pkt *stack.PacketBuffer Proto tcpip.NetworkProtocolNumber GSO *stack.GSO Route stack.Route } +// Notification is the interface for receiving notification from the packet +// queue. +type Notification interface { + // WriteNotify will be called when a write happens to the queue. + WriteNotify() +} + +// NotificationHandle is an opaque handle to the registered notification target. +// It can be used to unregister the notification when no longer interested. +// +// +stateify savable +type NotificationHandle struct { + n Notification +} + +type queue struct { + // c is the outbound packet channel. + c chan PacketInfo + // mu protects fields below. + mu sync.RWMutex + notify []*NotificationHandle +} + +func (q *queue) Close() { + close(q.c) +} + +func (q *queue) Read() (PacketInfo, bool) { + select { + case p := <-q.c: + return p, true + default: + return PacketInfo{}, false + } +} + +func (q *queue) ReadContext(ctx context.Context) (PacketInfo, bool) { + select { + case pkt := <-q.c: + return pkt, true + case <-ctx.Done(): + return PacketInfo{}, false + } +} + +func (q *queue) Write(p PacketInfo) bool { + wrote := false + select { + case q.c <- p: + wrote = true + default: + } + q.mu.Lock() + notify := q.notify + q.mu.Unlock() + + if wrote { + // Send notification outside of lock. + for _, h := range notify { + h.n.WriteNotify() + } + } + return wrote +} + +func (q *queue) Num() int { + return len(q.c) +} + +func (q *queue) AddNotify(notify Notification) *NotificationHandle { + q.mu.Lock() + defer q.mu.Unlock() + h := &NotificationHandle{n: notify} + q.notify = append(q.notify, h) + return h +} + +func (q *queue) RemoveNotify(handle *NotificationHandle) { + q.mu.Lock() + defer q.mu.Unlock() + // Make a copy, since we reads the array outside of lock when notifying. + notify := make([]*NotificationHandle, 0, len(q.notify)) + for _, h := range q.notify { + if h != handle { + notify = append(notify, h) + } + } + q.notify = notify +} + // Endpoint is link layer endpoint that stores outbound packets in a channel // and allows injection of inbound packets. type Endpoint struct { @@ -41,14 +132,16 @@ type Endpoint struct { linkAddr tcpip.LinkAddress LinkEPCapabilities stack.LinkEndpointCapabilities - // c is where outbound packets are queued. - c chan PacketInfo + // Outbound packet queue. + q *queue } // New creates a new channel endpoint. func New(size int, mtu uint32, linkAddr tcpip.LinkAddress) *Endpoint { return &Endpoint{ - c: make(chan PacketInfo, size), + q: &queue{ + c: make(chan PacketInfo, size), + }, mtu: mtu, linkAddr: linkAddr, } @@ -57,50 +150,43 @@ func New(size int, mtu uint32, linkAddr tcpip.LinkAddress) *Endpoint { // Close closes e. Further packet injections will panic. Reads continue to // succeed until all packets are read. func (e *Endpoint) Close() { - close(e.c) + e.q.Close() } -// Read does non-blocking read for one packet from the outbound packet queue. +// Read does non-blocking read one packet from the outbound packet queue. func (e *Endpoint) Read() (PacketInfo, bool) { - select { - case pkt := <-e.c: - return pkt, true - default: - return PacketInfo{}, false - } + return e.q.Read() } // ReadContext does blocking read for one packet from the outbound packet queue. // It can be cancelled by ctx, and in this case, it returns false. func (e *Endpoint) ReadContext(ctx context.Context) (PacketInfo, bool) { - select { - case pkt := <-e.c: - return pkt, true - case <-ctx.Done(): - return PacketInfo{}, false - } + return e.q.ReadContext(ctx) } // Drain removes all outbound packets from the channel and counts them. func (e *Endpoint) Drain() int { c := 0 for { - select { - case <-e.c: - c++ - default: + if _, ok := e.Read(); !ok { return c } + c++ } } +// NumQueued returns the number of packet queued for outbound. +func (e *Endpoint) NumQueued() int { + return e.q.Num() +} + // InjectInbound injects an inbound packet. -func (e *Endpoint) InjectInbound(protocol tcpip.NetworkProtocolNumber, pkt tcpip.PacketBuffer) { +func (e *Endpoint) InjectInbound(protocol tcpip.NetworkProtocolNumber, pkt stack.PacketBuffer) { e.InjectLinkAddr(protocol, "", pkt) } // InjectLinkAddr injects an inbound packet with a remote link address. -func (e *Endpoint) InjectLinkAddr(protocol tcpip.NetworkProtocolNumber, remote tcpip.LinkAddress, pkt tcpip.PacketBuffer) { +func (e *Endpoint) InjectLinkAddr(protocol tcpip.NetworkProtocolNumber, remote tcpip.LinkAddress, pkt stack.PacketBuffer) { e.dispatcher.DeliverNetworkPacket(e, remote, "" /* local */, protocol, pkt) } @@ -143,54 +229,42 @@ func (e *Endpoint) LinkAddress() tcpip.LinkAddress { } // WritePacket stores outbound packets into the channel. -func (e *Endpoint) WritePacket(r *stack.Route, gso *stack.GSO, protocol tcpip.NetworkProtocolNumber, pkt tcpip.PacketBuffer) *tcpip.Error { +func (e *Endpoint) WritePacket(r *stack.Route, gso *stack.GSO, protocol tcpip.NetworkProtocolNumber, pkt stack.PacketBuffer) *tcpip.Error { // Clone r then release its resource so we only get the relevant fields from // stack.Route without holding a reference to a NIC's endpoint. route := r.Clone() route.Release() p := PacketInfo{ - Pkt: pkt, + Pkt: &pkt, Proto: protocol, GSO: gso, Route: route, } - select { - case e.c <- p: - default: - } + e.q.Write(p) return nil } // WritePackets stores outbound packets into the channel. -func (e *Endpoint) WritePackets(r *stack.Route, gso *stack.GSO, pkts []tcpip.PacketBuffer, protocol tcpip.NetworkProtocolNumber) (int, *tcpip.Error) { +func (e *Endpoint) WritePackets(r *stack.Route, gso *stack.GSO, pkts stack.PacketBufferList, protocol tcpip.NetworkProtocolNumber) (int, *tcpip.Error) { // Clone r then release its resource so we only get the relevant fields from // stack.Route without holding a reference to a NIC's endpoint. route := r.Clone() route.Release() - payloadView := pkts[0].Data.ToView() n := 0 -packetLoop: - for _, pkt := range pkts { - off := pkt.DataOffset - size := pkt.DataSize + for pkt := pkts.Front(); pkt != nil; pkt = pkt.Next() { p := PacketInfo{ - Pkt: tcpip.PacketBuffer{ - Header: pkt.Header, - Data: buffer.NewViewFromBytes(payloadView[off : off+size]).ToVectorisedView(), - }, + Pkt: pkt, Proto: protocol, GSO: gso, Route: route, } - select { - case e.c <- p: - n++ - default: - break packetLoop + if !e.q.Write(p) { + break } + n++ } return n, nil @@ -199,18 +273,26 @@ packetLoop: // WriteRawPacket implements stack.LinkEndpoint.WriteRawPacket. func (e *Endpoint) WriteRawPacket(vv buffer.VectorisedView) *tcpip.Error { p := PacketInfo{ - Pkt: tcpip.PacketBuffer{Data: vv}, + Pkt: &stack.PacketBuffer{Data: vv}, Proto: 0, GSO: nil, } - select { - case e.c <- p: - default: - } + e.q.Write(p) return nil } // Wait implements stack.LinkEndpoint.Wait. func (*Endpoint) Wait() {} + +// AddNotify adds a notification target for receiving event about outgoing +// packets. +func (e *Endpoint) AddNotify(notify Notification) *NotificationHandle { + return e.q.AddNotify(notify) +} + +// RemoveNotify removes handle from the list of notification targets. +func (e *Endpoint) RemoveNotify(handle *NotificationHandle) { + e.q.RemoveNotify(handle) +} diff --git a/pkg/tcpip/link/fdbased/endpoint.go b/pkg/tcpip/link/fdbased/endpoint.go index b7f60178e..b857ce9d0 100644 --- a/pkg/tcpip/link/fdbased/endpoint.go +++ b/pkg/tcpip/link/fdbased/endpoint.go @@ -91,7 +91,7 @@ func (p PacketDispatchMode) String() string { case PacketMMap: return "PacketMMap" default: - return fmt.Sprintf("unknown packet dispatch mode %v", p) + return fmt.Sprintf("unknown packet dispatch mode '%d'", p) } } @@ -386,7 +386,7 @@ const ( // WritePacket writes outbound packets to the file descriptor. If it is not // currently writable, the packet is dropped. -func (e *endpoint) WritePacket(r *stack.Route, gso *stack.GSO, protocol tcpip.NetworkProtocolNumber, pkt tcpip.PacketBuffer) *tcpip.Error { +func (e *endpoint) WritePacket(r *stack.Route, gso *stack.GSO, protocol tcpip.NetworkProtocolNumber, pkt stack.PacketBuffer) *tcpip.Error { if e.hdrSize > 0 { // Add ethernet header if needed. eth := header.Ethernet(pkt.Header.Prepend(header.EthernetMinimumSize)) @@ -405,9 +405,9 @@ func (e *endpoint) WritePacket(r *stack.Route, gso *stack.GSO, protocol tcpip.Ne eth.Encode(ethHdr) } + fd := e.fds[pkt.Hash%uint32(len(e.fds))] if e.Capabilities()&stack.CapabilityHardwareGSO != 0 { vnetHdr := virtioNetHdr{} - vnetHdrBuf := vnetHdrToByteSlice(&vnetHdr) if gso != nil { vnetHdr.hdrLen = uint16(pkt.Header.UsedLength()) if gso.NeedsCsum { @@ -428,130 +428,120 @@ func (e *endpoint) WritePacket(r *stack.Route, gso *stack.GSO, protocol tcpip.Ne } } - return rawfile.NonBlockingWrite3(e.fds[0], vnetHdrBuf, pkt.Header.View(), pkt.Data.ToView()) + vnetHdrBuf := vnetHdrToByteSlice(&vnetHdr) + return rawfile.NonBlockingWrite3(fd, vnetHdrBuf, pkt.Header.View(), pkt.Data.ToView()) } if pkt.Data.Size() == 0 { - return rawfile.NonBlockingWrite(e.fds[0], pkt.Header.View()) + return rawfile.NonBlockingWrite(fd, pkt.Header.View()) } - return rawfile.NonBlockingWrite3(e.fds[0], pkt.Header.View(), pkt.Data.ToView(), nil) + return rawfile.NonBlockingWrite3(fd, pkt.Header.View(), pkt.Data.ToView(), nil) } // WritePackets writes outbound packets to the file descriptor. If it is not // currently writable, the packet is dropped. -func (e *endpoint) WritePackets(r *stack.Route, gso *stack.GSO, pkts []tcpip.PacketBuffer, protocol tcpip.NetworkProtocolNumber) (int, *tcpip.Error) { - var ethHdrBuf []byte - // hdr + data - iovLen := 2 - if e.hdrSize > 0 { - // Add ethernet header if needed. - ethHdrBuf = make([]byte, header.EthernetMinimumSize) - eth := header.Ethernet(ethHdrBuf) - ethHdr := &header.EthernetFields{ - DstAddr: r.RemoteLinkAddress, - Type: protocol, - } - - // Preserve the src address if it's set in the route. - if r.LocalLinkAddress != "" { - ethHdr.SrcAddr = r.LocalLinkAddress - } else { - ethHdr.SrcAddr = e.addr - } - eth.Encode(ethHdr) - iovLen++ - } +// +// NOTE: This API uses sendmmsg to batch packets. As a result the underlying FD +// picked to write the packet out has to be the same for all packets in the +// list. In other words all packets in the batch should belong to the same +// flow. +func (e *endpoint) WritePackets(r *stack.Route, gso *stack.GSO, pkts stack.PacketBufferList, protocol tcpip.NetworkProtocolNumber) (int, *tcpip.Error) { + n := pkts.Len() - n := len(pkts) - - views := pkts[0].Data.Views() - /* - * Each bondary in views can add one more iovec. - * - * payload | | | | - * ----------------------------- - * packets | | | | | | | - * ----------------------------- - * iovecs | | | | | | | | | - */ - iovec := make([]syscall.Iovec, n*iovLen+len(views)-1) mmsgHdrs := make([]rawfile.MMsgHdr, n) + i := 0 + for pkt := pkts.Front(); pkt != nil; pkt = pkt.Next() { + var ethHdrBuf []byte + iovLen := 0 + if e.hdrSize > 0 { + // Add ethernet header if needed. + ethHdrBuf = make([]byte, header.EthernetMinimumSize) + eth := header.Ethernet(ethHdrBuf) + ethHdr := &header.EthernetFields{ + DstAddr: r.RemoteLinkAddress, + Type: protocol, + } - iovecIdx := 0 - viewIdx := 0 - viewOff := 0 - off := 0 - nextOff := 0 - for i := range pkts { - // TODO(b/134618279): Different packets may have different data - // in the future. We should handle this. - if !viewsEqual(pkts[i].Data.Views(), views) { - panic("All packets in pkts should have the same Data.") + // Preserve the src address if it's set in the route. + if r.LocalLinkAddress != "" { + ethHdr.SrcAddr = r.LocalLinkAddress + } else { + ethHdr.SrcAddr = e.addr + } + eth.Encode(ethHdr) + iovLen++ } - prevIovecIdx := iovecIdx - mmsgHdr := &mmsgHdrs[i] - mmsgHdr.Msg.Iov = &iovec[iovecIdx] - packetSize := pkts[i].DataSize - hdr := &pkts[i].Header - - off = pkts[i].DataOffset - if off != nextOff { - // We stop in a different point last time. - size := packetSize - viewIdx = 0 - viewOff = 0 - for size > 0 { - if size >= len(views[viewIdx]) { - viewIdx++ - viewOff = 0 - size -= len(views[viewIdx]) - } else { - viewOff = size - size = 0 + var vnetHdrBuf []byte + vnetHdr := virtioNetHdr{} + if e.Capabilities()&stack.CapabilityHardwareGSO != 0 { + if gso != nil { + vnetHdr.hdrLen = uint16(pkt.Header.UsedLength()) + if gso.NeedsCsum { + vnetHdr.flags = _VIRTIO_NET_HDR_F_NEEDS_CSUM + vnetHdr.csumStart = header.EthernetMinimumSize + gso.L3HdrLen + vnetHdr.csumOffset = gso.CsumOffset + } + if gso.Type != stack.GSONone && uint16(pkt.Data.Size()) > gso.MSS { + switch gso.Type { + case stack.GSOTCPv4: + vnetHdr.gsoType = _VIRTIO_NET_HDR_GSO_TCPV4 + case stack.GSOTCPv6: + vnetHdr.gsoType = _VIRTIO_NET_HDR_GSO_TCPV6 + default: + panic(fmt.Sprintf("Unknown gso type: %v", gso.Type)) + } + vnetHdr.gsoSize = gso.MSS } } + vnetHdrBuf = vnetHdrToByteSlice(&vnetHdr) + iovLen++ } - nextOff = off + packetSize + iovecs := make([]syscall.Iovec, iovLen+1+len(pkt.Data.Views())) + mmsgHdr := &mmsgHdrs[i] + mmsgHdr.Msg.Iov = &iovecs[0] + iovecIdx := 0 + if vnetHdrBuf != nil { + v := &iovecs[iovecIdx] + v.Base = &vnetHdrBuf[0] + v.Len = uint64(len(vnetHdrBuf)) + iovecIdx++ + } if ethHdrBuf != nil { - v := &iovec[iovecIdx] + v := &iovecs[iovecIdx] v.Base = ðHdrBuf[0] v.Len = uint64(len(ethHdrBuf)) iovecIdx++ } - - v := &iovec[iovecIdx] + pktSize := uint64(0) + // Encode L3 Header + v := &iovecs[iovecIdx] + hdr := &pkt.Header hdrView := hdr.View() v.Base = &hdrView[0] v.Len = uint64(len(hdrView)) + pktSize += v.Len iovecIdx++ - for packetSize > 0 { - vec := &iovec[iovecIdx] + // Now encode the Transport Payload. + pktViews := pkt.Data.Views() + for i := range pktViews { + vec := &iovecs[iovecIdx] iovecIdx++ - - v := views[viewIdx] - vec.Base = &v[viewOff] - s := len(v) - viewOff - if s <= packetSize { - viewIdx++ - viewOff = 0 - } else { - s = packetSize - viewOff += s - } - vec.Len = uint64(s) - packetSize -= s + vec.Base = &pktViews[i][0] + vec.Len = uint64(len(pktViews[i])) + pktSize += vec.Len } - - mmsgHdr.Msg.Iovlen = uint64(iovecIdx - prevIovecIdx) + mmsgHdr.Msg.Iovlen = uint64(iovecIdx) + i++ } packets := 0 for packets < n { - sent, err := rawfile.NonBlockingSendMMsg(e.fds[0], mmsgHdrs) + fd := e.fds[pkts.Front().Hash%uint32(len(e.fds))] + sent, err := rawfile.NonBlockingSendMMsg(fd, mmsgHdrs) if err != nil { return packets, err } @@ -610,7 +600,7 @@ func (e *InjectableEndpoint) Attach(dispatcher stack.NetworkDispatcher) { } // InjectInbound injects an inbound packet. -func (e *InjectableEndpoint) InjectInbound(protocol tcpip.NetworkProtocolNumber, pkt tcpip.PacketBuffer) { +func (e *InjectableEndpoint) InjectInbound(protocol tcpip.NetworkProtocolNumber, pkt stack.PacketBuffer) { e.dispatcher.DeliverNetworkPacket(e, "" /* remote */, "" /* local */, protocol, pkt) } diff --git a/pkg/tcpip/link/fdbased/endpoint_test.go b/pkg/tcpip/link/fdbased/endpoint_test.go index 2066987eb..3bfb15a8e 100644 --- a/pkg/tcpip/link/fdbased/endpoint_test.go +++ b/pkg/tcpip/link/fdbased/endpoint_test.go @@ -45,40 +45,46 @@ const ( type packetInfo struct { raddr tcpip.LinkAddress proto tcpip.NetworkProtocolNumber - contents tcpip.PacketBuffer + contents stack.PacketBuffer } type context struct { - t *testing.T - fds [2]int - ep stack.LinkEndpoint - ch chan packetInfo - done chan struct{} + t *testing.T + readFDs []int + writeFDs []int + ep stack.LinkEndpoint + ch chan packetInfo + done chan struct{} } func newContext(t *testing.T, opt *Options) *context { - fds, err := syscall.Socketpair(syscall.AF_UNIX, syscall.SOCK_SEQPACKET, 0) + firstFDPair, err := syscall.Socketpair(syscall.AF_UNIX, syscall.SOCK_SEQPACKET, 0) + if err != nil { + t.Fatalf("Socketpair failed: %v", err) + } + secondFDPair, err := syscall.Socketpair(syscall.AF_UNIX, syscall.SOCK_SEQPACKET, 0) if err != nil { t.Fatalf("Socketpair failed: %v", err) } - done := make(chan struct{}, 1) + done := make(chan struct{}, 2) opt.ClosedFunc = func(*tcpip.Error) { done <- struct{}{} } - opt.FDs = []int{fds[1]} + opt.FDs = []int{firstFDPair[1], secondFDPair[1]} ep, err := New(opt) if err != nil { t.Fatalf("Failed to create FD endpoint: %v", err) } c := &context{ - t: t, - fds: fds, - ep: ep, - ch: make(chan packetInfo, 100), - done: done, + t: t, + readFDs: []int{firstFDPair[0], secondFDPair[0]}, + writeFDs: opt.FDs, + ep: ep, + ch: make(chan packetInfo, 100), + done: done, } ep.Attach(c) @@ -87,12 +93,17 @@ func newContext(t *testing.T, opt *Options) *context { } func (c *context) cleanup() { - syscall.Close(c.fds[0]) + for _, fd := range c.readFDs { + syscall.Close(fd) + } + <-c.done <-c.done - syscall.Close(c.fds[1]) + for _, fd := range c.writeFDs { + syscall.Close(fd) + } } -func (c *context) DeliverNetworkPacket(linkEP stack.LinkEndpoint, remote tcpip.LinkAddress, local tcpip.LinkAddress, protocol tcpip.NetworkProtocolNumber, pkt tcpip.PacketBuffer) { +func (c *context) DeliverNetworkPacket(linkEP stack.LinkEndpoint, remote tcpip.LinkAddress, local tcpip.LinkAddress, protocol tcpip.NetworkProtocolNumber, pkt stack.PacketBuffer) { c.ch <- packetInfo{remote, protocol, pkt} } @@ -136,7 +147,7 @@ func TestAddress(t *testing.T) { } } -func testWritePacket(t *testing.T, plen int, eth bool, gsoMaxSize uint32) { +func testWritePacket(t *testing.T, plen int, eth bool, gsoMaxSize uint32, hash uint32) { c := newContext(t, &Options{Address: laddr, MTU: mtu, EthernetHeader: eth, GSOMaxSize: gsoMaxSize}) defer c.cleanup() @@ -168,16 +179,18 @@ func testWritePacket(t *testing.T, plen int, eth bool, gsoMaxSize uint32) { L3HdrLen: header.IPv4MaximumHeaderSize, } } - if err := c.ep.WritePacket(r, gso, proto, tcpip.PacketBuffer{ + if err := c.ep.WritePacket(r, gso, proto, stack.PacketBuffer{ Header: hdr, Data: payload.ToVectorisedView(), + Hash: hash, }); err != nil { t.Fatalf("WritePacket failed: %v", err) } - // Read from fd, then compare with what we wrote. + // Read from the corresponding FD, then compare with what we wrote. b = make([]byte, mtu) - n, err := syscall.Read(c.fds[0], b) + fd := c.readFDs[hash%uint32(len(c.readFDs))] + n, err := syscall.Read(fd, b) if err != nil { t.Fatalf("Read failed: %v", err) } @@ -238,7 +251,7 @@ func TestWritePacket(t *testing.T) { t.Run( fmt.Sprintf("Eth=%v,PayloadLen=%v,GSOMaxSize=%v", eth, plen, gso), func(t *testing.T) { - testWritePacket(t, plen, eth, gso) + testWritePacket(t, plen, eth, gso, 0) }, ) } @@ -246,6 +259,27 @@ func TestWritePacket(t *testing.T) { } } +func TestHashedWritePacket(t *testing.T) { + lengths := []int{0, 100, 1000} + eths := []bool{true, false} + gsos := []uint32{0, 32768} + hashes := []uint32{0, 1} + for _, eth := range eths { + for _, plen := range lengths { + for _, gso := range gsos { + for _, hash := range hashes { + t.Run( + fmt.Sprintf("Eth=%v,PayloadLen=%v,GSOMaxSize=%v,Hash=%d", eth, plen, gso, hash), + func(t *testing.T) { + testWritePacket(t, plen, eth, gso, hash) + }, + ) + } + } + } + } +} + func TestPreserveSrcAddress(t *testing.T) { baddr := tcpip.LinkAddress("\xcc\xbb\xaa\x77\x88\x99") @@ -261,7 +295,7 @@ func TestPreserveSrcAddress(t *testing.T) { // WritePacket panics given a prependable with anything less than // the minimum size of the ethernet header. hdr := buffer.NewPrependable(header.EthernetMinimumSize) - if err := c.ep.WritePacket(r, nil /* gso */, proto, tcpip.PacketBuffer{ + if err := c.ep.WritePacket(r, nil /* gso */, proto, stack.PacketBuffer{ Header: hdr, Data: buffer.VectorisedView{}, }); err != nil { @@ -270,7 +304,7 @@ func TestPreserveSrcAddress(t *testing.T) { // Read from the FD, then compare with what we wrote. b := make([]byte, mtu) - n, err := syscall.Read(c.fds[0], b) + n, err := syscall.Read(c.readFDs[0], b) if err != nil { t.Fatalf("Read failed: %v", err) } @@ -314,7 +348,7 @@ func TestDeliverPacket(t *testing.T) { } // Write packet via the file descriptor. - if _, err := syscall.Write(c.fds[0], all); err != nil { + if _, err := syscall.Write(c.readFDs[0], all); err != nil { t.Fatalf("Write failed: %v", err) } @@ -324,7 +358,7 @@ func TestDeliverPacket(t *testing.T) { want := packetInfo{ raddr: raddr, proto: proto, - contents: tcpip.PacketBuffer{ + contents: stack.PacketBuffer{ Data: buffer.View(b).ToVectorisedView(), LinkHeader: buffer.View(hdr), }, diff --git a/pkg/tcpip/link/fdbased/endpoint_unsafe.go b/pkg/tcpip/link/fdbased/endpoint_unsafe.go index 97a477b61..d81858353 100644 --- a/pkg/tcpip/link/fdbased/endpoint_unsafe.go +++ b/pkg/tcpip/link/fdbased/endpoint_unsafe.go @@ -24,9 +24,10 @@ import ( const virtioNetHdrSize = int(unsafe.Sizeof(virtioNetHdr{})) func vnetHdrToByteSlice(hdr *virtioNetHdr) (slice []byte) { - sh := (*reflect.SliceHeader)(unsafe.Pointer(&slice)) - sh.Data = uintptr(unsafe.Pointer(hdr)) - sh.Len = virtioNetHdrSize - sh.Cap = virtioNetHdrSize + *(*reflect.SliceHeader)(unsafe.Pointer(&slice)) = reflect.SliceHeader{ + Data: uintptr((unsafe.Pointer(hdr))), + Len: virtioNetHdrSize, + Cap: virtioNetHdrSize, + } return } diff --git a/pkg/tcpip/link/fdbased/mmap.go b/pkg/tcpip/link/fdbased/mmap.go index 62ed1e569..fe2bf3b0b 100644 --- a/pkg/tcpip/link/fdbased/mmap.go +++ b/pkg/tcpip/link/fdbased/mmap.go @@ -25,6 +25,7 @@ import ( "gvisor.dev/gvisor/pkg/tcpip/buffer" "gvisor.dev/gvisor/pkg/tcpip/header" "gvisor.dev/gvisor/pkg/tcpip/link/rawfile" + "gvisor.dev/gvisor/pkg/tcpip/stack" ) const ( @@ -190,7 +191,7 @@ func (d *packetMMapDispatcher) dispatch() (bool, *tcpip.Error) { } pkt = pkt[d.e.hdrSize:] - d.e.dispatcher.DeliverNetworkPacket(d.e, remote, local, p, tcpip.PacketBuffer{ + d.e.dispatcher.DeliverNetworkPacket(d.e, remote, local, p, stack.PacketBuffer{ Data: buffer.View(pkt).ToVectorisedView(), LinkHeader: buffer.View(eth), }) diff --git a/pkg/tcpip/link/fdbased/packet_dispatchers.go b/pkg/tcpip/link/fdbased/packet_dispatchers.go index c67d684ce..cb4cbea69 100644 --- a/pkg/tcpip/link/fdbased/packet_dispatchers.go +++ b/pkg/tcpip/link/fdbased/packet_dispatchers.go @@ -139,7 +139,7 @@ func (d *readVDispatcher) dispatch() (bool, *tcpip.Error) { } used := d.capViews(n, BufConfig) - pkt := tcpip.PacketBuffer{ + pkt := stack.PacketBuffer{ Data: buffer.NewVectorisedView(n, append([]buffer.View(nil), d.views[:used]...)), LinkHeader: buffer.View(eth), } @@ -296,7 +296,7 @@ func (d *recvMMsgDispatcher) dispatch() (bool, *tcpip.Error) { } used := d.capViews(k, int(n), BufConfig) - pkt := tcpip.PacketBuffer{ + pkt := stack.PacketBuffer{ Data: buffer.NewVectorisedView(int(n), append([]buffer.View(nil), d.views[k][:used]...)), LinkHeader: buffer.View(eth), } diff --git a/pkg/tcpip/link/loopback/loopback.go b/pkg/tcpip/link/loopback/loopback.go index 499cc608f..1e2255bfa 100644 --- a/pkg/tcpip/link/loopback/loopback.go +++ b/pkg/tcpip/link/loopback/loopback.go @@ -76,7 +76,7 @@ func (*endpoint) Wait() {} // WritePacket implements stack.LinkEndpoint.WritePacket. It delivers outbound // packets to the network-layer dispatcher. -func (e *endpoint) WritePacket(_ *stack.Route, _ *stack.GSO, protocol tcpip.NetworkProtocolNumber, pkt tcpip.PacketBuffer) *tcpip.Error { +func (e *endpoint) WritePacket(_ *stack.Route, _ *stack.GSO, protocol tcpip.NetworkProtocolNumber, pkt stack.PacketBuffer) *tcpip.Error { views := make([]buffer.View, 1, 1+len(pkt.Data.Views())) views[0] = pkt.Header.View() views = append(views, pkt.Data.Views()...) @@ -84,7 +84,7 @@ func (e *endpoint) WritePacket(_ *stack.Route, _ *stack.GSO, protocol tcpip.Netw // Because we're immediately turning around and writing the packet back // to the rx path, we intentionally don't preserve the remote and local // link addresses from the stack.Route we're passed. - e.dispatcher.DeliverNetworkPacket(e, "" /* remote */, "" /* local */, protocol, tcpip.PacketBuffer{ + e.dispatcher.DeliverNetworkPacket(e, "" /* remote */, "" /* local */, protocol, stack.PacketBuffer{ Data: buffer.NewVectorisedView(len(views[0])+pkt.Data.Size(), views), }) @@ -92,7 +92,7 @@ func (e *endpoint) WritePacket(_ *stack.Route, _ *stack.GSO, protocol tcpip.Netw } // WritePackets implements stack.LinkEndpoint.WritePackets. -func (e *endpoint) WritePackets(*stack.Route, *stack.GSO, []tcpip.PacketBuffer, tcpip.NetworkProtocolNumber) (int, *tcpip.Error) { +func (e *endpoint) WritePackets(*stack.Route, *stack.GSO, stack.PacketBufferList, tcpip.NetworkProtocolNumber) (int, *tcpip.Error) { panic("not implemented") } @@ -106,7 +106,7 @@ func (e *endpoint) WriteRawPacket(vv buffer.VectorisedView) *tcpip.Error { // There should be an ethernet header at the beginning of vv. linkHeader := header.Ethernet(vv.First()[:header.EthernetMinimumSize]) vv.TrimFront(len(linkHeader)) - e.dispatcher.DeliverNetworkPacket(e, "" /* remote */, "" /* local */, linkHeader.Type(), tcpip.PacketBuffer{ + e.dispatcher.DeliverNetworkPacket(e, "" /* remote */, "" /* local */, linkHeader.Type(), stack.PacketBuffer{ Data: vv, LinkHeader: buffer.View(linkHeader), }) diff --git a/pkg/tcpip/link/muxed/injectable.go b/pkg/tcpip/link/muxed/injectable.go index 445b22c17..a5478ce17 100644 --- a/pkg/tcpip/link/muxed/injectable.go +++ b/pkg/tcpip/link/muxed/injectable.go @@ -80,14 +80,14 @@ func (m *InjectableEndpoint) IsAttached() bool { } // InjectInbound implements stack.InjectableLinkEndpoint. -func (m *InjectableEndpoint) InjectInbound(protocol tcpip.NetworkProtocolNumber, pkt tcpip.PacketBuffer) { +func (m *InjectableEndpoint) InjectInbound(protocol tcpip.NetworkProtocolNumber, pkt stack.PacketBuffer) { m.dispatcher.DeliverNetworkPacket(m, "" /* remote */, "" /* local */, protocol, pkt) } // WritePackets writes outbound packets to the appropriate // LinkInjectableEndpoint based on the RemoteAddress. HandleLocal only works if // r.RemoteAddress has a route registered in this endpoint. -func (m *InjectableEndpoint) WritePackets(r *stack.Route, gso *stack.GSO, pkts []tcpip.PacketBuffer, protocol tcpip.NetworkProtocolNumber) (int, *tcpip.Error) { +func (m *InjectableEndpoint) WritePackets(r *stack.Route, gso *stack.GSO, pkts stack.PacketBufferList, protocol tcpip.NetworkProtocolNumber) (int, *tcpip.Error) { endpoint, ok := m.routes[r.RemoteAddress] if !ok { return 0, tcpip.ErrNoRoute @@ -98,7 +98,7 @@ func (m *InjectableEndpoint) WritePackets(r *stack.Route, gso *stack.GSO, pkts [ // WritePacket writes outbound packets to the appropriate LinkInjectableEndpoint // based on the RemoteAddress. HandleLocal only works if r.RemoteAddress has a // route registered in this endpoint. -func (m *InjectableEndpoint) WritePacket(r *stack.Route, gso *stack.GSO, protocol tcpip.NetworkProtocolNumber, pkt tcpip.PacketBuffer) *tcpip.Error { +func (m *InjectableEndpoint) WritePacket(r *stack.Route, gso *stack.GSO, protocol tcpip.NetworkProtocolNumber, pkt stack.PacketBuffer) *tcpip.Error { if endpoint, ok := m.routes[r.RemoteAddress]; ok { return endpoint.WritePacket(r, gso, protocol, pkt) } diff --git a/pkg/tcpip/link/muxed/injectable_test.go b/pkg/tcpip/link/muxed/injectable_test.go index 63b249837..87c734c1f 100644 --- a/pkg/tcpip/link/muxed/injectable_test.go +++ b/pkg/tcpip/link/muxed/injectable_test.go @@ -50,7 +50,7 @@ func TestInjectableEndpointDispatch(t *testing.T) { hdr.Prepend(1)[0] = 0xFA packetRoute := stack.Route{RemoteAddress: dstIP} - endpoint.WritePacket(&packetRoute, nil /* gso */, ipv4.ProtocolNumber, tcpip.PacketBuffer{ + endpoint.WritePacket(&packetRoute, nil /* gso */, ipv4.ProtocolNumber, stack.PacketBuffer{ Header: hdr, Data: buffer.NewViewFromBytes([]byte{0xFB}).ToVectorisedView(), }) @@ -70,7 +70,7 @@ func TestInjectableEndpointDispatchHdrOnly(t *testing.T) { hdr := buffer.NewPrependable(1) hdr.Prepend(1)[0] = 0xFA packetRoute := stack.Route{RemoteAddress: dstIP} - endpoint.WritePacket(&packetRoute, nil /* gso */, ipv4.ProtocolNumber, tcpip.PacketBuffer{ + endpoint.WritePacket(&packetRoute, nil /* gso */, ipv4.ProtocolNumber, stack.PacketBuffer{ Header: hdr, Data: buffer.NewView(0).ToVectorisedView(), }) diff --git a/pkg/tcpip/link/sharedmem/sharedmem.go b/pkg/tcpip/link/sharedmem/sharedmem.go index 655e537c4..0796d717e 100644 --- a/pkg/tcpip/link/sharedmem/sharedmem.go +++ b/pkg/tcpip/link/sharedmem/sharedmem.go @@ -185,7 +185,7 @@ func (e *endpoint) LinkAddress() tcpip.LinkAddress { // WritePacket writes outbound packets to the file descriptor. If it is not // currently writable, the packet is dropped. -func (e *endpoint) WritePacket(r *stack.Route, _ *stack.GSO, protocol tcpip.NetworkProtocolNumber, pkt tcpip.PacketBuffer) *tcpip.Error { +func (e *endpoint) WritePacket(r *stack.Route, _ *stack.GSO, protocol tcpip.NetworkProtocolNumber, pkt stack.PacketBuffer) *tcpip.Error { // Add the ethernet header here. eth := header.Ethernet(pkt.Header.Prepend(header.EthernetMinimumSize)) pkt.LinkHeader = buffer.View(eth) @@ -214,7 +214,7 @@ func (e *endpoint) WritePacket(r *stack.Route, _ *stack.GSO, protocol tcpip.Netw } // WritePackets implements stack.LinkEndpoint.WritePackets. -func (e *endpoint) WritePackets(r *stack.Route, _ *stack.GSO, pkts []tcpip.PacketBuffer, protocol tcpip.NetworkProtocolNumber) (int, *tcpip.Error) { +func (e *endpoint) WritePackets(r *stack.Route, _ *stack.GSO, pkts stack.PacketBufferList, protocol tcpip.NetworkProtocolNumber) (int, *tcpip.Error) { panic("not implemented") } @@ -275,7 +275,7 @@ func (e *endpoint) dispatchLoop(d stack.NetworkDispatcher) { // Send packet up the stack. eth := header.Ethernet(b[:header.EthernetMinimumSize]) - d.DeliverNetworkPacket(e, eth.SourceAddress(), eth.DestinationAddress(), eth.Type(), tcpip.PacketBuffer{ + d.DeliverNetworkPacket(e, eth.SourceAddress(), eth.DestinationAddress(), eth.Type(), stack.PacketBuffer{ Data: buffer.View(b[header.EthernetMinimumSize:]).ToVectorisedView(), LinkHeader: buffer.View(eth), }) diff --git a/pkg/tcpip/link/sharedmem/sharedmem_test.go b/pkg/tcpip/link/sharedmem/sharedmem_test.go index 5c729a439..27ea3f531 100644 --- a/pkg/tcpip/link/sharedmem/sharedmem_test.go +++ b/pkg/tcpip/link/sharedmem/sharedmem_test.go @@ -131,7 +131,7 @@ func newTestContext(t *testing.T, mtu, bufferSize uint32, addr tcpip.LinkAddress return c } -func (c *testContext) DeliverNetworkPacket(_ stack.LinkEndpoint, remoteLinkAddr, localLinkAddr tcpip.LinkAddress, proto tcpip.NetworkProtocolNumber, pkt tcpip.PacketBuffer) { +func (c *testContext) DeliverNetworkPacket(_ stack.LinkEndpoint, remoteLinkAddr, localLinkAddr tcpip.LinkAddress, proto tcpip.NetworkProtocolNumber, pkt stack.PacketBuffer) { c.mu.Lock() c.packets = append(c.packets, packetInfo{ addr: remoteLinkAddr, @@ -273,7 +273,7 @@ func TestSimpleSend(t *testing.T) { randomFill(buf) proto := tcpip.NetworkProtocolNumber(rand.Intn(0x10000)) - if err := c.ep.WritePacket(&r, nil /* gso */, proto, tcpip.PacketBuffer{ + if err := c.ep.WritePacket(&r, nil /* gso */, proto, stack.PacketBuffer{ Header: hdr, Data: buf.ToVectorisedView(), }); err != nil { @@ -345,7 +345,7 @@ func TestPreserveSrcAddressInSend(t *testing.T) { hdr := buffer.NewPrependable(header.EthernetMinimumSize) proto := tcpip.NetworkProtocolNumber(rand.Intn(0x10000)) - if err := c.ep.WritePacket(&r, nil /* gso */, proto, tcpip.PacketBuffer{ + if err := c.ep.WritePacket(&r, nil /* gso */, proto, stack.PacketBuffer{ Header: hdr, }); err != nil { t.Fatalf("WritePacket failed: %v", err) @@ -401,7 +401,7 @@ func TestFillTxQueue(t *testing.T) { for i := queuePipeSize / 40; i > 0; i-- { hdr := buffer.NewPrependable(int(c.ep.MaxHeaderLength())) - if err := c.ep.WritePacket(&r, nil /* gso */, header.IPv4ProtocolNumber, tcpip.PacketBuffer{ + if err := c.ep.WritePacket(&r, nil /* gso */, header.IPv4ProtocolNumber, stack.PacketBuffer{ Header: hdr, Data: buf.ToVectorisedView(), }); err != nil { @@ -419,7 +419,7 @@ func TestFillTxQueue(t *testing.T) { // Next attempt to write must fail. hdr := buffer.NewPrependable(int(c.ep.MaxHeaderLength())) - if want, err := tcpip.ErrWouldBlock, c.ep.WritePacket(&r, nil /* gso */, header.IPv4ProtocolNumber, tcpip.PacketBuffer{ + if want, err := tcpip.ErrWouldBlock, c.ep.WritePacket(&r, nil /* gso */, header.IPv4ProtocolNumber, stack.PacketBuffer{ Header: hdr, Data: buf.ToVectorisedView(), }); err != want { @@ -447,7 +447,7 @@ func TestFillTxQueueAfterBadCompletion(t *testing.T) { // Send two packets so that the id slice has at least two slots. for i := 2; i > 0; i-- { hdr := buffer.NewPrependable(int(c.ep.MaxHeaderLength())) - if err := c.ep.WritePacket(&r, nil /* gso */, header.IPv4ProtocolNumber, tcpip.PacketBuffer{ + if err := c.ep.WritePacket(&r, nil /* gso */, header.IPv4ProtocolNumber, stack.PacketBuffer{ Header: hdr, Data: buf.ToVectorisedView(), }); err != nil { @@ -470,7 +470,7 @@ func TestFillTxQueueAfterBadCompletion(t *testing.T) { ids := make(map[uint64]struct{}) for i := queuePipeSize / 40; i > 0; i-- { hdr := buffer.NewPrependable(int(c.ep.MaxHeaderLength())) - if err := c.ep.WritePacket(&r, nil /* gso */, header.IPv4ProtocolNumber, tcpip.PacketBuffer{ + if err := c.ep.WritePacket(&r, nil /* gso */, header.IPv4ProtocolNumber, stack.PacketBuffer{ Header: hdr, Data: buf.ToVectorisedView(), }); err != nil { @@ -488,7 +488,7 @@ func TestFillTxQueueAfterBadCompletion(t *testing.T) { // Next attempt to write must fail. hdr := buffer.NewPrependable(int(c.ep.MaxHeaderLength())) - if want, err := tcpip.ErrWouldBlock, c.ep.WritePacket(&r, nil /* gso */, header.IPv4ProtocolNumber, tcpip.PacketBuffer{ + if want, err := tcpip.ErrWouldBlock, c.ep.WritePacket(&r, nil /* gso */, header.IPv4ProtocolNumber, stack.PacketBuffer{ Header: hdr, Data: buf.ToVectorisedView(), }); err != want { @@ -514,7 +514,7 @@ func TestFillTxMemory(t *testing.T) { ids := make(map[uint64]struct{}) for i := queueDataSize / bufferSize; i > 0; i-- { hdr := buffer.NewPrependable(int(c.ep.MaxHeaderLength())) - if err := c.ep.WritePacket(&r, nil /* gso */, header.IPv4ProtocolNumber, tcpip.PacketBuffer{ + if err := c.ep.WritePacket(&r, nil /* gso */, header.IPv4ProtocolNumber, stack.PacketBuffer{ Header: hdr, Data: buf.ToVectorisedView(), }); err != nil { @@ -533,7 +533,7 @@ func TestFillTxMemory(t *testing.T) { // Next attempt to write must fail. hdr := buffer.NewPrependable(int(c.ep.MaxHeaderLength())) - err := c.ep.WritePacket(&r, nil /* gso */, header.IPv4ProtocolNumber, tcpip.PacketBuffer{ + err := c.ep.WritePacket(&r, nil /* gso */, header.IPv4ProtocolNumber, stack.PacketBuffer{ Header: hdr, Data: buf.ToVectorisedView(), }) @@ -561,7 +561,7 @@ func TestFillTxMemoryWithMultiBuffer(t *testing.T) { // until there is only one buffer left. for i := queueDataSize/bufferSize - 1; i > 0; i-- { hdr := buffer.NewPrependable(int(c.ep.MaxHeaderLength())) - if err := c.ep.WritePacket(&r, nil /* gso */, header.IPv4ProtocolNumber, tcpip.PacketBuffer{ + if err := c.ep.WritePacket(&r, nil /* gso */, header.IPv4ProtocolNumber, stack.PacketBuffer{ Header: hdr, Data: buf.ToVectorisedView(), }); err != nil { @@ -577,7 +577,7 @@ func TestFillTxMemoryWithMultiBuffer(t *testing.T) { { hdr := buffer.NewPrependable(int(c.ep.MaxHeaderLength())) uu := buffer.NewView(bufferSize).ToVectorisedView() - if want, err := tcpip.ErrWouldBlock, c.ep.WritePacket(&r, nil /* gso */, header.IPv4ProtocolNumber, tcpip.PacketBuffer{ + if want, err := tcpip.ErrWouldBlock, c.ep.WritePacket(&r, nil /* gso */, header.IPv4ProtocolNumber, stack.PacketBuffer{ Header: hdr, Data: uu, }); err != want { @@ -588,7 +588,7 @@ func TestFillTxMemoryWithMultiBuffer(t *testing.T) { // Attempt to write the one-buffer packet again. It must succeed. { hdr := buffer.NewPrependable(int(c.ep.MaxHeaderLength())) - if err := c.ep.WritePacket(&r, nil /* gso */, header.IPv4ProtocolNumber, tcpip.PacketBuffer{ + if err := c.ep.WritePacket(&r, nil /* gso */, header.IPv4ProtocolNumber, stack.PacketBuffer{ Header: hdr, Data: buf.ToVectorisedView(), }); err != nil { diff --git a/pkg/tcpip/link/sniffer/sniffer.go b/pkg/tcpip/link/sniffer/sniffer.go index 3392b7edd..be2537a82 100644 --- a/pkg/tcpip/link/sniffer/sniffer.go +++ b/pkg/tcpip/link/sniffer/sniffer.go @@ -21,11 +21,9 @@ package sniffer import ( - "bytes" "encoding/binary" "fmt" "io" - "os" "sync/atomic" "time" @@ -42,12 +40,12 @@ import ( // LogPackets must be accessed atomically. var LogPackets uint32 = 1 -// LogPacketsToFile is a flag used to enable or disable logging packets to a -// pcap file. Valid values are 0 or 1. A file must have been specified when the +// LogPacketsToPCAP is a flag used to enable or disable logging packets to a +// pcap writer. Valid values are 0 or 1. A writer must have been specified when the // sniffer was created for this flag to have effect. // -// LogPacketsToFile must be accessed atomically. -var LogPacketsToFile uint32 = 1 +// LogPacketsToPCAP must be accessed atomically. +var LogPacketsToPCAP uint32 = 1 var transportProtocolMinSizes map[tcpip.TransportProtocolNumber]int = map[tcpip.TransportProtocolNumber]int{ header.ICMPv4ProtocolNumber: header.IPv4MinimumSize, @@ -59,7 +57,7 @@ var transportProtocolMinSizes map[tcpip.TransportProtocolNumber]int = map[tcpip. type endpoint struct { dispatcher stack.NetworkDispatcher lower stack.LinkEndpoint - file *os.File + writer io.Writer maxPCAPLen uint32 } @@ -99,23 +97,22 @@ func writePCAPHeader(w io.Writer, maxLen uint32) error { }) } -// NewWithFile creates a new sniffer link-layer endpoint. It wraps around -// another endpoint and logs packets and they traverse the endpoint. +// NewWithWriter creates a new sniffer link-layer endpoint. It wraps around +// another endpoint and logs packets as they traverse the endpoint. // -// Packets can be logged to file in the pcap format. A sniffer created -// with this function will not emit packets using the standard log -// package. +// Packets are logged to writer in the pcap format. A sniffer created with this +// function will not emit packets using the standard log package. // // snapLen is the maximum amount of a packet to be saved. Packets with a length -// less than or equal too snapLen will be saved in their entirety. Longer +// less than or equal to snapLen will be saved in their entirety. Longer // packets will be truncated to snapLen. -func NewWithFile(lower stack.LinkEndpoint, file *os.File, snapLen uint32) (stack.LinkEndpoint, error) { - if err := writePCAPHeader(file, snapLen); err != nil { +func NewWithWriter(lower stack.LinkEndpoint, writer io.Writer, snapLen uint32) (stack.LinkEndpoint, error) { + if err := writePCAPHeader(writer, snapLen); err != nil { return nil, err } return &endpoint{ lower: lower, - file: file, + writer: writer, maxPCAPLen: snapLen, }, nil } @@ -123,37 +120,8 @@ func NewWithFile(lower stack.LinkEndpoint, file *os.File, snapLen uint32) (stack // DeliverNetworkPacket implements the stack.NetworkDispatcher interface. It is // called by the link-layer endpoint being wrapped when a packet arrives, and // logs the packet before forwarding to the actual dispatcher. -func (e *endpoint) DeliverNetworkPacket(linkEP stack.LinkEndpoint, remote, local tcpip.LinkAddress, protocol tcpip.NetworkProtocolNumber, pkt tcpip.PacketBuffer) { - if atomic.LoadUint32(&LogPackets) == 1 && e.file == nil { - logPacket("recv", protocol, pkt.Data.First(), nil) - } - if e.file != nil && atomic.LoadUint32(&LogPacketsToFile) == 1 { - vs := pkt.Data.Views() - length := pkt.Data.Size() - if length > int(e.maxPCAPLen) { - length = int(e.maxPCAPLen) - } - - buf := bytes.NewBuffer(make([]byte, 0, pcapPacketHeaderLen+length)) - if err := binary.Write(buf, binary.BigEndian, newPCAPPacketHeader(uint32(length), uint32(pkt.Data.Size()))); err != nil { - panic(err) - } - for _, v := range vs { - if length == 0 { - break - } - if len(v) > length { - v = v[:length] - } - if _, err := buf.Write([]byte(v)); err != nil { - panic(err) - } - length -= len(v) - } - if _, err := e.file.Write(buf.Bytes()); err != nil { - panic(err) - } - } +func (e *endpoint) DeliverNetworkPacket(linkEP stack.LinkEndpoint, remote, local tcpip.LinkAddress, protocol tcpip.NetworkProtocolNumber, pkt stack.PacketBuffer) { + e.dumpPacket("recv", nil, protocol, &pkt) e.dispatcher.DeliverNetworkPacket(e, remote, local, protocol, pkt) } @@ -200,31 +168,43 @@ func (e *endpoint) GSOMaxSize() uint32 { return 0 } -func (e *endpoint) dumpPacket(gso *stack.GSO, protocol tcpip.NetworkProtocolNumber, pkt tcpip.PacketBuffer) { - if atomic.LoadUint32(&LogPackets) == 1 && e.file == nil { - logPacket("send", protocol, pkt.Header.View(), gso) +func (e *endpoint) dumpPacket(prefix string, gso *stack.GSO, protocol tcpip.NetworkProtocolNumber, pkt *stack.PacketBuffer) { + writer := e.writer + if writer == nil && atomic.LoadUint32(&LogPackets) == 1 { + first := pkt.Header.View() + if len(first) == 0 { + first = pkt.Data.First() + } + logPacket(prefix, protocol, first, gso) } - if e.file != nil && atomic.LoadUint32(&LogPacketsToFile) == 1 { - hdrBuf := pkt.Header.View() - length := len(hdrBuf) + pkt.Data.Size() - if length > int(e.maxPCAPLen) { - length = int(e.maxPCAPLen) + if writer != nil && atomic.LoadUint32(&LogPacketsToPCAP) == 1 { + totalLength := pkt.Header.UsedLength() + pkt.Data.Size() + length := totalLength + if max := int(e.maxPCAPLen); length > max { + length = max } - - buf := bytes.NewBuffer(make([]byte, 0, pcapPacketHeaderLen+length)) - if err := binary.Write(buf, binary.BigEndian, newPCAPPacketHeader(uint32(length), uint32(len(hdrBuf)+pkt.Data.Size()))); err != nil { + if err := binary.Write(writer, binary.BigEndian, newPCAPPacketHeader(uint32(length), uint32(totalLength))); err != nil { panic(err) } - if len(hdrBuf) > length { - hdrBuf = hdrBuf[:length] - } - if _, err := buf.Write(hdrBuf); err != nil { - panic(err) + write := func(b []byte) { + if len(b) > length { + b = b[:length] + } + for len(b) != 0 { + n, err := writer.Write(b) + if err != nil { + panic(err) + } + b = b[n:] + length -= n + } } - length -= len(hdrBuf) - logVectorisedView(pkt.Data, length, buf) - if _, err := e.file.Write(buf.Bytes()); err != nil { - panic(err) + write(pkt.Header.View()) + for _, view := range pkt.Data.Views() { + if length == 0 { + break + } + write(view) } } } @@ -232,69 +212,31 @@ func (e *endpoint) dumpPacket(gso *stack.GSO, protocol tcpip.NetworkProtocolNumb // WritePacket implements the stack.LinkEndpoint interface. It is called by // higher-level protocols to write packets; it just logs the packet and // forwards the request to the lower endpoint. -func (e *endpoint) WritePacket(r *stack.Route, gso *stack.GSO, protocol tcpip.NetworkProtocolNumber, pkt tcpip.PacketBuffer) *tcpip.Error { - e.dumpPacket(gso, protocol, pkt) +func (e *endpoint) WritePacket(r *stack.Route, gso *stack.GSO, protocol tcpip.NetworkProtocolNumber, pkt stack.PacketBuffer) *tcpip.Error { + e.dumpPacket("send", gso, protocol, &pkt) return e.lower.WritePacket(r, gso, protocol, pkt) } // WritePackets implements the stack.LinkEndpoint interface. It is called by // higher-level protocols to write packets; it just logs the packet and // forwards the request to the lower endpoint. -func (e *endpoint) WritePackets(r *stack.Route, gso *stack.GSO, pkts []tcpip.PacketBuffer, protocol tcpip.NetworkProtocolNumber) (int, *tcpip.Error) { - view := pkts[0].Data.ToView() - for _, pkt := range pkts { - e.dumpPacket(gso, protocol, tcpip.PacketBuffer{ - Header: pkt.Header, - Data: view[pkt.DataOffset:][:pkt.DataSize].ToVectorisedView(), - }) +func (e *endpoint) WritePackets(r *stack.Route, gso *stack.GSO, pkts stack.PacketBufferList, protocol tcpip.NetworkProtocolNumber) (int, *tcpip.Error) { + for pkt := pkts.Front(); pkt != nil; pkt = pkt.Next() { + e.dumpPacket("send", gso, protocol, pkt) } return e.lower.WritePackets(r, gso, pkts, protocol) } // WriteRawPacket implements stack.LinkEndpoint.WriteRawPacket. func (e *endpoint) WriteRawPacket(vv buffer.VectorisedView) *tcpip.Error { - if atomic.LoadUint32(&LogPackets) == 1 && e.file == nil { - logPacket("send", 0, buffer.View("[raw packet, no header available]"), nil /* gso */) - } - if e.file != nil && atomic.LoadUint32(&LogPacketsToFile) == 1 { - length := vv.Size() - if length > int(e.maxPCAPLen) { - length = int(e.maxPCAPLen) - } - - buf := bytes.NewBuffer(make([]byte, 0, pcapPacketHeaderLen+length)) - if err := binary.Write(buf, binary.BigEndian, newPCAPPacketHeader(uint32(length), uint32(vv.Size()))); err != nil { - panic(err) - } - logVectorisedView(vv, length, buf) - if _, err := e.file.Write(buf.Bytes()); err != nil { - panic(err) - } - } + e.dumpPacket("send", nil, 0, &stack.PacketBuffer{ + Data: vv, + }) return e.lower.WriteRawPacket(vv) } -func logVectorisedView(vv buffer.VectorisedView, length int, buf *bytes.Buffer) { - if length <= 0 { - return - } - for _, v := range vv.Views() { - if len(v) > length { - v = v[:length] - } - n, err := buf.Write(v) - if err != nil { - panic(err) - } - length -= n - if length == 0 { - return - } - } -} - // Wait implements stack.LinkEndpoint.Wait. -func (*endpoint) Wait() {} +func (e *endpoint) Wait() { e.lower.Wait() } func logPacket(prefix string, protocol tcpip.NetworkProtocolNumber, b buffer.View, gso *stack.GSO) { // Figure out the network layer info. diff --git a/pkg/tcpip/link/tun/BUILD b/pkg/tcpip/link/tun/BUILD index e5096ea38..e0db6cf54 100644 --- a/pkg/tcpip/link/tun/BUILD +++ b/pkg/tcpip/link/tun/BUILD @@ -4,6 +4,22 @@ package(licenses = ["notice"]) go_library( name = "tun", - srcs = ["tun_unsafe.go"], + srcs = [ + "device.go", + "protocol.go", + "tun_unsafe.go", + ], visibility = ["//visibility:public"], + deps = [ + "//pkg/abi/linux", + "//pkg/refs", + "//pkg/sync", + "//pkg/syserror", + "//pkg/tcpip", + "//pkg/tcpip/buffer", + "//pkg/tcpip/header", + "//pkg/tcpip/link/channel", + "//pkg/tcpip/stack", + "//pkg/waiter", + ], ) diff --git a/pkg/tcpip/link/tun/device.go b/pkg/tcpip/link/tun/device.go new file mode 100644 index 000000000..617446ea2 --- /dev/null +++ b/pkg/tcpip/link/tun/device.go @@ -0,0 +1,358 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package tun + +import ( + "fmt" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/refs" + "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/tcpip" + "gvisor.dev/gvisor/pkg/tcpip/buffer" + "gvisor.dev/gvisor/pkg/tcpip/header" + "gvisor.dev/gvisor/pkg/tcpip/link/channel" + "gvisor.dev/gvisor/pkg/tcpip/stack" + "gvisor.dev/gvisor/pkg/waiter" +) + +const ( + // drivers/net/tun.c:tun_net_init() + defaultDevMtu = 1500 + + // Queue length for outbound packet, arriving at fd side for read. Overflow + // causes packet drops. gVisor implementation-specific. + defaultDevOutQueueLen = 1024 +) + +var zeroMAC [6]byte + +// Device is an opened /dev/net/tun device. +// +// +stateify savable +type Device struct { + waiter.Queue + + mu sync.RWMutex `state:"nosave"` + endpoint *tunEndpoint + notifyHandle *channel.NotificationHandle + flags uint16 +} + +// beforeSave is invoked by stateify. +func (d *Device) beforeSave() { + d.mu.Lock() + defer d.mu.Unlock() + // TODO(b/110961832): Restore the device to stack. At this moment, the stack + // is not savable. + if d.endpoint != nil { + panic("/dev/net/tun does not support save/restore when a device is associated with it.") + } +} + +// Release implements fs.FileOperations.Release. +func (d *Device) Release() { + d.mu.Lock() + defer d.mu.Unlock() + + // Decrease refcount if there is an endpoint associated with this file. + if d.endpoint != nil { + d.endpoint.RemoveNotify(d.notifyHandle) + d.endpoint.DecRef() + d.endpoint = nil + } +} + +// SetIff services TUNSETIFF ioctl(2) request. +func (d *Device) SetIff(s *stack.Stack, name string, flags uint16) error { + d.mu.Lock() + defer d.mu.Unlock() + + if d.endpoint != nil { + return syserror.EINVAL + } + + // Input validations. + isTun := flags&linux.IFF_TUN != 0 + isTap := flags&linux.IFF_TAP != 0 + supportedFlags := uint16(linux.IFF_TUN | linux.IFF_TAP | linux.IFF_NO_PI) + if isTap && isTun || !isTap && !isTun || flags&^supportedFlags != 0 { + return syserror.EINVAL + } + + prefix := "tun" + if isTap { + prefix = "tap" + } + + linkCaps := stack.CapabilityNone + if isTap { + linkCaps |= stack.CapabilityResolutionRequired + } + + endpoint, err := attachOrCreateNIC(s, name, prefix, linkCaps) + if err != nil { + return syserror.EINVAL + } + + d.endpoint = endpoint + d.notifyHandle = d.endpoint.AddNotify(d) + d.flags = flags + return nil +} + +func attachOrCreateNIC(s *stack.Stack, name, prefix string, linkCaps stack.LinkEndpointCapabilities) (*tunEndpoint, error) { + for { + // 1. Try to attach to an existing NIC. + if name != "" { + if nic, found := s.GetNICByName(name); found { + endpoint, ok := nic.LinkEndpoint().(*tunEndpoint) + if !ok { + // Not a NIC created by tun device. + return nil, syserror.EOPNOTSUPP + } + if !endpoint.TryIncRef() { + // Race detected: NIC got deleted in between. + continue + } + return endpoint, nil + } + } + + // 2. Creating a new NIC. + id := tcpip.NICID(s.UniqueID()) + endpoint := &tunEndpoint{ + Endpoint: channel.New(defaultDevOutQueueLen, defaultDevMtu, ""), + stack: s, + nicID: id, + name: name, + } + endpoint.Endpoint.LinkEPCapabilities = linkCaps + if endpoint.name == "" { + endpoint.name = fmt.Sprintf("%s%d", prefix, id) + } + err := s.CreateNICWithOptions(endpoint.nicID, endpoint, stack.NICOptions{ + Name: endpoint.name, + }) + switch err { + case nil: + return endpoint, nil + case tcpip.ErrDuplicateNICID: + // Race detected: A NIC has been created in between. + continue + default: + return nil, syserror.EINVAL + } + } +} + +// Write inject one inbound packet to the network interface. +func (d *Device) Write(data []byte) (int64, error) { + d.mu.RLock() + endpoint := d.endpoint + d.mu.RUnlock() + if endpoint == nil { + return 0, syserror.EBADFD + } + if !endpoint.IsAttached() { + return 0, syserror.EIO + } + + dataLen := int64(len(data)) + + // Packet information. + var pktInfoHdr PacketInfoHeader + if !d.hasFlags(linux.IFF_NO_PI) { + if len(data) < PacketInfoHeaderSize { + // Ignore bad packet. + return dataLen, nil + } + pktInfoHdr = PacketInfoHeader(data[:PacketInfoHeaderSize]) + data = data[PacketInfoHeaderSize:] + } + + // Ethernet header (TAP only). + var ethHdr header.Ethernet + if d.hasFlags(linux.IFF_TAP) { + if len(data) < header.EthernetMinimumSize { + // Ignore bad packet. + return dataLen, nil + } + ethHdr = header.Ethernet(data[:header.EthernetMinimumSize]) + data = data[header.EthernetMinimumSize:] + } + + // Try to determine network protocol number, default zero. + var protocol tcpip.NetworkProtocolNumber + switch { + case pktInfoHdr != nil: + protocol = pktInfoHdr.Protocol() + case ethHdr != nil: + protocol = ethHdr.Type() + } + + // Try to determine remote link address, default zero. + var remote tcpip.LinkAddress + switch { + case ethHdr != nil: + remote = ethHdr.SourceAddress() + default: + remote = tcpip.LinkAddress(zeroMAC[:]) + } + + pkt := stack.PacketBuffer{ + Data: buffer.View(data).ToVectorisedView(), + } + if ethHdr != nil { + pkt.LinkHeader = buffer.View(ethHdr) + } + endpoint.InjectLinkAddr(protocol, remote, pkt) + return dataLen, nil +} + +// Read reads one outgoing packet from the network interface. +func (d *Device) Read() ([]byte, error) { + d.mu.RLock() + endpoint := d.endpoint + d.mu.RUnlock() + if endpoint == nil { + return nil, syserror.EBADFD + } + + for { + info, ok := endpoint.Read() + if !ok { + return nil, syserror.ErrWouldBlock + } + + v, ok := d.encodePkt(&info) + if !ok { + // Ignore unsupported packet. + continue + } + return v, nil + } +} + +// encodePkt encodes packet for fd side. +func (d *Device) encodePkt(info *channel.PacketInfo) (buffer.View, bool) { + var vv buffer.VectorisedView + + // Packet information. + if !d.hasFlags(linux.IFF_NO_PI) { + hdr := make(PacketInfoHeader, PacketInfoHeaderSize) + hdr.Encode(&PacketInfoFields{ + Protocol: info.Proto, + }) + vv.AppendView(buffer.View(hdr)) + } + + // If the packet does not already have link layer header, and the route + // does not exist, we can't compute it. This is possibly a raw packet, tun + // device doesn't support this at the moment. + if info.Pkt.LinkHeader == nil && info.Route.RemoteLinkAddress == "" { + return nil, false + } + + // Ethernet header (TAP only). + if d.hasFlags(linux.IFF_TAP) { + // Add ethernet header if not provided. + if info.Pkt.LinkHeader == nil { + hdr := &header.EthernetFields{ + SrcAddr: info.Route.LocalLinkAddress, + DstAddr: info.Route.RemoteLinkAddress, + Type: info.Proto, + } + if hdr.SrcAddr == "" { + hdr.SrcAddr = d.endpoint.LinkAddress() + } + + eth := make(header.Ethernet, header.EthernetMinimumSize) + eth.Encode(hdr) + vv.AppendView(buffer.View(eth)) + } else { + vv.AppendView(info.Pkt.LinkHeader) + } + } + + // Append upper headers. + vv.AppendView(buffer.View(info.Pkt.Header.View()[len(info.Pkt.LinkHeader):])) + // Append data payload. + vv.Append(info.Pkt.Data) + + return vv.ToView(), true +} + +// Name returns the name of the attached network interface. Empty string if +// unattached. +func (d *Device) Name() string { + d.mu.RLock() + defer d.mu.RUnlock() + if d.endpoint != nil { + return d.endpoint.name + } + return "" +} + +// Flags returns the flags set for d. Zero value if unset. +func (d *Device) Flags() uint16 { + d.mu.RLock() + defer d.mu.RUnlock() + return d.flags +} + +func (d *Device) hasFlags(flags uint16) bool { + return d.flags&flags == flags +} + +// Readiness implements watier.Waitable.Readiness. +func (d *Device) Readiness(mask waiter.EventMask) waiter.EventMask { + if mask&waiter.EventIn != 0 { + d.mu.RLock() + endpoint := d.endpoint + d.mu.RUnlock() + if endpoint != nil && endpoint.NumQueued() == 0 { + mask &= ^waiter.EventIn + } + } + return mask & (waiter.EventIn | waiter.EventOut) +} + +// WriteNotify implements channel.Notification.WriteNotify. +func (d *Device) WriteNotify() { + d.Notify(waiter.EventIn) +} + +// tunEndpoint is the link endpoint for the NIC created by the tun device. +// +// It is ref-counted as multiple opening files can attach to the same NIC. +// The last owner is responsible for deleting the NIC. +type tunEndpoint struct { + *channel.Endpoint + + refs.AtomicRefCount + + stack *stack.Stack + nicID tcpip.NICID + name string +} + +// DecRef decrements refcount of e, removes NIC if refcount goes to 0. +func (e *tunEndpoint) DecRef() { + e.DecRefWithDestructor(func() { + e.stack.RemoveNIC(e.nicID) + }) +} diff --git a/pkg/tcpip/link/tun/protocol.go b/pkg/tcpip/link/tun/protocol.go new file mode 100644 index 000000000..89d9d91a9 --- /dev/null +++ b/pkg/tcpip/link/tun/protocol.go @@ -0,0 +1,56 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package tun + +import ( + "encoding/binary" + + "gvisor.dev/gvisor/pkg/tcpip" +) + +const ( + // PacketInfoHeaderSize is the size of the packet information header. + PacketInfoHeaderSize = 4 + + offsetFlags = 0 + offsetProtocol = 2 +) + +// PacketInfoFields contains fields sent through the wire if IFF_NO_PI flag is +// not set. +type PacketInfoFields struct { + Flags uint16 + Protocol tcpip.NetworkProtocolNumber +} + +// PacketInfoHeader is the wire representation of the packet information sent if +// IFF_NO_PI flag is not set. +type PacketInfoHeader []byte + +// Encode encodes f into h. +func (h PacketInfoHeader) Encode(f *PacketInfoFields) { + binary.BigEndian.PutUint16(h[offsetFlags:][:2], f.Flags) + binary.BigEndian.PutUint16(h[offsetProtocol:][:2], uint16(f.Protocol)) +} + +// Flags returns the flag field in h. +func (h PacketInfoHeader) Flags() uint16 { + return binary.BigEndian.Uint16(h[offsetFlags:]) +} + +// Protocol returns the protocol field in h. +func (h PacketInfoHeader) Protocol() tcpip.NetworkProtocolNumber { + return tcpip.NetworkProtocolNumber(binary.BigEndian.Uint16(h[offsetProtocol:])) +} diff --git a/pkg/tcpip/link/waitable/waitable.go b/pkg/tcpip/link/waitable/waitable.go index a8de38979..2b3741276 100644 --- a/pkg/tcpip/link/waitable/waitable.go +++ b/pkg/tcpip/link/waitable/waitable.go @@ -50,7 +50,7 @@ func New(lower stack.LinkEndpoint) *Endpoint { // It is called by the link-layer endpoint being wrapped when a packet arrives, // and only forwards to the actual dispatcher if Wait or WaitDispatch haven't // been called. -func (e *Endpoint) DeliverNetworkPacket(linkEP stack.LinkEndpoint, remote, local tcpip.LinkAddress, protocol tcpip.NetworkProtocolNumber, pkt tcpip.PacketBuffer) { +func (e *Endpoint) DeliverNetworkPacket(linkEP stack.LinkEndpoint, remote, local tcpip.LinkAddress, protocol tcpip.NetworkProtocolNumber, pkt stack.PacketBuffer) { if !e.dispatchGate.Enter() { return } @@ -99,7 +99,7 @@ func (e *Endpoint) LinkAddress() tcpip.LinkAddress { // WritePacket implements stack.LinkEndpoint.WritePacket. It is called by // higher-level protocols to write packets. It only forwards packets to the // lower endpoint if Wait or WaitWrite haven't been called. -func (e *Endpoint) WritePacket(r *stack.Route, gso *stack.GSO, protocol tcpip.NetworkProtocolNumber, pkt tcpip.PacketBuffer) *tcpip.Error { +func (e *Endpoint) WritePacket(r *stack.Route, gso *stack.GSO, protocol tcpip.NetworkProtocolNumber, pkt stack.PacketBuffer) *tcpip.Error { if !e.writeGate.Enter() { return nil } @@ -112,9 +112,9 @@ func (e *Endpoint) WritePacket(r *stack.Route, gso *stack.GSO, protocol tcpip.Ne // WritePackets implements stack.LinkEndpoint.WritePackets. It is called by // higher-level protocols to write packets. It only forwards packets to the // lower endpoint if Wait or WaitWrite haven't been called. -func (e *Endpoint) WritePackets(r *stack.Route, gso *stack.GSO, pkts []tcpip.PacketBuffer, protocol tcpip.NetworkProtocolNumber) (int, *tcpip.Error) { +func (e *Endpoint) WritePackets(r *stack.Route, gso *stack.GSO, pkts stack.PacketBufferList, protocol tcpip.NetworkProtocolNumber) (int, *tcpip.Error) { if !e.writeGate.Enter() { - return len(pkts), nil + return pkts.Len(), nil } n, err := e.lower.WritePackets(r, gso, pkts, protocol) diff --git a/pkg/tcpip/link/waitable/waitable_test.go b/pkg/tcpip/link/waitable/waitable_test.go index 31b11a27a..54eb5322b 100644 --- a/pkg/tcpip/link/waitable/waitable_test.go +++ b/pkg/tcpip/link/waitable/waitable_test.go @@ -35,7 +35,7 @@ type countedEndpoint struct { dispatcher stack.NetworkDispatcher } -func (e *countedEndpoint) DeliverNetworkPacket(linkEP stack.LinkEndpoint, remote, local tcpip.LinkAddress, protocol tcpip.NetworkProtocolNumber, pkt tcpip.PacketBuffer) { +func (e *countedEndpoint) DeliverNetworkPacket(linkEP stack.LinkEndpoint, remote, local tcpip.LinkAddress, protocol tcpip.NetworkProtocolNumber, pkt stack.PacketBuffer) { e.dispatchCount++ } @@ -65,15 +65,15 @@ func (e *countedEndpoint) LinkAddress() tcpip.LinkAddress { return e.linkAddr } -func (e *countedEndpoint) WritePacket(r *stack.Route, _ *stack.GSO, protocol tcpip.NetworkProtocolNumber, pkt tcpip.PacketBuffer) *tcpip.Error { +func (e *countedEndpoint) WritePacket(r *stack.Route, _ *stack.GSO, protocol tcpip.NetworkProtocolNumber, pkt stack.PacketBuffer) *tcpip.Error { e.writeCount++ return nil } // WritePackets implements stack.LinkEndpoint.WritePackets. -func (e *countedEndpoint) WritePackets(r *stack.Route, _ *stack.GSO, pkts []tcpip.PacketBuffer, protocol tcpip.NetworkProtocolNumber) (int, *tcpip.Error) { - e.writeCount += len(pkts) - return len(pkts), nil +func (e *countedEndpoint) WritePackets(r *stack.Route, _ *stack.GSO, pkts stack.PacketBufferList, protocol tcpip.NetworkProtocolNumber) (int, *tcpip.Error) { + e.writeCount += pkts.Len() + return pkts.Len(), nil } func (e *countedEndpoint) WriteRawPacket(buffer.VectorisedView) *tcpip.Error { @@ -89,21 +89,21 @@ func TestWaitWrite(t *testing.T) { wep := New(ep) // Write and check that it goes through. - wep.WritePacket(nil, nil /* gso */, 0, tcpip.PacketBuffer{}) + wep.WritePacket(nil, nil /* gso */, 0, stack.PacketBuffer{}) if want := 1; ep.writeCount != want { t.Fatalf("Unexpected writeCount: got=%v, want=%v", ep.writeCount, want) } // Wait on dispatches, then try to write. It must go through. wep.WaitDispatch() - wep.WritePacket(nil, nil /* gso */, 0, tcpip.PacketBuffer{}) + wep.WritePacket(nil, nil /* gso */, 0, stack.PacketBuffer{}) if want := 2; ep.writeCount != want { t.Fatalf("Unexpected writeCount: got=%v, want=%v", ep.writeCount, want) } // Wait on writes, then try to write. It must not go through. wep.WaitWrite() - wep.WritePacket(nil, nil /* gso */, 0, tcpip.PacketBuffer{}) + wep.WritePacket(nil, nil /* gso */, 0, stack.PacketBuffer{}) if want := 2; ep.writeCount != want { t.Fatalf("Unexpected writeCount: got=%v, want=%v", ep.writeCount, want) } @@ -120,21 +120,21 @@ func TestWaitDispatch(t *testing.T) { } // Dispatch and check that it goes through. - ep.dispatcher.DeliverNetworkPacket(ep, "", "", 0, tcpip.PacketBuffer{}) + ep.dispatcher.DeliverNetworkPacket(ep, "", "", 0, stack.PacketBuffer{}) if want := 1; ep.dispatchCount != want { t.Fatalf("Unexpected dispatchCount: got=%v, want=%v", ep.dispatchCount, want) } // Wait on writes, then try to dispatch. It must go through. wep.WaitWrite() - ep.dispatcher.DeliverNetworkPacket(ep, "", "", 0, tcpip.PacketBuffer{}) + ep.dispatcher.DeliverNetworkPacket(ep, "", "", 0, stack.PacketBuffer{}) if want := 2; ep.dispatchCount != want { t.Fatalf("Unexpected dispatchCount: got=%v, want=%v", ep.dispatchCount, want) } // Wait on dispatches, then try to dispatch. It must not go through. wep.WaitDispatch() - ep.dispatcher.DeliverNetworkPacket(ep, "", "", 0, tcpip.PacketBuffer{}) + ep.dispatcher.DeliverNetworkPacket(ep, "", "", 0, stack.PacketBuffer{}) if want := 2; ep.dispatchCount != want { t.Fatalf("Unexpected dispatchCount: got=%v, want=%v", ep.dispatchCount, want) } diff --git a/pkg/tcpip/network/arp/arp.go b/pkg/tcpip/network/arp/arp.go index 4da13c5df..7acbfa0a8 100644 --- a/pkg/tcpip/network/arp/arp.go +++ b/pkg/tcpip/network/arp/arp.go @@ -79,20 +79,20 @@ func (e *endpoint) MaxHeaderLength() uint16 { func (e *endpoint) Close() {} -func (e *endpoint) WritePacket(*stack.Route, *stack.GSO, stack.NetworkHeaderParams, tcpip.PacketBuffer) *tcpip.Error { +func (e *endpoint) WritePacket(*stack.Route, *stack.GSO, stack.NetworkHeaderParams, stack.PacketBuffer) *tcpip.Error { return tcpip.ErrNotSupported } // WritePackets implements stack.NetworkEndpoint.WritePackets. -func (e *endpoint) WritePackets(*stack.Route, *stack.GSO, []tcpip.PacketBuffer, stack.NetworkHeaderParams) (int, *tcpip.Error) { +func (e *endpoint) WritePackets(*stack.Route, *stack.GSO, stack.PacketBufferList, stack.NetworkHeaderParams) (int, *tcpip.Error) { return 0, tcpip.ErrNotSupported } -func (e *endpoint) WriteHeaderIncludedPacket(r *stack.Route, pkt tcpip.PacketBuffer) *tcpip.Error { +func (e *endpoint) WriteHeaderIncludedPacket(r *stack.Route, pkt stack.PacketBuffer) *tcpip.Error { return tcpip.ErrNotSupported } -func (e *endpoint) HandlePacket(r *stack.Route, pkt tcpip.PacketBuffer) { +func (e *endpoint) HandlePacket(r *stack.Route, pkt stack.PacketBuffer) { v := pkt.Data.First() h := header.ARP(v) if !h.IsValid() { @@ -113,7 +113,7 @@ func (e *endpoint) HandlePacket(r *stack.Route, pkt tcpip.PacketBuffer) { copy(packet.ProtocolAddressSender(), h.ProtocolAddressTarget()) copy(packet.HardwareAddressTarget(), h.HardwareAddressSender()) copy(packet.ProtocolAddressTarget(), h.ProtocolAddressSender()) - e.linkEP.WritePacket(r, nil /* gso */, ProtocolNumber, tcpip.PacketBuffer{ + e.linkEP.WritePacket(r, nil /* gso */, ProtocolNumber, stack.PacketBuffer{ Header: hdr, }) fallthrough // also fill the cache from requests @@ -148,12 +148,12 @@ func (p *protocol) NewEndpoint(nicID tcpip.NICID, addrWithPrefix tcpip.AddressWi }, nil } -// LinkAddressProtocol implements stack.LinkAddressResolver. +// LinkAddressProtocol implements stack.LinkAddressResolver.LinkAddressProtocol. func (*protocol) LinkAddressProtocol() tcpip.NetworkProtocolNumber { return header.IPv4ProtocolNumber } -// LinkAddressRequest implements stack.LinkAddressResolver. +// LinkAddressRequest implements stack.LinkAddressResolver.LinkAddressRequest. func (*protocol) LinkAddressRequest(addr, localAddr tcpip.Address, linkEP stack.LinkEndpoint) *tcpip.Error { r := &stack.Route{ RemoteLinkAddress: broadcastMAC, @@ -167,12 +167,12 @@ func (*protocol) LinkAddressRequest(addr, localAddr tcpip.Address, linkEP stack. copy(h.ProtocolAddressSender(), localAddr) copy(h.ProtocolAddressTarget(), addr) - return linkEP.WritePacket(r, nil /* gso */, ProtocolNumber, tcpip.PacketBuffer{ + return linkEP.WritePacket(r, nil /* gso */, ProtocolNumber, stack.PacketBuffer{ Header: hdr, }) } -// ResolveStaticAddress implements stack.LinkAddressResolver. +// ResolveStaticAddress implements stack.LinkAddressResolver.ResolveStaticAddress. func (*protocol) ResolveStaticAddress(addr tcpip.Address) (tcpip.LinkAddress, bool) { if addr == header.IPv4Broadcast { return broadcastMAC, true @@ -183,16 +183,22 @@ func (*protocol) ResolveStaticAddress(addr tcpip.Address) (tcpip.LinkAddress, bo return tcpip.LinkAddress([]byte(nil)), false } -// SetOption implements NetworkProtocol. -func (p *protocol) SetOption(option interface{}) *tcpip.Error { +// SetOption implements stack.NetworkProtocol.SetOption. +func (*protocol) SetOption(option interface{}) *tcpip.Error { return tcpip.ErrUnknownProtocolOption } -// Option implements NetworkProtocol. -func (p *protocol) Option(option interface{}) *tcpip.Error { +// Option implements stack.NetworkProtocol.Option. +func (*protocol) Option(option interface{}) *tcpip.Error { return tcpip.ErrUnknownProtocolOption } +// Close implements stack.TransportProtocol.Close. +func (*protocol) Close() {} + +// Wait implements stack.TransportProtocol.Wait. +func (*protocol) Wait() {} + var broadcastMAC = tcpip.LinkAddress([]byte{0xff, 0xff, 0xff, 0xff, 0xff, 0xff}) // NewProtocol returns an ARP network protocol. diff --git a/pkg/tcpip/network/arp/arp_test.go b/pkg/tcpip/network/arp/arp_test.go index 03cf03b6d..1646d9cde 100644 --- a/pkg/tcpip/network/arp/arp_test.go +++ b/pkg/tcpip/network/arp/arp_test.go @@ -103,7 +103,7 @@ func TestDirectRequest(t *testing.T) { inject := func(addr tcpip.Address) { copy(h.ProtocolAddressTarget(), addr) - c.linkEP.InjectInbound(arp.ProtocolNumber, tcpip.PacketBuffer{ + c.linkEP.InjectInbound(arp.ProtocolNumber, stack.PacketBuffer{ Data: v.ToVectorisedView(), }) } @@ -138,7 +138,8 @@ func TestDirectRequest(t *testing.T) { // Sleep tests are gross, but this will only potentially flake // if there's a bug. If there is no bug this will reliably // succeed. - ctx, _ := context.WithTimeout(context.Background(), 100*time.Millisecond) + ctx, cancel := context.WithTimeout(context.Background(), 100*time.Millisecond) + defer cancel() if pkt, ok := c.linkEP.ReadContext(ctx); ok { t.Errorf("stackAddrBad: unexpected packet sent, Proto=%v", pkt.Proto) } diff --git a/pkg/tcpip/network/fragmentation/fragmentation.go b/pkg/tcpip/network/fragmentation/fragmentation.go index 92f2aa13a..f42abc4bb 100644 --- a/pkg/tcpip/network/fragmentation/fragmentation.go +++ b/pkg/tcpip/network/fragmentation/fragmentation.go @@ -115,10 +115,12 @@ func (f *Fragmentation) Process(id uint32, first, last uint16, more bool, vv buf // Evict reassemblers if we are consuming more memory than highLimit until // we reach lowLimit. if f.size > f.highLimit { - tail := f.rList.Back() - for f.size > f.lowLimit && tail != nil { + for f.size > f.lowLimit { + tail := f.rList.Back() + if tail == nil { + break + } f.release(tail) - tail = tail.Prev() } } f.mu.Unlock() diff --git a/pkg/tcpip/network/hash/hash.go b/pkg/tcpip/network/hash/hash.go index 6a215938b..8f65713c5 100644 --- a/pkg/tcpip/network/hash/hash.go +++ b/pkg/tcpip/network/hash/hash.go @@ -80,12 +80,12 @@ func IPv4FragmentHash(h header.IPv4) uint32 { // RFC 2640 (sec 4.5) is not very sharp on this aspect. // As a reference, also Linux ignores the protocol to compute // the hash (inet6_hash_frag). -func IPv6FragmentHash(h header.IPv6, f header.IPv6Fragment) uint32 { +func IPv6FragmentHash(h header.IPv6, id uint32) uint32 { t := h.SourceAddress() y := uint32(t[0]) | uint32(t[1])<<8 | uint32(t[2])<<16 | uint32(t[3])<<24 t = h.DestinationAddress() z := uint32(t[0]) | uint32(t[1])<<8 | uint32(t[2])<<16 | uint32(t[3])<<24 - return Hash3Words(f.ID(), y, z, hashIV) + return Hash3Words(id, y, z, hashIV) } func rol32(v, shift uint32) uint32 { diff --git a/pkg/tcpip/network/ip_test.go b/pkg/tcpip/network/ip_test.go index f4d78f8c6..4c20301c6 100644 --- a/pkg/tcpip/network/ip_test.go +++ b/pkg/tcpip/network/ip_test.go @@ -96,7 +96,7 @@ func (t *testObject) checkValues(protocol tcpip.TransportProtocolNumber, vv buff // DeliverTransportPacket is called by network endpoints after parsing incoming // packets. This is used by the test object to verify that the results of the // parsing are expected. -func (t *testObject) DeliverTransportPacket(r *stack.Route, protocol tcpip.TransportProtocolNumber, pkt tcpip.PacketBuffer) { +func (t *testObject) DeliverTransportPacket(r *stack.Route, protocol tcpip.TransportProtocolNumber, pkt stack.PacketBuffer) { t.checkValues(protocol, pkt.Data, r.RemoteAddress, r.LocalAddress) t.dataCalls++ } @@ -104,7 +104,7 @@ func (t *testObject) DeliverTransportPacket(r *stack.Route, protocol tcpip.Trans // DeliverTransportControlPacket is called by network endpoints after parsing // incoming control (ICMP) packets. This is used by the test object to verify // that the results of the parsing are expected. -func (t *testObject) DeliverTransportControlPacket(local, remote tcpip.Address, net tcpip.NetworkProtocolNumber, trans tcpip.TransportProtocolNumber, typ stack.ControlType, extra uint32, pkt tcpip.PacketBuffer) { +func (t *testObject) DeliverTransportControlPacket(local, remote tcpip.Address, net tcpip.NetworkProtocolNumber, trans tcpip.TransportProtocolNumber, typ stack.ControlType, extra uint32, pkt stack.PacketBuffer) { t.checkValues(trans, pkt.Data, remote, local) if typ != t.typ { t.t.Errorf("typ = %v, want %v", typ, t.typ) @@ -150,7 +150,7 @@ func (*testObject) Wait() {} // WritePacket is called by network endpoints after producing a packet and // writing it to the link endpoint. This is used by the test object to verify // that the produced packet is as expected. -func (t *testObject) WritePacket(_ *stack.Route, _ *stack.GSO, protocol tcpip.NetworkProtocolNumber, pkt tcpip.PacketBuffer) *tcpip.Error { +func (t *testObject) WritePacket(_ *stack.Route, _ *stack.GSO, protocol tcpip.NetworkProtocolNumber, pkt stack.PacketBuffer) *tcpip.Error { var prot tcpip.TransportProtocolNumber var srcAddr tcpip.Address var dstAddr tcpip.Address @@ -172,7 +172,7 @@ func (t *testObject) WritePacket(_ *stack.Route, _ *stack.GSO, protocol tcpip.Ne } // WritePackets implements stack.LinkEndpoint.WritePackets. -func (t *testObject) WritePackets(_ *stack.Route, _ *stack.GSO, pkt []tcpip.PacketBuffer, protocol tcpip.NetworkProtocolNumber) (int, *tcpip.Error) { +func (t *testObject) WritePackets(_ *stack.Route, _ *stack.GSO, pkt stack.PacketBufferList, protocol tcpip.NetworkProtocolNumber) (int, *tcpip.Error) { panic("not implemented") } @@ -246,7 +246,7 @@ func TestIPv4Send(t *testing.T) { if err != nil { t.Fatalf("could not find route: %v", err) } - if err := ep.WritePacket(&r, nil /* gso */, stack.NetworkHeaderParams{Protocol: 123, TTL: 123, TOS: stack.DefaultTOS}, tcpip.PacketBuffer{ + if err := ep.WritePacket(&r, nil /* gso */, stack.NetworkHeaderParams{Protocol: 123, TTL: 123, TOS: stack.DefaultTOS}, stack.PacketBuffer{ Header: hdr, Data: payload.ToVectorisedView(), }); err != nil { @@ -289,7 +289,7 @@ func TestIPv4Receive(t *testing.T) { if err != nil { t.Fatalf("could not find route: %v", err) } - ep.HandlePacket(&r, tcpip.PacketBuffer{ + ep.HandlePacket(&r, stack.PacketBuffer{ Data: view.ToVectorisedView(), }) if o.dataCalls != 1 { @@ -379,7 +379,7 @@ func TestIPv4ReceiveControl(t *testing.T) { o.extra = c.expectedExtra vv := view[:len(view)-c.trunc].ToVectorisedView() - ep.HandlePacket(&r, tcpip.PacketBuffer{ + ep.HandlePacket(&r, stack.PacketBuffer{ Data: vv, }) if want := c.expectedCount; o.controlCalls != want { @@ -444,7 +444,7 @@ func TestIPv4FragmentationReceive(t *testing.T) { } // Send first segment. - ep.HandlePacket(&r, tcpip.PacketBuffer{ + ep.HandlePacket(&r, stack.PacketBuffer{ Data: frag1.ToVectorisedView(), }) if o.dataCalls != 0 { @@ -452,7 +452,7 @@ func TestIPv4FragmentationReceive(t *testing.T) { } // Send second segment. - ep.HandlePacket(&r, tcpip.PacketBuffer{ + ep.HandlePacket(&r, stack.PacketBuffer{ Data: frag2.ToVectorisedView(), }) if o.dataCalls != 1 { @@ -487,7 +487,7 @@ func TestIPv6Send(t *testing.T) { if err != nil { t.Fatalf("could not find route: %v", err) } - if err := ep.WritePacket(&r, nil /* gso */, stack.NetworkHeaderParams{Protocol: 123, TTL: 123, TOS: stack.DefaultTOS}, tcpip.PacketBuffer{ + if err := ep.WritePacket(&r, nil /* gso */, stack.NetworkHeaderParams{Protocol: 123, TTL: 123, TOS: stack.DefaultTOS}, stack.PacketBuffer{ Header: hdr, Data: payload.ToVectorisedView(), }); err != nil { @@ -530,7 +530,7 @@ func TestIPv6Receive(t *testing.T) { t.Fatalf("could not find route: %v", err) } - ep.HandlePacket(&r, tcpip.PacketBuffer{ + ep.HandlePacket(&r, stack.PacketBuffer{ Data: view.ToVectorisedView(), }) if o.dataCalls != 1 { @@ -644,7 +644,7 @@ func TestIPv6ReceiveControl(t *testing.T) { // Set ICMPv6 checksum. icmp.SetChecksum(header.ICMPv6Checksum(icmp, outerSrcAddr, localIpv6Addr, buffer.VectorisedView{})) - ep.HandlePacket(&r, tcpip.PacketBuffer{ + ep.HandlePacket(&r, stack.PacketBuffer{ Data: view[:len(view)-c.trunc].ToVectorisedView(), }) if want := c.expectedCount; o.controlCalls != want { diff --git a/pkg/tcpip/network/ipv4/BUILD b/pkg/tcpip/network/ipv4/BUILD index 0fef2b1f1..880ea7de2 100644 --- a/pkg/tcpip/network/ipv4/BUILD +++ b/pkg/tcpip/network/ipv4/BUILD @@ -13,7 +13,6 @@ go_library( "//pkg/tcpip", "//pkg/tcpip/buffer", "//pkg/tcpip/header", - "//pkg/tcpip/iptables", "//pkg/tcpip/network/fragmentation", "//pkg/tcpip/network/hash", "//pkg/tcpip/stack", diff --git a/pkg/tcpip/network/ipv4/icmp.go b/pkg/tcpip/network/ipv4/icmp.go index 32bf39e43..c4bf1ba5c 100644 --- a/pkg/tcpip/network/ipv4/icmp.go +++ b/pkg/tcpip/network/ipv4/icmp.go @@ -15,7 +15,6 @@ package ipv4 import ( - "gvisor.dev/gvisor/pkg/tcpip" "gvisor.dev/gvisor/pkg/tcpip/buffer" "gvisor.dev/gvisor/pkg/tcpip/header" "gvisor.dev/gvisor/pkg/tcpip/stack" @@ -25,7 +24,7 @@ import ( // the original packet that caused the ICMP one to be sent. This information is // used to find out which transport endpoint must be notified about the ICMP // packet. -func (e *endpoint) handleControl(typ stack.ControlType, extra uint32, pkt tcpip.PacketBuffer) { +func (e *endpoint) handleControl(typ stack.ControlType, extra uint32, pkt stack.PacketBuffer) { h := header.IPv4(pkt.Data.First()) // We don't use IsValid() here because ICMP only requires that the IP @@ -53,7 +52,7 @@ func (e *endpoint) handleControl(typ stack.ControlType, extra uint32, pkt tcpip. e.dispatcher.DeliverTransportControlPacket(e.id.LocalAddress, h.DestinationAddress(), ProtocolNumber, p, typ, extra, pkt) } -func (e *endpoint) handleICMP(r *stack.Route, pkt tcpip.PacketBuffer) { +func (e *endpoint) handleICMP(r *stack.Route, pkt stack.PacketBuffer) { stats := r.Stats() received := stats.ICMP.V4PacketsReceived v := pkt.Data.First() @@ -85,7 +84,7 @@ func (e *endpoint) handleICMP(r *stack.Route, pkt tcpip.PacketBuffer) { // It's possible that a raw socket expects to receive this. h.SetChecksum(wantChecksum) - e.dispatcher.DeliverTransportPacket(r, header.ICMPv4ProtocolNumber, tcpip.PacketBuffer{ + e.dispatcher.DeliverTransportPacket(r, header.ICMPv4ProtocolNumber, stack.PacketBuffer{ Data: pkt.Data.Clone(nil), NetworkHeader: append(buffer.View(nil), pkt.NetworkHeader...), }) @@ -99,7 +98,7 @@ func (e *endpoint) handleICMP(r *stack.Route, pkt tcpip.PacketBuffer) { pkt.SetChecksum(0) pkt.SetChecksum(^header.Checksum(pkt, header.ChecksumVV(vv, 0))) sent := stats.ICMP.V4PacketsSent - if err := r.WritePacket(nil /* gso */, stack.NetworkHeaderParams{Protocol: header.ICMPv4ProtocolNumber, TTL: r.DefaultTTL(), TOS: stack.DefaultTOS}, tcpip.PacketBuffer{ + if err := r.WritePacket(nil /* gso */, stack.NetworkHeaderParams{Protocol: header.ICMPv4ProtocolNumber, TTL: r.DefaultTTL(), TOS: stack.DefaultTOS}, stack.PacketBuffer{ Header: hdr, Data: vv, TransportHeader: buffer.View(pkt), diff --git a/pkg/tcpip/network/ipv4/ipv4.go b/pkg/tcpip/network/ipv4/ipv4.go index 6597e6781..104aafbed 100644 --- a/pkg/tcpip/network/ipv4/ipv4.go +++ b/pkg/tcpip/network/ipv4/ipv4.go @@ -26,7 +26,6 @@ import ( "gvisor.dev/gvisor/pkg/tcpip" "gvisor.dev/gvisor/pkg/tcpip/buffer" "gvisor.dev/gvisor/pkg/tcpip/header" - "gvisor.dev/gvisor/pkg/tcpip/iptables" "gvisor.dev/gvisor/pkg/tcpip/network/fragmentation" "gvisor.dev/gvisor/pkg/tcpip/network/hash" "gvisor.dev/gvisor/pkg/tcpip/stack" @@ -125,7 +124,7 @@ func (e *endpoint) GSOMaxSize() uint32 { // packet's stated length matches the length of the header+payload. mtu // includes the IP header and options. This does not support the DontFragment // IP flag. -func (e *endpoint) writePacketFragments(r *stack.Route, gso *stack.GSO, mtu int, pkt tcpip.PacketBuffer) *tcpip.Error { +func (e *endpoint) writePacketFragments(r *stack.Route, gso *stack.GSO, mtu int, pkt stack.PacketBuffer) *tcpip.Error { // This packet is too big, it needs to be fragmented. ip := header.IPv4(pkt.Header.View()) flags := ip.Flags() @@ -165,7 +164,7 @@ func (e *endpoint) writePacketFragments(r *stack.Route, gso *stack.GSO, mtu int, if i > 0 { newPayload := pkt.Data.Clone(nil) newPayload.CapLength(innerMTU) - if err := e.linkEP.WritePacket(r, gso, ProtocolNumber, tcpip.PacketBuffer{ + if err := e.linkEP.WritePacket(r, gso, ProtocolNumber, stack.PacketBuffer{ Header: pkt.Header, Data: newPayload, NetworkHeader: buffer.View(h), @@ -184,7 +183,7 @@ func (e *endpoint) writePacketFragments(r *stack.Route, gso *stack.GSO, mtu int, newPayload := pkt.Data.Clone(nil) newPayloadLength := outerMTU - pkt.Header.UsedLength() newPayload.CapLength(newPayloadLength) - if err := e.linkEP.WritePacket(r, gso, ProtocolNumber, tcpip.PacketBuffer{ + if err := e.linkEP.WritePacket(r, gso, ProtocolNumber, stack.PacketBuffer{ Header: pkt.Header, Data: newPayload, NetworkHeader: buffer.View(h), @@ -198,7 +197,7 @@ func (e *endpoint) writePacketFragments(r *stack.Route, gso *stack.GSO, mtu int, startOfHdr := pkt.Header startOfHdr.TrimBack(pkt.Header.UsedLength() - outerMTU) emptyVV := buffer.NewVectorisedView(0, []buffer.View{}) - if err := e.linkEP.WritePacket(r, gso, ProtocolNumber, tcpip.PacketBuffer{ + if err := e.linkEP.WritePacket(r, gso, ProtocolNumber, stack.PacketBuffer{ Header: startOfHdr, Data: emptyVV, NetworkHeader: buffer.View(h), @@ -241,10 +240,18 @@ func (e *endpoint) addIPHeader(r *stack.Route, hdr *buffer.Prependable, payloadS } // WritePacket writes a packet to the given destination address and protocol. -func (e *endpoint) WritePacket(r *stack.Route, gso *stack.GSO, params stack.NetworkHeaderParams, pkt tcpip.PacketBuffer) *tcpip.Error { +func (e *endpoint) WritePacket(r *stack.Route, gso *stack.GSO, params stack.NetworkHeaderParams, pkt stack.PacketBuffer) *tcpip.Error { ip := e.addIPHeader(r, &pkt.Header, pkt.Data.Size(), params) pkt.NetworkHeader = buffer.View(ip) + // iptables filtering. All packets that reach here are locally + // generated. + ipt := e.stack.IPTables() + if ok := ipt.Check(stack.Output, pkt); !ok { + // iptables is telling us to drop the packet. + return nil + } + if r.Loop&stack.PacketLoop != 0 { // The inbound path expects the network header to still be in // the PacketBuffer's Data field. @@ -253,7 +260,7 @@ func (e *endpoint) WritePacket(r *stack.Route, gso *stack.GSO, params stack.Netw views = append(views, pkt.Data.Views()...) loopedR := r.MakeLoopedRoute() - e.HandlePacket(&loopedR, tcpip.PacketBuffer{ + e.HandlePacket(&loopedR, stack.PacketBuffer{ Data: buffer.NewVectorisedView(len(views[0])+pkt.Data.Size(), views), }) @@ -273,26 +280,52 @@ func (e *endpoint) WritePacket(r *stack.Route, gso *stack.GSO, params stack.Netw } // WritePackets implements stack.NetworkEndpoint.WritePackets. -func (e *endpoint) WritePackets(r *stack.Route, gso *stack.GSO, pkts []tcpip.PacketBuffer, params stack.NetworkHeaderParams) (int, *tcpip.Error) { +func (e *endpoint) WritePackets(r *stack.Route, gso *stack.GSO, pkts stack.PacketBufferList, params stack.NetworkHeaderParams) (int, *tcpip.Error) { if r.Loop&stack.PacketLoop != 0 { panic("multiple packets in local loop") } if r.Loop&stack.PacketOut == 0 { - return len(pkts), nil + return pkts.Len(), nil + } + + for pkt := pkts.Front(); pkt != nil; { + ip := e.addIPHeader(r, &pkt.Header, pkt.Data.Size(), params) + pkt.NetworkHeader = buffer.View(ip) + pkt = pkt.Next() } - for i := range pkts { - ip := e.addIPHeader(r, &pkts[i].Header, pkts[i].DataSize, params) - pkts[i].NetworkHeader = buffer.View(ip) + // iptables filtering. All packets that reach here are locally + // generated. + ipt := e.stack.IPTables() + dropped := ipt.CheckPackets(stack.Output, pkts) + if len(dropped) == 0 { + // Fast path: If no packets are to be dropped then we can just invoke the + // faster WritePackets API directly. + n, err := e.linkEP.WritePackets(r, gso, pkts, ProtocolNumber) + r.Stats().IP.PacketsSent.IncrementBy(uint64(n)) + return n, err + } + + // Slow Path as we are dropping some packets in the batch degrade to + // emitting one packet at a time. + n := 0 + for pkt := pkts.Front(); pkt != nil; pkt = pkt.Next() { + if _, ok := dropped[pkt]; ok { + continue + } + if err := e.linkEP.WritePacket(r, gso, ProtocolNumber, *pkt); err != nil { + r.Stats().IP.PacketsSent.IncrementBy(uint64(n)) + return n, err + } + n++ } - n, err := e.linkEP.WritePackets(r, gso, pkts, ProtocolNumber) r.Stats().IP.PacketsSent.IncrementBy(uint64(n)) - return n, err + return n, nil } // WriteHeaderIncludedPacket writes a packet already containing a network // header through the given route. -func (e *endpoint) WriteHeaderIncludedPacket(r *stack.Route, pkt tcpip.PacketBuffer) *tcpip.Error { +func (e *endpoint) WriteHeaderIncludedPacket(r *stack.Route, pkt stack.PacketBuffer) *tcpip.Error { // The packet already has an IP header, but there are a few required // checks. ip := header.IPv4(pkt.Data.First()) @@ -344,7 +377,7 @@ func (e *endpoint) WriteHeaderIncludedPacket(r *stack.Route, pkt tcpip.PacketBuf // HandlePacket is called by the link layer when new ipv4 packets arrive for // this endpoint. -func (e *endpoint) HandlePacket(r *stack.Route, pkt tcpip.PacketBuffer) { +func (e *endpoint) HandlePacket(r *stack.Route, pkt stack.PacketBuffer) { headerView := pkt.Data.First() h := header.IPv4(headerView) if !h.IsValid(pkt.Data.Size()) { @@ -361,7 +394,7 @@ func (e *endpoint) HandlePacket(r *stack.Route, pkt tcpip.PacketBuffer) { // iptables filtering. All packets that reach here are intended for // this machine and will not be forwarded. ipt := e.stack.IPTables() - if ok := ipt.Check(iptables.Input, pkt); !ok { + if ok := ipt.Check(stack.Input, pkt); !ok { // iptables is telling us to drop the packet. return } @@ -473,6 +506,12 @@ func (p *protocol) DefaultTTL() uint8 { return uint8(atomic.LoadUint32(&p.defaultTTL)) } +// Close implements stack.TransportProtocol.Close. +func (*protocol) Close() {} + +// Wait implements stack.TransportProtocol.Wait. +func (*protocol) Wait() {} + // calculateMTU calculates the network-layer payload MTU based on the link-layer // payload mtu. func calculateMTU(mtu uint32) uint32 { diff --git a/pkg/tcpip/network/ipv4/ipv4_test.go b/pkg/tcpip/network/ipv4/ipv4_test.go index e900f1b45..5a864d832 100644 --- a/pkg/tcpip/network/ipv4/ipv4_test.go +++ b/pkg/tcpip/network/ipv4/ipv4_test.go @@ -113,7 +113,7 @@ func makeHdrAndPayload(hdrLength int, extraLength int, viewSizes []int) (buffer. // comparePayloads compared the contents of all the packets against the contents // of the source packet. -func compareFragments(t *testing.T, packets []tcpip.PacketBuffer, sourcePacketInfo tcpip.PacketBuffer, mtu uint32) { +func compareFragments(t *testing.T, packets []stack.PacketBuffer, sourcePacketInfo stack.PacketBuffer, mtu uint32) { t.Helper() // Make a complete array of the sourcePacketInfo packet. source := header.IPv4(packets[0].Header.View()[:header.IPv4MinimumSize]) @@ -173,7 +173,7 @@ func compareFragments(t *testing.T, packets []tcpip.PacketBuffer, sourcePacketIn type errorChannel struct { *channel.Endpoint - Ch chan tcpip.PacketBuffer + Ch chan stack.PacketBuffer packetCollectorErrors []*tcpip.Error } @@ -183,7 +183,7 @@ type errorChannel struct { func newErrorChannel(size int, mtu uint32, linkAddr tcpip.LinkAddress, packetCollectorErrors []*tcpip.Error) *errorChannel { return &errorChannel{ Endpoint: channel.New(size, mtu, linkAddr), - Ch: make(chan tcpip.PacketBuffer, size), + Ch: make(chan stack.PacketBuffer, size), packetCollectorErrors: packetCollectorErrors, } } @@ -202,7 +202,7 @@ func (e *errorChannel) Drain() int { } // WritePacket stores outbound packets into the channel. -func (e *errorChannel) WritePacket(r *stack.Route, gso *stack.GSO, protocol tcpip.NetworkProtocolNumber, pkt tcpip.PacketBuffer) *tcpip.Error { +func (e *errorChannel) WritePacket(r *stack.Route, gso *stack.GSO, protocol tcpip.NetworkProtocolNumber, pkt stack.PacketBuffer) *tcpip.Error { select { case e.Ch <- pkt: default: @@ -281,13 +281,13 @@ func TestFragmentation(t *testing.T) { for _, ft := range fragTests { t.Run(ft.description, func(t *testing.T) { hdr, payload := makeHdrAndPayload(ft.hdrLength, ft.extraLength, ft.payloadViewsSizes) - source := tcpip.PacketBuffer{ + source := stack.PacketBuffer{ Header: hdr, // Save the source payload because WritePacket will modify it. Data: payload.Clone(nil), } c := buildContext(t, nil, ft.mtu) - err := c.Route.WritePacket(ft.gso, stack.NetworkHeaderParams{Protocol: tcp.ProtocolNumber, TTL: 42, TOS: stack.DefaultTOS}, tcpip.PacketBuffer{ + err := c.Route.WritePacket(ft.gso, stack.NetworkHeaderParams{Protocol: tcp.ProtocolNumber, TTL: 42, TOS: stack.DefaultTOS}, stack.PacketBuffer{ Header: hdr, Data: payload, }) @@ -295,7 +295,7 @@ func TestFragmentation(t *testing.T) { t.Errorf("err got %v, want %v", err, nil) } - var results []tcpip.PacketBuffer + var results []stack.PacketBuffer L: for { select { @@ -337,7 +337,7 @@ func TestFragmentationErrors(t *testing.T) { t.Run(ft.description, func(t *testing.T) { hdr, payload := makeHdrAndPayload(ft.hdrLength, header.IPv4MinimumSize, ft.payloadViewsSizes) c := buildContext(t, ft.packetCollectorErrors, ft.mtu) - err := c.Route.WritePacket(&stack.GSO{}, stack.NetworkHeaderParams{Protocol: tcp.ProtocolNumber, TTL: 42, TOS: stack.DefaultTOS}, tcpip.PacketBuffer{ + err := c.Route.WritePacket(&stack.GSO{}, stack.NetworkHeaderParams{Protocol: tcp.ProtocolNumber, TTL: 42, TOS: stack.DefaultTOS}, stack.PacketBuffer{ Header: hdr, Data: payload, }) @@ -459,7 +459,7 @@ func TestInvalidFragments(t *testing.T) { s.CreateNIC(nicID, sniffer.New(ep)) for _, pkt := range tc.packets { - ep.InjectLinkAddr(header.IPv4ProtocolNumber, remoteLinkAddr, tcpip.PacketBuffer{ + ep.InjectLinkAddr(header.IPv4ProtocolNumber, remoteLinkAddr, stack.PacketBuffer{ Data: buffer.NewVectorisedView(len(pkt), []buffer.View{pkt}), }) } diff --git a/pkg/tcpip/network/ipv6/BUILD b/pkg/tcpip/network/ipv6/BUILD index fb11874c6..3f71fc520 100644 --- a/pkg/tcpip/network/ipv6/BUILD +++ b/pkg/tcpip/network/ipv6/BUILD @@ -13,6 +13,8 @@ go_library( "//pkg/tcpip", "//pkg/tcpip/buffer", "//pkg/tcpip/header", + "//pkg/tcpip/network/fragmentation", + "//pkg/tcpip/network/hash", "//pkg/tcpip/stack", ], ) @@ -29,6 +31,7 @@ go_test( deps = [ "//pkg/tcpip", "//pkg/tcpip/buffer", + "//pkg/tcpip/checker", "//pkg/tcpip/header", "//pkg/tcpip/link/channel", "//pkg/tcpip/link/sniffer", @@ -36,5 +39,6 @@ go_test( "//pkg/tcpip/transport/icmp", "//pkg/tcpip/transport/udp", "//pkg/waiter", + "@com_github_google_go-cmp//cmp:go_default_library", ], ) diff --git a/pkg/tcpip/network/ipv6/icmp.go b/pkg/tcpip/network/ipv6/icmp.go index 45dc757c7..b68983d10 100644 --- a/pkg/tcpip/network/ipv6/icmp.go +++ b/pkg/tcpip/network/ipv6/icmp.go @@ -15,7 +15,7 @@ package ipv6 import ( - "log" + "fmt" "gvisor.dev/gvisor/pkg/tcpip" "gvisor.dev/gvisor/pkg/tcpip/buffer" @@ -27,7 +27,7 @@ import ( // the original packet that caused the ICMP one to be sent. This information is // used to find out which transport endpoint must be notified about the ICMP // packet. -func (e *endpoint) handleControl(typ stack.ControlType, extra uint32, pkt tcpip.PacketBuffer) { +func (e *endpoint) handleControl(typ stack.ControlType, extra uint32, pkt stack.PacketBuffer) { h := header.IPv6(pkt.Data.First()) // We don't use IsValid() here because ICMP only requires that up to @@ -62,7 +62,7 @@ func (e *endpoint) handleControl(typ stack.ControlType, extra uint32, pkt tcpip. e.dispatcher.DeliverTransportControlPacket(e.id.LocalAddress, h.DestinationAddress(), ProtocolNumber, p, typ, extra, pkt) } -func (e *endpoint) handleICMP(r *stack.Route, netHeader buffer.View, pkt tcpip.PacketBuffer) { +func (e *endpoint) handleICMP(r *stack.Route, netHeader buffer.View, pkt stack.PacketBuffer, hasFragmentHeader bool) { stats := r.Stats().ICMP sent := stats.V6PacketsSent received := stats.V6PacketsReceived @@ -79,32 +79,22 @@ func (e *endpoint) handleICMP(r *stack.Route, netHeader buffer.View, pkt tcpip.P // Only the first view in vv is accounted for by h. To account for the // rest of vv, a shallow copy is made and the first view is removed. // This copy is used as extra payload during the checksum calculation. - payload := pkt.Data + payload := pkt.Data.Clone(nil) payload.RemoveFirst() if got, want := h.Checksum(), header.ICMPv6Checksum(h, iph.SourceAddress(), iph.DestinationAddress(), payload); got != want { received.Invalid.Increment() return } - // As per RFC 4861 sections 4.1 - 4.5, 6.1.1, 6.1.2, 7.1.1, 7.1.2 and - // 8.1, nodes MUST silently drop NDP packets where the Hop Limit field - // in the IPv6 header is not set to 255, or the ICMPv6 Code field is not - // set to 0. - switch h.Type() { - case header.ICMPv6NeighborSolicit, - header.ICMPv6NeighborAdvert, - header.ICMPv6RouterSolicit, - header.ICMPv6RouterAdvert, - header.ICMPv6RedirectMsg: - if iph.HopLimit() != header.NDPHopLimit { - received.Invalid.Increment() - return - } - - if h.Code() != 0 { - received.Invalid.Increment() - return - } + isNDPValid := func() bool { + // As per RFC 4861 sections 4.1 - 4.5, 6.1.1, 6.1.2, 7.1.1, 7.1.2 and + // 8.1, nodes MUST silently drop NDP packets where the Hop Limit field + // in the IPv6 header is not set to 255, or the ICMPv6 Code field is not + // set to 0. + // + // As per RFC 6980 section 5, nodes MUST silently drop NDP messages if the + // packet includes a fragmentation header. + return !hasFragmentHeader && iph.HopLimit() == header.NDPHopLimit && h.Code() == 0 } // TODO(b/112892170): Meaningfully handle all ICMP types. @@ -133,7 +123,7 @@ func (e *endpoint) handleICMP(r *stack.Route, netHeader buffer.View, pkt tcpip.P case header.ICMPv6NeighborSolicit: received.NeighborSolicit.Increment() - if len(v) < header.ICMPv6NeighborSolicitMinimumSize { + if len(v) < header.ICMPv6NeighborSolicitMinimumSize || !isNDPValid() { received.Invalid.Increment() return } @@ -148,58 +138,53 @@ func (e *endpoint) handleICMP(r *stack.Route, netHeader buffer.View, pkt tcpip.P targetAddr := ns.TargetAddress() s := r.Stack() - rxNICID := r.NICID() - if isTentative, err := s.IsAddrTentative(rxNICID, targetAddr); err != nil { - // We will only get an error if rxNICID is unrecognized, - // which should not happen. For now short-circuit this - // packet. + if isTentative, err := s.IsAddrTentative(e.nicID, targetAddr); err != nil { + // We will only get an error if the NIC is unrecognized, which should not + // happen. For now, drop this packet. // // TODO(b/141002840): Handle this better? return } else if isTentative { - // If the target address is tentative and the source - // of the packet is a unicast (specified) address, then - // the source of the packet is attempting to perform - // address resolution on the target. In this case, the - // solicitation is silently ignored, as per RFC 4862 - // section 5.4.3. + // If the target address is tentative and the source of the packet is a + // unicast (specified) address, then the source of the packet is + // attempting to perform address resolution on the target. In this case, + // the solicitation is silently ignored, as per RFC 4862 section 5.4.3. // - // If the target address is tentative and the source of - // the packet is the unspecified address (::), then we - // know another node is also performing DAD for the - // same address (since targetAddr is tentative for us, - // we know we are also performing DAD on it). In this - // case we let the stack know so it can handle such a - // scenario and do nothing further with the NDP NS. - if iph.SourceAddress() == header.IPv6Any { - s.DupTentativeAddrDetected(rxNICID, targetAddr) + // If the target address is tentative and the source of the packet is the + // unspecified address (::), then we know another node is also performing + // DAD for the same address (since the target address is tentative for us, + // we know we are also performing DAD on it). In this case we let the + // stack know so it can handle such a scenario and do nothing further with + // the NS. + if r.RemoteAddress == header.IPv6Any { + s.DupTentativeAddrDetected(e.nicID, targetAddr) } - // Do not handle neighbor solicitations targeted - // to an address that is tentative on the received - // NIC any further. + // Do not handle neighbor solicitations targeted to an address that is + // tentative on the NIC any further. return } - // At this point we know that targetAddr is not tentative on - // rxNICID so the packet is processed as defined in RFC 4861, - // as per RFC 4862 section 5.4.3. + // At this point we know that the target address is not tentative on the NIC + // so the packet is processed as defined in RFC 4861, as per RFC 4862 + // section 5.4.3. + // Is the NS targetting us? if e.linkAddrCache.CheckLocalAddress(e.nicID, ProtocolNumber, targetAddr) == 0 { - // We don't have a useful answer; the best we can do is ignore the request. return } - // If the NS message has the source link layer option, update the link - // address cache with the link address for the sender of the message. + // If the NS message contains the Source Link-Layer Address option, update + // the link address cache with the value of the option. // // TODO(b/148429853): Properly process the NS message and do Neighbor // Unreachability Detection. + var sourceLinkAddr tcpip.LinkAddress for { opt, done, err := it.Next() if err != nil { // This should never happen as Iter(true) above did not return an error. - log.Fatalf("unexpected error when iterating over NDP options: %s", err) + panic(fmt.Sprintf("unexpected error when iterating over NDP options: %s", err)) } if done { break @@ -207,22 +192,36 @@ func (e *endpoint) handleICMP(r *stack.Route, netHeader buffer.View, pkt tcpip.P switch opt := opt.(type) { case header.NDPSourceLinkLayerAddressOption: - e.linkAddrCache.AddLinkAddress(e.nicID, r.RemoteAddress, opt.EthernetAddress()) + // No RFCs define what to do when an NS message has multiple Source + // Link-Layer Address options. Since no interface can have multiple + // link-layer addresses, we consider such messages invalid. + if len(sourceLinkAddr) != 0 { + received.Invalid.Increment() + return + } + + sourceLinkAddr = opt.EthernetAddress() } } - optsSerializer := header.NDPOptionsSerializer{ - header.NDPTargetLinkLayerAddressOption(r.LocalLinkAddress[:]), + unspecifiedSource := r.RemoteAddress == header.IPv6Any + + // As per RFC 4861 section 4.3, the Source Link-Layer Address Option MUST + // NOT be included when the source IP address is the unspecified address. + // Otherwise, on link layers that have addresses this option MUST be + // included in multicast solicitations and SHOULD be included in unicast + // solicitations. + if len(sourceLinkAddr) == 0 { + if header.IsV6MulticastAddress(r.LocalAddress) && !unspecifiedSource { + received.Invalid.Increment() + return + } + } else if unspecifiedSource { + received.Invalid.Increment() + return + } else { + e.linkAddrCache.AddLinkAddress(e.nicID, r.RemoteAddress, sourceLinkAddr) } - hdr := buffer.NewPrependable(int(r.MaxHeaderLength()) + header.ICMPv6NeighborAdvertMinimumSize + int(optsSerializer.Length())) - packet := header.ICMPv6(hdr.Prepend(header.ICMPv6NeighborAdvertSize)) - packet.SetType(header.ICMPv6NeighborAdvert) - na := header.NDPNeighborAdvert(packet.NDPPayload()) - na.SetSolicitedFlag(true) - na.SetOverrideFlag(true) - na.SetTargetAddress(targetAddr) - opts := na.Options() - opts.Serialize(optsSerializer) // ICMPv6 Neighbor Solicit messages are always sent to // specially crafted IPv6 multicast addresses. As a result, the @@ -235,6 +234,40 @@ func (e *endpoint) handleICMP(r *stack.Route, netHeader buffer.View, pkt tcpip.P r := r.Clone() defer r.Release() r.LocalAddress = targetAddr + + // As per RFC 4861 section 7.2.4, if the the source of the solicitation is + // the unspecified address, the node MUST set the Solicited flag to zero and + // multicast the advertisement to the all-nodes address. + solicited := true + if unspecifiedSource { + solicited = false + r.RemoteAddress = header.IPv6AllNodesMulticastAddress + } + + // If the NS has a source link-layer option, use the link address it + // specifies as the remote link address for the response instead of the + // source link address of the packet. + // + // TODO(#2401): As per RFC 4861 section 7.2.4 we should consult our link + // address cache for the right destination link address instead of manually + // patching the route with the remote link address if one is specified in a + // Source Link-Layer Address option. + if len(sourceLinkAddr) != 0 { + r.RemoteLinkAddress = sourceLinkAddr + } + + optsSerializer := header.NDPOptionsSerializer{ + header.NDPTargetLinkLayerAddressOption(r.LocalLinkAddress), + } + hdr := buffer.NewPrependable(int(r.MaxHeaderLength()) + header.ICMPv6NeighborAdvertMinimumSize + int(optsSerializer.Length())) + packet := header.ICMPv6(hdr.Prepend(header.ICMPv6NeighborAdvertSize)) + packet.SetType(header.ICMPv6NeighborAdvert) + na := header.NDPNeighborAdvert(packet.NDPPayload()) + na.SetSolicitedFlag(solicited) + na.SetOverrideFlag(true) + na.SetTargetAddress(targetAddr) + opts := na.Options() + opts.Serialize(optsSerializer) packet.SetChecksum(header.ICMPv6Checksum(packet, r.LocalAddress, r.RemoteAddress, buffer.VectorisedView{})) // RFC 4861 Neighbor Discovery for IP version 6 (IPv6) @@ -243,7 +276,7 @@ func (e *endpoint) handleICMP(r *stack.Route, netHeader buffer.View, pkt tcpip.P // // The IP Hop Limit field has a value of 255, i.e., the packet // could not possibly have been forwarded by a router. - if err := r.WritePacket(nil /* gso */, stack.NetworkHeaderParams{Protocol: header.ICMPv6ProtocolNumber, TTL: header.NDPHopLimit, TOS: stack.DefaultTOS}, tcpip.PacketBuffer{ + if err := r.WritePacket(nil /* gso */, stack.NetworkHeaderParams{Protocol: header.ICMPv6ProtocolNumber, TTL: header.NDPHopLimit, TOS: stack.DefaultTOS}, stack.PacketBuffer{ Header: hdr, }); err != nil { sent.Dropped.Increment() @@ -253,7 +286,7 @@ func (e *endpoint) handleICMP(r *stack.Route, netHeader buffer.View, pkt tcpip.P case header.ICMPv6NeighborAdvert: received.NeighborAdvert.Increment() - if len(v) < header.ICMPv6NeighborAdvertSize { + if len(v) < header.ICMPv6NeighborAdvertSize || !isNDPValid() { received.Invalid.Increment() return } @@ -268,45 +301,43 @@ func (e *endpoint) handleICMP(r *stack.Route, netHeader buffer.View, pkt tcpip.P targetAddr := na.TargetAddress() stack := r.Stack() - rxNICID := r.NICID() - if isTentative, err := stack.IsAddrTentative(rxNICID, targetAddr); err != nil { - // We will only get an error if rxNICID is unrecognized, - // which should not happen. For now short-circuit this - // packet. + if isTentative, err := stack.IsAddrTentative(e.nicID, targetAddr); err != nil { + // We will only get an error if the NIC is unrecognized, which should not + // happen. For now short-circuit this packet. // // TODO(b/141002840): Handle this better? return } else if isTentative { - // We just got an NA from a node that owns an address we - // are performing DAD on, implying the address is not - // unique. In this case we let the stack know so it can - // handle such a scenario and do nothing furthur with + // We just got an NA from a node that owns an address we are performing + // DAD on, implying the address is not unique. In this case we let the + // stack know so it can handle such a scenario and do nothing furthur with // the NDP NA. - stack.DupTentativeAddrDetected(rxNICID, targetAddr) + stack.DupTentativeAddrDetected(e.nicID, targetAddr) return } - // At this point we know that the targetAddress is not tentative - // on rxNICID. However, targetAddr may still be assigned to - // rxNICID but not tentative (it could be permanent). Such a - // scenario is beyond the scope of RFC 4862. As such, we simply - // ignore such a scenario for now and proceed as normal. + // At this point we know that the target address is not tentative on the + // NIC. However, the target address may still be assigned to the NIC but not + // tentative (it could be permanent). Such a scenario is beyond the scope of + // RFC 4862. As such, we simply ignore such a scenario for now and proceed + // as normal. // + // TODO(b/143147598): Handle the scenario described above. Also inform the + // netstack integration that a duplicate address was detected outside of + // DAD. + // If the NA message has the target link layer option, update the link // address cache with the link address for the target of the message. // - // TODO(b/143147598): Handle the scenario described above. Also - // inform the netstack integration that a duplicate address was - // detected outside of DAD. - // // TODO(b/148429853): Properly process the NA message and do Neighbor // Unreachability Detection. + var targetLinkAddr tcpip.LinkAddress for { opt, done, err := it.Next() if err != nil { // This should never happen as Iter(true) above did not return an error. - log.Fatalf("unexpected error when iterating over NDP options: %s", err) + panic(fmt.Sprintf("unexpected error when iterating over NDP options: %s", err)) } if done { break @@ -314,10 +345,22 @@ func (e *endpoint) handleICMP(r *stack.Route, netHeader buffer.View, pkt tcpip.P switch opt := opt.(type) { case header.NDPTargetLinkLayerAddressOption: - e.linkAddrCache.AddLinkAddress(e.nicID, targetAddr, opt.EthernetAddress()) + // No RFCs define what to do when an NA message has multiple Target + // Link-Layer Address options. Since no interface can have multiple + // link-layer addresses, we consider such messages invalid. + if len(targetLinkAddr) != 0 { + received.Invalid.Increment() + return + } + + targetLinkAddr = opt.EthernetAddress() } } + if len(targetLinkAddr) != 0 { + e.linkAddrCache.AddLinkAddress(e.nicID, targetAddr, targetLinkAddr) + } + case header.ICMPv6EchoRequest: received.EchoRequest.Increment() if len(v) < header.ICMPv6EchoMinimumSize { @@ -330,7 +373,7 @@ func (e *endpoint) handleICMP(r *stack.Route, netHeader buffer.View, pkt tcpip.P copy(packet, h) packet.SetType(header.ICMPv6EchoReply) packet.SetChecksum(header.ICMPv6Checksum(packet, r.LocalAddress, r.RemoteAddress, pkt.Data)) - if err := r.WritePacket(nil /* gso */, stack.NetworkHeaderParams{Protocol: header.ICMPv6ProtocolNumber, TTL: r.DefaultTTL(), TOS: stack.DefaultTOS}, tcpip.PacketBuffer{ + if err := r.WritePacket(nil /* gso */, stack.NetworkHeaderParams{Protocol: header.ICMPv6ProtocolNumber, TTL: r.DefaultTTL(), TOS: stack.DefaultTOS}, stack.PacketBuffer{ Header: hdr, Data: pkt.Data, }); err != nil { @@ -355,8 +398,20 @@ func (e *endpoint) handleICMP(r *stack.Route, netHeader buffer.View, pkt tcpip.P case header.ICMPv6RouterSolicit: received.RouterSolicit.Increment() + if !isNDPValid() { + received.Invalid.Increment() + return + } case header.ICMPv6RouterAdvert: + received.RouterAdvert.Increment() + + p := h.NDPPayload() + if len(p) < header.NDPRAMinimumSize || !isNDPValid() { + received.Invalid.Increment() + return + } + routerAddr := iph.SourceAddress() // @@ -370,16 +425,6 @@ func (e *endpoint) handleICMP(r *stack.Route, netHeader buffer.View, pkt tcpip.P return } - p := h.NDPPayload() - - // Is the NDP payload of sufficient size to hold a Router - // Advertisement? - if len(p) < header.NDPRAMinimumSize { - // ...No, silently drop the packet. - received.Invalid.Increment() - return - } - ra := header.NDPRouterAdvert(p) opts := ra.Options() @@ -395,8 +440,6 @@ func (e *endpoint) handleICMP(r *stack.Route, netHeader buffer.View, pkt tcpip.P // as RFC 4861 section 6.1.2 is concerned. // - received.RouterAdvert.Increment() - // Tell the NIC to handle the RA. stack := r.Stack() rxNICID := r.NICID() @@ -404,6 +447,10 @@ func (e *endpoint) handleICMP(r *stack.Route, netHeader buffer.View, pkt tcpip.P case header.ICMPv6RedirectMsg: received.RedirectMsg.Increment() + if !isNDPValid() { + received.Invalid.Increment() + return + } default: received.Invalid.Increment() @@ -463,7 +510,7 @@ func (*protocol) LinkAddressRequest(addr, localAddr tcpip.Address, linkEP stack. }) // TODO(stijlist): count this in ICMP stats. - return linkEP.WritePacket(r, nil /* gso */, ProtocolNumber, tcpip.PacketBuffer{ + return linkEP.WritePacket(r, nil /* gso */, ProtocolNumber, stack.PacketBuffer{ Header: hdr, }) } diff --git a/pkg/tcpip/network/ipv6/icmp_test.go b/pkg/tcpip/network/ipv6/icmp_test.go index 50c4b6474..bd099a7f8 100644 --- a/pkg/tcpip/network/ipv6/icmp_test.go +++ b/pkg/tcpip/network/ipv6/icmp_test.go @@ -32,7 +32,8 @@ import ( const ( linkAddr0 = tcpip.LinkAddress("\x02\x02\x03\x04\x05\x06") - linkAddr1 = tcpip.LinkAddress("\x0a\x0b\x0c\x0d\x0e\x0f") + linkAddr1 = tcpip.LinkAddress("\x0a\x0b\x0c\x0d\x0e\x0e") + linkAddr2 = tcpip.LinkAddress("\x0a\x0b\x0c\x0d\x0e\x0f") ) var ( @@ -56,7 +57,7 @@ func (*stubLinkEndpoint) LinkAddress() tcpip.LinkAddress { return "" } -func (*stubLinkEndpoint) WritePacket(*stack.Route, *stack.GSO, tcpip.NetworkProtocolNumber, tcpip.PacketBuffer) *tcpip.Error { +func (*stubLinkEndpoint) WritePacket(*stack.Route, *stack.GSO, tcpip.NetworkProtocolNumber, stack.PacketBuffer) *tcpip.Error { return nil } @@ -66,7 +67,7 @@ type stubDispatcher struct { stack.TransportDispatcher } -func (*stubDispatcher) DeliverTransportPacket(*stack.Route, tcpip.TransportProtocolNumber, tcpip.PacketBuffer) { +func (*stubDispatcher) DeliverTransportPacket(*stack.Route, tcpip.TransportProtocolNumber, stack.PacketBuffer) { } type stubLinkAddressCache struct { @@ -187,7 +188,7 @@ func TestICMPCounts(t *testing.T) { SrcAddr: r.LocalAddress, DstAddr: r.RemoteAddress, }) - ep.HandlePacket(&r, tcpip.PacketBuffer{ + ep.HandlePacket(&r, stack.PacketBuffer{ Data: hdr.View().ToVectorisedView(), }) } @@ -326,7 +327,7 @@ func routeICMPv6Packet(t *testing.T, args routeArgs, fn func(*testing.T, header. views := []buffer.View{pi.Pkt.Header.View(), pi.Pkt.Data.ToView()} size := pi.Pkt.Header.UsedLength() + pi.Pkt.Data.Size() vv := buffer.NewVectorisedView(size, views) - args.dst.InjectLinkAddr(pi.Proto, args.dst.LinkAddress(), tcpip.PacketBuffer{ + args.dst.InjectLinkAddr(pi.Proto, args.dst.LinkAddress(), stack.PacketBuffer{ Data: vv, }) } @@ -561,7 +562,7 @@ func TestICMPChecksumValidationSimple(t *testing.T) { SrcAddr: lladdr1, DstAddr: lladdr0, }) - e.InjectInbound(ProtocolNumber, tcpip.PacketBuffer{ + e.InjectInbound(ProtocolNumber, stack.PacketBuffer{ Data: hdr.View().ToVectorisedView(), }) } @@ -738,7 +739,7 @@ func TestICMPChecksumValidationWithPayload(t *testing.T) { SrcAddr: lladdr1, DstAddr: lladdr0, }) - e.InjectInbound(ProtocolNumber, tcpip.PacketBuffer{ + e.InjectInbound(ProtocolNumber, stack.PacketBuffer{ Data: hdr.View().ToVectorisedView(), }) } @@ -916,7 +917,7 @@ func TestICMPChecksumValidationWithPayloadMultipleViews(t *testing.T) { SrcAddr: lladdr1, DstAddr: lladdr0, }) - e.InjectInbound(ProtocolNumber, tcpip.PacketBuffer{ + e.InjectInbound(ProtocolNumber, stack.PacketBuffer{ Data: buffer.NewVectorisedView(header.IPv6MinimumSize+size+payloadSize, []buffer.View{hdr.View(), payload}), }) } diff --git a/pkg/tcpip/network/ipv6/ipv6.go b/pkg/tcpip/network/ipv6/ipv6.go index 180a480fd..331b0817b 100644 --- a/pkg/tcpip/network/ipv6/ipv6.go +++ b/pkg/tcpip/network/ipv6/ipv6.go @@ -21,11 +21,14 @@ package ipv6 import ( + "fmt" "sync/atomic" "gvisor.dev/gvisor/pkg/tcpip" "gvisor.dev/gvisor/pkg/tcpip/buffer" "gvisor.dev/gvisor/pkg/tcpip/header" + "gvisor.dev/gvisor/pkg/tcpip/network/fragmentation" + "gvisor.dev/gvisor/pkg/tcpip/network/hash" "gvisor.dev/gvisor/pkg/tcpip/stack" ) @@ -49,6 +52,7 @@ type endpoint struct { linkEP stack.LinkEndpoint linkAddrCache stack.LinkAddressCache dispatcher stack.TransportDispatcher + fragmentation *fragmentation.Fragmentation protocol *protocol } @@ -112,7 +116,7 @@ func (e *endpoint) addIPHeader(r *stack.Route, hdr *buffer.Prependable, payloadS } // WritePacket writes a packet to the given destination address and protocol. -func (e *endpoint) WritePacket(r *stack.Route, gso *stack.GSO, params stack.NetworkHeaderParams, pkt tcpip.PacketBuffer) *tcpip.Error { +func (e *endpoint) WritePacket(r *stack.Route, gso *stack.GSO, params stack.NetworkHeaderParams, pkt stack.PacketBuffer) *tcpip.Error { ip := e.addIPHeader(r, &pkt.Header, pkt.Data.Size(), params) pkt.NetworkHeader = buffer.View(ip) @@ -124,7 +128,7 @@ func (e *endpoint) WritePacket(r *stack.Route, gso *stack.GSO, params stack.Netw views = append(views, pkt.Data.Views()...) loopedR := r.MakeLoopedRoute() - e.HandlePacket(&loopedR, tcpip.PacketBuffer{ + e.HandlePacket(&loopedR, stack.PacketBuffer{ Data: buffer.NewVectorisedView(len(views[0])+pkt.Data.Size(), views), }) @@ -139,19 +143,17 @@ func (e *endpoint) WritePacket(r *stack.Route, gso *stack.GSO, params stack.Netw } // WritePackets implements stack.LinkEndpoint.WritePackets. -func (e *endpoint) WritePackets(r *stack.Route, gso *stack.GSO, pkts []tcpip.PacketBuffer, params stack.NetworkHeaderParams) (int, *tcpip.Error) { +func (e *endpoint) WritePackets(r *stack.Route, gso *stack.GSO, pkts stack.PacketBufferList, params stack.NetworkHeaderParams) (int, *tcpip.Error) { if r.Loop&stack.PacketLoop != 0 { panic("not implemented") } if r.Loop&stack.PacketOut == 0 { - return len(pkts), nil + return pkts.Len(), nil } - for i := range pkts { - hdr := &pkts[i].Header - size := pkts[i].DataSize - ip := e.addIPHeader(r, hdr, size, params) - pkts[i].NetworkHeader = buffer.View(ip) + for pb := pkts.Front(); pb != nil; pb = pb.Next() { + ip := e.addIPHeader(r, &pb.Header, pb.Data.Size(), params) + pb.NetworkHeader = buffer.View(ip) } n, err := e.linkEP.WritePackets(r, gso, pkts, ProtocolNumber) @@ -161,17 +163,18 @@ func (e *endpoint) WritePackets(r *stack.Route, gso *stack.GSO, pkts []tcpip.Pac // WriteHeaderIncludedPacker implements stack.NetworkEndpoint. It is not yet // supported by IPv6. -func (*endpoint) WriteHeaderIncludedPacket(r *stack.Route, pkt tcpip.PacketBuffer) *tcpip.Error { +func (*endpoint) WriteHeaderIncludedPacket(r *stack.Route, pkt stack.PacketBuffer) *tcpip.Error { // TODO(b/146666412): Support IPv6 header-included packets. return tcpip.ErrNotSupported } // HandlePacket is called by the link layer when new ipv6 packets arrive for // this endpoint. -func (e *endpoint) HandlePacket(r *stack.Route, pkt tcpip.PacketBuffer) { +func (e *endpoint) HandlePacket(r *stack.Route, pkt stack.PacketBuffer) { headerView := pkt.Data.First() h := header.IPv6(headerView) if !h.IsValid(pkt.Data.Size()) { + r.Stats().IP.MalformedPacketsReceived.Increment() return } @@ -179,14 +182,235 @@ func (e *endpoint) HandlePacket(r *stack.Route, pkt tcpip.PacketBuffer) { pkt.Data.TrimFront(header.IPv6MinimumSize) pkt.Data.CapLength(int(h.PayloadLength())) - p := h.TransportProtocol() - if p == header.ICMPv6ProtocolNumber { - e.handleICMP(r, headerView, pkt) - return + it := header.MakeIPv6PayloadIterator(header.IPv6ExtensionHeaderIdentifier(h.NextHeader()), pkt.Data) + hasFragmentHeader := false + + for firstHeader := true; ; firstHeader = false { + extHdr, done, err := it.Next() + if err != nil { + r.Stats().IP.MalformedPacketsReceived.Increment() + return + } + if done { + break + } + + switch extHdr := extHdr.(type) { + case header.IPv6HopByHopOptionsExtHdr: + // As per RFC 8200 section 4.1, the Hop By Hop extension header is + // restricted to appear immediately after an IPv6 fixed header. + // + // TODO(b/152019344): Send an ICMPv6 Parameter Problem, Code 1 + // (unrecognized next header) error in response to an extension header's + // Next Header field with the Hop By Hop extension header identifier. + if !firstHeader { + return + } + + optsIt := extHdr.Iter() + + for { + opt, done, err := optsIt.Next() + if err != nil { + r.Stats().IP.MalformedPacketsReceived.Increment() + return + } + if done { + break + } + + // We currently do not support any IPv6 Hop By Hop extension header + // options. + switch opt.UnknownAction() { + case header.IPv6OptionUnknownActionSkip: + case header.IPv6OptionUnknownActionDiscard: + return + case header.IPv6OptionUnknownActionDiscardSendICMP: + // TODO(b/152019344): Send an ICMPv6 Parameter Problem Code 2 for + // unrecognized IPv6 extension header options. + return + case header.IPv6OptionUnknownActionDiscardSendICMPNoMulticastDest: + // TODO(b/152019344): Send an ICMPv6 Parameter Problem Code 2 for + // unrecognized IPv6 extension header options. + return + default: + panic(fmt.Sprintf("unrecognized action for an unrecognized Hop By Hop extension header option = %d", opt)) + } + } + + case header.IPv6RoutingExtHdr: + // As per RFC 8200 section 4.4, if a node encounters a routing header with + // an unrecognized routing type value, with a non-zero Segments Left + // value, the node must discard the packet and send an ICMP Parameter + // Problem, Code 0. If the Segments Left is 0, the node must ignore the + // Routing extension header and process the next header in the packet. + // + // Note, the stack does not yet handle any type of routing extension + // header, so we just make sure Segments Left is zero before processing + // the next extension header. + // + // TODO(b/152019344): Send an ICMPv6 Parameter Problem Code 0 for + // unrecognized routing types with a non-zero Segments Left value. + if extHdr.SegmentsLeft() != 0 { + return + } + + case header.IPv6FragmentExtHdr: + hasFragmentHeader = true + + fragmentOffset := extHdr.FragmentOffset() + more := extHdr.More() + if !more && fragmentOffset == 0 { + // This fragment extension header indicates that this packet is an + // atomic fragment. An atomic fragment is a fragment that contains + // all the data required to reassemble a full packet. As per RFC 6946, + // atomic fragments must not interfere with "normal" fragmented traffic + // so we skip processing the fragment instead of feeding it through the + // reassembly process below. + continue + } + + // Don't consume the iterator if we have the first fragment because we + // will use it to validate that the first fragment holds the upper layer + // header. + rawPayload := it.AsRawHeader(fragmentOffset != 0 /* consume */) + + if fragmentOffset == 0 { + // Check that the iterator ends with a raw payload as the first fragment + // should include all headers up to and including any upper layer + // headers, as per RFC 8200 section 4.5; only upper layer data + // (non-headers) should follow the fragment extension header. + var lastHdr header.IPv6PayloadHeader + + for { + it, done, err := it.Next() + if err != nil { + r.Stats().IP.MalformedPacketsReceived.Increment() + r.Stats().IP.MalformedPacketsReceived.Increment() + return + } + if done { + break + } + + lastHdr = it + } + + // If the last header is a raw header, then the last portion of the IPv6 + // payload is not a known IPv6 extension header. Note, this does not + // mean that the last portion is an upper layer header or not an + // extension header because: + // 1) we do not yet support all extension headers + // 2) we do not validate the upper layer header before reassembling. + // + // This check makes sure that a known IPv6 extension header is not + // present after the Fragment extension header in a non-initial + // fragment. + // + // TODO(#2196): Support IPv6 Authentication and Encapsulated + // Security Payload extension headers. + // TODO(#2333): Validate that the upper layer header is valid. + switch lastHdr.(type) { + case header.IPv6RawPayloadHeader: + default: + r.Stats().IP.MalformedPacketsReceived.Increment() + r.Stats().IP.MalformedFragmentsReceived.Increment() + return + } + } + + fragmentPayloadLen := rawPayload.Buf.Size() + if fragmentPayloadLen == 0 { + // Drop the packet as it's marked as a fragment but has no payload. + r.Stats().IP.MalformedPacketsReceived.Increment() + r.Stats().IP.MalformedFragmentsReceived.Increment() + return + } + + // The packet is a fragment, let's try to reassemble it. + start := fragmentOffset * header.IPv6FragmentExtHdrFragmentOffsetBytesPerUnit + last := start + uint16(fragmentPayloadLen) - 1 + + // Drop the packet if the fragmentOffset is incorrect. i.e the + // combination of fragmentOffset and pkt.Data.size() causes a + // wrap around resulting in last being less than the offset. + if last < start { + r.Stats().IP.MalformedPacketsReceived.Increment() + r.Stats().IP.MalformedFragmentsReceived.Increment() + return + } + + var ready bool + pkt.Data, ready, err = e.fragmentation.Process(hash.IPv6FragmentHash(h, extHdr.ID()), start, last, more, rawPayload.Buf) + if err != nil { + r.Stats().IP.MalformedPacketsReceived.Increment() + r.Stats().IP.MalformedFragmentsReceived.Increment() + return + } + + if ready { + // We create a new iterator with the reassembled packet because we could + // have more extension headers in the reassembled payload, as per RFC + // 8200 section 4.5. + it = header.MakeIPv6PayloadIterator(rawPayload.Identifier, pkt.Data) + } + + case header.IPv6DestinationOptionsExtHdr: + optsIt := extHdr.Iter() + + for { + opt, done, err := optsIt.Next() + if err != nil { + r.Stats().IP.MalformedPacketsReceived.Increment() + return + } + if done { + break + } + + // We currently do not support any IPv6 Destination extension header + // options. + switch opt.UnknownAction() { + case header.IPv6OptionUnknownActionSkip: + case header.IPv6OptionUnknownActionDiscard: + return + case header.IPv6OptionUnknownActionDiscardSendICMP: + // TODO(b/152019344): Send an ICMPv6 Parameter Problem Code 2 for + // unrecognized IPv6 extension header options. + return + case header.IPv6OptionUnknownActionDiscardSendICMPNoMulticastDest: + // TODO(b/152019344): Send an ICMPv6 Parameter Problem Code 2 for + // unrecognized IPv6 extension header options. + return + default: + panic(fmt.Sprintf("unrecognized action for an unrecognized Destination extension header option = %d", opt)) + } + } + + case header.IPv6RawPayloadHeader: + // If the last header in the payload isn't a known IPv6 extension header, + // handle it as if it is transport layer data. + pkt.Data = extHdr.Buf + + if p := tcpip.TransportProtocolNumber(extHdr.Identifier); p == header.ICMPv6ProtocolNumber { + e.handleICMP(r, headerView, pkt, hasFragmentHeader) + } else { + r.Stats().IP.PacketsDelivered.Increment() + // TODO(b/152019344): Send an ICMPv6 Parameter Problem, Code 1 error + // in response to unrecognized next header values. + e.dispatcher.DeliverTransportPacket(r, p, pkt) + } + + default: + // If we receive a packet for an extension header we do not yet handle, + // drop the packet for now. + // + // TODO(b/152019344): Send an ICMPv6 Parameter Problem, Code 1 error + // in response to unrecognized next header values. + r.Stats().UnknownProtocolRcvdPackets.Increment() + return + } } - - r.Stats().IP.PacketsDelivered.Increment() - e.dispatcher.DeliverTransportPacket(r, p, pkt) } // Close cleans up resources associated with the endpoint. @@ -229,6 +453,7 @@ func (p *protocol) NewEndpoint(nicID tcpip.NICID, addrWithPrefix tcpip.AddressWi linkEP: linkEP, linkAddrCache: linkAddrCache, dispatcher: dispatcher, + fragmentation: fragmentation.NewFragmentation(fragmentation.HighFragThreshold, fragmentation.LowFragThreshold, fragmentation.DefaultReassembleTimeout), protocol: p, }, nil } @@ -265,6 +490,12 @@ func (p *protocol) DefaultTTL() uint8 { return uint8(atomic.LoadUint32(&p.defaultTTL)) } +// Close implements stack.TransportProtocol.Close. +func (*protocol) Close() {} + +// Wait implements stack.TransportProtocol.Wait. +func (*protocol) Wait() {} + // calculateMTU calculates the network-layer payload MTU based on the link-layer // payload mtu. func calculateMTU(mtu uint32) uint32 { diff --git a/pkg/tcpip/network/ipv6/ipv6_test.go b/pkg/tcpip/network/ipv6/ipv6_test.go index 1cbfa7278..841a0cb7a 100644 --- a/pkg/tcpip/network/ipv6/ipv6_test.go +++ b/pkg/tcpip/network/ipv6/ipv6_test.go @@ -17,6 +17,7 @@ package ipv6 import ( "testing" + "github.com/google/go-cmp/cmp" "gvisor.dev/gvisor/pkg/tcpip" "gvisor.dev/gvisor/pkg/tcpip/buffer" "gvisor.dev/gvisor/pkg/tcpip/header" @@ -33,6 +34,15 @@ const ( // The least significant 3 bytes are the same as addr2 so both addr2 and // addr3 will have the same solicited-node address. addr3 = "\x0a\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x01\x00\x00\x02" + addr4 = "\x0a\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x01\x00\x00\x03" + + // Tests use the extension header identifier values as uint8 instead of + // header.IPv6ExtensionHeaderIdentifier. + hopByHopExtHdrID = uint8(header.IPv6HopByHopOptionsExtHdrIdentifier) + routingExtHdrID = uint8(header.IPv6RoutingExtHdrIdentifier) + fragmentExtHdrID = uint8(header.IPv6FragmentExtHdrIdentifier) + destinationExtHdrID = uint8(header.IPv6DestinationOptionsExtHdrIdentifier) + noNextHdrID = uint8(header.IPv6NoNextHeaderIdentifier) ) // testReceiveICMP tests receiving an ICMP packet from src to dst. want is the @@ -55,7 +65,7 @@ func testReceiveICMP(t *testing.T, s *stack.Stack, e *channel.Endpoint, src, dst DstAddr: dst, }) - e.InjectInbound(ProtocolNumber, tcpip.PacketBuffer{ + e.InjectInbound(ProtocolNumber, stack.PacketBuffer{ Data: hdr.View().ToVectorisedView(), }) @@ -113,7 +123,7 @@ func testReceiveUDP(t *testing.T, s *stack.Stack, e *channel.Endpoint, src, dst DstAddr: dst, }) - e.InjectInbound(ProtocolNumber, tcpip.PacketBuffer{ + e.InjectInbound(ProtocolNumber, stack.PacketBuffer{ Data: hdr.View().ToVectorisedView(), }) @@ -158,6 +168,8 @@ func TestReceiveOnAllNodesMulticastAddr(t *testing.T) { // packets destined to the IPv6 solicited-node address of an assigned IPv6 // address. func TestReceiveOnSolicitedNodeAddr(t *testing.T) { + const nicID = 1 + tests := []struct { name string protocolFactory stack.TransportProtocol @@ -175,50 +187,61 @@ func TestReceiveOnSolicitedNodeAddr(t *testing.T) { NetworkProtocols: []stack.NetworkProtocol{NewProtocol()}, TransportProtocols: []stack.TransportProtocol{test.protocolFactory}, }) - e := channel.New(10, 1280, linkAddr1) - if err := s.CreateNIC(1, e); err != nil { - t.Fatalf("CreateNIC(_) = %s", err) + e := channel.New(1, 1280, linkAddr1) + if err := s.CreateNIC(nicID, e); err != nil { + t.Fatalf("CreateNIC(%d, _) = %s", nicID, err) } - // Should not receive a packet destined to the solicited - // node address of addr2/addr3 yet as we haven't added - // those addresses. + s.SetRouteTable([]tcpip.Route{ + tcpip.Route{ + Destination: header.IPv6EmptySubnet, + NIC: nicID, + }, + }) + + // Should not receive a packet destined to the solicited node address of + // addr2/addr3 yet as we haven't added those addresses. test.rxf(t, s, e, addr1, snmc, 0) - if err := s.AddAddress(1, ProtocolNumber, addr2); err != nil { - t.Fatalf("AddAddress(_, %d, %s) = %s", ProtocolNumber, addr2, err) + if err := s.AddAddress(nicID, ProtocolNumber, addr2); err != nil { + t.Fatalf("AddAddress(%d, %d, %s) = %s", nicID, ProtocolNumber, addr2, err) } - // Should receive a packet destined to the solicited - // node address of addr2/addr3 now that we have added - // added addr2. + // Should receive a packet destined to the solicited node address of + // addr2/addr3 now that we have added added addr2. test.rxf(t, s, e, addr1, snmc, 1) - if err := s.AddAddress(1, ProtocolNumber, addr3); err != nil { - t.Fatalf("AddAddress(_, %d, %s) = %s", ProtocolNumber, addr3, err) + if err := s.AddAddress(nicID, ProtocolNumber, addr3); err != nil { + t.Fatalf("AddAddress(%d, %d, %s) = %s", nicID, ProtocolNumber, addr3, err) } - // Should still receive a packet destined to the - // solicited node address of addr2/addr3 now that we - // have added addr3. + // Should still receive a packet destined to the solicited node address of + // addr2/addr3 now that we have added addr3. test.rxf(t, s, e, addr1, snmc, 2) - if err := s.RemoveAddress(1, addr2); err != nil { - t.Fatalf("RemoveAddress(_, %s) = %s", addr2, err) + if err := s.RemoveAddress(nicID, addr2); err != nil { + t.Fatalf("RemoveAddress(%d, %s) = %s", nicID, addr2, err) } - // Should still receive a packet destined to the - // solicited node address of addr2/addr3 now that we - // have removed addr2. + // Should still receive a packet destined to the solicited node address of + // addr2/addr3 now that we have removed addr2. test.rxf(t, s, e, addr1, snmc, 3) - if err := s.RemoveAddress(1, addr3); err != nil { - t.Fatalf("RemoveAddress(_, %s) = %s", addr3, err) + // Make sure addr3's endpoint does not get removed from the NIC by + // incrementing its reference count with a route. + r, err := s.FindRoute(nicID, addr3, addr4, ProtocolNumber, false) + if err != nil { + t.Fatalf("FindRoute(%d, %s, %s, %d, false): %s", nicID, addr3, addr4, ProtocolNumber, err) + } + defer r.Release() + + if err := s.RemoveAddress(nicID, addr3); err != nil { + t.Fatalf("RemoveAddress(%d, %s) = %s", nicID, addr3, err) } - // Should not receive a packet destined to the solicited - // node address of addr2/addr3 yet as both of them got - // removed. + // Should not receive a packet destined to the solicited node address of + // addr2/addr3 yet as both of them got removed, even though a route using + // addr3 exists. test.rxf(t, s, e, addr1, snmc, 3) }) } @@ -268,3 +291,975 @@ func TestAddIpv6Address(t *testing.T) { }) } } + +func TestReceiveIPv6ExtHdrs(t *testing.T) { + const nicID = 1 + + tests := []struct { + name string + extHdr func(nextHdr uint8) ([]byte, uint8) + shouldAccept bool + }{ + { + name: "None", + extHdr: func(nextHdr uint8) ([]byte, uint8) { return []byte{}, nextHdr }, + shouldAccept: true, + }, + { + name: "hopbyhop with unknown option skippable action", + extHdr: func(nextHdr uint8) ([]byte, uint8) { + return []byte{ + nextHdr, 1, + + // Skippable unknown. + 63, 4, 1, 2, 3, 4, + + // Skippable unknown. + 62, 6, 1, 2, 3, 4, 5, 6, + }, hopByHopExtHdrID + }, + shouldAccept: true, + }, + { + name: "hopbyhop with unknown option discard action", + extHdr: func(nextHdr uint8) ([]byte, uint8) { + return []byte{ + nextHdr, 1, + + // Skippable unknown. + 63, 4, 1, 2, 3, 4, + + // Discard unknown. + 127, 6, 1, 2, 3, 4, 5, 6, + }, hopByHopExtHdrID + }, + shouldAccept: false, + }, + { + name: "hopbyhop with unknown option discard and send icmp action", + extHdr: func(nextHdr uint8) ([]byte, uint8) { + return []byte{ + nextHdr, 1, + + // Skippable unknown. + 63, 4, 1, 2, 3, 4, + + // Discard & send ICMP if option is unknown. + 191, 6, 1, 2, 3, 4, 5, 6, + }, hopByHopExtHdrID + }, + shouldAccept: false, + }, + { + name: "hopbyhop with unknown option discard and send icmp action unless multicast dest", + extHdr: func(nextHdr uint8) ([]byte, uint8) { + return []byte{ + nextHdr, 1, + + // Skippable unknown. + 63, 4, 1, 2, 3, 4, + + // Discard & send ICMP unless packet is for multicast destination if + // option is unknown. + 255, 6, 1, 2, 3, 4, 5, 6, + }, hopByHopExtHdrID + }, + shouldAccept: false, + }, + { + name: "routing with zero segments left", + extHdr: func(nextHdr uint8) ([]byte, uint8) { return []byte{nextHdr, 0, 1, 0, 2, 3, 4, 5}, routingExtHdrID }, + shouldAccept: true, + }, + { + name: "routing with non-zero segments left", + extHdr: func(nextHdr uint8) ([]byte, uint8) { return []byte{nextHdr, 0, 1, 1, 2, 3, 4, 5}, routingExtHdrID }, + shouldAccept: false, + }, + { + name: "atomic fragment with zero ID", + extHdr: func(nextHdr uint8) ([]byte, uint8) { return []byte{nextHdr, 0, 0, 0, 0, 0, 0, 0}, fragmentExtHdrID }, + shouldAccept: true, + }, + { + name: "atomic fragment with non-zero ID", + extHdr: func(nextHdr uint8) ([]byte, uint8) { return []byte{nextHdr, 0, 0, 0, 1, 2, 3, 4}, fragmentExtHdrID }, + shouldAccept: true, + }, + { + name: "fragment", + extHdr: func(nextHdr uint8) ([]byte, uint8) { return []byte{nextHdr, 0, 1, 0, 1, 2, 3, 4}, fragmentExtHdrID }, + shouldAccept: false, + }, + { + name: "No next header", + extHdr: func(nextHdr uint8) ([]byte, uint8) { return []byte{}, noNextHdrID }, + shouldAccept: false, + }, + { + name: "destination with unknown option skippable action", + extHdr: func(nextHdr uint8) ([]byte, uint8) { + return []byte{ + nextHdr, 1, + + // Skippable unknown. + 63, 4, 1, 2, 3, 4, + + // Skippable unknown. + 62, 6, 1, 2, 3, 4, 5, 6, + }, destinationExtHdrID + }, + shouldAccept: true, + }, + { + name: "destination with unknown option discard action", + extHdr: func(nextHdr uint8) ([]byte, uint8) { + return []byte{ + nextHdr, 1, + + // Skippable unknown. + 63, 4, 1, 2, 3, 4, + + // Discard unknown. + 127, 6, 1, 2, 3, 4, 5, 6, + }, destinationExtHdrID + }, + shouldAccept: false, + }, + { + name: "destination with unknown option discard and send icmp action", + extHdr: func(nextHdr uint8) ([]byte, uint8) { + return []byte{ + nextHdr, 1, + + // Skippable unknown. + 63, 4, 1, 2, 3, 4, + + // Discard & send ICMP if option is unknown. + 191, 6, 1, 2, 3, 4, 5, 6, + }, destinationExtHdrID + }, + shouldAccept: false, + }, + { + name: "destination with unknown option discard and send icmp action unless multicast dest", + extHdr: func(nextHdr uint8) ([]byte, uint8) { + return []byte{ + nextHdr, 1, + + // Skippable unknown. + 63, 4, 1, 2, 3, 4, + + // Discard & send ICMP unless packet is for multicast destination if + // option is unknown. + 255, 6, 1, 2, 3, 4, 5, 6, + }, destinationExtHdrID + }, + shouldAccept: false, + }, + { + name: "routing - atomic fragment", + extHdr: func(nextHdr uint8) ([]byte, uint8) { + return []byte{ + // Routing extension header. + fragmentExtHdrID, 0, 1, 0, 2, 3, 4, 5, + + // Fragment extension header. + nextHdr, 0, 0, 0, 1, 2, 3, 4, + }, routingExtHdrID + }, + shouldAccept: true, + }, + { + name: "atomic fragment - routing", + extHdr: func(nextHdr uint8) ([]byte, uint8) { + return []byte{ + // Fragment extension header. + routingExtHdrID, 0, 0, 0, 1, 2, 3, 4, + + // Routing extension header. + nextHdr, 0, 1, 0, 2, 3, 4, 5, + }, fragmentExtHdrID + }, + shouldAccept: true, + }, + { + name: "hop by hop (with skippable unknown) - routing", + extHdr: func(nextHdr uint8) ([]byte, uint8) { + return []byte{ + // Hop By Hop extension header with skippable unknown option. + routingExtHdrID, 0, 62, 4, 1, 2, 3, 4, + + // Routing extension header. + nextHdr, 0, 1, 0, 2, 3, 4, 5, + }, hopByHopExtHdrID + }, + shouldAccept: true, + }, + { + name: "routing - hop by hop (with skippable unknown)", + extHdr: func(nextHdr uint8) ([]byte, uint8) { + return []byte{ + // Routing extension header. + hopByHopExtHdrID, 0, 1, 0, 2, 3, 4, 5, + + // Hop By Hop extension header with skippable unknown option. + nextHdr, 0, 62, 4, 1, 2, 3, 4, + }, routingExtHdrID + }, + shouldAccept: false, + }, + { + name: "No next header", + extHdr: func(nextHdr uint8) ([]byte, uint8) { return []byte{}, noNextHdrID }, + shouldAccept: false, + }, + { + name: "hopbyhop (with skippable unknown) - routing - atomic fragment - destination (with skippable unknown)", + extHdr: func(nextHdr uint8) ([]byte, uint8) { + return []byte{ + // Hop By Hop extension header with skippable unknown option. + routingExtHdrID, 0, 62, 4, 1, 2, 3, 4, + + // Routing extension header. + fragmentExtHdrID, 0, 1, 0, 2, 3, 4, 5, + + // Fragment extension header. + destinationExtHdrID, 0, 0, 0, 1, 2, 3, 4, + + // Destination extension header with skippable unknown option. + nextHdr, 0, 63, 4, 1, 2, 3, 4, + }, hopByHopExtHdrID + }, + shouldAccept: true, + }, + { + name: "hopbyhop (with discard unknown) - routing - atomic fragment - destination (with skippable unknown)", + extHdr: func(nextHdr uint8) ([]byte, uint8) { + return []byte{ + // Hop By Hop extension header with discard action for unknown option. + routingExtHdrID, 0, 65, 4, 1, 2, 3, 4, + + // Routing extension header. + fragmentExtHdrID, 0, 1, 0, 2, 3, 4, 5, + + // Fragment extension header. + destinationExtHdrID, 0, 0, 0, 1, 2, 3, 4, + + // Destination extension header with skippable unknown option. + nextHdr, 0, 63, 4, 1, 2, 3, 4, + }, hopByHopExtHdrID + }, + shouldAccept: false, + }, + { + name: "hopbyhop (with skippable unknown) - routing - atomic fragment - destination (with discard unknown)", + extHdr: func(nextHdr uint8) ([]byte, uint8) { + return []byte{ + // Hop By Hop extension header with skippable unknown option. + routingExtHdrID, 0, 62, 4, 1, 2, 3, 4, + + // Routing extension header. + fragmentExtHdrID, 0, 1, 0, 2, 3, 4, 5, + + // Fragment extension header. + destinationExtHdrID, 0, 0, 0, 1, 2, 3, 4, + + // Destination extension header with discard action for unknown + // option. + nextHdr, 0, 65, 4, 1, 2, 3, 4, + }, hopByHopExtHdrID + }, + shouldAccept: false, + }, + } + + for _, test := range tests { + t.Run(test.name, func(t *testing.T) { + s := stack.New(stack.Options{ + NetworkProtocols: []stack.NetworkProtocol{NewProtocol()}, + TransportProtocols: []stack.TransportProtocol{udp.NewProtocol()}, + }) + e := channel.New(0, 1280, linkAddr1) + if err := s.CreateNIC(nicID, e); err != nil { + t.Fatalf("CreateNIC(%d, _) = %s", nicID, err) + } + if err := s.AddAddress(nicID, ProtocolNumber, addr2); err != nil { + t.Fatalf("AddAddress(%d, %d, %s) = %s", nicID, ProtocolNumber, addr2, err) + } + + wq := waiter.Queue{} + we, ch := waiter.NewChannelEntry(nil) + wq.EventRegister(&we, waiter.EventIn) + defer wq.EventUnregister(&we) + defer close(ch) + ep, err := s.NewEndpoint(udp.ProtocolNumber, ProtocolNumber, &wq) + if err != nil { + t.Fatalf("NewEndpoint(%d, %d, _): %s", udp.ProtocolNumber, ProtocolNumber, err) + } + defer ep.Close() + + bindAddr := tcpip.FullAddress{Addr: addr2, Port: 80} + if err := ep.Bind(bindAddr); err != nil { + t.Fatalf("Bind(%+v): %s", bindAddr, err) + } + + udpPayload := []byte{1, 2, 3, 4, 5, 6, 7, 8} + udpLength := header.UDPMinimumSize + len(udpPayload) + extHdrBytes, ipv6NextHdr := test.extHdr(uint8(header.UDPProtocolNumber)) + extHdrLen := len(extHdrBytes) + hdr := buffer.NewPrependable(header.IPv6MinimumSize + extHdrLen + udpLength) + + // Serialize UDP message. + u := header.UDP(hdr.Prepend(udpLength)) + u.Encode(&header.UDPFields{ + SrcPort: 5555, + DstPort: 80, + Length: uint16(udpLength), + }) + copy(u.Payload(), udpPayload) + sum := header.PseudoHeaderChecksum(udp.ProtocolNumber, addr1, addr2, uint16(udpLength)) + sum = header.Checksum(udpPayload, sum) + u.SetChecksum(^u.CalculateChecksum(sum)) + + // Copy extension header bytes between the UDP message and the IPv6 + // fixed header. + copy(hdr.Prepend(extHdrLen), extHdrBytes) + + // Serialize IPv6 fixed header. + payloadLength := hdr.UsedLength() + ip := header.IPv6(hdr.Prepend(header.IPv6MinimumSize)) + ip.Encode(&header.IPv6Fields{ + PayloadLength: uint16(payloadLength), + NextHeader: ipv6NextHdr, + HopLimit: 255, + SrcAddr: addr1, + DstAddr: addr2, + }) + + e.InjectInbound(ProtocolNumber, stack.PacketBuffer{ + Data: hdr.View().ToVectorisedView(), + }) + + stats := s.Stats().UDP.PacketsReceived + + if !test.shouldAccept { + if got := stats.Value(); got != 0 { + t.Errorf("got UDP Rx Packets = %d, want = 0", got) + } + + return + } + + // Expect a UDP packet. + if got := stats.Value(); got != 1 { + t.Errorf("got UDP Rx Packets = %d, want = 1", got) + } + gotPayload, _, err := ep.Read(nil) + if err != nil { + t.Fatalf("Read(nil): %s", err) + } + if diff := cmp.Diff(buffer.View(udpPayload), gotPayload); diff != "" { + t.Errorf("got UDP payload mismatch (-want +got):\n%s", diff) + } + + // Should not have any more UDP packets. + if gotPayload, _, err := ep.Read(nil); err != tcpip.ErrWouldBlock { + t.Fatalf("got Read(nil) = (%x, _, %v), want = (_, _, %s)", gotPayload, err, tcpip.ErrWouldBlock) + } + }) + } +} + +// fragmentData holds the IPv6 payload for a fragmented IPv6 packet. +type fragmentData struct { + nextHdr uint8 + data buffer.VectorisedView +} + +func TestReceiveIPv6Fragments(t *testing.T) { + const nicID = 1 + const udpPayload1Length = 256 + const udpPayload2Length = 128 + const fragmentExtHdrLen = 8 + // Note, not all routing extension headers will be 8 bytes but this test + // uses 8 byte routing extension headers for most sub tests. + const routingExtHdrLen = 8 + + udpGen := func(payload []byte, multiplier uint8) buffer.View { + payloadLen := len(payload) + for i := 0; i < payloadLen; i++ { + payload[i] = uint8(i) * multiplier + } + + udpLength := header.UDPMinimumSize + payloadLen + + hdr := buffer.NewPrependable(udpLength) + u := header.UDP(hdr.Prepend(udpLength)) + u.Encode(&header.UDPFields{ + SrcPort: 5555, + DstPort: 80, + Length: uint16(udpLength), + }) + copy(u.Payload(), payload) + sum := header.PseudoHeaderChecksum(udp.ProtocolNumber, addr1, addr2, uint16(udpLength)) + sum = header.Checksum(payload, sum) + u.SetChecksum(^u.CalculateChecksum(sum)) + return hdr.View() + } + + var udpPayload1Buf [udpPayload1Length]byte + udpPayload1 := udpPayload1Buf[:] + ipv6Payload1 := udpGen(udpPayload1, 1) + + var udpPayload2Buf [udpPayload2Length]byte + udpPayload2 := udpPayload2Buf[:] + ipv6Payload2 := udpGen(udpPayload2, 2) + + tests := []struct { + name string + expectedPayload []byte + fragments []fragmentData + expectedPayloads [][]byte + }{ + { + name: "No fragmentation", + fragments: []fragmentData{ + { + nextHdr: uint8(header.UDPProtocolNumber), + data: ipv6Payload1.ToVectorisedView(), + }, + }, + expectedPayloads: [][]byte{udpPayload1}, + }, + { + name: "Atomic fragment", + fragments: []fragmentData{ + { + nextHdr: fragmentExtHdrID, + data: buffer.NewVectorisedView( + fragmentExtHdrLen+len(ipv6Payload1), + []buffer.View{ + // Fragment extension header. + buffer.View([]byte{uint8(header.UDPProtocolNumber), 0, 0, 0, 0, 0, 0, 0}), + + ipv6Payload1, + }, + ), + }, + }, + expectedPayloads: [][]byte{udpPayload1}, + }, + { + name: "Two fragments", + fragments: []fragmentData{ + { + nextHdr: fragmentExtHdrID, + data: buffer.NewVectorisedView( + fragmentExtHdrLen+64, + []buffer.View{ + // Fragment extension header. + // + // Fragment offset = 0, More = true, ID = 1 + buffer.View([]byte{uint8(header.UDPProtocolNumber), 0, 0, 1, 0, 0, 0, 1}), + + ipv6Payload1[:64], + }, + ), + }, + { + nextHdr: fragmentExtHdrID, + data: buffer.NewVectorisedView( + fragmentExtHdrLen+len(ipv6Payload1)-64, + []buffer.View{ + // Fragment extension header. + // + // Fragment offset = 8, More = false, ID = 1 + buffer.View([]byte{uint8(header.UDPProtocolNumber), 0, 0, 64, 0, 0, 0, 1}), + + ipv6Payload1[64:], + }, + ), + }, + }, + expectedPayloads: [][]byte{udpPayload1}, + }, + { + name: "Two fragments with different IDs", + fragments: []fragmentData{ + { + nextHdr: fragmentExtHdrID, + data: buffer.NewVectorisedView( + fragmentExtHdrLen+64, + []buffer.View{ + // Fragment extension header. + // + // Fragment offset = 0, More = true, ID = 1 + buffer.View([]byte{uint8(header.UDPProtocolNumber), 0, 0, 1, 0, 0, 0, 1}), + + ipv6Payload1[:64], + }, + ), + }, + { + nextHdr: fragmentExtHdrID, + data: buffer.NewVectorisedView( + fragmentExtHdrLen+len(ipv6Payload1)-64, + []buffer.View{ + // Fragment extension header. + // + // Fragment offset = 8, More = false, ID = 2 + buffer.View([]byte{uint8(header.UDPProtocolNumber), 0, 0, 64, 0, 0, 0, 2}), + + ipv6Payload1[64:], + }, + ), + }, + }, + expectedPayloads: nil, + }, + { + name: "Two fragments with per-fragment routing header with zero segments left", + fragments: []fragmentData{ + { + nextHdr: routingExtHdrID, + data: buffer.NewVectorisedView( + routingExtHdrLen+fragmentExtHdrLen+64, + []buffer.View{ + // Routing extension header. + // + // Segments left = 0. + buffer.View([]byte{fragmentExtHdrID, 0, 1, 0, 2, 3, 4, 5}), + + // Fragment extension header. + // + // Fragment offset = 0, More = true, ID = 1 + buffer.View([]byte{uint8(header.UDPProtocolNumber), 0, 0, 1, 0, 0, 0, 1}), + + ipv6Payload1[:64], + }, + ), + }, + { + nextHdr: routingExtHdrID, + data: buffer.NewVectorisedView( + routingExtHdrLen+fragmentExtHdrLen+len(ipv6Payload1)-64, + []buffer.View{ + // Routing extension header. + // + // Segments left = 0. + buffer.View([]byte{fragmentExtHdrID, 0, 1, 0, 2, 3, 4, 5}), + + // Fragment extension header. + // + // Fragment offset = 8, More = false, ID = 1 + buffer.View([]byte{uint8(header.UDPProtocolNumber), 0, 0, 64, 0, 0, 0, 1}), + + ipv6Payload1[64:], + }, + ), + }, + }, + expectedPayloads: [][]byte{udpPayload1}, + }, + { + name: "Two fragments with per-fragment routing header with non-zero segments left", + fragments: []fragmentData{ + { + nextHdr: routingExtHdrID, + data: buffer.NewVectorisedView( + routingExtHdrLen+fragmentExtHdrLen+64, + []buffer.View{ + // Routing extension header. + // + // Segments left = 1. + buffer.View([]byte{fragmentExtHdrID, 0, 1, 1, 2, 3, 4, 5}), + + // Fragment extension header. + // + // Fragment offset = 0, More = true, ID = 1 + buffer.View([]byte{uint8(header.UDPProtocolNumber), 0, 0, 1, 0, 0, 0, 1}), + + ipv6Payload1[:64], + }, + ), + }, + { + nextHdr: routingExtHdrID, + data: buffer.NewVectorisedView( + routingExtHdrLen+fragmentExtHdrLen+len(ipv6Payload1)-64, + []buffer.View{ + // Routing extension header. + // + // Segments left = 1. + buffer.View([]byte{fragmentExtHdrID, 0, 1, 1, 2, 3, 4, 5}), + + // Fragment extension header. + // + // Fragment offset = 9, More = false, ID = 1 + buffer.View([]byte{uint8(header.UDPProtocolNumber), 0, 0, 72, 0, 0, 0, 1}), + + ipv6Payload1[64:], + }, + ), + }, + }, + expectedPayloads: nil, + }, + { + name: "Two fragments with routing header with zero segments left", + fragments: []fragmentData{ + { + nextHdr: fragmentExtHdrID, + data: buffer.NewVectorisedView( + routingExtHdrLen+fragmentExtHdrLen+64, + []buffer.View{ + // Fragment extension header. + // + // Fragment offset = 0, More = true, ID = 1 + buffer.View([]byte{routingExtHdrID, 0, 0, 1, 0, 0, 0, 1}), + + // Routing extension header. + // + // Segments left = 0. + buffer.View([]byte{uint8(header.UDPProtocolNumber), 0, 1, 0, 2, 3, 4, 5}), + + ipv6Payload1[:64], + }, + ), + }, + { + nextHdr: fragmentExtHdrID, + data: buffer.NewVectorisedView( + fragmentExtHdrLen+len(ipv6Payload1)-64, + []buffer.View{ + // Fragment extension header. + // + // Fragment offset = 9, More = false, ID = 1 + buffer.View([]byte{routingExtHdrID, 0, 0, 72, 0, 0, 0, 1}), + + ipv6Payload1[64:], + }, + ), + }, + }, + expectedPayloads: [][]byte{udpPayload1}, + }, + { + name: "Two fragments with routing header with non-zero segments left", + fragments: []fragmentData{ + { + nextHdr: fragmentExtHdrID, + data: buffer.NewVectorisedView( + routingExtHdrLen+fragmentExtHdrLen+64, + []buffer.View{ + // Fragment extension header. + // + // Fragment offset = 0, More = true, ID = 1 + buffer.View([]byte{routingExtHdrID, 0, 0, 1, 0, 0, 0, 1}), + + // Routing extension header. + // + // Segments left = 1. + buffer.View([]byte{uint8(header.UDPProtocolNumber), 0, 1, 1, 2, 3, 4, 5}), + + ipv6Payload1[:64], + }, + ), + }, + { + nextHdr: fragmentExtHdrID, + data: buffer.NewVectorisedView( + fragmentExtHdrLen+len(ipv6Payload1)-64, + []buffer.View{ + // Fragment extension header. + // + // Fragment offset = 9, More = false, ID = 1 + buffer.View([]byte{routingExtHdrID, 0, 0, 72, 0, 0, 0, 1}), + + ipv6Payload1[64:], + }, + ), + }, + }, + expectedPayloads: nil, + }, + { + name: "Two fragments with routing header with zero segments left across fragments", + fragments: []fragmentData{ + { + nextHdr: fragmentExtHdrID, + data: buffer.NewVectorisedView( + // The length of this payload is fragmentExtHdrLen+8 because the + // first 8 bytes of the 16 byte routing extension header is in + // this fragment. + fragmentExtHdrLen+8, + []buffer.View{ + // Fragment extension header. + // + // Fragment offset = 0, More = true, ID = 1 + buffer.View([]byte{routingExtHdrID, 0, 0, 1, 0, 0, 0, 1}), + + // Routing extension header (part 1) + // + // Segments left = 0. + buffer.View([]byte{uint8(header.UDPProtocolNumber), 1, 1, 0, 2, 3, 4, 5}), + }, + ), + }, + { + nextHdr: fragmentExtHdrID, + data: buffer.NewVectorisedView( + // The length of this payload is + // fragmentExtHdrLen+8+len(ipv6Payload1) because the last 8 bytes of + // the 16 byte routing extension header is in this fagment. + fragmentExtHdrLen+8+len(ipv6Payload1), + []buffer.View{ + // Fragment extension header. + // + // Fragment offset = 1, More = false, ID = 1 + buffer.View([]byte{routingExtHdrID, 0, 0, 8, 0, 0, 0, 1}), + + // Routing extension header (part 2) + buffer.View([]byte{6, 7, 8, 9, 10, 11, 12, 13}), + + ipv6Payload1, + }, + ), + }, + }, + expectedPayloads: nil, + }, + { + name: "Two fragments with routing header with non-zero segments left across fragments", + fragments: []fragmentData{ + { + nextHdr: fragmentExtHdrID, + data: buffer.NewVectorisedView( + // The length of this payload is fragmentExtHdrLen+8 because the + // first 8 bytes of the 16 byte routing extension header is in + // this fragment. + fragmentExtHdrLen+8, + []buffer.View{ + // Fragment extension header. + // + // Fragment offset = 0, More = true, ID = 1 + buffer.View([]byte{routingExtHdrID, 0, 0, 1, 0, 0, 0, 1}), + + // Routing extension header (part 1) + // + // Segments left = 1. + buffer.View([]byte{uint8(header.UDPProtocolNumber), 1, 1, 1, 2, 3, 4, 5}), + }, + ), + }, + { + nextHdr: fragmentExtHdrID, + data: buffer.NewVectorisedView( + // The length of this payload is + // fragmentExtHdrLen+8+len(ipv6Payload1) because the last 8 bytes of + // the 16 byte routing extension header is in this fagment. + fragmentExtHdrLen+8+len(ipv6Payload1), + []buffer.View{ + // Fragment extension header. + // + // Fragment offset = 1, More = false, ID = 1 + buffer.View([]byte{routingExtHdrID, 0, 0, 8, 0, 0, 0, 1}), + + // Routing extension header (part 2) + buffer.View([]byte{6, 7, 8, 9, 10, 11, 12, 13}), + + ipv6Payload1, + }, + ), + }, + }, + expectedPayloads: nil, + }, + // As per RFC 6946, IPv6 atomic fragments MUST NOT interfere with "normal" + // fragmented traffic. + { + name: "Two fragments with atomic", + fragments: []fragmentData{ + { + nextHdr: fragmentExtHdrID, + data: buffer.NewVectorisedView( + fragmentExtHdrLen+64, + []buffer.View{ + // Fragment extension header. + // + // Fragment offset = 0, More = true, ID = 1 + buffer.View([]byte{uint8(header.UDPProtocolNumber), 0, 0, 1, 0, 0, 0, 1}), + + ipv6Payload1[:64], + }, + ), + }, + // This fragment has the same ID as the other fragments but is an atomic + // fragment. It should not interfere with the other fragments. + { + nextHdr: fragmentExtHdrID, + data: buffer.NewVectorisedView( + fragmentExtHdrLen+len(ipv6Payload2), + []buffer.View{ + // Fragment extension header. + // + // Fragment offset = 0, More = false, ID = 1 + buffer.View([]byte{uint8(header.UDPProtocolNumber), 0, 0, 0, 0, 0, 0, 1}), + + ipv6Payload2, + }, + ), + }, + { + nextHdr: fragmentExtHdrID, + data: buffer.NewVectorisedView( + fragmentExtHdrLen+len(ipv6Payload1)-64, + []buffer.View{ + // Fragment extension header. + // + // Fragment offset = 8, More = false, ID = 1 + buffer.View([]byte{uint8(header.UDPProtocolNumber), 0, 0, 64, 0, 0, 0, 1}), + + ipv6Payload1[64:], + }, + ), + }, + }, + expectedPayloads: [][]byte{udpPayload2, udpPayload1}, + }, + { + name: "Two interleaved fragmented packets", + fragments: []fragmentData{ + { + nextHdr: fragmentExtHdrID, + data: buffer.NewVectorisedView( + fragmentExtHdrLen+64, + []buffer.View{ + // Fragment extension header. + // + // Fragment offset = 0, More = true, ID = 1 + buffer.View([]byte{uint8(header.UDPProtocolNumber), 0, 0, 1, 0, 0, 0, 1}), + + ipv6Payload1[:64], + }, + ), + }, + { + nextHdr: fragmentExtHdrID, + data: buffer.NewVectorisedView( + fragmentExtHdrLen+32, + []buffer.View{ + // Fragment extension header. + // + // Fragment offset = 0, More = true, ID = 2 + buffer.View([]byte{uint8(header.UDPProtocolNumber), 0, 0, 1, 0, 0, 0, 2}), + + ipv6Payload2[:32], + }, + ), + }, + { + nextHdr: fragmentExtHdrID, + data: buffer.NewVectorisedView( + fragmentExtHdrLen+len(ipv6Payload1)-64, + []buffer.View{ + // Fragment extension header. + // + // Fragment offset = 8, More = false, ID = 1 + buffer.View([]byte{uint8(header.UDPProtocolNumber), 0, 0, 64, 0, 0, 0, 1}), + + ipv6Payload1[64:], + }, + ), + }, + { + nextHdr: fragmentExtHdrID, + data: buffer.NewVectorisedView( + fragmentExtHdrLen+len(ipv6Payload2)-32, + []buffer.View{ + // Fragment extension header. + // + // Fragment offset = 4, More = false, ID = 2 + buffer.View([]byte{uint8(header.UDPProtocolNumber), 0, 0, 32, 0, 0, 0, 2}), + + ipv6Payload2[32:], + }, + ), + }, + }, + expectedPayloads: [][]byte{udpPayload1, udpPayload2}, + }, + } + + for _, test := range tests { + t.Run(test.name, func(t *testing.T) { + s := stack.New(stack.Options{ + NetworkProtocols: []stack.NetworkProtocol{NewProtocol()}, + TransportProtocols: []stack.TransportProtocol{udp.NewProtocol()}, + }) + e := channel.New(0, 1280, linkAddr1) + if err := s.CreateNIC(nicID, e); err != nil { + t.Fatalf("CreateNIC(%d, _) = %s", nicID, err) + } + if err := s.AddAddress(nicID, ProtocolNumber, addr2); err != nil { + t.Fatalf("AddAddress(%d, %d, %s) = %s", nicID, ProtocolNumber, addr2, err) + } + + wq := waiter.Queue{} + we, ch := waiter.NewChannelEntry(nil) + wq.EventRegister(&we, waiter.EventIn) + defer wq.EventUnregister(&we) + defer close(ch) + ep, err := s.NewEndpoint(udp.ProtocolNumber, ProtocolNumber, &wq) + if err != nil { + t.Fatalf("NewEndpoint(%d, %d, _): %s", udp.ProtocolNumber, ProtocolNumber, err) + } + defer ep.Close() + + bindAddr := tcpip.FullAddress{Addr: addr2, Port: 80} + if err := ep.Bind(bindAddr); err != nil { + t.Fatalf("Bind(%+v): %s", bindAddr, err) + } + + for _, f := range test.fragments { + hdr := buffer.NewPrependable(header.IPv6MinimumSize) + + // Serialize IPv6 fixed header. + ip := header.IPv6(hdr.Prepend(header.IPv6MinimumSize)) + ip.Encode(&header.IPv6Fields{ + PayloadLength: uint16(f.data.Size()), + NextHeader: f.nextHdr, + HopLimit: 255, + SrcAddr: addr1, + DstAddr: addr2, + }) + + vv := hdr.View().ToVectorisedView() + vv.Append(f.data) + + e.InjectInbound(ProtocolNumber, stack.PacketBuffer{ + Data: vv, + }) + } + + if got, want := s.Stats().UDP.PacketsReceived.Value(), uint64(len(test.expectedPayloads)); got != want { + t.Errorf("got UDP Rx Packets = %d, want = %d", got, want) + } + + for i, p := range test.expectedPayloads { + gotPayload, _, err := ep.Read(nil) + if err != nil { + t.Fatalf("(i=%d) Read(nil): %s", i, err) + } + if diff := cmp.Diff(buffer.View(p), gotPayload); diff != "" { + t.Errorf("(i=%d) got UDP payload mismatch (-want +got):\n%s", i, diff) + } + } + + if gotPayload, _, err := ep.Read(nil); err != tcpip.ErrWouldBlock { + t.Fatalf("(last) got Read(nil) = (%x, _, %v), want = (_, _, %s)", gotPayload, err, tcpip.ErrWouldBlock) + } + }) + } +} diff --git a/pkg/tcpip/network/ipv6/ndp_test.go b/pkg/tcpip/network/ipv6/ndp_test.go index c9395de52..12b70f7e9 100644 --- a/pkg/tcpip/network/ipv6/ndp_test.go +++ b/pkg/tcpip/network/ipv6/ndp_test.go @@ -20,6 +20,7 @@ import ( "gvisor.dev/gvisor/pkg/tcpip" "gvisor.dev/gvisor/pkg/tcpip/buffer" + "gvisor.dev/gvisor/pkg/tcpip/checker" "gvisor.dev/gvisor/pkg/tcpip/header" "gvisor.dev/gvisor/pkg/tcpip/link/channel" "gvisor.dev/gvisor/pkg/tcpip/stack" @@ -135,7 +136,7 @@ func TestNeighorSolicitationWithSourceLinkLayerOption(t *testing.T) { t.Fatalf("got invalid = %d, want = 0", got) } - e.InjectInbound(ProtocolNumber, tcpip.PacketBuffer{ + e.InjectInbound(ProtocolNumber, stack.PacketBuffer{ Data: hdr.View().ToVectorisedView(), }) @@ -173,6 +174,257 @@ func TestNeighorSolicitationWithSourceLinkLayerOption(t *testing.T) { } } +func TestNeighorSolicitationResponse(t *testing.T) { + const nicID = 1 + nicAddr := lladdr0 + remoteAddr := lladdr1 + nicAddrSNMC := header.SolicitedNodeAddr(nicAddr) + nicLinkAddr := linkAddr0 + remoteLinkAddr0 := linkAddr1 + remoteLinkAddr1 := linkAddr2 + + tests := []struct { + name string + nsOpts header.NDPOptionsSerializer + nsSrcLinkAddr tcpip.LinkAddress + nsSrc tcpip.Address + nsDst tcpip.Address + nsInvalid bool + naDstLinkAddr tcpip.LinkAddress + naSolicited bool + naSrc tcpip.Address + naDst tcpip.Address + }{ + { + name: "Unspecified source to multicast destination", + nsOpts: nil, + nsSrcLinkAddr: remoteLinkAddr0, + nsSrc: header.IPv6Any, + nsDst: nicAddrSNMC, + nsInvalid: false, + naDstLinkAddr: remoteLinkAddr0, + naSolicited: false, + naSrc: nicAddr, + naDst: header.IPv6AllNodesMulticastAddress, + }, + { + name: "Unspecified source with source ll option to multicast destination", + nsOpts: header.NDPOptionsSerializer{ + header.NDPSourceLinkLayerAddressOption(remoteLinkAddr0[:]), + }, + nsSrcLinkAddr: remoteLinkAddr0, + nsSrc: header.IPv6Any, + nsDst: nicAddrSNMC, + nsInvalid: true, + }, + { + name: "Unspecified source to unicast destination", + nsOpts: nil, + nsSrcLinkAddr: remoteLinkAddr0, + nsSrc: header.IPv6Any, + nsDst: nicAddr, + nsInvalid: false, + naDstLinkAddr: remoteLinkAddr0, + naSolicited: false, + naSrc: nicAddr, + naDst: header.IPv6AllNodesMulticastAddress, + }, + { + name: "Unspecified source with source ll option to unicast destination", + nsOpts: header.NDPOptionsSerializer{ + header.NDPSourceLinkLayerAddressOption(remoteLinkAddr0[:]), + }, + nsSrcLinkAddr: remoteLinkAddr0, + nsSrc: header.IPv6Any, + nsDst: nicAddr, + nsInvalid: true, + }, + + { + name: "Specified source with 1 source ll to multicast destination", + nsOpts: header.NDPOptionsSerializer{ + header.NDPSourceLinkLayerAddressOption(remoteLinkAddr0[:]), + }, + nsSrcLinkAddr: remoteLinkAddr0, + nsSrc: remoteAddr, + nsDst: nicAddrSNMC, + nsInvalid: false, + naDstLinkAddr: remoteLinkAddr0, + naSolicited: true, + naSrc: nicAddr, + naDst: remoteAddr, + }, + { + name: "Specified source with 1 source ll different from route to multicast destination", + nsOpts: header.NDPOptionsSerializer{ + header.NDPSourceLinkLayerAddressOption(remoteLinkAddr1[:]), + }, + nsSrcLinkAddr: remoteLinkAddr0, + nsSrc: remoteAddr, + nsDst: nicAddrSNMC, + nsInvalid: false, + naDstLinkAddr: remoteLinkAddr1, + naSolicited: true, + naSrc: nicAddr, + naDst: remoteAddr, + }, + { + name: "Specified source to multicast destination", + nsOpts: nil, + nsSrcLinkAddr: remoteLinkAddr0, + nsSrc: remoteAddr, + nsDst: nicAddrSNMC, + nsInvalid: true, + }, + { + name: "Specified source with 2 source ll to multicast destination", + nsOpts: header.NDPOptionsSerializer{ + header.NDPSourceLinkLayerAddressOption(remoteLinkAddr0[:]), + header.NDPSourceLinkLayerAddressOption(remoteLinkAddr1[:]), + }, + nsSrcLinkAddr: remoteLinkAddr0, + nsSrc: remoteAddr, + nsDst: nicAddrSNMC, + nsInvalid: true, + }, + + { + name: "Specified source to unicast destination", + nsOpts: nil, + nsSrcLinkAddr: remoteLinkAddr0, + nsSrc: remoteAddr, + nsDst: nicAddr, + nsInvalid: false, + naDstLinkAddr: remoteLinkAddr0, + naSolicited: true, + naSrc: nicAddr, + naDst: remoteAddr, + }, + { + name: "Specified source with 1 source ll to unicast destination", + nsOpts: header.NDPOptionsSerializer{ + header.NDPSourceLinkLayerAddressOption(remoteLinkAddr0[:]), + }, + nsSrcLinkAddr: remoteLinkAddr0, + nsSrc: remoteAddr, + nsDst: nicAddr, + nsInvalid: false, + naDstLinkAddr: remoteLinkAddr0, + naSolicited: true, + naSrc: nicAddr, + naDst: remoteAddr, + }, + { + name: "Specified source with 1 source ll different from route to unicast destination", + nsOpts: header.NDPOptionsSerializer{ + header.NDPSourceLinkLayerAddressOption(remoteLinkAddr1[:]), + }, + nsSrcLinkAddr: remoteLinkAddr0, + nsSrc: remoteAddr, + nsDst: nicAddr, + nsInvalid: false, + naDstLinkAddr: remoteLinkAddr1, + naSolicited: true, + naSrc: nicAddr, + naDst: remoteAddr, + }, + { + name: "Specified source with 2 source ll to unicast destination", + nsOpts: header.NDPOptionsSerializer{ + header.NDPSourceLinkLayerAddressOption(remoteLinkAddr0[:]), + header.NDPSourceLinkLayerAddressOption(remoteLinkAddr1[:]), + }, + nsSrcLinkAddr: remoteLinkAddr0, + nsSrc: remoteAddr, + nsDst: nicAddr, + nsInvalid: true, + }, + } + + for _, test := range tests { + t.Run(test.name, func(t *testing.T) { + s := stack.New(stack.Options{ + NetworkProtocols: []stack.NetworkProtocol{NewProtocol()}, + }) + e := channel.New(1, 1280, nicLinkAddr) + if err := s.CreateNIC(nicID, e); err != nil { + t.Fatalf("CreateNIC(%d, _) = %s", nicID, err) + } + if err := s.AddAddress(nicID, ProtocolNumber, nicAddr); err != nil { + t.Fatalf("AddAddress(%d, %d, %s) = %s", nicID, ProtocolNumber, nicAddr, err) + } + + ndpNSSize := header.ICMPv6NeighborSolicitMinimumSize + test.nsOpts.Length() + hdr := buffer.NewPrependable(header.IPv6MinimumSize + ndpNSSize) + pkt := header.ICMPv6(hdr.Prepend(ndpNSSize)) + pkt.SetType(header.ICMPv6NeighborSolicit) + ns := header.NDPNeighborSolicit(pkt.NDPPayload()) + ns.SetTargetAddress(nicAddr) + opts := ns.Options() + opts.Serialize(test.nsOpts) + pkt.SetChecksum(header.ICMPv6Checksum(pkt, test.nsSrc, test.nsDst, buffer.VectorisedView{})) + payloadLength := hdr.UsedLength() + ip := header.IPv6(hdr.Prepend(header.IPv6MinimumSize)) + ip.Encode(&header.IPv6Fields{ + PayloadLength: uint16(payloadLength), + NextHeader: uint8(header.ICMPv6ProtocolNumber), + HopLimit: 255, + SrcAddr: test.nsSrc, + DstAddr: test.nsDst, + }) + + invalid := s.Stats().ICMP.V6PacketsReceived.Invalid + + // Invalid count should initially be 0. + if got := invalid.Value(); got != 0 { + t.Fatalf("got invalid = %d, want = 0", got) + } + + e.InjectLinkAddr(ProtocolNumber, test.nsSrcLinkAddr, stack.PacketBuffer{ + Data: hdr.View().ToVectorisedView(), + }) + + if test.nsInvalid { + if got := invalid.Value(); got != 1 { + t.Fatalf("got invalid = %d, want = 1", got) + } + + if p, got := e.Read(); got { + t.Fatalf("unexpected response to an invalid NS = %+v", p.Pkt) + } + + // If we expected the NS to be invalid, we have nothing else to check. + return + } + + if got := invalid.Value(); got != 0 { + t.Fatalf("got invalid = %d, want = 0", got) + } + + p, got := e.Read() + if !got { + t.Fatal("expected an NDP NA response") + } + + if p.Route.RemoteLinkAddress != test.naDstLinkAddr { + t.Errorf("got p.Route.RemoteLinkAddress = %s, want = %s", p.Route.RemoteLinkAddress, test.naDstLinkAddr) + } + + checker.IPv6(t, p.Pkt.Header.View(), + checker.SrcAddr(test.naSrc), + checker.DstAddr(test.naDst), + checker.TTL(header.NDPHopLimit), + checker.NDPNA( + checker.NDPNASolicitedFlag(test.naSolicited), + checker.NDPNATargetAddress(nicAddr), + checker.NDPNAOptions([]header.NDPOption{ + header.NDPTargetLinkLayerAddressOption(nicLinkAddr[:]), + }), + )) + }) + } +} + // TestNeighorAdvertisementWithTargetLinkLayerOption tests that receiving a // valid NDP NA message with the Target Link Layer Address option results in a // new entry in the link address cache for the target of the message. @@ -197,6 +449,13 @@ func TestNeighorAdvertisementWithTargetLinkLayerOption(t *testing.T) { name: "Invalid Length", optsBuf: []byte{2, 2, 2, 3, 4, 5, 6, 7}, }, + { + name: "Multiple", + optsBuf: []byte{ + 2, 1, 2, 3, 4, 5, 6, 7, + 2, 1, 2, 3, 4, 5, 6, 8, + }, + }, } for _, test := range tests { @@ -238,7 +497,7 @@ func TestNeighorAdvertisementWithTargetLinkLayerOption(t *testing.T) { t.Fatalf("got invalid = %d, want = 0", got) } - e.InjectInbound(ProtocolNumber, tcpip.PacketBuffer{ + e.InjectInbound(ProtocolNumber, stack.PacketBuffer{ Data: hdr.View().ToVectorisedView(), }) @@ -276,9 +535,7 @@ func TestNeighorAdvertisementWithTargetLinkLayerOption(t *testing.T) { } } -// TestHopLimitValidation is a test that makes sure that NDP packets are only -// received if their IP header's hop limit is set to 255. -func TestHopLimitValidation(t *testing.T) { +func TestNDPValidation(t *testing.T) { setup := func(t *testing.T) (*stack.Stack, stack.NetworkEndpoint, stack.Route) { t.Helper() @@ -294,17 +551,24 @@ func TestHopLimitValidation(t *testing.T) { return s, ep, r } - handleIPv6Payload := func(hdr buffer.Prependable, hopLimit uint8, ep stack.NetworkEndpoint, r *stack.Route) { + handleIPv6Payload := func(hdr buffer.Prependable, hopLimit uint8, atomicFragment bool, ep stack.NetworkEndpoint, r *stack.Route) { + nextHdr := uint8(header.ICMPv6ProtocolNumber) + if atomicFragment { + bytes := hdr.Prepend(header.IPv6FragmentExtHdrLength) + bytes[0] = nextHdr + nextHdr = uint8(header.IPv6FragmentExtHdrIdentifier) + } + payloadLength := hdr.UsedLength() ip := header.IPv6(hdr.Prepend(header.IPv6MinimumSize)) ip.Encode(&header.IPv6Fields{ PayloadLength: uint16(payloadLength), - NextHeader: uint8(header.ICMPv6ProtocolNumber), + NextHeader: nextHdr, HopLimit: hopLimit, SrcAddr: r.LocalAddress, DstAddr: r.RemoteAddress, }) - ep.HandlePacket(r, tcpip.PacketBuffer{ + ep.HandlePacket(r, stack.PacketBuffer{ Data: hdr.View().ToVectorisedView(), }) } @@ -364,61 +628,93 @@ func TestHopLimitValidation(t *testing.T) { }, } + subTests := []struct { + name string + atomicFragment bool + hopLimit uint8 + code uint8 + valid bool + }{ + { + name: "Valid", + atomicFragment: false, + hopLimit: header.NDPHopLimit, + code: 0, + valid: true, + }, + { + name: "Fragmented", + atomicFragment: true, + hopLimit: header.NDPHopLimit, + code: 0, + valid: false, + }, + { + name: "Invalid hop limit", + atomicFragment: false, + hopLimit: header.NDPHopLimit - 1, + code: 0, + valid: false, + }, + { + name: "Invalid ICMPv6 code", + atomicFragment: false, + hopLimit: header.NDPHopLimit, + code: 1, + valid: false, + }, + } + for _, typ := range types { t.Run(typ.name, func(t *testing.T) { - s, ep, r := setup(t) - defer r.Release() - - stats := s.Stats().ICMP.V6PacketsReceived - invalid := stats.Invalid - typStat := typ.statCounter(stats) - - extraDataLen := len(typ.extraData) - hdr := buffer.NewPrependable(header.IPv6MinimumSize + typ.size + extraDataLen) - extraData := buffer.View(hdr.Prepend(extraDataLen)) - copy(extraData, typ.extraData) - pkt := header.ICMPv6(hdr.Prepend(typ.size)) - pkt.SetType(typ.typ) - pkt.SetChecksum(header.ICMPv6Checksum(pkt, r.LocalAddress, r.RemoteAddress, extraData.ToVectorisedView())) - - // Invalid count should initially be 0. - if got := invalid.Value(); got != 0 { - t.Fatalf("got invalid = %d, want = 0", got) - } - - // Should not have received any ICMPv6 packets with - // type = typ.typ. - if got := typStat.Value(); got != 0 { - t.Fatalf("got %s = %d, want = 0", typ.name, got) - } - - // Receive the NDP packet with an invalid hop limit - // value. - handleIPv6Payload(hdr, header.NDPHopLimit-1, ep, &r) - - // Invalid count should have increased. - if got := invalid.Value(); got != 1 { - t.Fatalf("got invalid = %d, want = 1", got) - } - - // Rx count of NDP packet of type typ.typ should not - // have increased. - if got := typStat.Value(); got != 0 { - t.Fatalf("got %s = %d, want = 0", typ.name, got) - } - - // Receive the NDP packet with a valid hop limit value. - handleIPv6Payload(hdr, header.NDPHopLimit, ep, &r) - - // Rx count of NDP packet of type typ.typ should have - // increased. - if got := typStat.Value(); got != 1 { - t.Fatalf("got %s = %d, want = 1", typ.name, got) - } - - // Invalid count should not have increased again. - if got := invalid.Value(); got != 1 { - t.Fatalf("got invalid = %d, want = 1", got) + for _, test := range subTests { + t.Run(test.name, func(t *testing.T) { + s, ep, r := setup(t) + defer r.Release() + + stats := s.Stats().ICMP.V6PacketsReceived + invalid := stats.Invalid + typStat := typ.statCounter(stats) + + extraDataLen := len(typ.extraData) + hdr := buffer.NewPrependable(header.IPv6MinimumSize + typ.size + extraDataLen + header.IPv6FragmentExtHdrLength) + extraData := buffer.View(hdr.Prepend(extraDataLen)) + copy(extraData, typ.extraData) + pkt := header.ICMPv6(hdr.Prepend(typ.size)) + pkt.SetType(typ.typ) + pkt.SetCode(test.code) + pkt.SetChecksum(header.ICMPv6Checksum(pkt, r.LocalAddress, r.RemoteAddress, extraData.ToVectorisedView())) + + // Rx count of the NDP message should initially be 0. + if got := typStat.Value(); got != 0 { + t.Errorf("got %s = %d, want = 0", typ.name, got) + } + + // Invalid count should initially be 0. + if got := invalid.Value(); got != 0 { + t.Errorf("got invalid = %d, want = 0", got) + } + + if t.Failed() { + t.FailNow() + } + + handleIPv6Payload(hdr, test.hopLimit, test.atomicFragment, ep, &r) + + // Rx count of the NDP packet should have increased. + if got := typStat.Value(); got != 1 { + t.Errorf("got %s = %d, want = 1", typ.name, got) + } + + want := uint64(0) + if !test.valid { + // Invalid count should have increased. + want = 1 + } + if got := invalid.Value(); got != want { + t.Errorf("got invalid = %d, want = %d", got, want) + } + }) } }) } @@ -588,25 +884,22 @@ func TestRouterAdvertValidation(t *testing.T) { t.Fatalf("got rxRA = %d, want = 0", got) } - e.InjectInbound(header.IPv6ProtocolNumber, tcpip.PacketBuffer{ + e.InjectInbound(header.IPv6ProtocolNumber, stack.PacketBuffer{ Data: hdr.View().ToVectorisedView(), }) + if got := rxRA.Value(); got != 1 { + t.Fatalf("got rxRA = %d, want = 1", got) + } + if test.expectedSuccess { if got := invalid.Value(); got != 0 { t.Fatalf("got invalid = %d, want = 0", got) } - if got := rxRA.Value(); got != 1 { - t.Fatalf("got rxRA = %d, want = 1", got) - } - } else { if got := invalid.Value(); got != 1 { t.Fatalf("got invalid = %d, want = 1", got) } - if got := rxRA.Value(); got != 0 { - t.Fatalf("got rxRA = %d, want = 0", got) - } } }) } diff --git a/pkg/tcpip/seqnum/seqnum.go b/pkg/tcpip/seqnum/seqnum.go index b40a3c212..d3bea7de4 100644 --- a/pkg/tcpip/seqnum/seqnum.go +++ b/pkg/tcpip/seqnum/seqnum.go @@ -46,11 +46,6 @@ func (v Value) InWindow(first Value, size Size) bool { return v.InRange(first, first.Add(size)) } -// Overlap checks if the window [a,a+b) overlaps with the window [x, x+y). -func Overlap(a Value, b Size, x Value, y Size) bool { - return a.LessThan(x.Add(y)) && x.LessThan(a.Add(b)) -} - // Add calculates the sequence number following the [v, v+s) window. func (v Value) Add(s Size) Value { return v + Value(s) diff --git a/pkg/tcpip/stack/BUILD b/pkg/tcpip/stack/BUILD index 705cf01ee..5e963a4af 100644 --- a/pkg/tcpip/stack/BUILD +++ b/pkg/tcpip/stack/BUILD @@ -15,14 +15,34 @@ go_template_instance( }, ) +go_template_instance( + name = "packet_buffer_list", + out = "packet_buffer_list.go", + package = "stack", + prefix = "PacketBuffer", + template = "//pkg/ilist:generic_list", + types = { + "Element": "*PacketBuffer", + "Linker": "*PacketBuffer", + }, +) + go_library( name = "stack", srcs = [ + "dhcpv6configurationfromndpra_string.go", + "forwarder.go", "icmp_rate_limit.go", + "iptables.go", + "iptables_targets.go", + "iptables_types.go", "linkaddrcache.go", "linkaddrentry_list.go", "ndp.go", "nic.go", + "packet_buffer.go", + "packet_buffer_list.go", + "rand.go", "registration.go", "route.go", "stack.go", @@ -32,6 +52,7 @@ go_library( visibility = ["//visibility:public"], deps = [ "//pkg/ilist", + "//pkg/log", "//pkg/rand", "//pkg/sleep", "//pkg/sync", @@ -39,7 +60,6 @@ go_library( "//pkg/tcpip/buffer", "//pkg/tcpip/hash/jenkins", "//pkg/tcpip/header", - "//pkg/tcpip/iptables", "//pkg/tcpip/ports", "//pkg/tcpip/seqnum", "//pkg/waiter", @@ -63,7 +83,6 @@ go_test( "//pkg/tcpip/buffer", "//pkg/tcpip/checker", "//pkg/tcpip/header", - "//pkg/tcpip/iptables", "//pkg/tcpip/link/channel", "//pkg/tcpip/link/loopback", "//pkg/tcpip/network/ipv4", @@ -79,6 +98,7 @@ go_test( name = "stack_test", size = "small", srcs = [ + "forwarder_test.go", "linkaddrcache_test.go", "nic_test.go", ], diff --git a/pkg/tcpip/stack/dhcpv6configurationfromndpra_string.go b/pkg/tcpip/stack/dhcpv6configurationfromndpra_string.go new file mode 100644 index 000000000..8b4213eec --- /dev/null +++ b/pkg/tcpip/stack/dhcpv6configurationfromndpra_string.go @@ -0,0 +1,39 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// Code generated by "stringer -type=DHCPv6ConfigurationFromNDPRA"; DO NOT EDIT. + +package stack + +import "strconv" + +func _() { + // An "invalid array index" compiler error signifies that the constant values have changed. + // Re-run the stringer command to generate them again. + var x [1]struct{} + _ = x[DHCPv6NoConfiguration-0] + _ = x[DHCPv6ManagedAddress-1] + _ = x[DHCPv6OtherConfigurations-2] +} + +const _DHCPv6ConfigurationFromNDPRA_name = "DHCPv6NoConfigurationDHCPv6ManagedAddressDHCPv6OtherConfigurations" + +var _DHCPv6ConfigurationFromNDPRA_index = [...]uint8{0, 21, 41, 66} + +func (i DHCPv6ConfigurationFromNDPRA) String() string { + if i < 0 || i >= DHCPv6ConfigurationFromNDPRA(len(_DHCPv6ConfigurationFromNDPRA_index)-1) { + return "DHCPv6ConfigurationFromNDPRA(" + strconv.FormatInt(int64(i), 10) + ")" + } + return _DHCPv6ConfigurationFromNDPRA_name[_DHCPv6ConfigurationFromNDPRA_index[i]:_DHCPv6ConfigurationFromNDPRA_index[i+1]] +} diff --git a/pkg/tcpip/stack/forwarder.go b/pkg/tcpip/stack/forwarder.go new file mode 100644 index 000000000..6b64cd37f --- /dev/null +++ b/pkg/tcpip/stack/forwarder.go @@ -0,0 +1,131 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package stack + +import ( + "fmt" + + "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/pkg/tcpip" +) + +const ( + // maxPendingResolutions is the maximum number of pending link-address + // resolutions. + maxPendingResolutions = 64 + maxPendingPacketsPerResolution = 256 +) + +type pendingPacket struct { + nic *NIC + route *Route + proto tcpip.NetworkProtocolNumber + pkt PacketBuffer +} + +type forwardQueue struct { + sync.Mutex + + // The packets to send once the resolver completes. + packets map[<-chan struct{}][]*pendingPacket + + // FIFO of channels used to cancel the oldest goroutine waiting for + // link-address resolution. + cancelChans []chan struct{} +} + +func newForwardQueue() *forwardQueue { + return &forwardQueue{packets: make(map[<-chan struct{}][]*pendingPacket)} +} + +func (f *forwardQueue) enqueue(ch <-chan struct{}, n *NIC, r *Route, protocol tcpip.NetworkProtocolNumber, pkt PacketBuffer) { + shouldWait := false + + f.Lock() + packets, ok := f.packets[ch] + if !ok { + shouldWait = true + } + for len(packets) == maxPendingPacketsPerResolution { + p := packets[0] + packets = packets[1:] + p.nic.stack.stats.IP.OutgoingPacketErrors.Increment() + p.route.Release() + } + if l := len(packets); l >= maxPendingPacketsPerResolution { + panic(fmt.Sprintf("max pending packets for resolution reached; got %d packets, max = %d", l, maxPendingPacketsPerResolution)) + } + f.packets[ch] = append(packets, &pendingPacket{ + nic: n, + route: r, + proto: protocol, + pkt: pkt, + }) + f.Unlock() + + if !shouldWait { + return + } + + // Wait for the link-address resolution to complete. + // Start a goroutine with a forwarding-cancel channel so that we can + // limit the maximum number of goroutines running concurrently. + cancel := f.newCancelChannel() + go func() { + cancelled := false + select { + case <-ch: + case <-cancel: + cancelled = true + } + + f.Lock() + packets := f.packets[ch] + delete(f.packets, ch) + f.Unlock() + + for _, p := range packets { + if cancelled { + p.nic.stack.stats.IP.OutgoingPacketErrors.Increment() + } else if _, err := p.route.Resolve(nil); err != nil { + p.nic.stack.stats.IP.OutgoingPacketErrors.Increment() + } else { + p.nic.forwardPacket(p.route, p.proto, p.pkt) + } + p.route.Release() + } + }() +} + +// newCancelChannel creates a channel that can cancel a pending forwarding +// activity. The oldest channel is closed if the number of open channels would +// exceed maxPendingResolutions. +func (f *forwardQueue) newCancelChannel() chan struct{} { + f.Lock() + defer f.Unlock() + + if len(f.cancelChans) == maxPendingResolutions { + ch := f.cancelChans[0] + f.cancelChans = f.cancelChans[1:] + close(ch) + } + if l := len(f.cancelChans); l >= maxPendingResolutions { + panic(fmt.Sprintf("max pending resolutions reached; got %d active resolutions, max = %d", l, maxPendingResolutions)) + } + + ch := make(chan struct{}) + f.cancelChans = append(f.cancelChans, ch) + return ch +} diff --git a/pkg/tcpip/stack/forwarder_test.go b/pkg/tcpip/stack/forwarder_test.go new file mode 100644 index 000000000..e9c652042 --- /dev/null +++ b/pkg/tcpip/stack/forwarder_test.go @@ -0,0 +1,635 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package stack + +import ( + "encoding/binary" + "math" + "testing" + "time" + + "gvisor.dev/gvisor/pkg/tcpip" + "gvisor.dev/gvisor/pkg/tcpip/buffer" +) + +const ( + fwdTestNetNumber tcpip.NetworkProtocolNumber = math.MaxUint32 + fwdTestNetHeaderLen = 12 + fwdTestNetDefaultPrefixLen = 8 + + // fwdTestNetDefaultMTU is the MTU, in bytes, used throughout the tests, + // except where another value is explicitly used. It is chosen to match + // the MTU of loopback interfaces on linux systems. + fwdTestNetDefaultMTU = 65536 +) + +// fwdTestNetworkEndpoint is a network-layer protocol endpoint. +// Headers of this protocol are fwdTestNetHeaderLen bytes, but we currently only +// use the first three: destination address, source address, and transport +// protocol. They're all one byte fields to simplify parsing. +type fwdTestNetworkEndpoint struct { + nicID tcpip.NICID + id NetworkEndpointID + prefixLen int + proto *fwdTestNetworkProtocol + dispatcher TransportDispatcher + ep LinkEndpoint +} + +func (f *fwdTestNetworkEndpoint) MTU() uint32 { + return f.ep.MTU() - uint32(f.MaxHeaderLength()) +} + +func (f *fwdTestNetworkEndpoint) NICID() tcpip.NICID { + return f.nicID +} + +func (f *fwdTestNetworkEndpoint) PrefixLen() int { + return f.prefixLen +} + +func (*fwdTestNetworkEndpoint) DefaultTTL() uint8 { + return 123 +} + +func (f *fwdTestNetworkEndpoint) ID() *NetworkEndpointID { + return &f.id +} + +func (f *fwdTestNetworkEndpoint) HandlePacket(r *Route, pkt PacketBuffer) { + // Consume the network header. + b := pkt.Data.First() + pkt.Data.TrimFront(fwdTestNetHeaderLen) + + // Dispatch the packet to the transport protocol. + f.dispatcher.DeliverTransportPacket(r, tcpip.TransportProtocolNumber(b[2]), pkt) +} + +func (f *fwdTestNetworkEndpoint) MaxHeaderLength() uint16 { + return f.ep.MaxHeaderLength() + fwdTestNetHeaderLen +} + +func (f *fwdTestNetworkEndpoint) PseudoHeaderChecksum(protocol tcpip.TransportProtocolNumber, dstAddr tcpip.Address) uint16 { + return 0 +} + +func (f *fwdTestNetworkEndpoint) Capabilities() LinkEndpointCapabilities { + return f.ep.Capabilities() +} + +func (f *fwdTestNetworkEndpoint) WritePacket(r *Route, gso *GSO, params NetworkHeaderParams, pkt PacketBuffer) *tcpip.Error { + // Add the protocol's header to the packet and send it to the link + // endpoint. + b := pkt.Header.Prepend(fwdTestNetHeaderLen) + b[0] = r.RemoteAddress[0] + b[1] = f.id.LocalAddress[0] + b[2] = byte(params.Protocol) + + return f.ep.WritePacket(r, gso, fwdTestNetNumber, pkt) +} + +// WritePackets implements LinkEndpoint.WritePackets. +func (f *fwdTestNetworkEndpoint) WritePackets(r *Route, gso *GSO, pkts PacketBufferList, params NetworkHeaderParams) (int, *tcpip.Error) { + panic("not implemented") +} + +func (*fwdTestNetworkEndpoint) WriteHeaderIncludedPacket(r *Route, pkt PacketBuffer) *tcpip.Error { + return tcpip.ErrNotSupported +} + +func (*fwdTestNetworkEndpoint) Close() {} + +// fwdTestNetworkProtocol is a network-layer protocol that implements Address +// resolution. +type fwdTestNetworkProtocol struct { + addrCache *linkAddrCache + addrResolveDelay time.Duration + onLinkAddressResolved func(cache *linkAddrCache, addr tcpip.Address) + onResolveStaticAddress func(tcpip.Address) (tcpip.LinkAddress, bool) +} + +func (f *fwdTestNetworkProtocol) Number() tcpip.NetworkProtocolNumber { + return fwdTestNetNumber +} + +func (f *fwdTestNetworkProtocol) MinimumPacketSize() int { + return fwdTestNetHeaderLen +} + +func (f *fwdTestNetworkProtocol) DefaultPrefixLen() int { + return fwdTestNetDefaultPrefixLen +} + +func (*fwdTestNetworkProtocol) ParseAddresses(v buffer.View) (src, dst tcpip.Address) { + return tcpip.Address(v[1:2]), tcpip.Address(v[0:1]) +} + +func (f *fwdTestNetworkProtocol) NewEndpoint(nicID tcpip.NICID, addrWithPrefix tcpip.AddressWithPrefix, linkAddrCache LinkAddressCache, dispatcher TransportDispatcher, ep LinkEndpoint, _ *Stack) (NetworkEndpoint, *tcpip.Error) { + return &fwdTestNetworkEndpoint{ + nicID: nicID, + id: NetworkEndpointID{LocalAddress: addrWithPrefix.Address}, + prefixLen: addrWithPrefix.PrefixLen, + proto: f, + dispatcher: dispatcher, + ep: ep, + }, nil +} + +func (f *fwdTestNetworkProtocol) SetOption(option interface{}) *tcpip.Error { + return tcpip.ErrUnknownProtocolOption +} + +func (f *fwdTestNetworkProtocol) Option(option interface{}) *tcpip.Error { + return tcpip.ErrUnknownProtocolOption +} + +func (f *fwdTestNetworkProtocol) Close() {} + +func (f *fwdTestNetworkProtocol) Wait() {} + +func (f *fwdTestNetworkProtocol) LinkAddressRequest(addr, localAddr tcpip.Address, linkEP LinkEndpoint) *tcpip.Error { + if f.addrCache != nil && f.onLinkAddressResolved != nil { + time.AfterFunc(f.addrResolveDelay, func() { + f.onLinkAddressResolved(f.addrCache, addr) + }) + } + return nil +} + +func (f *fwdTestNetworkProtocol) ResolveStaticAddress(addr tcpip.Address) (tcpip.LinkAddress, bool) { + if f.onResolveStaticAddress != nil { + return f.onResolveStaticAddress(addr) + } + return "", false +} + +func (f *fwdTestNetworkProtocol) LinkAddressProtocol() tcpip.NetworkProtocolNumber { + return fwdTestNetNumber +} + +// fwdTestPacketInfo holds all the information about an outbound packet. +type fwdTestPacketInfo struct { + RemoteLinkAddress tcpip.LinkAddress + LocalLinkAddress tcpip.LinkAddress + Pkt PacketBuffer +} + +type fwdTestLinkEndpoint struct { + dispatcher NetworkDispatcher + mtu uint32 + linkAddr tcpip.LinkAddress + + // C is where outbound packets are queued. + C chan fwdTestPacketInfo +} + +// InjectInbound injects an inbound packet. +func (e *fwdTestLinkEndpoint) InjectInbound(protocol tcpip.NetworkProtocolNumber, pkt PacketBuffer) { + e.InjectLinkAddr(protocol, "", pkt) +} + +// InjectLinkAddr injects an inbound packet with a remote link address. +func (e *fwdTestLinkEndpoint) InjectLinkAddr(protocol tcpip.NetworkProtocolNumber, remote tcpip.LinkAddress, pkt PacketBuffer) { + e.dispatcher.DeliverNetworkPacket(e, remote, "" /* local */, protocol, pkt) +} + +// Attach saves the stack network-layer dispatcher for use later when packets +// are injected. +func (e *fwdTestLinkEndpoint) Attach(dispatcher NetworkDispatcher) { + e.dispatcher = dispatcher +} + +// IsAttached implements stack.LinkEndpoint.IsAttached. +func (e *fwdTestLinkEndpoint) IsAttached() bool { + return e.dispatcher != nil +} + +// MTU implements stack.LinkEndpoint.MTU. It returns the value initialized +// during construction. +func (e *fwdTestLinkEndpoint) MTU() uint32 { + return e.mtu +} + +// Capabilities implements stack.LinkEndpoint.Capabilities. +func (e fwdTestLinkEndpoint) Capabilities() LinkEndpointCapabilities { + caps := LinkEndpointCapabilities(0) + return caps | CapabilityResolutionRequired +} + +// GSOMaxSize returns the maximum GSO packet size. +func (*fwdTestLinkEndpoint) GSOMaxSize() uint32 { + return 1 << 15 +} + +// MaxHeaderLength returns the maximum size of the link layer header. Given it +// doesn't have a header, it just returns 0. +func (*fwdTestLinkEndpoint) MaxHeaderLength() uint16 { + return 0 +} + +// LinkAddress returns the link address of this endpoint. +func (e *fwdTestLinkEndpoint) LinkAddress() tcpip.LinkAddress { + return e.linkAddr +} + +func (e fwdTestLinkEndpoint) WritePacket(r *Route, gso *GSO, protocol tcpip.NetworkProtocolNumber, pkt PacketBuffer) *tcpip.Error { + p := fwdTestPacketInfo{ + RemoteLinkAddress: r.RemoteLinkAddress, + LocalLinkAddress: r.LocalLinkAddress, + Pkt: pkt, + } + + select { + case e.C <- p: + default: + } + + return nil +} + +// WritePackets stores outbound packets into the channel. +func (e *fwdTestLinkEndpoint) WritePackets(r *Route, gso *GSO, pkts PacketBufferList, protocol tcpip.NetworkProtocolNumber) (int, *tcpip.Error) { + n := 0 + for pkt := pkts.Front(); pkt != nil; pkt = pkt.Next() { + e.WritePacket(r, gso, protocol, *pkt) + n++ + } + + return n, nil +} + +// WriteRawPacket implements stack.LinkEndpoint.WriteRawPacket. +func (e *fwdTestLinkEndpoint) WriteRawPacket(vv buffer.VectorisedView) *tcpip.Error { + p := fwdTestPacketInfo{ + Pkt: PacketBuffer{Data: vv}, + } + + select { + case e.C <- p: + default: + } + + return nil +} + +// Wait implements stack.LinkEndpoint.Wait. +func (*fwdTestLinkEndpoint) Wait() {} + +func fwdTestNetFactory(t *testing.T, proto *fwdTestNetworkProtocol) (ep1, ep2 *fwdTestLinkEndpoint) { + // Create a stack with the network protocol and two NICs. + s := New(Options{ + NetworkProtocols: []NetworkProtocol{proto}, + }) + + proto.addrCache = s.linkAddrCache + + // Enable forwarding. + s.SetForwarding(true) + + // NIC 1 has the link address "a", and added the network address 1. + ep1 = &fwdTestLinkEndpoint{ + C: make(chan fwdTestPacketInfo, 300), + mtu: fwdTestNetDefaultMTU, + linkAddr: "a", + } + if err := s.CreateNIC(1, ep1); err != nil { + t.Fatal("CreateNIC #1 failed:", err) + } + if err := s.AddAddress(1, fwdTestNetNumber, "\x01"); err != nil { + t.Fatal("AddAddress #1 failed:", err) + } + + // NIC 2 has the link address "b", and added the network address 2. + ep2 = &fwdTestLinkEndpoint{ + C: make(chan fwdTestPacketInfo, 300), + mtu: fwdTestNetDefaultMTU, + linkAddr: "b", + } + if err := s.CreateNIC(2, ep2); err != nil { + t.Fatal("CreateNIC #2 failed:", err) + } + if err := s.AddAddress(2, fwdTestNetNumber, "\x02"); err != nil { + t.Fatal("AddAddress #2 failed:", err) + } + + // Route all packets to NIC 2. + { + subnet, err := tcpip.NewSubnet("\x00", "\x00") + if err != nil { + t.Fatal(err) + } + s.SetRouteTable([]tcpip.Route{{Destination: subnet, NIC: 2}}) + } + + return ep1, ep2 +} + +func TestForwardingWithStaticResolver(t *testing.T) { + // Create a network protocol with a static resolver. + proto := &fwdTestNetworkProtocol{ + onResolveStaticAddress: + // The network address 3 is resolved to the link address "c". + func(addr tcpip.Address) (tcpip.LinkAddress, bool) { + if addr == "\x03" { + return "c", true + } + return "", false + }, + } + + ep1, ep2 := fwdTestNetFactory(t, proto) + + // Inject an inbound packet to address 3 on NIC 1, and see if it is + // forwarded to NIC 2. + buf := buffer.NewView(30) + buf[0] = 3 + ep1.InjectInbound(fwdTestNetNumber, PacketBuffer{ + Data: buf.ToVectorisedView(), + }) + + var p fwdTestPacketInfo + + select { + case p = <-ep2.C: + default: + t.Fatal("packet not forwarded") + } + + // Test that the static address resolution happened correctly. + if p.RemoteLinkAddress != "c" { + t.Fatalf("got p.RemoteLinkAddress = %s, want = c", p.RemoteLinkAddress) + } + if p.LocalLinkAddress != "b" { + t.Fatalf("got p.LocalLinkAddress = %s, want = b", p.LocalLinkAddress) + } +} + +func TestForwardingWithFakeResolver(t *testing.T) { + // Create a network protocol with a fake resolver. + proto := &fwdTestNetworkProtocol{ + addrResolveDelay: 500 * time.Millisecond, + onLinkAddressResolved: func(cache *linkAddrCache, addr tcpip.Address) { + // Any address will be resolved to the link address "c". + cache.add(tcpip.FullAddress{NIC: 2, Addr: addr}, "c") + }, + } + + ep1, ep2 := fwdTestNetFactory(t, proto) + + // Inject an inbound packet to address 3 on NIC 1, and see if it is + // forwarded to NIC 2. + buf := buffer.NewView(30) + buf[0] = 3 + ep1.InjectInbound(fwdTestNetNumber, PacketBuffer{ + Data: buf.ToVectorisedView(), + }) + + var p fwdTestPacketInfo + + select { + case p = <-ep2.C: + case <-time.After(time.Second): + t.Fatal("packet not forwarded") + } + + // Test that the address resolution happened correctly. + if p.RemoteLinkAddress != "c" { + t.Fatalf("got p.RemoteLinkAddress = %s, want = c", p.RemoteLinkAddress) + } + if p.LocalLinkAddress != "b" { + t.Fatalf("got p.LocalLinkAddress = %s, want = b", p.LocalLinkAddress) + } +} + +func TestForwardingWithNoResolver(t *testing.T) { + // Create a network protocol without a resolver. + proto := &fwdTestNetworkProtocol{} + + ep1, ep2 := fwdTestNetFactory(t, proto) + + // inject an inbound packet to address 3 on NIC 1, and see if it is + // forwarded to NIC 2. + buf := buffer.NewView(30) + buf[0] = 3 + ep1.InjectInbound(fwdTestNetNumber, PacketBuffer{ + Data: buf.ToVectorisedView(), + }) + + select { + case <-ep2.C: + t.Fatal("Packet should not be forwarded") + case <-time.After(time.Second): + } +} + +func TestForwardingWithFakeResolverPartialTimeout(t *testing.T) { + // Create a network protocol with a fake resolver. + proto := &fwdTestNetworkProtocol{ + addrResolveDelay: 500 * time.Millisecond, + onLinkAddressResolved: func(cache *linkAddrCache, addr tcpip.Address) { + // Only packets to address 3 will be resolved to the + // link address "c". + if addr == "\x03" { + cache.add(tcpip.FullAddress{NIC: 2, Addr: addr}, "c") + } + }, + } + + ep1, ep2 := fwdTestNetFactory(t, proto) + + // Inject an inbound packet to address 4 on NIC 1. This packet should + // not be forwarded. + buf := buffer.NewView(30) + buf[0] = 4 + ep1.InjectInbound(fwdTestNetNumber, PacketBuffer{ + Data: buf.ToVectorisedView(), + }) + + // Inject an inbound packet to address 3 on NIC 1, and see if it is + // forwarded to NIC 2. + buf = buffer.NewView(30) + buf[0] = 3 + ep1.InjectInbound(fwdTestNetNumber, PacketBuffer{ + Data: buf.ToVectorisedView(), + }) + + var p fwdTestPacketInfo + + select { + case p = <-ep2.C: + case <-time.After(time.Second): + t.Fatal("packet not forwarded") + } + + b := p.Pkt.Header.View() + if b[0] != 3 { + t.Fatalf("got b[0] = %d, want = 3", b[0]) + } + + // Test that the address resolution happened correctly. + if p.RemoteLinkAddress != "c" { + t.Fatalf("got p.RemoteLinkAddress = %s, want = c", p.RemoteLinkAddress) + } + if p.LocalLinkAddress != "b" { + t.Fatalf("got p.LocalLinkAddress = %s, want = b", p.LocalLinkAddress) + } +} + +func TestForwardingWithFakeResolverTwoPackets(t *testing.T) { + // Create a network protocol with a fake resolver. + proto := &fwdTestNetworkProtocol{ + addrResolveDelay: 500 * time.Millisecond, + onLinkAddressResolved: func(cache *linkAddrCache, addr tcpip.Address) { + // Any packets will be resolved to the link address "c". + cache.add(tcpip.FullAddress{NIC: 2, Addr: addr}, "c") + }, + } + + ep1, ep2 := fwdTestNetFactory(t, proto) + + // Inject two inbound packets to address 3 on NIC 1. + for i := 0; i < 2; i++ { + buf := buffer.NewView(30) + buf[0] = 3 + ep1.InjectInbound(fwdTestNetNumber, PacketBuffer{ + Data: buf.ToVectorisedView(), + }) + } + + for i := 0; i < 2; i++ { + var p fwdTestPacketInfo + + select { + case p = <-ep2.C: + case <-time.After(time.Second): + t.Fatal("packet not forwarded") + } + + b := p.Pkt.Header.View() + if b[0] != 3 { + t.Fatalf("got b[0] = %d, want = 3", b[0]) + } + + // Test that the address resolution happened correctly. + if p.RemoteLinkAddress != "c" { + t.Fatalf("got p.RemoteLinkAddress = %s, want = c", p.RemoteLinkAddress) + } + if p.LocalLinkAddress != "b" { + t.Fatalf("got p.LocalLinkAddress = %s, want = b", p.LocalLinkAddress) + } + } +} + +func TestForwardingWithFakeResolverManyPackets(t *testing.T) { + // Create a network protocol with a fake resolver. + proto := &fwdTestNetworkProtocol{ + addrResolveDelay: 500 * time.Millisecond, + onLinkAddressResolved: func(cache *linkAddrCache, addr tcpip.Address) { + // Any packets will be resolved to the link address "c". + cache.add(tcpip.FullAddress{NIC: 2, Addr: addr}, "c") + }, + } + + ep1, ep2 := fwdTestNetFactory(t, proto) + + for i := 0; i < maxPendingPacketsPerResolution+5; i++ { + // Inject inbound 'maxPendingPacketsPerResolution + 5' packets on NIC 1. + buf := buffer.NewView(30) + buf[0] = 3 + // Set the packet sequence number. + binary.BigEndian.PutUint16(buf[fwdTestNetHeaderLen:], uint16(i)) + ep1.InjectInbound(fwdTestNetNumber, PacketBuffer{ + Data: buf.ToVectorisedView(), + }) + } + + for i := 0; i < maxPendingPacketsPerResolution; i++ { + var p fwdTestPacketInfo + + select { + case p = <-ep2.C: + case <-time.After(time.Second): + t.Fatal("packet not forwarded") + } + + b := p.Pkt.Header.View() + if b[0] != 3 { + t.Fatalf("got b[0] = %d, want = 3", b[0]) + } + // The first 5 packets should not be forwarded so the the + // sequemnce number should start with 5. + want := uint16(i + 5) + if n := binary.BigEndian.Uint16(b[fwdTestNetHeaderLen:]); n != want { + t.Fatalf("got the packet #%d, want = #%d", n, want) + } + + // Test that the address resolution happened correctly. + if p.RemoteLinkAddress != "c" { + t.Fatalf("got p.RemoteLinkAddress = %s, want = c", p.RemoteLinkAddress) + } + if p.LocalLinkAddress != "b" { + t.Fatalf("got p.LocalLinkAddress = %s, want = b", p.LocalLinkAddress) + } + } +} + +func TestForwardingWithFakeResolverManyResolutions(t *testing.T) { + // Create a network protocol with a fake resolver. + proto := &fwdTestNetworkProtocol{ + addrResolveDelay: 500 * time.Millisecond, + onLinkAddressResolved: func(cache *linkAddrCache, addr tcpip.Address) { + // Any packets will be resolved to the link address "c". + cache.add(tcpip.FullAddress{NIC: 2, Addr: addr}, "c") + }, + } + + ep1, ep2 := fwdTestNetFactory(t, proto) + + for i := 0; i < maxPendingResolutions+5; i++ { + // Inject inbound 'maxPendingResolutions + 5' packets on NIC 1. + // Each packet has a different destination address (3 to + // maxPendingResolutions + 7). + buf := buffer.NewView(30) + buf[0] = byte(3 + i) + ep1.InjectInbound(fwdTestNetNumber, PacketBuffer{ + Data: buf.ToVectorisedView(), + }) + } + + for i := 0; i < maxPendingResolutions; i++ { + var p fwdTestPacketInfo + + select { + case p = <-ep2.C: + case <-time.After(time.Second): + t.Fatal("packet not forwarded") + } + + // The first 5 packets (address 3 to 7) should not be forwarded + // because their address resolutions are interrupted. + b := p.Pkt.Header.View() + if b[0] < 8 { + t.Fatalf("got b[0] = %d, want b[0] >= 8", b[0]) + } + + // Test that the address resolution happened correctly. + if p.RemoteLinkAddress != "c" { + t.Fatalf("got p.RemoteLinkAddress = %s, want = c", p.RemoteLinkAddress) + } + if p.LocalLinkAddress != "b" { + t.Fatalf("got p.LocalLinkAddress = %s, want = b", p.LocalLinkAddress) + } + } +} diff --git a/pkg/tcpip/stack/iptables.go b/pkg/tcpip/stack/iptables.go new file mode 100644 index 000000000..6c0a4b24d --- /dev/null +++ b/pkg/tcpip/stack/iptables.go @@ -0,0 +1,328 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package stack + +import ( + "fmt" + + "gvisor.dev/gvisor/pkg/tcpip/header" +) + +// Table names. +const ( + TablenameNat = "nat" + TablenameMangle = "mangle" + TablenameFilter = "filter" +) + +// Chain names as defined by net/ipv4/netfilter/ip_tables.c. +const ( + ChainNamePrerouting = "PREROUTING" + ChainNameInput = "INPUT" + ChainNameForward = "FORWARD" + ChainNameOutput = "OUTPUT" + ChainNamePostrouting = "POSTROUTING" +) + +// HookUnset indicates that there is no hook set for an entrypoint or +// underflow. +const HookUnset = -1 + +// DefaultTables returns a default set of tables. Each chain is set to accept +// all packets. +func DefaultTables() IPTables { + // TODO(gvisor.dev/issue/170): We may be able to swap out some strings for + // iotas. + return IPTables{ + Tables: map[string]Table{ + TablenameNat: Table{ + Rules: []Rule{ + Rule{Target: AcceptTarget{}}, + Rule{Target: AcceptTarget{}}, + Rule{Target: AcceptTarget{}}, + Rule{Target: AcceptTarget{}}, + Rule{Target: ErrorTarget{}}, + }, + BuiltinChains: map[Hook]int{ + Prerouting: 0, + Input: 1, + Output: 2, + Postrouting: 3, + }, + Underflows: map[Hook]int{ + Prerouting: 0, + Input: 1, + Output: 2, + Postrouting: 3, + }, + UserChains: map[string]int{}, + }, + TablenameMangle: Table{ + Rules: []Rule{ + Rule{Target: AcceptTarget{}}, + Rule{Target: AcceptTarget{}}, + Rule{Target: ErrorTarget{}}, + }, + BuiltinChains: map[Hook]int{ + Prerouting: 0, + Output: 1, + }, + Underflows: map[Hook]int{ + Prerouting: 0, + Output: 1, + }, + UserChains: map[string]int{}, + }, + TablenameFilter: Table{ + Rules: []Rule{ + Rule{Target: AcceptTarget{}}, + Rule{Target: AcceptTarget{}}, + Rule{Target: AcceptTarget{}}, + Rule{Target: ErrorTarget{}}, + }, + BuiltinChains: map[Hook]int{ + Input: 0, + Forward: 1, + Output: 2, + }, + Underflows: map[Hook]int{ + Input: 0, + Forward: 1, + Output: 2, + }, + UserChains: map[string]int{}, + }, + }, + Priorities: map[Hook][]string{ + Input: []string{TablenameNat, TablenameFilter}, + Prerouting: []string{TablenameMangle, TablenameNat}, + Output: []string{TablenameMangle, TablenameNat, TablenameFilter}, + }, + } +} + +// EmptyFilterTable returns a Table with no rules and the filter table chains +// mapped to HookUnset. +func EmptyFilterTable() Table { + return Table{ + Rules: []Rule{}, + BuiltinChains: map[Hook]int{ + Input: HookUnset, + Forward: HookUnset, + Output: HookUnset, + }, + Underflows: map[Hook]int{ + Input: HookUnset, + Forward: HookUnset, + Output: HookUnset, + }, + UserChains: map[string]int{}, + } +} + +// EmptyNatTable returns a Table with no rules and the filter table chains +// mapped to HookUnset. +func EmptyNatTable() Table { + return Table{ + Rules: []Rule{}, + BuiltinChains: map[Hook]int{ + Prerouting: HookUnset, + Input: HookUnset, + Output: HookUnset, + Postrouting: HookUnset, + }, + Underflows: map[Hook]int{ + Prerouting: HookUnset, + Input: HookUnset, + Output: HookUnset, + Postrouting: HookUnset, + }, + UserChains: map[string]int{}, + } +} + +// A chainVerdict is what a table decides should be done with a packet. +type chainVerdict int + +const ( + // chainAccept indicates the packet should continue through netstack. + chainAccept chainVerdict = iota + + // chainAccept indicates the packet should be dropped. + chainDrop + + // chainReturn indicates the packet should return to the calling chain + // or the underflow rule of a builtin chain. + chainReturn +) + +// Check runs pkt through the rules for hook. It returns true when the packet +// should continue traversing the network stack and false when it should be +// dropped. +// +// Precondition: pkt.NetworkHeader is set. +func (it *IPTables) Check(hook Hook, pkt PacketBuffer) bool { + // Go through each table containing the hook. + for _, tablename := range it.Priorities[hook] { + table := it.Tables[tablename] + ruleIdx := table.BuiltinChains[hook] + switch verdict := it.checkChain(hook, pkt, table, ruleIdx); verdict { + // If the table returns Accept, move on to the next table. + case chainAccept: + continue + // The Drop verdict is final. + case chainDrop: + return false + case chainReturn: + // Any Return from a built-in chain means we have to + // call the underflow. + underflow := table.Rules[table.Underflows[hook]] + switch v, _ := underflow.Target.Action(pkt); v { + case RuleAccept: + continue + case RuleDrop: + return false + case RuleJump, RuleReturn: + panic("Underflows should only return RuleAccept or RuleDrop.") + default: + panic(fmt.Sprintf("Unknown verdict: %d", v)) + } + + default: + panic(fmt.Sprintf("Unknown verdict %v.", verdict)) + } + } + + // Every table returned Accept. + return true +} + +// CheckPackets runs pkts through the rules for hook and returns a map of packets that +// should not go forward. +// +// NOTE: unlike the Check API the returned map contains packets that should be +// dropped. +func (it *IPTables) CheckPackets(hook Hook, pkts PacketBufferList) (drop map[*PacketBuffer]struct{}) { + for pkt := pkts.Front(); pkt != nil; pkt = pkt.Next() { + if ok := it.Check(hook, *pkt); !ok { + if drop == nil { + drop = make(map[*PacketBuffer]struct{}) + } + drop[pkt] = struct{}{} + } + } + return drop +} + +// Precondition: pkt.NetworkHeader is set. +func (it *IPTables) checkChain(hook Hook, pkt PacketBuffer, table Table, ruleIdx int) chainVerdict { + // Start from ruleIdx and walk the list of rules until a rule gives us + // a verdict. + for ruleIdx < len(table.Rules) { + switch verdict, jumpTo := it.checkRule(hook, pkt, table, ruleIdx); verdict { + case RuleAccept: + return chainAccept + + case RuleDrop: + return chainDrop + + case RuleReturn: + return chainReturn + + case RuleJump: + // "Jumping" to the next rule just means we're + // continuing on down the list. + if jumpTo == ruleIdx+1 { + ruleIdx++ + continue + } + switch verdict := it.checkChain(hook, pkt, table, jumpTo); verdict { + case chainAccept: + return chainAccept + case chainDrop: + return chainDrop + case chainReturn: + ruleIdx++ + continue + default: + panic(fmt.Sprintf("Unknown verdict: %d", verdict)) + } + + default: + panic(fmt.Sprintf("Unknown verdict: %d", verdict)) + } + + } + + // We got through the entire table without a decision. Default to DROP + // for safety. + return chainDrop +} + +// Precondition: pk.NetworkHeader is set. +func (it *IPTables) checkRule(hook Hook, pkt PacketBuffer, table Table, ruleIdx int) (RuleVerdict, int) { + rule := table.Rules[ruleIdx] + + // If pkt.NetworkHeader hasn't been set yet, it will be contained in + // pkt.Data.First(). + if pkt.NetworkHeader == nil { + pkt.NetworkHeader = pkt.Data.First() + } + + // Check whether the packet matches the IP header filter. + if !filterMatch(rule.Filter, header.IPv4(pkt.NetworkHeader)) { + // Continue on to the next rule. + return RuleJump, ruleIdx + 1 + } + + // Go through each rule matcher. If they all match, run + // the rule target. + for _, matcher := range rule.Matchers { + matches, hotdrop := matcher.Match(hook, pkt, "") + if hotdrop { + return RuleDrop, 0 + } + if !matches { + // Continue on to the next rule. + return RuleJump, ruleIdx + 1 + } + } + + // All the matchers matched, so run the target. + return rule.Target.Action(pkt) +} + +func filterMatch(filter IPHeaderFilter, hdr header.IPv4) bool { + // TODO(gvisor.dev/issue/170): Support other fields of the filter. + // Check the transport protocol. + if filter.Protocol != 0 && filter.Protocol != hdr.TransportProtocol() { + return false + } + + // Check the destination IP. + dest := hdr.DestinationAddress() + matches := true + for i := range filter.Dst { + if dest[i]&filter.DstMask[i] != filter.Dst[i] { + matches = false + break + } + } + if matches == filter.DstInvert { + return false + } + + return true +} diff --git a/pkg/tcpip/stack/iptables_targets.go b/pkg/tcpip/stack/iptables_targets.go new file mode 100644 index 000000000..7b4543caf --- /dev/null +++ b/pkg/tcpip/stack/iptables_targets.go @@ -0,0 +1,144 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package stack + +import ( + "gvisor.dev/gvisor/pkg/log" + "gvisor.dev/gvisor/pkg/tcpip" + "gvisor.dev/gvisor/pkg/tcpip/header" +) + +// AcceptTarget accepts packets. +type AcceptTarget struct{} + +// Action implements Target.Action. +func (AcceptTarget) Action(packet PacketBuffer) (RuleVerdict, int) { + return RuleAccept, 0 +} + +// DropTarget drops packets. +type DropTarget struct{} + +// Action implements Target.Action. +func (DropTarget) Action(packet PacketBuffer) (RuleVerdict, int) { + return RuleDrop, 0 +} + +// ErrorTarget logs an error and drops the packet. It represents a target that +// should be unreachable. +type ErrorTarget struct{} + +// Action implements Target.Action. +func (ErrorTarget) Action(packet PacketBuffer) (RuleVerdict, int) { + log.Debugf("ErrorTarget triggered.") + return RuleDrop, 0 +} + +// UserChainTarget marks a rule as the beginning of a user chain. +type UserChainTarget struct { + Name string +} + +// Action implements Target.Action. +func (UserChainTarget) Action(PacketBuffer) (RuleVerdict, int) { + panic("UserChainTarget should never be called.") +} + +// ReturnTarget returns from the current chain. If the chain is a built-in, the +// hook's underflow should be called. +type ReturnTarget struct{} + +// Action implements Target.Action. +func (ReturnTarget) Action(PacketBuffer) (RuleVerdict, int) { + return RuleReturn, 0 +} + +// RedirectTarget redirects the packet by modifying the destination port/IP. +// Min and Max values for IP and Ports in the struct indicate the range of +// values which can be used to redirect. +type RedirectTarget struct { + // TODO(gvisor.dev/issue/170): Other flags need to be added after + // we support them. + // RangeProtoSpecified flag indicates single port is specified to + // redirect. + RangeProtoSpecified bool + + // Min address used to redirect. + MinIP tcpip.Address + + // Max address used to redirect. + MaxIP tcpip.Address + + // Min port used to redirect. + MinPort uint16 + + // Max port used to redirect. + MaxPort uint16 +} + +// Action implements Target.Action. +// TODO(gvisor.dev/issue/170): Parse headers without copying. The current +// implementation only works for PREROUTING and calls pkt.Clone(), neither +// of which should be the case. +func (rt RedirectTarget) Action(pkt PacketBuffer) (RuleVerdict, int) { + newPkt := pkt.Clone() + + // Set network header. + headerView := newPkt.Data.First() + netHeader := header.IPv4(headerView) + newPkt.NetworkHeader = headerView[:header.IPv4MinimumSize] + + hlen := int(netHeader.HeaderLength()) + tlen := int(netHeader.TotalLength()) + newPkt.Data.TrimFront(hlen) + newPkt.Data.CapLength(tlen - hlen) + + // TODO(gvisor.dev/issue/170): Change destination address to + // loopback or interface address on which the packet was + // received. + + // TODO(gvisor.dev/issue/170): Check Flags in RedirectTarget if + // we need to change dest address (for OUTPUT chain) or ports. + switch protocol := netHeader.TransportProtocol(); protocol { + case header.UDPProtocolNumber: + var udpHeader header.UDP + if newPkt.TransportHeader != nil { + udpHeader = header.UDP(newPkt.TransportHeader) + } else { + if len(pkt.Data.First()) < header.UDPMinimumSize { + return RuleDrop, 0 + } + udpHeader = header.UDP(newPkt.Data.First()) + } + udpHeader.SetDestinationPort(rt.MinPort) + case header.TCPProtocolNumber: + var tcpHeader header.TCP + if newPkt.TransportHeader != nil { + tcpHeader = header.TCP(newPkt.TransportHeader) + } else { + if len(pkt.Data.First()) < header.TCPMinimumSize { + return RuleDrop, 0 + } + tcpHeader = header.TCP(newPkt.TransportHeader) + } + // TODO(gvisor.dev/issue/170): Need to recompute checksum + // and implement nat connection tracking to support TCP. + tcpHeader.SetDestinationPort(rt.MinPort) + default: + return RuleDrop, 0 + } + + return RuleAccept, 0 +} diff --git a/pkg/tcpip/iptables/types.go b/pkg/tcpip/stack/iptables_types.go index 7d593c35c..2ffb55f2a 100644 --- a/pkg/tcpip/iptables/types.go +++ b/pkg/tcpip/stack/iptables_types.go @@ -12,7 +12,7 @@ // See the License for the specific language governing permissions and // limitations under the License. -package iptables +package stack import ( "gvisor.dev/gvisor/pkg/tcpip" @@ -56,44 +56,21 @@ const ( NumHooks ) -// A Verdict is returned by a rule's target to indicate how traversal of rules -// should (or should not) continue. -type Verdict int +// A RuleVerdict is what a rule decides should be done with a packet. +type RuleVerdict int const ( - // Invalid indicates an unkonwn or erroneous verdict. - Invalid Verdict = iota + // RuleAccept indicates the packet should continue through netstack. + RuleAccept RuleVerdict = iota - // Accept indicates the packet should continue traversing netstack as - // normal. - Accept + // RuleDrop indicates the packet should be dropped. + RuleDrop - // Drop inicates the packet should be dropped, stopping traversing - // netstack. - Drop + // RuleJump indicates the packet should jump to another chain. + RuleJump - // Stolen indicates the packet was co-opted by the target and should - // stop traversing netstack. - Stolen - - // Queue indicates the packet should be queued for userspace processing. - Queue - - // Repeat indicates the packet should re-traverse the chains for the - // current hook. - Repeat - - // None indicates no verdict was reached. - None - - // Jump indicates a jump to another chain. - Jump - - // Continue indicates that traversal should continue at the next rule. - Continue - - // Return indicates that traversal should return to the calling chain. - Return + // RuleReturn indicates the packet should return to the previous chain. + RuleReturn ) // IPTables holds all the tables for a netstack. @@ -167,6 +144,18 @@ type Rule struct { type IPHeaderFilter struct { // Protocol matches the transport protocol. Protocol tcpip.TransportProtocolNumber + + // Dst matches the destination IP address. + Dst tcpip.Address + + // DstMask masks bits of the destination IP address when comparing with + // Dst. + DstMask tcpip.Address + + // DstInvert inverts the meaning of the destination IP check, i.e. when + // true the filter will match packets that fail the destination + // comparison. + DstInvert bool } // A Matcher is the interface for matching packets. @@ -179,13 +168,13 @@ type Matcher interface { // used for suspicious packets. // // Precondition: packet.NetworkHeader is set. - Match(hook Hook, packet tcpip.PacketBuffer, interfaceName string) (matches bool, hotdrop bool) + Match(hook Hook, packet PacketBuffer, interfaceName string) (matches bool, hotdrop bool) } // A Target is the interface for taking an action for a packet. type Target interface { // Action takes an action on the packet and returns a verdict on how // traversal should (or should not) continue. If the return value is - // Jump, it also returns the name of the chain to jump to. - Action(packet tcpip.PacketBuffer) (Verdict, string) + // Jump, it also returns the index of the rule to jump to. + Action(packet PacketBuffer) (RuleVerdict, int) } diff --git a/pkg/tcpip/stack/ndp.go b/pkg/tcpip/stack/ndp.go index 045409bda..193a9dfde 100644 --- a/pkg/tcpip/stack/ndp.go +++ b/pkg/tcpip/stack/ndp.go @@ -15,6 +15,7 @@ package stack import ( + "fmt" "log" "math/rand" "time" @@ -240,6 +241,16 @@ type NDPDispatcher interface { // call functions on the stack itself. OnRecursiveDNSServerOption(nicID tcpip.NICID, addrs []tcpip.Address, lifetime time.Duration) + // OnDNSSearchListOption will be called when an NDP option with a DNS + // search list has been received. + // + // It is up to the caller to use the domain names in the search list + // for only their valid lifetime. OnDNSSearchListOption may be called + // with new or already known domain names. If called with known domain + // names, their valid lifetimes must be refreshed to lifetime (it may + // be increased, decreased or completely invalidated when lifetime = 0. + OnDNSSearchListOption(nicID tcpip.NICID, domainNames []string, lifetime time.Duration) + // OnDHCPv6Configuration will be called with an updated configuration that is // available via DHCPv6 for a specified NIC. // @@ -304,6 +315,15 @@ type NDPConfigurations struct { // lifetime(s) of the generated address changes; this option only // affects the generation of new addresses as part of SLAAC. AutoGenGlobalAddresses bool + + // AutoGenAddressConflictRetries determines how many times to attempt to retry + // generation of a permanent auto-generated address in response to DAD + // conflicts. + // + // If the method used to generate the address does not support creating + // alternative addresses (e.g. IIDs based on the modified EUI64 of a NIC's + // MAC address), then no attempt will be made to resolve the conflict. + AutoGenAddressConflictRetries uint8 } // DefaultNDPConfigurations returns an NDPConfigurations populated with @@ -361,16 +381,16 @@ type ndpState struct { // The default routers discovered through Router Advertisements. defaultRouters map[tcpip.Address]defaultRouterState + // The timer used to send the next router solicitation message. + rtrSolicitTimer *time.Timer + // The on-link prefixes discovered through Router Advertisements' Prefix // Information option. onLinkPrefixes map[tcpip.Subnet]onLinkPrefixState - // The timer used to send the next router solicitation message. - // If routers are being solicited, rtrSolicitTimer MUST NOT be nil. - rtrSolicitTimer *time.Timer - - // The addresses generated by SLAAC. - autoGenAddresses map[tcpip.Address]autoGenAddressState + // The SLAAC prefixes discovered through Router Advertisements' Prefix + // Information option. + slaacPrefixes map[tcpip.Subnet]slaacPrefixState // The last learned DHCPv6 configuration from an NDP RA. dhcpv6Configuration DHCPv6ConfigurationFromNDPRA @@ -402,18 +422,31 @@ type onLinkPrefixState struct { invalidationTimer tcpip.CancellableTimer } -// autoGenAddressState holds data associated with an address generated via -// SLAAC. -type autoGenAddressState struct { - // A reference to the referencedNetworkEndpoint that this autoGenAddressState - // is holding state for. - ref *referencedNetworkEndpoint - +// slaacPrefixState holds state associated with a SLAAC prefix. +type slaacPrefixState struct { deprecationTimer tcpip.CancellableTimer invalidationTimer tcpip.CancellableTimer // Nonzero only when the address is not valid forever. validUntil time.Time + + // Nonzero only when the address is not preferred forever. + preferredUntil time.Time + + // The prefix's permanent address endpoint. + // + // May only be nil when a SLAAC address is being (re-)generated. Otherwise, + // must not be nil as all SLAAC prefixes must have a SLAAC address. + ref *referencedNetworkEndpoint + + // The number of times a permanent address has been generated for the prefix. + // + // Addresses may be regenerated in reseponse to a DAD conflicts. + generationAttempts uint8 + + // The maximum number of times to attempt regeneration of a permanent SLAAC + // address in response to DAD conflicts. + maxGenerationAttempts uint8 } // startDuplicateAddressDetection performs Duplicate Address Detection. @@ -430,7 +463,7 @@ func (ndp *ndpState) startDuplicateAddressDetection(addr tcpip.Address, ref *ref if ref.getKind() != permanentTentative { // The endpoint should be marked as tentative since we are starting DAD. - log.Fatalf("ndpdad: addr %s is not tentative on NIC(%d)", addr, ndp.nic.ID()) + panic(fmt.Sprintf("ndpdad: addr %s is not tentative on NIC(%d)", addr, ndp.nic.ID())) } // Should not attempt to perform DAD on an address that is currently in the @@ -442,7 +475,7 @@ func (ndp *ndpState) startDuplicateAddressDetection(addr tcpip.Address, ref *ref // address, or its reference count would have been increased without doing // the work that would have been done for an address that was brand new. // See NIC.addAddressLocked. - log.Fatalf("ndpdad: already performing DAD for addr %s on NIC(%d)", addr, ndp.nic.ID()) + panic(fmt.Sprintf("ndpdad: already performing DAD for addr %s on NIC(%d)", addr, ndp.nic.ID())) } remaining := ndp.configs.DupAddrDetectTransmits @@ -478,7 +511,7 @@ func (ndp *ndpState) startDuplicateAddressDetection(addr tcpip.Address, ref *ref if ref.getKind() != permanentTentative { // The endpoint should still be marked as tentative since we are still // performing DAD on it. - log.Fatalf("ndpdad: addr %s is no longer tentative on NIC(%d)", addr, ndp.nic.ID()) + panic(fmt.Sprintf("ndpdad: addr %s is no longer tentative on NIC(%d)", addr, ndp.nic.ID())) } dadDone := remaining == 0 @@ -548,9 +581,9 @@ func (ndp *ndpState) sendDADPacket(addr tcpip.Address) *tcpip.Error { // Route should resolve immediately since snmc is a multicast address so a // remote link address can be calculated without a resolution process. if c, err := r.Resolve(nil); err != nil { - log.Fatalf("ndp: error when resolving route to send NDP NS for DAD (%s -> %s on NIC(%d)): %s", header.IPv6Any, snmc, ndp.nic.ID(), err) + panic(fmt.Sprintf("ndp: error when resolving route to send NDP NS for DAD (%s -> %s on NIC(%d)): %s", header.IPv6Any, snmc, ndp.nic.ID(), err)) } else if c != nil { - log.Fatalf("ndp: route resolution not immediate for route to send NDP NS for DAD (%s -> %s on NIC(%d))", header.IPv6Any, snmc, ndp.nic.ID()) + panic(fmt.Sprintf("ndp: route resolution not immediate for route to send NDP NS for DAD (%s -> %s on NIC(%d))", header.IPv6Any, snmc, ndp.nic.ID())) } hdr := buffer.NewPrependable(int(r.MaxHeaderLength()) + header.ICMPv6NeighborSolicitMinimumSize) @@ -566,7 +599,7 @@ func (ndp *ndpState) sendDADPacket(addr tcpip.Address) *tcpip.Error { Protocol: header.ICMPv6ProtocolNumber, TTL: header.NDPHopLimit, TOS: DefaultTOS, - }, tcpip.PacketBuffer{Header: hdr}, + }, PacketBuffer{Header: hdr}, ); err != nil { sent.Dropped.Increment() return err @@ -688,7 +721,16 @@ func (ndp *ndpState) handleRA(ip tcpip.Address, ra header.NDPRouterAdvert) { continue } - ndp.nic.stack.ndpDisp.OnRecursiveDNSServerOption(ndp.nic.ID(), opt.Addresses(), opt.Lifetime()) + addrs, _ := opt.Addresses() + ndp.nic.stack.ndpDisp.OnRecursiveDNSServerOption(ndp.nic.ID(), addrs, opt.Lifetime()) + + case header.NDPDNSSearchList: + if ndp.nic.stack.ndpDisp == nil { + continue + } + + domainNames, _ := opt.DomainNames() + ndp.nic.stack.ndpDisp.OnDNSSearchListOption(ndp.nic.ID(), domainNames, opt.Lifetime()) case header.NDPPrefixInformation: prefix := opt.Subnet() @@ -899,23 +941,15 @@ func (ndp *ndpState) handleAutonomousPrefixInformation(pi header.NDPPrefixInform prefix := pi.Subnet() - // Check if we already have an auto-generated address for prefix. - for addr, addrState := range ndp.autoGenAddresses { - refAddrWithPrefix := tcpip.AddressWithPrefix{Address: addr, PrefixLen: addrState.ref.ep.PrefixLen()} - if refAddrWithPrefix.Subnet() != prefix { - continue - } - - // At this point, we know we are refreshing a SLAAC generated IPv6 address - // with the prefix prefix. Do the work as outlined by RFC 4862 section - // 5.5.3.e. - ndp.refreshAutoGenAddressLifetimes(addr, pl, vl) + // Check if we already maintain SLAAC state for prefix. + if _, ok := ndp.slaacPrefixes[prefix]; ok { + // As per RFC 4862 section 5.5.3.e, refresh prefix's SLAAC lifetimes. + ndp.refreshSLAACPrefixLifetimes(prefix, pl, vl) return } - // We do not already have an address with the prefix prefix. Do the - // work as outlined by RFC 4862 section 5.5.3.d if n is configured - // to auto-generate global addresses by SLAAC. + // prefix is a new SLAAC prefix. Do the work as outlined by RFC 4862 section + // 5.5.3.d if ndp is configured to auto-generate new addresses via SLAAC. if !ndp.configs.AutoGenGlobalAddresses { return } @@ -927,6 +961,8 @@ func (ndp *ndpState) handleAutonomousPrefixInformation(pi header.NDPPrefixInform // for prefix. // // pl is the new preferred lifetime. vl is the new valid lifetime. +// +// The NIC that ndp belongs to MUST be locked. func (ndp *ndpState) doSLAAC(prefix tcpip.Subnet, pl, vl time.Duration) { // If we do not already have an address for this prefix and the valid // lifetime is 0, no need to do anything further, as per RFC 4862 @@ -942,10 +978,83 @@ func (ndp *ndpState) doSLAAC(prefix tcpip.Subnet, pl, vl time.Duration) { return } + state := slaacPrefixState{ + deprecationTimer: tcpip.MakeCancellableTimer(&ndp.nic.mu, func() { + state, ok := ndp.slaacPrefixes[prefix] + if !ok { + panic(fmt.Sprintf("ndp: must have a slaacPrefixes entry for the deprecated SLAAC prefix %s", prefix)) + } + + ndp.deprecateSLAACAddress(state.ref) + }), + invalidationTimer: tcpip.MakeCancellableTimer(&ndp.nic.mu, func() { + state, ok := ndp.slaacPrefixes[prefix] + if !ok { + panic(fmt.Sprintf("ndp: must have a slaacPrefixes entry for the invalidated SLAAC prefix %s", prefix)) + } + + ndp.invalidateSLAACPrefix(prefix, state) + }), + maxGenerationAttempts: ndp.configs.AutoGenAddressConflictRetries + 1, + } + + now := time.Now() + + // The time an address is preferred until is needed to properly generate the + // address. + if pl < header.NDPInfiniteLifetime { + state.preferredUntil = now.Add(pl) + } + + if !ndp.generateSLAACAddr(prefix, &state) { + // We were unable to generate an address for the prefix, we do not nothing + // further as there is no reason to maintain state or timers for a prefix we + // do not have an address for. + return + } + + // Setup the initial timers to deprecate and invalidate prefix. + + if pl < header.NDPInfiniteLifetime && pl != 0 { + state.deprecationTimer.Reset(pl) + } + + if vl < header.NDPInfiniteLifetime { + state.invalidationTimer.Reset(vl) + state.validUntil = now.Add(vl) + } + + ndp.slaacPrefixes[prefix] = state +} + +// generateSLAACAddr generates a SLAAC address for prefix. +// +// Returns true if an address was successfully generated. +// +// Panics if the prefix is not a SLAAC prefix or it already has an address. +// +// The NIC that ndp belongs to MUST be locked. +func (ndp *ndpState) generateSLAACAddr(prefix tcpip.Subnet, state *slaacPrefixState) bool { + if r := state.ref; r != nil { + panic(fmt.Sprintf("ndp: SLAAC prefix %s already has a permenant address %s", prefix, r.addrWithPrefix())) + } + + // If we have already reached the maximum address generation attempts for the + // prefix, do not generate another address. + if state.generationAttempts == state.maxGenerationAttempts { + return false + } + addrBytes := []byte(prefix.ID()) if oIID := ndp.nic.stack.opaqueIIDOpts; oIID.NICNameFromID != nil { - addrBytes = header.AppendOpaqueInterfaceIdentifier(addrBytes[:header.IIDOffsetInIPv6Address], prefix, oIID.NICNameFromID(ndp.nic.ID(), ndp.nic.name), 0 /* dadCounter */, oIID.SecretKey) - } else { + addrBytes = header.AppendOpaqueInterfaceIdentifier( + addrBytes[:header.IIDOffsetInIPv6Address], + prefix, + oIID.NICNameFromID(ndp.nic.ID(), ndp.nic.name), + state.generationAttempts, + oIID.SecretKey, + ) + } else if state.generationAttempts == 0 { // Only attempt to generate an interface-specific IID if we have a valid // link address. // @@ -953,137 +1062,140 @@ func (ndp *ndpState) doSLAAC(prefix tcpip.Subnet, pl, vl time.Duration) { // LinkEndpoint.LinkAddress) before reaching this point. linkAddr := ndp.nic.linkEP.LinkAddress() if !header.IsValidUnicastEthernetAddress(linkAddr) { - return + return false } // Generate an address within prefix from the modified EUI-64 of ndp's NIC's // Ethernet MAC address. header.EthernetAdddressToModifiedEUI64IntoBuf(linkAddr, addrBytes[header.IIDOffsetInIPv6Address:]) + } else { + // We have no way to regenerate an address when addresses are not generated + // with opaque IIDs. + return false } - addr := tcpip.Address(addrBytes) - addrWithPrefix := tcpip.AddressWithPrefix{ - Address: addr, - PrefixLen: validPrefixLenForAutoGen, + + generatedAddr := tcpip.ProtocolAddress{ + Protocol: header.IPv6ProtocolNumber, + AddressWithPrefix: tcpip.AddressWithPrefix{ + Address: tcpip.Address(addrBytes), + PrefixLen: validPrefixLenForAutoGen, + }, } // If the nic already has this address, do nothing further. - if ndp.nic.hasPermanentAddrLocked(addr) { - return + if ndp.nic.hasPermanentAddrLocked(generatedAddr.AddressWithPrefix.Address) { + return false } // Inform the integrator that we have a new SLAAC address. ndpDisp := ndp.nic.stack.ndpDisp if ndpDisp == nil { - return + return false } - if !ndpDisp.OnAutoGenAddress(ndp.nic.ID(), addrWithPrefix) { + + if !ndpDisp.OnAutoGenAddress(ndp.nic.ID(), generatedAddr.AddressWithPrefix) { // Informed by the integrator not to add the address. - return + return false } - protocolAddr := tcpip.ProtocolAddress{ - Protocol: header.IPv6ProtocolNumber, - AddressWithPrefix: addrWithPrefix, - } - // If the preferred lifetime is zero, then the address should be considered - // deprecated. - deprecated := pl == 0 - ref, err := ndp.nic.addAddressLocked(protocolAddr, FirstPrimaryEndpoint, permanent, slaac, deprecated) + deprecated := time.Since(state.preferredUntil) >= 0 + ref, err := ndp.nic.addAddressLocked(generatedAddr, FirstPrimaryEndpoint, permanent, slaac, deprecated) if err != nil { - log.Fatalf("ndp: error when adding address %s: %s", protocolAddr, err) + panic(fmt.Sprintf("ndp: error when adding address %+v: %s", generatedAddr, err)) } - state := autoGenAddressState{ - ref: ref, - deprecationTimer: tcpip.MakeCancellableTimer(&ndp.nic.mu, func() { - addrState, ok := ndp.autoGenAddresses[addr] - if !ok { - log.Fatalf("ndp: must have an autoGenAddressess entry for the SLAAC generated IPv6 address %s", addr) - } - addrState.ref.deprecated = true - ndp.notifyAutoGenAddressDeprecated(addr) - }), - invalidationTimer: tcpip.MakeCancellableTimer(&ndp.nic.mu, func() { - ndp.invalidateAutoGenAddress(addr) - }), - } - - // Setup the initial timers to deprecate and invalidate this newly generated - // address. + state.generationAttempts++ + state.ref = ref + return true +} - if !deprecated && pl < header.NDPInfiniteLifetime { - state.deprecationTimer.Reset(pl) +// regenerateSLAACAddr regenerates an address for a SLAAC prefix. +// +// If generating a new address for the prefix fails, the prefix will be +// invalidated. +// +// The NIC that ndp belongs to MUST be locked. +func (ndp *ndpState) regenerateSLAACAddr(prefix tcpip.Subnet) { + state, ok := ndp.slaacPrefixes[prefix] + if !ok { + panic(fmt.Sprintf("ndp: SLAAC prefix state not found to regenerate address for %s", prefix)) } - if vl < header.NDPInfiniteLifetime { - state.invalidationTimer.Reset(vl) - state.validUntil = time.Now().Add(vl) + if ndp.generateSLAACAddr(prefix, &state) { + ndp.slaacPrefixes[prefix] = state + return } - ndp.autoGenAddresses[addr] = state + // We were unable to generate a permanent address for the SLAAC prefix so + // invalidate the prefix as there is no reason to maintain state for a + // SLAAC prefix we do not have an address for. + ndp.invalidateSLAACPrefix(prefix, state) } -// refreshAutoGenAddressLifetimes refreshes the lifetime of a SLAAC generated -// address addr. +// refreshSLAACPrefixLifetimes refreshes the lifetimes of a SLAAC prefix. // // pl is the new preferred lifetime. vl is the new valid lifetime. -func (ndp *ndpState) refreshAutoGenAddressLifetimes(addr tcpip.Address, pl, vl time.Duration) { - addrState, ok := ndp.autoGenAddresses[addr] +// +// The NIC that ndp belongs to MUST be locked. +func (ndp *ndpState) refreshSLAACPrefixLifetimes(prefix tcpip.Subnet, pl, vl time.Duration) { + prefixState, ok := ndp.slaacPrefixes[prefix] if !ok { - log.Fatalf("ndp: SLAAC state not found to refresh lifetimes for %s", addr) + panic(fmt.Sprintf("ndp: SLAAC prefix state not found to refresh lifetimes for %s", prefix)) } - defer func() { ndp.autoGenAddresses[addr] = addrState }() + defer func() { ndp.slaacPrefixes[prefix] = prefixState }() - // If the preferred lifetime is zero, then the address should be considered - // deprecated. + // If the preferred lifetime is zero, then the prefix should be deprecated. deprecated := pl == 0 - wasDeprecated := addrState.ref.deprecated - addrState.ref.deprecated = deprecated - - // Only send the deprecation event if the deprecated status for addr just - // changed from non-deprecated to deprecated. - if !wasDeprecated && deprecated { - ndp.notifyAutoGenAddressDeprecated(addr) + if deprecated { + ndp.deprecateSLAACAddress(prefixState.ref) + } else { + prefixState.ref.deprecated = false } - // If addr was preferred for some finite lifetime before, stop the deprecation - // timer so it can be reset. - addrState.deprecationTimer.StopLocked() + // If prefix was preferred for some finite lifetime before, stop the + // deprecation timer so it can be reset. + prefixState.deprecationTimer.StopLocked() + + now := time.Now() - // Reset the deprecation timer if addr has a finite preferred lifetime. - if !deprecated && pl < header.NDPInfiniteLifetime { - addrState.deprecationTimer.Reset(pl) + // Reset the deprecation timer if prefix has a finite preferred lifetime. + if pl < header.NDPInfiniteLifetime { + if !deprecated { + prefixState.deprecationTimer.Reset(pl) + } + prefixState.preferredUntil = now.Add(pl) + } else { + prefixState.preferredUntil = time.Time{} } - // As per RFC 4862 section 5.5.3.e, the valid lifetime of the address - // + // As per RFC 4862 section 5.5.3.e, update the valid lifetime for prefix: // // 1) If the received Valid Lifetime is greater than 2 hours or greater than - // RemainingLifetime, set the valid lifetime of the address to the + // RemainingLifetime, set the valid lifetime of the prefix to the // advertised Valid Lifetime. // // 2) If RemainingLifetime is less than or equal to 2 hours, ignore the // advertised Valid Lifetime. // - // 3) Otherwise, reset the valid lifetime of the address to 2 hours. + // 3) Otherwise, reset the valid lifetime of the prefix to 2 hours. // Handle the infinite valid lifetime separately as we do not keep a timer in // this case. if vl >= header.NDPInfiniteLifetime { - addrState.invalidationTimer.StopLocked() - addrState.validUntil = time.Time{} + prefixState.invalidationTimer.StopLocked() + prefixState.validUntil = time.Time{} return } var effectiveVl time.Duration var rl time.Duration - // If the address was originally set to be valid forever, assume the remaining + // If the prefix was originally set to be valid forever, assume the remaining // time to be the maximum possible value. - if addrState.validUntil == (time.Time{}) { + if prefixState.validUntil == (time.Time{}) { rl = header.NDPInfiniteLifetime } else { - rl = time.Until(addrState.validUntil) + rl = time.Until(prefixState.validUntil) } if vl > MinPrefixInformationValidLifetimeForUpdate || vl > rl { @@ -1094,85 +1206,110 @@ func (ndp *ndpState) refreshAutoGenAddressLifetimes(addr tcpip.Address, pl, vl t effectiveVl = MinPrefixInformationValidLifetimeForUpdate } - addrState.invalidationTimer.StopLocked() - addrState.invalidationTimer.Reset(effectiveVl) - addrState.validUntil = time.Now().Add(effectiveVl) + prefixState.invalidationTimer.StopLocked() + prefixState.invalidationTimer.Reset(effectiveVl) + prefixState.validUntil = now.Add(effectiveVl) } -// notifyAutoGenAddressDeprecated notifies the stack's NDP dispatcher that addr -// has been deprecated. -func (ndp *ndpState) notifyAutoGenAddressDeprecated(addr tcpip.Address) { +// deprecateSLAACAddress marks ref as deprecated and notifies the stack's NDP +// dispatcher that ref has been deprecated. +// +// deprecateSLAACAddress does nothing if ref is already deprecated. +// +// The NIC that ndp belongs to MUST be locked. +func (ndp *ndpState) deprecateSLAACAddress(ref *referencedNetworkEndpoint) { + if ref.deprecated { + return + } + + ref.deprecated = true if ndpDisp := ndp.nic.stack.ndpDisp; ndpDisp != nil { - ndpDisp.OnAutoGenAddressDeprecated(ndp.nic.ID(), tcpip.AddressWithPrefix{ - Address: addr, - PrefixLen: validPrefixLenForAutoGen, - }) + ndpDisp.OnAutoGenAddressDeprecated(ndp.nic.ID(), ref.addrWithPrefix()) } } -// invalidateAutoGenAddress invalidates an auto-generated address. +// invalidateSLAACPrefix invalidates a SLAAC prefix. // // The NIC that ndp belongs to MUST be locked. -func (ndp *ndpState) invalidateAutoGenAddress(addr tcpip.Address) { - if !ndp.cleanupAutoGenAddrResourcesAndNotify(addr) { - return +func (ndp *ndpState) invalidateSLAACPrefix(prefix tcpip.Subnet, state slaacPrefixState) { + if r := state.ref; r != nil { + // Since we are already invalidating the prefix, do not invalidate the + // prefix when removing the address. + if err := ndp.nic.removePermanentIPv6EndpointLocked(r, false /* allowSLAACPrefixInvalidation */); err != nil { + panic(fmt.Sprintf("ndp: removePermanentIPv6EndpointLocked(%s, false): %s", r.addrWithPrefix(), err)) + } } - ndp.nic.removePermanentAddressLocked(addr) + ndp.cleanupSLAACPrefixResources(prefix, state) } -// cleanupAutoGenAddrResourcesAndNotify cleans up an invalidated auto-generated -// address's resources from ndp. If the stack has an NDP dispatcher, it will -// be notified that addr has been invalidated. -// -// Returns true if ndp had resources for addr to cleanup. +// cleanupSLAACAddrResourcesAndNotify cleans up an invalidated SLAAC address's +// resources. // // The NIC that ndp belongs to MUST be locked. -func (ndp *ndpState) cleanupAutoGenAddrResourcesAndNotify(addr tcpip.Address) bool { - state, ok := ndp.autoGenAddresses[addr] - if !ok { - return false +func (ndp *ndpState) cleanupSLAACAddrResourcesAndNotify(addr tcpip.AddressWithPrefix, invalidatePrefix bool) { + if ndpDisp := ndp.nic.stack.ndpDisp; ndpDisp != nil { + ndpDisp.OnAutoGenAddressInvalidated(ndp.nic.ID(), addr) } - state.deprecationTimer.StopLocked() - state.invalidationTimer.StopLocked() - delete(ndp.autoGenAddresses, addr) + prefix := addr.Subnet() + state, ok := ndp.slaacPrefixes[prefix] + if !ok || state.ref == nil || addr.Address != state.ref.ep.ID().LocalAddress { + return + } - if ndpDisp := ndp.nic.stack.ndpDisp; ndpDisp != nil { - ndpDisp.OnAutoGenAddressInvalidated(ndp.nic.ID(), tcpip.AddressWithPrefix{ - Address: addr, - PrefixLen: validPrefixLenForAutoGen, - }) + if !invalidatePrefix { + // If the prefix is not being invalidated, disassociate the address from the + // prefix and do nothing further. + state.ref = nil + ndp.slaacPrefixes[prefix] = state + return } - return true + ndp.cleanupSLAACPrefixResources(prefix, state) +} + +// cleanupSLAACPrefixResources cleansup a SLAAC prefix's timers and entry. +// +// Panics if the SLAAC prefix is not known. +// +// The NIC that ndp belongs to MUST be locked. +func (ndp *ndpState) cleanupSLAACPrefixResources(prefix tcpip.Subnet, state slaacPrefixState) { + state.deprecationTimer.StopLocked() + state.invalidationTimer.StopLocked() + delete(ndp.slaacPrefixes, prefix) } -// cleanupHostOnlyState cleans up any state that is only useful for hosts. +// cleanupState cleans up ndp's state. +// +// If hostOnly is true, then only host-specific state will be cleaned up. +// +// cleanupState MUST be called with hostOnly set to true when ndp's NIC is +// transitioning from a host to a router. This function will invalidate all +// discovered on-link prefixes, discovered routers, and auto-generated +// addresses. // -// cleanupHostOnlyState MUST be called when ndp's NIC is transitioning from a -// host to a router. This function will invalidate all discovered on-link -// prefixes, discovered routers, and auto-generated addresses as routers do not -// normally process Router Advertisements to discover default routers and -// on-link prefixes, and auto-generate addresses via SLAAC. +// If hostOnly is true, then the link-local auto-generated address will not be +// invalidated as routers are also expected to generate a link-local address. // // The NIC that ndp belongs to MUST be locked. -func (ndp *ndpState) cleanupHostOnlyState() { +func (ndp *ndpState) cleanupState(hostOnly bool) { linkLocalSubnet := header.IPv6LinkLocalPrefix.Subnet() - linkLocalAddrs := 0 - for addr := range ndp.autoGenAddresses { + linkLocalPrefixes := 0 + for prefix, state := range ndp.slaacPrefixes { // RFC 4862 section 5 states that routers are also expected to generate a - // link-local address so we do not invalidate them. - if linkLocalSubnet.Contains(addr) { - linkLocalAddrs++ + // link-local address so we do not invalidate them if we are cleaning up + // host-only state. + if hostOnly && prefix == linkLocalSubnet { + linkLocalPrefixes++ continue } - ndp.invalidateAutoGenAddress(addr) + ndp.invalidateSLAACPrefix(prefix, state) } - if got := len(ndp.autoGenAddresses); got != linkLocalAddrs { - log.Fatalf("ndp: still have non-linklocal auto-generated addresses after cleaning up; found = %d prefixes, of which %d are link-local", got, linkLocalAddrs) + if got := len(ndp.slaacPrefixes); got != linkLocalPrefixes { + panic(fmt.Sprintf("ndp: still have non-linklocal SLAAC prefixes after cleaning up; found = %d prefixes, of which %d are link-local", got, linkLocalPrefixes)) } for prefix := range ndp.onLinkPrefixes { @@ -1180,7 +1317,7 @@ func (ndp *ndpState) cleanupHostOnlyState() { } if got := len(ndp.onLinkPrefixes); got != 0 { - log.Fatalf("ndp: still have discovered on-link prefixes after cleaning up; found = %d", got) + panic(fmt.Sprintf("ndp: still have discovered on-link prefixes after cleaning up; found = %d", got)) } for router := range ndp.defaultRouters { @@ -1188,7 +1325,7 @@ func (ndp *ndpState) cleanupHostOnlyState() { } if got := len(ndp.defaultRouters); got != 0 { - log.Fatalf("ndp: still have discovered default routers after cleaning up; found = %d", got) + panic(fmt.Sprintf("ndp: still have discovered default routers after cleaning up; found = %d", got)) } } @@ -1215,24 +1352,45 @@ func (ndp *ndpState) startSolicitingRouters() { } ndp.rtrSolicitTimer = time.AfterFunc(delay, func() { - // Send an RS message with the unspecified source address. - ref := ndp.nic.getRefOrCreateTemp(header.IPv6ProtocolNumber, header.IPv6Any, NeverPrimaryEndpoint, forceSpoofing) - r := makeRoute(header.IPv6ProtocolNumber, header.IPv6Any, header.IPv6AllRoutersMulticastAddress, ndp.nic.linkEP.LinkAddress(), ref, false, false) + // As per RFC 4861 section 4.1, the source of the RS is an address assigned + // to the sending interface, or the unspecified address if no address is + // assigned to the sending interface. + ref := ndp.nic.primaryIPv6Endpoint(header.IPv6AllRoutersMulticastAddress) + if ref == nil { + ref = ndp.nic.getRefOrCreateTemp(header.IPv6ProtocolNumber, header.IPv6Any, NeverPrimaryEndpoint, forceSpoofing) + } + localAddr := ref.ep.ID().LocalAddress + r := makeRoute(header.IPv6ProtocolNumber, localAddr, header.IPv6AllRoutersMulticastAddress, ndp.nic.linkEP.LinkAddress(), ref, false, false) defer r.Release() // Route should resolve immediately since // header.IPv6AllRoutersMulticastAddress is a multicast address so a // remote link address can be calculated without a resolution process. if c, err := r.Resolve(nil); err != nil { - log.Fatalf("ndp: error when resolving route to send NDP RS (%s -> %s on NIC(%d)): %s", header.IPv6Any, header.IPv6AllRoutersMulticastAddress, ndp.nic.ID(), err) + panic(fmt.Sprintf("ndp: error when resolving route to send NDP RS (%s -> %s on NIC(%d)): %s", header.IPv6Any, header.IPv6AllRoutersMulticastAddress, ndp.nic.ID(), err)) } else if c != nil { - log.Fatalf("ndp: route resolution not immediate for route to send NDP RS (%s -> %s on NIC(%d))", header.IPv6Any, header.IPv6AllRoutersMulticastAddress, ndp.nic.ID()) + panic(fmt.Sprintf("ndp: route resolution not immediate for route to send NDP RS (%s -> %s on NIC(%d))", header.IPv6Any, header.IPv6AllRoutersMulticastAddress, ndp.nic.ID())) } - payloadSize := header.ICMPv6HeaderSize + header.NDPRSMinimumSize - hdr := buffer.NewPrependable(header.IPv6MinimumSize + payloadSize) + // As per RFC 4861 section 4.1, an NDP RS SHOULD include the source + // link-layer address option if the source address of the NDP RS is + // specified. This option MUST NOT be included if the source address is + // unspecified. + // + // TODO(b/141011931): Validate a LinkEndpoint's link address (provided by + // LinkEndpoint.LinkAddress) before reaching this point. + var optsSerializer header.NDPOptionsSerializer + if localAddr != header.IPv6Any && header.IsValidUnicastEthernetAddress(r.LocalLinkAddress) { + optsSerializer = header.NDPOptionsSerializer{ + header.NDPSourceLinkLayerAddressOption(r.LocalLinkAddress), + } + } + payloadSize := header.ICMPv6HeaderSize + header.NDPRSMinimumSize + int(optsSerializer.Length()) + hdr := buffer.NewPrependable(int(r.MaxHeaderLength()) + payloadSize) pkt := header.ICMPv6(hdr.Prepend(payloadSize)) pkt.SetType(header.ICMPv6RouterSolicit) + rs := header.NDPRouterSolicit(pkt.NDPPayload()) + rs.Options().Serialize(optsSerializer) pkt.SetChecksum(header.ICMPv6Checksum(pkt, r.LocalAddress, r.RemoteAddress, buffer.VectorisedView{})) sent := r.Stats().ICMP.V6PacketsSent @@ -1241,7 +1399,7 @@ func (ndp *ndpState) startSolicitingRouters() { Protocol: header.ICMPv6ProtocolNumber, TTL: header.NDPHopLimit, TOS: DefaultTOS, - }, tcpip.PacketBuffer{Header: hdr}, + }, PacketBuffer{Header: hdr}, ); err != nil { sent.Dropped.Increment() log.Printf("startSolicitingRouters: error writing NDP router solicit message on NIC(%d); err = %s", ndp.nic.ID(), err) diff --git a/pkg/tcpip/stack/ndp_test.go b/pkg/tcpip/stack/ndp_test.go index 1f6f77439..6dd460984 100644 --- a/pkg/tcpip/stack/ndp_test.go +++ b/pkg/tcpip/stack/ndp_test.go @@ -133,6 +133,12 @@ type ndpRDNSSEvent struct { rdnss ndpRDNSS } +type ndpDNSSLEvent struct { + nicID tcpip.NICID + domainNames []string + lifetime time.Duration +} + type ndpDHCPv6Event struct { nicID tcpip.NICID configuration stack.DHCPv6ConfigurationFromNDPRA @@ -150,6 +156,8 @@ type ndpDispatcher struct { rememberPrefix bool autoGenAddrC chan ndpAutoGenAddrEvent rdnssC chan ndpRDNSSEvent + dnsslC chan ndpDNSSLEvent + routeTable []tcpip.Route dhcpv6ConfigurationC chan ndpDHCPv6Event } @@ -257,6 +265,17 @@ func (n *ndpDispatcher) OnRecursiveDNSServerOption(nicID tcpip.NICID, addrs []tc } } +// Implements stack.NDPDispatcher.OnDNSSearchListOption. +func (n *ndpDispatcher) OnDNSSearchListOption(nicID tcpip.NICID, domainNames []string, lifetime time.Duration) { + if n.dnsslC != nil { + n.dnsslC <- ndpDNSSLEvent{ + nicID, + domainNames, + lifetime, + } + } +} + // Implements stack.NDPDispatcher.OnDHCPv6Configuration. func (n *ndpDispatcher) OnDHCPv6Configuration(nicID tcpip.NICID, configuration stack.DHCPv6ConfigurationFromNDPRA) { if c := n.dhcpv6ConfigurationC; c != nil { @@ -267,6 +286,17 @@ func (n *ndpDispatcher) OnDHCPv6Configuration(nicID tcpip.NICID, configuration s } } +// channelLinkWithHeaderLength is a channel.Endpoint with a configurable +// header length. +type channelLinkWithHeaderLength struct { + *channel.Endpoint + headerLength uint16 +} + +func (l *channelLinkWithHeaderLength) MaxHeaderLength() uint16 { + return l.headerLength +} + // Check e to make sure that the event is for addr on nic with ID 1, and the // resolved flag set to resolved with the specified err. func checkDADEvent(e ndpDADEvent, nicID tcpip.NICID, addr tcpip.Address, resolved bool, err *tcpip.Error) string { @@ -323,21 +353,46 @@ func TestDADDisabled(t *testing.T) { // DAD for various values of DupAddrDetectTransmits and RetransmitTimer. // Included in the subtests is a test to make sure that an invalid // RetransmitTimer (<1ms) values get fixed to the default RetransmitTimer of 1s. +// This tests also validates the NDP NS packet that is transmitted. func TestDADResolve(t *testing.T) { const nicID = 1 tests := []struct { name string + linkHeaderLen uint16 dupAddrDetectTransmits uint8 retransTimer time.Duration expectedRetransmitTimer time.Duration }{ - {"1:1s:1s", 1, time.Second, time.Second}, - {"2:1s:1s", 2, time.Second, time.Second}, - {"1:2s:2s", 1, 2 * time.Second, 2 * time.Second}, + { + name: "1:1s:1s", + dupAddrDetectTransmits: 1, + retransTimer: time.Second, + expectedRetransmitTimer: time.Second, + }, + { + name: "2:1s:1s", + linkHeaderLen: 1, + dupAddrDetectTransmits: 2, + retransTimer: time.Second, + expectedRetransmitTimer: time.Second, + }, + { + name: "1:2s:2s", + linkHeaderLen: 2, + dupAddrDetectTransmits: 1, + retransTimer: 2 * time.Second, + expectedRetransmitTimer: 2 * time.Second, + }, // 0s is an invalid RetransmitTimer timer and will be fixed to // the default RetransmitTimer value of 1s. - {"1:0s:1s", 1, 0, time.Second}, + { + name: "1:0s:1s", + linkHeaderLen: 3, + dupAddrDetectTransmits: 1, + retransTimer: 0, + expectedRetransmitTimer: time.Second, + }, } for _, test := range tests { @@ -356,10 +411,13 @@ func TestDADResolve(t *testing.T) { opts.NDPConfigs.RetransmitTimer = test.retransTimer opts.NDPConfigs.DupAddrDetectTransmits = test.dupAddrDetectTransmits - e := channel.New(int(test.dupAddrDetectTransmits), 1280, linkAddr1) - e.LinkEPCapabilities |= stack.CapabilityResolutionRequired + e := channelLinkWithHeaderLength{ + Endpoint: channel.New(int(test.dupAddrDetectTransmits), 1280, linkAddr1), + headerLength: test.linkHeaderLen, + } + e.Endpoint.LinkEPCapabilities |= stack.CapabilityResolutionRequired s := stack.New(opts) - if err := s.CreateNIC(nicID, e); err != nil { + if err := s.CreateNIC(nicID, &e); err != nil { t.Fatalf("CreateNIC(%d, _) = %s", nicID, err) } @@ -367,8 +425,7 @@ func TestDADResolve(t *testing.T) { t.Fatalf("AddAddress(%d, %d, %s) = %s", nicID, header.IPv6ProtocolNumber, addr1, err) } - // Address should not be considered bound to the NIC yet - // (DAD ongoing). + // Address should not be considered bound to the NIC yet (DAD ongoing). addr, err := s.GetMainNICAddress(nicID, header.IPv6ProtocolNumber) if err != nil { t.Fatalf("got stack.GetMainNICAddress(%d, %d) = (_, %v), want = (_, nil)", nicID, header.IPv6ProtocolNumber, err) @@ -377,10 +434,9 @@ func TestDADResolve(t *testing.T) { t.Fatalf("got stack.GetMainNICAddress(%d, %d) = (%s, nil), want = (%s, nil)", nicID, header.IPv6ProtocolNumber, addr, want) } - // Wait for the remaining time - some delta (500ms), to - // make sure the address is still not resolved. - const delta = 500 * time.Millisecond - time.Sleep(test.expectedRetransmitTimer*time.Duration(test.dupAddrDetectTransmits) - delta) + // Make sure the address does not resolve before the resolution time has + // passed. + time.Sleep(test.expectedRetransmitTimer*time.Duration(test.dupAddrDetectTransmits) - defaultAsyncEventTimeout) addr, err = s.GetMainNICAddress(nicID, header.IPv6ProtocolNumber) if err != nil { t.Fatalf("got stack.GetMainNICAddress(%d, %d) = (_, %v), want = (_, nil)", nicID, header.IPv6ProtocolNumber, err) @@ -391,13 +447,7 @@ func TestDADResolve(t *testing.T) { // Wait for DAD to resolve. select { - case <-time.After(2 * delta): - // We should get a resolution event after 500ms - // (delta) since we wait for 500ms less than the - // expected resolution time above to make sure - // that the address did not yet resolve. Waiting - // for 1s (2x delta) without a resolution event - // means something is wrong. + case <-time.After(2 * defaultAsyncEventTimeout): t.Fatal("timed out waiting for DAD resolution") case e := <-ndpDisp.dadC: if diff := checkDADEvent(e, nicID, addr1, true, nil); diff != "" { @@ -437,7 +487,7 @@ func TestDADResolve(t *testing.T) { // As per RFC 4861 section 4.3, a possible option is the Source Link // Layer option, but this option MUST NOT be included when the source // address of the packet is the unspecified address. - checker.IPv6(t, p.Pkt.Header.View().ToVectorisedView().First(), + checker.IPv6(t, p.Pkt.Header.View(), checker.SrcAddr(header.IPv6Any), checker.DstAddr(snmc), checker.TTL(header.NDPHopLimit), @@ -445,6 +495,10 @@ func TestDADResolve(t *testing.T) { checker.NDPNSTargetAddress(addr1), checker.NDPNSOptions(nil), )) + + if l, want := p.Pkt.Header.AvailableLength(), int(test.linkHeaderLen); l != want { + t.Errorf("got p.Pkt.Header.AvailableLength() = %d; want = %d", l, want) + } } }) } @@ -559,7 +613,7 @@ func TestDADFail(t *testing.T) { // Receive a packet to simulate multiple nodes owning or // attempting to own the same address. hdr := test.makeBuf(addr1) - e.InjectInbound(header.IPv6ProtocolNumber, tcpip.PacketBuffer{ + e.InjectInbound(header.IPv6ProtocolNumber, stack.PacketBuffer{ Data: hdr.View().ToVectorisedView(), }) @@ -588,74 +642,121 @@ func TestDADFail(t *testing.T) { if want := (tcpip.AddressWithPrefix{}); addr != want { t.Fatalf("got stack.GetMainNICAddress(%d, %d) = (%s, nil), want = (%s, nil)", nicID, header.IPv6ProtocolNumber, addr, want) } + + // Attempting to add the address again should not fail if the address's + // state was cleaned up when DAD failed. + if err := s.AddAddress(nicID, header.IPv6ProtocolNumber, addr1); err != nil { + t.Fatalf("AddAddress(%d, %d, %s) = %s", nicID, header.IPv6ProtocolNumber, addr1, err) + } }) } } -// TestDADStop tests to make sure that the DAD process stops when an address is -// removed. func TestDADStop(t *testing.T) { const nicID = 1 - ndpDisp := ndpDispatcher{ - dadC: make(chan ndpDADEvent, 1), - } - ndpConfigs := stack.NDPConfigurations{ - RetransmitTimer: time.Second, - DupAddrDetectTransmits: 2, - } - opts := stack.Options{ - NetworkProtocols: []stack.NetworkProtocol{ipv6.NewProtocol()}, - NDPDisp: &ndpDisp, - NDPConfigs: ndpConfigs, - } + tests := []struct { + name string + stopFn func(t *testing.T, s *stack.Stack) + skipFinalAddrCheck bool + }{ + // Tests to make sure that DAD stops when an address is removed. + { + name: "Remove address", + stopFn: func(t *testing.T, s *stack.Stack) { + if err := s.RemoveAddress(nicID, addr1); err != nil { + t.Fatalf("RemoveAddress(%d, %s): %s", nicID, addr1, err) + } + }, + }, - e := channel.New(0, 1280, linkAddr1) - s := stack.New(opts) - if err := s.CreateNIC(nicID, e); err != nil { - t.Fatalf("CreateNIC(%d, _) = %s", nicID, err) - } + // Tests to make sure that DAD stops when the NIC is disabled. + { + name: "Disable NIC", + stopFn: func(t *testing.T, s *stack.Stack) { + if err := s.DisableNIC(nicID); err != nil { + t.Fatalf("DisableNIC(%d): %s", nicID, err) + } + }, + }, - if err := s.AddAddress(nicID, header.IPv6ProtocolNumber, addr1); err != nil { - t.Fatalf("AddAddress(%d, %d, %s) = %s", nicID, header.IPv6ProtocolNumber, addr1, err) + // Tests to make sure that DAD stops when the NIC is removed. + { + name: "Remove NIC", + stopFn: func(t *testing.T, s *stack.Stack) { + if err := s.RemoveNIC(nicID); err != nil { + t.Fatalf("RemoveNIC(%d): %s", nicID, err) + } + }, + // The NIC is removed so we can't check its addresses after calling + // stopFn. + skipFinalAddrCheck: true, + }, } - // Address should not be considered bound to the NIC yet (DAD ongoing). - addr, err := s.GetMainNICAddress(nicID, header.IPv6ProtocolNumber) - if err != nil { - t.Fatalf("got stack.GetMainNICAddress(%d, %d) = (_, %v), want = (_, nil)", nicID, header.IPv6ProtocolNumber, err) - } - if want := (tcpip.AddressWithPrefix{}); addr != want { - t.Fatalf("got stack.GetMainNICAddress(%d, %d) = (%s, nil), want = (%s, nil)", nicID, header.IPv6ProtocolNumber, addr, want) - } + for _, test := range tests { + t.Run(test.name, func(t *testing.T) { + ndpDisp := ndpDispatcher{ + dadC: make(chan ndpDADEvent, 1), + } + ndpConfigs := stack.NDPConfigurations{ + RetransmitTimer: time.Second, + DupAddrDetectTransmits: 2, + } + opts := stack.Options{ + NetworkProtocols: []stack.NetworkProtocol{ipv6.NewProtocol()}, + NDPDisp: &ndpDisp, + NDPConfigs: ndpConfigs, + } - // Remove the address. This should stop DAD. - if err := s.RemoveAddress(nicID, addr1); err != nil { - t.Fatalf("RemoveAddress(%d, %s) = %s", nicID, addr1, err) - } + e := channel.New(0, 1280, linkAddr1) + s := stack.New(opts) + if err := s.CreateNIC(nicID, e); err != nil { + t.Fatalf("CreateNIC(%d, _): %s", nicID, err) + } - // Wait for DAD to fail (since the address was removed during DAD). - select { - case <-time.After(time.Duration(ndpConfigs.DupAddrDetectTransmits)*ndpConfigs.RetransmitTimer + time.Second): - // If we don't get a failure event after the expected resolution - // time + extra 1s buffer, something is wrong. - t.Fatal("timed out waiting for DAD failure") - case e := <-ndpDisp.dadC: - if diff := checkDADEvent(e, nicID, addr1, false, nil); diff != "" { - t.Errorf("dad event mismatch (-want +got):\n%s", diff) - } - } - addr, err = s.GetMainNICAddress(nicID, header.IPv6ProtocolNumber) - if err != nil { - t.Fatalf("got stack.GetMainNICAddress(%d, %d) = (_, %v), want = (_, nil)", nicID, header.IPv6ProtocolNumber, err) - } - if want := (tcpip.AddressWithPrefix{}); addr != want { - t.Fatalf("got stack.GetMainNICAddress(%d, %d) = (%s, nil), want = (%s, nil)", nicID, header.IPv6ProtocolNumber, addr, want) - } + if err := s.AddAddress(nicID, header.IPv6ProtocolNumber, addr1); err != nil { + t.Fatalf("AddAddress(%d, %d, %s): %s", nicID, header.IPv6ProtocolNumber, addr1, err) + } + + // Address should not be considered bound to the NIC yet (DAD ongoing). + addr, err := s.GetMainNICAddress(nicID, header.IPv6ProtocolNumber) + if err != nil { + t.Fatalf("got stack.GetMainNICAddress(%d, %d) = (_, %v), want = (_, nil)", nicID, header.IPv6ProtocolNumber, err) + } + if want := (tcpip.AddressWithPrefix{}); addr != want { + t.Fatalf("got stack.GetMainNICAddress(%d, %d) = (%s, nil), want = (%s, nil)", nicID, header.IPv6ProtocolNumber, addr, want) + } + + test.stopFn(t, s) + + // Wait for DAD to fail (since the address was removed during DAD). + select { + case <-time.After(time.Duration(ndpConfigs.DupAddrDetectTransmits)*ndpConfigs.RetransmitTimer + time.Second): + // If we don't get a failure event after the expected resolution + // time + extra 1s buffer, something is wrong. + t.Fatal("timed out waiting for DAD failure") + case e := <-ndpDisp.dadC: + if diff := checkDADEvent(e, nicID, addr1, false, nil); diff != "" { + t.Errorf("dad event mismatch (-want +got):\n%s", diff) + } + } + + if !test.skipFinalAddrCheck { + addr, err := s.GetMainNICAddress(nicID, header.IPv6ProtocolNumber) + if err != nil { + t.Fatalf("got stack.GetMainNICAddress(%d, %d) = (_, %v), want = (_, nil)", nicID, header.IPv6ProtocolNumber, err) + } + if want := (tcpip.AddressWithPrefix{}); addr != want { + t.Errorf("got stack.GetMainNICAddress(%d, %d) = (%s, nil), want = (%s, nil)", nicID, header.IPv6ProtocolNumber, addr, want) + } + } - // Should not have sent more than 1 NS message. - if got := s.Stats().ICMP.V6PacketsSent.NeighborSolicit.Value(); got > 1 { - t.Fatalf("got NeighborSolicit = %d, want <= 1", got) + // Should not have sent more than 1 NS message. + if got := s.Stats().ICMP.V6PacketsSent.NeighborSolicit.Value(); got > 1 { + t.Errorf("got NeighborSolicit = %d, want <= 1", got) + } + }) } } @@ -834,7 +935,7 @@ func TestSetNDPConfigurations(t *testing.T) { // raBufWithOptsAndDHCPv6 returns a valid NDP Router Advertisement with options // and DHCPv6 configurations specified. -func raBufWithOptsAndDHCPv6(ip tcpip.Address, rl uint16, managedAddress, otherConfigurations bool, optSer header.NDPOptionsSerializer) tcpip.PacketBuffer { +func raBufWithOptsAndDHCPv6(ip tcpip.Address, rl uint16, managedAddress, otherConfigurations bool, optSer header.NDPOptionsSerializer) stack.PacketBuffer { icmpSize := header.ICMPv6HeaderSize + header.NDPRAMinimumSize + int(optSer.Length()) hdr := buffer.NewPrependable(header.IPv6MinimumSize + icmpSize) pkt := header.ICMPv6(hdr.Prepend(icmpSize)) @@ -869,14 +970,14 @@ func raBufWithOptsAndDHCPv6(ip tcpip.Address, rl uint16, managedAddress, otherCo DstAddr: header.IPv6AllNodesMulticastAddress, }) - return tcpip.PacketBuffer{Data: hdr.View().ToVectorisedView()} + return stack.PacketBuffer{Data: hdr.View().ToVectorisedView()} } // raBufWithOpts returns a valid NDP Router Advertisement with options. // // Note, raBufWithOpts does not populate any of the RA fields other than the // Router Lifetime. -func raBufWithOpts(ip tcpip.Address, rl uint16, optSer header.NDPOptionsSerializer) tcpip.PacketBuffer { +func raBufWithOpts(ip tcpip.Address, rl uint16, optSer header.NDPOptionsSerializer) stack.PacketBuffer { return raBufWithOptsAndDHCPv6(ip, rl, false, false, optSer) } @@ -885,7 +986,7 @@ func raBufWithOpts(ip tcpip.Address, rl uint16, optSer header.NDPOptionsSerializ // // Note, raBufWithDHCPv6 does not populate any of the RA fields other than the // DHCPv6 related ones. -func raBufWithDHCPv6(ip tcpip.Address, managedAddresses, otherConfiguratiosns bool) tcpip.PacketBuffer { +func raBufWithDHCPv6(ip tcpip.Address, managedAddresses, otherConfiguratiosns bool) stack.PacketBuffer { return raBufWithOptsAndDHCPv6(ip, 0, managedAddresses, otherConfiguratiosns, header.NDPOptionsSerializer{}) } @@ -893,7 +994,7 @@ func raBufWithDHCPv6(ip tcpip.Address, managedAddresses, otherConfiguratiosns bo // // Note, raBuf does not populate any of the RA fields other than the // Router Lifetime. -func raBuf(ip tcpip.Address, rl uint16) tcpip.PacketBuffer { +func raBuf(ip tcpip.Address, rl uint16) stack.PacketBuffer { return raBufWithOpts(ip, rl, header.NDPOptionsSerializer{}) } @@ -902,7 +1003,7 @@ func raBuf(ip tcpip.Address, rl uint16) tcpip.PacketBuffer { // // Note, raBufWithPI does not populate any of the RA fields other than the // Router Lifetime. -func raBufWithPI(ip tcpip.Address, rl uint16, prefix tcpip.AddressWithPrefix, onLink, auto bool, vl, pl uint32) tcpip.PacketBuffer { +func raBufWithPI(ip tcpip.Address, rl uint16, prefix tcpip.AddressWithPrefix, onLink, auto bool, vl, pl uint32) stack.PacketBuffer { flags := uint8(0) if onLink { // The OnLink flag is the 7th bit in the flags byte. @@ -950,8 +1051,6 @@ func TestNoRouterDiscovery(t *testing.T) { forwarding := i&4 == 0 t.Run(fmt.Sprintf("HandleRAs(%t), DiscoverDefaultRouters(%t), Forwarding(%t)", handle, discover, forwarding), func(t *testing.T) { - t.Parallel() - ndpDisp := ndpDispatcher{ routerC: make(chan ndpRouterEvent, 1), } @@ -990,8 +1089,6 @@ func checkRouterEvent(e ndpRouterEvent, addr tcpip.Address, discovered bool) str // TestRouterDiscoveryDispatcherNoRemember tests that the stack does not // remember a discovered router when the dispatcher asks it not to. func TestRouterDiscoveryDispatcherNoRemember(t *testing.T) { - t.Parallel() - ndpDisp := ndpDispatcher{ routerC: make(chan ndpRouterEvent, 1), } @@ -1032,8 +1129,6 @@ func TestRouterDiscoveryDispatcherNoRemember(t *testing.T) { } func TestRouterDiscovery(t *testing.T) { - t.Parallel() - ndpDisp := ndpDispatcher{ routerC: make(chan ndpRouterEvent, 1), rememberRouter: true, @@ -1135,8 +1230,6 @@ func TestRouterDiscovery(t *testing.T) { // TestRouterDiscoveryMaxRouters tests that only // stack.MaxDiscoveredDefaultRouters discovered routers are remembered. func TestRouterDiscoveryMaxRouters(t *testing.T) { - t.Parallel() - ndpDisp := ndpDispatcher{ routerC: make(chan ndpRouterEvent, 1), rememberRouter: true, @@ -1203,8 +1296,6 @@ func TestNoPrefixDiscovery(t *testing.T) { forwarding := i&4 == 0 t.Run(fmt.Sprintf("HandleRAs(%t), DiscoverOnLinkPrefixes(%t), Forwarding(%t)", handle, discover, forwarding), func(t *testing.T) { - t.Parallel() - ndpDisp := ndpDispatcher{ prefixC: make(chan ndpPrefixEvent, 1), } @@ -1244,8 +1335,6 @@ func checkPrefixEvent(e ndpPrefixEvent, prefix tcpip.Subnet, discovered bool) st // TestPrefixDiscoveryDispatcherNoRemember tests that the stack does not // remember a discovered on-link prefix when the dispatcher asks it not to. func TestPrefixDiscoveryDispatcherNoRemember(t *testing.T) { - t.Parallel() - prefix, subnet, _ := prefixSubnetAddr(0, "") ndpDisp := ndpDispatcher{ @@ -1289,8 +1378,6 @@ func TestPrefixDiscoveryDispatcherNoRemember(t *testing.T) { } func TestPrefixDiscovery(t *testing.T) { - t.Parallel() - prefix1, subnet1, _ := prefixSubnetAddr(0, "") prefix2, subnet2, _ := prefixSubnetAddr(1, "") prefix3, subnet3, _ := prefixSubnetAddr(2, "") @@ -1479,8 +1566,6 @@ func TestPrefixDiscoveryWithInfiniteLifetime(t *testing.T) { // TestPrefixDiscoveryMaxRouters tests that only // stack.MaxDiscoveredOnLinkPrefixes discovered on-link prefixes are remembered. func TestPrefixDiscoveryMaxOnLinkPrefixes(t *testing.T) { - t.Parallel() - ndpDisp := ndpDispatcher{ prefixC: make(chan ndpPrefixEvent, stack.MaxDiscoveredOnLinkPrefixes+3), rememberPrefix: true, @@ -1575,8 +1660,6 @@ func TestNoAutoGenAddr(t *testing.T) { forwarding := i&4 == 0 t.Run(fmt.Sprintf("HandleRAs(%t), AutoGenAddr(%t), Forwarding(%t)", handle, autogen, forwarding), func(t *testing.T) { - t.Parallel() - ndpDisp := ndpDispatcher{ autoGenAddrC: make(chan ndpAutoGenAddrEvent, 1), } @@ -1901,7 +1984,7 @@ func TestAutoGenAddrDeprecateFromPI(t *testing.T) { // addr2 is deprecated but if explicitly requested, it should be used. fullAddr2 := tcpip.FullAddress{Addr: addr2.Address, NIC: nicID} if got := addrForNewConnectionWithAddr(t, s, fullAddr2); got != addr2.Address { - t.Errorf("got addrForNewConnectionWithAddr(_, _, %+v) = %s, want = %s", got, addr2.Address) + t.Errorf("got addrForNewConnectionWithAddr(_, _, %+v) = %s, want = %s", fullAddr2, got, addr2.Address) } // Another PI w/ 0 preferred lifetime should not result in a deprecation @@ -1914,7 +1997,7 @@ func TestAutoGenAddrDeprecateFromPI(t *testing.T) { } expectPrimaryAddr(addr1) if got := addrForNewConnectionWithAddr(t, s, fullAddr2); got != addr2.Address { - t.Errorf("got addrForNewConnectionWithAddr(_, _, %+v) = %s, want = %s", got, addr2.Address) + t.Errorf("got addrForNewConnectionWithAddr(_, _, %+v) = %s, want = %s", fullAddr2, got, addr2.Address) } // Refresh lifetimes of addr generated from prefix2. @@ -2026,7 +2109,7 @@ func TestAutoGenAddrTimerDeprecation(t *testing.T) { // addr1 is deprecated but if explicitly requested, it should be used. fullAddr1 := tcpip.FullAddress{Addr: addr1.Address, NIC: nicID} if got := addrForNewConnectionWithAddr(t, s, fullAddr1); got != addr1.Address { - t.Errorf("got addrForNewConnectionWithAddr(_, _, %+v) = %s, want = %s", got, addr1.Address) + t.Errorf("got addrForNewConnectionWithAddr(_, _, %+v) = %s, want = %s", fullAddr1, got, addr1.Address) } // Refresh valid lifetime for addr of prefix1, w/ 0 preferred lifetime to make @@ -2039,7 +2122,7 @@ func TestAutoGenAddrTimerDeprecation(t *testing.T) { } expectPrimaryAddr(addr2) if got := addrForNewConnectionWithAddr(t, s, fullAddr1); got != addr1.Address { - t.Errorf("got addrForNewConnectionWithAddr(_, _, %+v) = %s, want = %s", got, addr1.Address) + t.Errorf("got addrForNewConnectionWithAddr(_, _, %+v) = %s, want = %s", fullAddr1, got, addr1.Address) } // Refresh lifetimes for addr of prefix1. @@ -2063,7 +2146,7 @@ func TestAutoGenAddrTimerDeprecation(t *testing.T) { // addr2 should be the primary endpoint now since it is not deprecated. expectPrimaryAddr(addr2) if got := addrForNewConnectionWithAddr(t, s, fullAddr1); got != addr1.Address { - t.Errorf("got addrForNewConnectionWithAddr(_, _, %+v) = %s, want = %s", got, addr1.Address) + t.Errorf("got addrForNewConnectionWithAddr(_, _, %+v) = %s, want = %s", fullAddr1, got, addr1.Address) } // Wait for addr of prefix1 to be invalidated. @@ -2326,8 +2409,6 @@ func TestAutoGenAddrValidLifetimeUpdates(t *testing.T) { }, } - const delta = 500 * time.Millisecond - // This Run will not return until the parallel tests finish. // // We need this because we need to do some teardown work after the @@ -2380,24 +2461,21 @@ func TestAutoGenAddrValidLifetimeUpdates(t *testing.T) { // to test.evl. // - // Make sure we do not get any invalidation - // events until atleast 500ms (delta) before - // test.evl. + // The address should not be invalidated until the effective valid + // lifetime has passed. select { case <-ndpDisp.autoGenAddrC: t.Fatal("unexpectedly received an auto gen addr event") - case <-time.After(time.Duration(test.evl)*time.Second - delta): + case <-time.After(time.Duration(test.evl)*time.Second - defaultAsyncEventTimeout): } - // Wait for another second (2x delta), but now - // we expect the invalidation event. + // Wait for the invalidation event. select { case e := <-ndpDisp.autoGenAddrC: if diff := checkAutoGenAddrEvent(e, addr, invalidatedAddr); diff != "" { t.Errorf("auto-gen addr event mismatch (-want +got):\n%s", diff) } - - case <-time.After(2 * delta): + case <-time.After(2 * defaultAsyncEventTimeout): t.Fatal("timeout waiting for addr auto gen event") } }) @@ -2409,8 +2487,6 @@ func TestAutoGenAddrValidLifetimeUpdates(t *testing.T) { // by the user, its resources will be cleaned up and an invalidation event will // be sent to the integrator. func TestAutoGenAddrRemoval(t *testing.T) { - t.Parallel() - prefix, _, addr := prefixSubnetAddr(0, linkAddr1) ndpDisp := ndpDispatcher{ @@ -2467,8 +2543,6 @@ func TestAutoGenAddrRemoval(t *testing.T) { // TestAutoGenAddrAfterRemoval tests adding a SLAAC address that was previously // assigned to the NIC but is in the permanentExpired state. func TestAutoGenAddrAfterRemoval(t *testing.T) { - t.Parallel() - const nicID = 1 prefix1, _, addr1 := prefixSubnetAddr(0, linkAddr1) @@ -2515,7 +2589,7 @@ func TestAutoGenAddrAfterRemoval(t *testing.T) { AddressWithPrefix: addr2, } if err := s.AddProtocolAddressWithOptions(nicID, protoAddr2, stack.FirstPrimaryEndpoint); err != nil { - t.Fatalf("AddProtocolAddressWithOptions(%d, %+v, %d, %s) = %s", nicID, protoAddr2, stack.FirstPrimaryEndpoint, err) + t.Fatalf("AddProtocolAddressWithOptions(%d, %+v, %d) = %s", nicID, protoAddr2, stack.FirstPrimaryEndpoint, err) } // addr2 should be more preferred now since it is at the front of the primary // list. @@ -2580,8 +2654,6 @@ func TestAutoGenAddrAfterRemoval(t *testing.T) { // TestAutoGenAddrStaticConflict tests that if SLAAC generates an address that // is already assigned to the NIC, the static address remains. func TestAutoGenAddrStaticConflict(t *testing.T) { - t.Parallel() - prefix, _, addr := prefixSubnetAddr(0, linkAddr1) ndpDisp := ndpDispatcher{ @@ -2637,8 +2709,6 @@ func TestAutoGenAddrStaticConflict(t *testing.T) { // TestAutoGenAddrWithOpaqueIID tests that SLAAC generated addresses will use // opaque interface identifiers when configured to do so. func TestAutoGenAddrWithOpaqueIID(t *testing.T) { - t.Parallel() - const nicID = 1 const nicName = "nic1" var secretKeyBuf [header.OpaqueIIDSecretKeyMinBytes]byte @@ -2738,12 +2808,465 @@ func TestAutoGenAddrWithOpaqueIID(t *testing.T) { } } +// TestAutoGenAddrWithOpaqueIIDDADRetries tests the regeneration of an +// auto-generated IPv6 address in response to a DAD conflict. +func TestAutoGenAddrWithOpaqueIIDDADRetries(t *testing.T) { + const nicID = 1 + const nicName = "nic" + const dadTransmits = 1 + const retransmitTimer = time.Second + const maxMaxRetries = 3 + const lifetimeSeconds = 10 + + var secretKeyBuf [header.OpaqueIIDSecretKeyMinBytes]byte + secretKey := secretKeyBuf[:] + n, err := rand.Read(secretKey) + if err != nil { + t.Fatalf("rand.Read(_): %s", err) + } + if n != header.OpaqueIIDSecretKeyMinBytes { + t.Fatalf("got rand.Read(_) = (%d, _), want = (%d, _)", n, header.OpaqueIIDSecretKeyMinBytes) + } + + prefix, subnet, _ := prefixSubnetAddr(0, linkAddr1) + + for maxRetries := uint8(0); maxRetries <= maxMaxRetries; maxRetries++ { + for numFailures := uint8(0); numFailures <= maxRetries+1; numFailures++ { + addrTypes := []struct { + name string + ndpConfigs stack.NDPConfigurations + autoGenLinkLocal bool + subnet tcpip.Subnet + triggerSLAACFn func(e *channel.Endpoint) + }{ + { + name: "Global address", + ndpConfigs: stack.NDPConfigurations{ + DupAddrDetectTransmits: dadTransmits, + RetransmitTimer: retransmitTimer, + HandleRAs: true, + AutoGenGlobalAddresses: true, + AutoGenAddressConflictRetries: maxRetries, + }, + subnet: subnet, + triggerSLAACFn: func(e *channel.Endpoint) { + // Receive an RA with prefix1 in a PI. + e.InjectInbound(header.IPv6ProtocolNumber, raBufWithPI(llAddr2, 0, prefix, true, true, lifetimeSeconds, lifetimeSeconds)) + + }, + }, + { + name: "LinkLocal address", + ndpConfigs: stack.NDPConfigurations{ + DupAddrDetectTransmits: dadTransmits, + RetransmitTimer: retransmitTimer, + AutoGenAddressConflictRetries: maxRetries, + }, + autoGenLinkLocal: true, + subnet: header.IPv6LinkLocalPrefix.Subnet(), + triggerSLAACFn: func(e *channel.Endpoint) {}, + }, + } + + for _, addrType := range addrTypes { + maxRetries := maxRetries + numFailures := numFailures + addrType := addrType + + t.Run(fmt.Sprintf("%s with %d max retries and %d failures", addrType.name, maxRetries, numFailures), func(t *testing.T) { + t.Parallel() + + ndpDisp := ndpDispatcher{ + dadC: make(chan ndpDADEvent, 1), + autoGenAddrC: make(chan ndpAutoGenAddrEvent, 2), + } + e := channel.New(0, 1280, linkAddr1) + s := stack.New(stack.Options{ + NetworkProtocols: []stack.NetworkProtocol{ipv6.NewProtocol()}, + AutoGenIPv6LinkLocal: addrType.autoGenLinkLocal, + NDPConfigs: addrType.ndpConfigs, + NDPDisp: &ndpDisp, + OpaqueIIDOpts: stack.OpaqueInterfaceIdentifierOptions{ + NICNameFromID: func(_ tcpip.NICID, nicName string) string { + return nicName + }, + SecretKey: secretKey, + }, + }) + opts := stack.NICOptions{Name: nicName} + if err := s.CreateNICWithOptions(nicID, e, opts); err != nil { + t.Fatalf("CreateNICWithOptions(%d, _, %+v) = %s", nicID, opts, err) + } + + expectAutoGenAddrEvent := func(addr tcpip.AddressWithPrefix, eventType ndpAutoGenAddrEventType) { + t.Helper() + + select { + case e := <-ndpDisp.autoGenAddrC: + if diff := checkAutoGenAddrEvent(e, addr, eventType); diff != "" { + t.Errorf("auto-gen addr event mismatch (-want +got):\n%s", diff) + } + default: + t.Fatal("expected addr auto gen event") + } + } + + addrType.triggerSLAACFn(e) + + // Simulate DAD conflicts so the address is regenerated. + for i := uint8(0); i < numFailures; i++ { + addrBytes := []byte(addrType.subnet.ID()) + addr := tcpip.AddressWithPrefix{ + Address: tcpip.Address(header.AppendOpaqueInterfaceIdentifier(addrBytes[:header.IIDOffsetInIPv6Address], addrType.subnet, nicName, i, secretKey)), + PrefixLen: 64, + } + expectAutoGenAddrEvent(addr, newAddr) + + // Should not have any addresses assigned to the NIC. + mainAddr, err := s.GetMainNICAddress(nicID, header.IPv6ProtocolNumber) + if err != nil { + t.Fatalf("stack.GetMainNICAddress(%d, _) err = %s", nicID, err) + } + if want := (tcpip.AddressWithPrefix{}); mainAddr != want { + t.Fatalf("got stack.GetMainNICAddress(_, _) = (%s, nil), want = (%s, nil)", mainAddr, want) + } + + // Simulate a DAD conflict. + if err := s.DupTentativeAddrDetected(nicID, addr.Address); err != nil { + t.Fatalf("s.DupTentativeAddrDetected(%d, %s): %s", nicID, addr.Address, err) + } + expectAutoGenAddrEvent(addr, invalidatedAddr) + select { + case e := <-ndpDisp.dadC: + if diff := checkDADEvent(e, nicID, addr.Address, false, nil); diff != "" { + t.Errorf("dad event mismatch (-want +got):\n%s", diff) + } + default: + t.Fatal("expected DAD event") + } + + // Attempting to add the address manually should not fail if the + // address's state was cleaned up when DAD failed. + if err := s.AddAddress(nicID, header.IPv6ProtocolNumber, addr.Address); err != nil { + t.Fatalf("AddAddress(%d, %d, %s) = %s", nicID, header.IPv6ProtocolNumber, addr.Address, err) + } + if err := s.RemoveAddress(nicID, addr.Address); err != nil { + t.Fatalf("RemoveAddress(%d, %s) = %s", nicID, addr.Address, err) + } + select { + case e := <-ndpDisp.dadC: + if diff := checkDADEvent(e, nicID, addr.Address, false, nil); diff != "" { + t.Errorf("dad event mismatch (-want +got):\n%s", diff) + } + default: + t.Fatal("expected DAD event") + } + } + + // Should not have any addresses assigned to the NIC. + mainAddr, err := s.GetMainNICAddress(nicID, header.IPv6ProtocolNumber) + if err != nil { + t.Fatalf("stack.GetMainNICAddress(%d, _) err = %s", nicID, err) + } + if want := (tcpip.AddressWithPrefix{}); mainAddr != want { + t.Fatalf("got stack.GetMainNICAddress(_, _) = (%s, nil), want = (%s, nil)", mainAddr, want) + } + + // If we had less failures than generation attempts, we should have an + // address after DAD resolves. + if maxRetries+1 > numFailures { + addrBytes := []byte(addrType.subnet.ID()) + addr := tcpip.AddressWithPrefix{ + Address: tcpip.Address(header.AppendOpaqueInterfaceIdentifier(addrBytes[:header.IIDOffsetInIPv6Address], addrType.subnet, nicName, numFailures, secretKey)), + PrefixLen: 64, + } + expectAutoGenAddrEvent(addr, newAddr) + + select { + case e := <-ndpDisp.dadC: + if diff := checkDADEvent(e, nicID, addr.Address, true, nil); diff != "" { + t.Errorf("dad event mismatch (-want +got):\n%s", diff) + } + case <-time.After(dadTransmits*retransmitTimer + defaultAsyncEventTimeout): + t.Fatal("timed out waiting for DAD event") + } + + mainAddr, err := s.GetMainNICAddress(nicID, header.IPv6ProtocolNumber) + if err != nil { + t.Fatalf("stack.GetMainNICAddress(%d, _) err = %s", nicID, err) + } + if mainAddr != addr { + t.Fatalf("got stack.GetMainNICAddress(_, _) = (%s, nil), want = (%s, nil)", mainAddr, addr) + } + } + + // Should not attempt address regeneration again. + select { + case e := <-ndpDisp.autoGenAddrC: + t.Fatalf("unexpectedly got an auto-generated address event = %+v", e) + case <-time.After(defaultAsyncEventTimeout): + } + }) + } + } + } +} + +// TestAutoGenAddrWithEUI64IIDNoDADRetries tests that a regeneration attempt is +// not made for SLAAC addresses generated with an IID based on the NIC's link +// address. +func TestAutoGenAddrWithEUI64IIDNoDADRetries(t *testing.T) { + const nicID = 1 + const dadTransmits = 1 + const retransmitTimer = time.Second + const maxRetries = 3 + const lifetimeSeconds = 10 + + prefix, subnet, _ := prefixSubnetAddr(0, linkAddr1) + + addrTypes := []struct { + name string + ndpConfigs stack.NDPConfigurations + autoGenLinkLocal bool + subnet tcpip.Subnet + triggerSLAACFn func(e *channel.Endpoint) + }{ + { + name: "Global address", + ndpConfigs: stack.NDPConfigurations{ + DupAddrDetectTransmits: dadTransmits, + RetransmitTimer: retransmitTimer, + HandleRAs: true, + AutoGenGlobalAddresses: true, + AutoGenAddressConflictRetries: maxRetries, + }, + subnet: subnet, + triggerSLAACFn: func(e *channel.Endpoint) { + // Receive an RA with prefix1 in a PI. + e.InjectInbound(header.IPv6ProtocolNumber, raBufWithPI(llAddr2, 0, prefix, true, true, lifetimeSeconds, lifetimeSeconds)) + + }, + }, + { + name: "LinkLocal address", + ndpConfigs: stack.NDPConfigurations{ + DupAddrDetectTransmits: dadTransmits, + RetransmitTimer: retransmitTimer, + AutoGenAddressConflictRetries: maxRetries, + }, + autoGenLinkLocal: true, + subnet: header.IPv6LinkLocalPrefix.Subnet(), + triggerSLAACFn: func(e *channel.Endpoint) {}, + }, + } + + for _, addrType := range addrTypes { + addrType := addrType + + t.Run(addrType.name, func(t *testing.T) { + t.Parallel() + + ndpDisp := ndpDispatcher{ + dadC: make(chan ndpDADEvent, 1), + autoGenAddrC: make(chan ndpAutoGenAddrEvent, 2), + } + e := channel.New(0, 1280, linkAddr1) + s := stack.New(stack.Options{ + NetworkProtocols: []stack.NetworkProtocol{ipv6.NewProtocol()}, + AutoGenIPv6LinkLocal: addrType.autoGenLinkLocal, + NDPConfigs: addrType.ndpConfigs, + NDPDisp: &ndpDisp, + }) + if err := s.CreateNIC(nicID, e); err != nil { + t.Fatalf("CreateNIC(%d, _) = %s", nicID, err) + } + + expectAutoGenAddrEvent := func(addr tcpip.AddressWithPrefix, eventType ndpAutoGenAddrEventType) { + t.Helper() + + select { + case e := <-ndpDisp.autoGenAddrC: + if diff := checkAutoGenAddrEvent(e, addr, eventType); diff != "" { + t.Errorf("auto-gen addr event mismatch (-want +got):\n%s", diff) + } + default: + t.Fatal("expected addr auto gen event") + } + } + + addrType.triggerSLAACFn(e) + + addrBytes := []byte(addrType.subnet.ID()) + header.EthernetAdddressToModifiedEUI64IntoBuf(linkAddr1, addrBytes[header.IIDOffsetInIPv6Address:]) + addr := tcpip.AddressWithPrefix{ + Address: tcpip.Address(addrBytes), + PrefixLen: 64, + } + expectAutoGenAddrEvent(addr, newAddr) + + // Simulate a DAD conflict. + if err := s.DupTentativeAddrDetected(nicID, addr.Address); err != nil { + t.Fatalf("s.DupTentativeAddrDetected(%d, %s): %s", nicID, addr.Address, err) + } + expectAutoGenAddrEvent(addr, invalidatedAddr) + select { + case e := <-ndpDisp.dadC: + if diff := checkDADEvent(e, nicID, addr.Address, false, nil); diff != "" { + t.Errorf("dad event mismatch (-want +got):\n%s", diff) + } + default: + t.Fatal("expected DAD event") + } + + // Should not attempt address regeneration. + select { + case e := <-ndpDisp.autoGenAddrC: + t.Fatalf("unexpectedly got an auto-generated address event = %+v", e) + case <-time.After(defaultAsyncEventTimeout): + } + }) + } +} + +// TestAutoGenAddrContinuesLifetimesAfterRetry tests that retrying address +// generation in response to DAD conflicts does not refresh the lifetimes. +func TestAutoGenAddrContinuesLifetimesAfterRetry(t *testing.T) { + const nicID = 1 + const nicName = "nic" + const dadTransmits = 1 + const retransmitTimer = 2 * time.Second + const failureTimer = time.Second + const maxRetries = 1 + const lifetimeSeconds = 5 + + var secretKeyBuf [header.OpaqueIIDSecretKeyMinBytes]byte + secretKey := secretKeyBuf[:] + n, err := rand.Read(secretKey) + if err != nil { + t.Fatalf("rand.Read(_): %s", err) + } + if n != header.OpaqueIIDSecretKeyMinBytes { + t.Fatalf("got rand.Read(_) = (%d, _), want = (%d, _)", n, header.OpaqueIIDSecretKeyMinBytes) + } + + prefix, subnet, _ := prefixSubnetAddr(0, linkAddr1) + + ndpDisp := ndpDispatcher{ + dadC: make(chan ndpDADEvent, 1), + autoGenAddrC: make(chan ndpAutoGenAddrEvent, 2), + } + e := channel.New(0, 1280, linkAddr1) + s := stack.New(stack.Options{ + NetworkProtocols: []stack.NetworkProtocol{ipv6.NewProtocol()}, + NDPConfigs: stack.NDPConfigurations{ + DupAddrDetectTransmits: dadTransmits, + RetransmitTimer: retransmitTimer, + HandleRAs: true, + AutoGenGlobalAddresses: true, + AutoGenAddressConflictRetries: maxRetries, + }, + NDPDisp: &ndpDisp, + OpaqueIIDOpts: stack.OpaqueInterfaceIdentifierOptions{ + NICNameFromID: func(_ tcpip.NICID, nicName string) string { + return nicName + }, + SecretKey: secretKey, + }, + }) + opts := stack.NICOptions{Name: nicName} + if err := s.CreateNICWithOptions(nicID, e, opts); err != nil { + t.Fatalf("CreateNICWithOptions(%d, _, %+v) = %s", nicID, opts, err) + } + + expectAutoGenAddrEvent := func(addr tcpip.AddressWithPrefix, eventType ndpAutoGenAddrEventType) { + t.Helper() + + select { + case e := <-ndpDisp.autoGenAddrC: + if diff := checkAutoGenAddrEvent(e, addr, eventType); diff != "" { + t.Errorf("auto-gen addr event mismatch (-want +got):\n%s", diff) + } + default: + t.Fatal("expected addr auto gen event") + } + } + + // Receive an RA with prefix in a PI. + e.InjectInbound(header.IPv6ProtocolNumber, raBufWithPI(llAddr2, 0, prefix, true, true, lifetimeSeconds, lifetimeSeconds)) + + addrBytes := []byte(subnet.ID()) + addr := tcpip.AddressWithPrefix{ + Address: tcpip.Address(header.AppendOpaqueInterfaceIdentifier(addrBytes[:header.IIDOffsetInIPv6Address], subnet, nicName, 0, secretKey)), + PrefixLen: 64, + } + expectAutoGenAddrEvent(addr, newAddr) + + // Simulate a DAD conflict after some time has passed. + time.Sleep(failureTimer) + if err := s.DupTentativeAddrDetected(nicID, addr.Address); err != nil { + t.Fatalf("s.DupTentativeAddrDetected(%d, %s): %s", nicID, addr.Address, err) + } + expectAutoGenAddrEvent(addr, invalidatedAddr) + select { + case e := <-ndpDisp.dadC: + if diff := checkDADEvent(e, nicID, addr.Address, false, nil); diff != "" { + t.Errorf("dad event mismatch (-want +got):\n%s", diff) + } + default: + t.Fatal("expected DAD event") + } + + // Let the next address resolve. + addr.Address = tcpip.Address(header.AppendOpaqueInterfaceIdentifier(addrBytes[:header.IIDOffsetInIPv6Address], subnet, nicName, 1, secretKey)) + expectAutoGenAddrEvent(addr, newAddr) + select { + case e := <-ndpDisp.dadC: + if diff := checkDADEvent(e, nicID, addr.Address, true, nil); diff != "" { + t.Errorf("dad event mismatch (-want +got):\n%s", diff) + } + case <-time.After(dadTransmits*retransmitTimer + defaultAsyncEventTimeout): + t.Fatal("timed out waiting for DAD event") + } + + // Address should be deprecated/invalidated after the lifetime expires. + // + // Note, the remaining lifetime is calculated from when the PI was first + // processed. Since we wait for some time before simulating a DAD conflict + // and more time for the new address to resolve, the new address is only + // expected to be valid for the remaining time. The DAD conflict should + // not have reset the lifetimes. + // + // We expect either just the invalidation event or the deprecation event + // followed by the invalidation event. + select { + case e := <-ndpDisp.autoGenAddrC: + if e.eventType == deprecatedAddr { + if diff := checkAutoGenAddrEvent(e, addr, deprecatedAddr); diff != "" { + t.Errorf("auto-gen addr event mismatch (-want +got):\n%s", diff) + } + + select { + case e := <-ndpDisp.autoGenAddrC: + if diff := checkAutoGenAddrEvent(e, addr, invalidatedAddr); diff != "" { + t.Errorf("auto-gen addr event mismatch (-want +got):\n%s", diff) + } + case <-time.After(defaultAsyncEventTimeout): + t.Fatal("timed out waiting for invalidated auto gen addr event after deprecation") + } + } else { + if diff := checkAutoGenAddrEvent(e, addr, invalidatedAddr); diff != "" { + t.Errorf("auto-gen addr event mismatch (-want +got):\n%s", diff) + } + } + case <-time.After(lifetimeSeconds*time.Second - failureTimer - dadTransmits*retransmitTimer + defaultAsyncEventTimeout): + t.Fatal("timed out waiting for auto gen addr event") + } +} + // TestNDPRecursiveDNSServerDispatch tests that we properly dispatch an event // to the integrator when an RA is received with the NDP Recursive DNS Server // option with at least one valid address. func TestNDPRecursiveDNSServerDispatch(t *testing.T) { - t.Parallel() - tests := []struct { name string opt header.NDPRecursiveDNSServer @@ -2835,11 +3358,7 @@ func TestNDPRecursiveDNSServerDispatch(t *testing.T) { } for _, test := range tests { - test := test - t.Run(test.name, func(t *testing.T) { - t.Parallel() - ndpDisp := ndpDispatcher{ // We do not expect more than a single RDNSS // event at any time for this test. @@ -2886,17 +3405,120 @@ func TestNDPRecursiveDNSServerDispatch(t *testing.T) { } } -// TestCleanupHostOnlyStateOnBecomingRouter tests that all discovered routers -// and prefixes, and non-linklocal auto-generated addresses are invalidated when -// a NIC becomes a router. -func TestCleanupHostOnlyStateOnBecomingRouter(t *testing.T) { - t.Parallel() +// TestNDPDNSSearchListDispatch tests that the integrator is informed when an +// NDP DNS Search List option is received with at least one domain name in the +// search list. +func TestNDPDNSSearchListDispatch(t *testing.T) { + const nicID = 1 + + ndpDisp := ndpDispatcher{ + dnsslC: make(chan ndpDNSSLEvent, 3), + } + e := channel.New(0, 1280, linkAddr1) + s := stack.New(stack.Options{ + NetworkProtocols: []stack.NetworkProtocol{ipv6.NewProtocol()}, + NDPConfigs: stack.NDPConfigurations{ + HandleRAs: true, + }, + NDPDisp: &ndpDisp, + }) + if err := s.CreateNIC(nicID, e); err != nil { + t.Fatalf("CreateNIC(%d, _) = %s", nicID, err) + } + + optSer := header.NDPOptionsSerializer{ + header.NDPDNSSearchList([]byte{ + 0, 0, + 0, 0, 0, 0, + 2, 'h', 'i', + 0, + }), + header.NDPDNSSearchList([]byte{ + 0, 0, + 0, 0, 0, 1, + 1, 'i', + 0, + 2, 'a', 'm', + 2, 'm', 'e', + 0, + }), + header.NDPDNSSearchList([]byte{ + 0, 0, + 0, 0, 1, 0, + 3, 'x', 'y', 'z', + 0, + 5, 'h', 'e', 'l', 'l', 'o', + 5, 'w', 'o', 'r', 'l', 'd', + 0, + 4, 't', 'h', 'i', 's', + 2, 'i', 's', + 1, 'a', + 4, 't', 'e', 's', 't', + 0, + }), + } + expected := []struct { + domainNames []string + lifetime time.Duration + }{ + { + domainNames: []string{ + "hi", + }, + lifetime: 0, + }, + { + domainNames: []string{ + "i", + "am.me", + }, + lifetime: time.Second, + }, + { + domainNames: []string{ + "xyz", + "hello.world", + "this.is.a.test", + }, + lifetime: 256 * time.Second, + }, + } + + e.InjectInbound(header.IPv6ProtocolNumber, raBufWithOpts(llAddr1, 0, optSer)) + + for i, expected := range expected { + select { + case dnssl := <-ndpDisp.dnsslC: + if dnssl.nicID != nicID { + t.Errorf("got %d-th dnssl nicID = %d, want = %d", i, dnssl.nicID, nicID) + } + if diff := cmp.Diff(dnssl.domainNames, expected.domainNames); diff != "" { + t.Errorf("%d-th dnssl domain names mismatch (-want +got):\n%s", i, diff) + } + if dnssl.lifetime != expected.lifetime { + t.Errorf("got %d-th dnssl lifetime = %s, want = %s", i, dnssl.lifetime, expected.lifetime) + } + default: + t.Fatal("expected a DNSSL event") + } + } + + // Should have no more DNSSL options. + select { + case <-ndpDisp.dnsslC: + t.Fatal("unexpectedly got a DNSSL event") + default: + } +} +// TestCleanupNDPState tests that all discovered routers and prefixes, and +// auto-generated addresses are invalidated when a NIC becomes a router. +func TestCleanupNDPState(t *testing.T) { const ( - lifetimeSeconds = 5 - maxEvents = 4 - nicID1 = 1 - nicID2 = 2 + lifetimeSeconds = 5 + maxRouterAndPrefixEvents = 4 + nicID1 = 1 + nicID2 = 2 ) prefix1, subnet1, e1Addr1 := prefixSubnetAddr(0, linkAddr1) @@ -2912,254 +3534,331 @@ func TestCleanupHostOnlyStateOnBecomingRouter(t *testing.T) { PrefixLen: 64, } - ndpDisp := ndpDispatcher{ - routerC: make(chan ndpRouterEvent, maxEvents), - rememberRouter: true, - prefixC: make(chan ndpPrefixEvent, maxEvents), - rememberPrefix: true, - autoGenAddrC: make(chan ndpAutoGenAddrEvent, maxEvents), - } - s := stack.New(stack.Options{ - NetworkProtocols: []stack.NetworkProtocol{ipv6.NewProtocol()}, - AutoGenIPv6LinkLocal: true, - NDPConfigs: stack.NDPConfigurations{ - HandleRAs: true, - DiscoverDefaultRouters: true, - DiscoverOnLinkPrefixes: true, - AutoGenGlobalAddresses: true, + tests := []struct { + name string + cleanupFn func(t *testing.T, s *stack.Stack) + keepAutoGenLinkLocal bool + maxAutoGenAddrEvents int + skipFinalAddrCheck bool + }{ + // A NIC should still keep its auto-generated link-local address when + // becoming a router. + { + name: "Enable forwarding", + cleanupFn: func(t *testing.T, s *stack.Stack) { + t.Helper() + s.SetForwarding(true) + }, + keepAutoGenLinkLocal: true, + maxAutoGenAddrEvents: 4, }, - NDPDisp: &ndpDisp, - }) - expectRouterEvent := func() (bool, ndpRouterEvent) { - select { - case e := <-ndpDisp.routerC: - return true, e - default: - } + // A NIC should cleanup all NDP state when it is disabled. + { + name: "Disable NIC", + cleanupFn: func(t *testing.T, s *stack.Stack) { + t.Helper() - return false, ndpRouterEvent{} - } + if err := s.DisableNIC(nicID1); err != nil { + t.Fatalf("s.DisableNIC(%d): %s", nicID1, err) + } + if err := s.DisableNIC(nicID2); err != nil { + t.Fatalf("s.DisableNIC(%d): %s", nicID2, err) + } + }, + keepAutoGenLinkLocal: false, + maxAutoGenAddrEvents: 6, + }, - expectPrefixEvent := func() (bool, ndpPrefixEvent) { - select { - case e := <-ndpDisp.prefixC: - return true, e - default: - } + // A NIC should cleanup all NDP state when it is removed. + { + name: "Remove NIC", + cleanupFn: func(t *testing.T, s *stack.Stack) { + t.Helper() - return false, ndpPrefixEvent{} + if err := s.RemoveNIC(nicID1); err != nil { + t.Fatalf("s.RemoveNIC(%d): %s", nicID1, err) + } + if err := s.RemoveNIC(nicID2); err != nil { + t.Fatalf("s.RemoveNIC(%d): %s", nicID2, err) + } + }, + keepAutoGenLinkLocal: false, + maxAutoGenAddrEvents: 6, + // The NICs are removed so we can't check their addresses after calling + // stopFn. + skipFinalAddrCheck: true, + }, } - expectAutoGenAddrEvent := func() (bool, ndpAutoGenAddrEvent) { - select { - case e := <-ndpDisp.autoGenAddrC: - return true, e - default: - } + for _, test := range tests { + t.Run(test.name, func(t *testing.T) { + ndpDisp := ndpDispatcher{ + routerC: make(chan ndpRouterEvent, maxRouterAndPrefixEvents), + rememberRouter: true, + prefixC: make(chan ndpPrefixEvent, maxRouterAndPrefixEvents), + rememberPrefix: true, + autoGenAddrC: make(chan ndpAutoGenAddrEvent, test.maxAutoGenAddrEvents), + } + s := stack.New(stack.Options{ + NetworkProtocols: []stack.NetworkProtocol{ipv6.NewProtocol()}, + AutoGenIPv6LinkLocal: true, + NDPConfigs: stack.NDPConfigurations{ + HandleRAs: true, + DiscoverDefaultRouters: true, + DiscoverOnLinkPrefixes: true, + AutoGenGlobalAddresses: true, + }, + NDPDisp: &ndpDisp, + }) - return false, ndpAutoGenAddrEvent{} - } + expectRouterEvent := func() (bool, ndpRouterEvent) { + select { + case e := <-ndpDisp.routerC: + return true, e + default: + } - e1 := channel.New(0, 1280, linkAddr1) - if err := s.CreateNIC(nicID1, e1); err != nil { - t.Fatalf("CreateNIC(%d, _) = %s", nicID1, err) - } - // We have other tests that make sure we receive the *correct* events - // on normal discovery of routers/prefixes, and auto-generated - // addresses. Here we just make sure we get an event and let other tests - // handle the correctness check. - expectAutoGenAddrEvent() + return false, ndpRouterEvent{} + } - e2 := channel.New(0, 1280, linkAddr2) - if err := s.CreateNIC(nicID2, e2); err != nil { - t.Fatalf("CreateNIC(%d, _) = %s", nicID2, err) - } - expectAutoGenAddrEvent() + expectPrefixEvent := func() (bool, ndpPrefixEvent) { + select { + case e := <-ndpDisp.prefixC: + return true, e + default: + } - // Receive RAs on NIC(1) and NIC(2) from default routers (llAddr3 and - // llAddr4) w/ PI (for prefix1 in RA from llAddr3 and prefix2 in RA from - // llAddr4) to discover multiple routers and prefixes, and auto-gen - // multiple addresses. + return false, ndpPrefixEvent{} + } - e1.InjectInbound(header.IPv6ProtocolNumber, raBufWithPI(llAddr3, lifetimeSeconds, prefix1, true, true, lifetimeSeconds, lifetimeSeconds)) - if ok, _ := expectRouterEvent(); !ok { - t.Errorf("expected router event for %s on NIC(%d)", llAddr3, nicID1) - } - if ok, _ := expectPrefixEvent(); !ok { - t.Errorf("expected prefix event for %s on NIC(%d)", prefix1, nicID1) - } - if ok, _ := expectAutoGenAddrEvent(); !ok { - t.Errorf("expected auto-gen addr event for %s on NIC(%d)", e1Addr1, nicID1) - } + expectAutoGenAddrEvent := func() (bool, ndpAutoGenAddrEvent) { + select { + case e := <-ndpDisp.autoGenAddrC: + return true, e + default: + } - e1.InjectInbound(header.IPv6ProtocolNumber, raBufWithPI(llAddr4, lifetimeSeconds, prefix2, true, true, lifetimeSeconds, lifetimeSeconds)) - if ok, _ := expectRouterEvent(); !ok { - t.Errorf("expected router event for %s on NIC(%d)", llAddr4, nicID1) - } - if ok, _ := expectPrefixEvent(); !ok { - t.Errorf("expected prefix event for %s on NIC(%d)", prefix2, nicID1) - } - if ok, _ := expectAutoGenAddrEvent(); !ok { - t.Errorf("expected auto-gen addr event for %s on NIC(%d)", e1Addr2, nicID1) - } + return false, ndpAutoGenAddrEvent{} + } - e2.InjectInbound(header.IPv6ProtocolNumber, raBufWithPI(llAddr3, lifetimeSeconds, prefix1, true, true, lifetimeSeconds, lifetimeSeconds)) - if ok, _ := expectRouterEvent(); !ok { - t.Errorf("expected router event for %s on NIC(%d)", llAddr3, nicID2) - } - if ok, _ := expectPrefixEvent(); !ok { - t.Errorf("expected prefix event for %s on NIC(%d)", prefix1, nicID2) - } - if ok, _ := expectAutoGenAddrEvent(); !ok { - t.Errorf("expected auto-gen addr event for %s on NIC(%d)", e1Addr2, nicID2) - } + e1 := channel.New(0, 1280, linkAddr1) + if err := s.CreateNIC(nicID1, e1); err != nil { + t.Fatalf("CreateNIC(%d, _) = %s", nicID1, err) + } + // We have other tests that make sure we receive the *correct* events + // on normal discovery of routers/prefixes, and auto-generated + // addresses. Here we just make sure we get an event and let other tests + // handle the correctness check. + expectAutoGenAddrEvent() + + e2 := channel.New(0, 1280, linkAddr2) + if err := s.CreateNIC(nicID2, e2); err != nil { + t.Fatalf("CreateNIC(%d, _) = %s", nicID2, err) + } + expectAutoGenAddrEvent() - e2.InjectInbound(header.IPv6ProtocolNumber, raBufWithPI(llAddr4, lifetimeSeconds, prefix2, true, true, lifetimeSeconds, lifetimeSeconds)) - if ok, _ := expectRouterEvent(); !ok { - t.Errorf("expected router event for %s on NIC(%d)", llAddr4, nicID2) - } - if ok, _ := expectPrefixEvent(); !ok { - t.Errorf("expected prefix event for %s on NIC(%d)", prefix2, nicID2) - } - if ok, _ := expectAutoGenAddrEvent(); !ok { - t.Errorf("expected auto-gen addr event for %s on NIC(%d)", e2Addr2, nicID2) - } + // Receive RAs on NIC(1) and NIC(2) from default routers (llAddr3 and + // llAddr4) w/ PI (for prefix1 in RA from llAddr3 and prefix2 in RA from + // llAddr4) to discover multiple routers and prefixes, and auto-gen + // multiple addresses. - // We should have the auto-generated addresses added. - nicinfo := s.NICInfo() - nic1Addrs := nicinfo[nicID1].ProtocolAddresses - nic2Addrs := nicinfo[nicID2].ProtocolAddresses - if !containsV6Addr(nic1Addrs, llAddrWithPrefix1) { - t.Errorf("missing %s from the list of addresses for NIC(%d): %+v", llAddrWithPrefix1, nicID1, nic1Addrs) - } - if !containsV6Addr(nic1Addrs, e1Addr1) { - t.Errorf("missing %s from the list of addresses for NIC(%d): %+v", e1Addr1, nicID1, nic1Addrs) - } - if !containsV6Addr(nic1Addrs, e1Addr2) { - t.Errorf("missing %s from the list of addresses for NIC(%d): %+v", e1Addr2, nicID1, nic1Addrs) - } - if !containsV6Addr(nic2Addrs, llAddrWithPrefix2) { - t.Errorf("missing %s from the list of addresses for NIC(%d): %+v", llAddrWithPrefix2, nicID2, nic2Addrs) - } - if !containsV6Addr(nic2Addrs, e2Addr1) { - t.Errorf("missing %s from the list of addresses for NIC(%d): %+v", e2Addr1, nicID2, nic2Addrs) - } - if !containsV6Addr(nic2Addrs, e2Addr2) { - t.Errorf("missing %s from the list of addresses for NIC(%d): %+v", e2Addr2, nicID2, nic2Addrs) - } + e1.InjectInbound(header.IPv6ProtocolNumber, raBufWithPI(llAddr3, lifetimeSeconds, prefix1, true, true, lifetimeSeconds, lifetimeSeconds)) + if ok, _ := expectRouterEvent(); !ok { + t.Errorf("expected router event for %s on NIC(%d)", llAddr3, nicID1) + } + if ok, _ := expectPrefixEvent(); !ok { + t.Errorf("expected prefix event for %s on NIC(%d)", prefix1, nicID1) + } + if ok, _ := expectAutoGenAddrEvent(); !ok { + t.Errorf("expected auto-gen addr event for %s on NIC(%d)", e1Addr1, nicID1) + } - // We can't proceed any further if we already failed the test (missing - // some discovery/auto-generated address events or addresses). - if t.Failed() { - t.FailNow() - } + e1.InjectInbound(header.IPv6ProtocolNumber, raBufWithPI(llAddr4, lifetimeSeconds, prefix2, true, true, lifetimeSeconds, lifetimeSeconds)) + if ok, _ := expectRouterEvent(); !ok { + t.Errorf("expected router event for %s on NIC(%d)", llAddr4, nicID1) + } + if ok, _ := expectPrefixEvent(); !ok { + t.Errorf("expected prefix event for %s on NIC(%d)", prefix2, nicID1) + } + if ok, _ := expectAutoGenAddrEvent(); !ok { + t.Errorf("expected auto-gen addr event for %s on NIC(%d)", e1Addr2, nicID1) + } - s.SetForwarding(true) + e2.InjectInbound(header.IPv6ProtocolNumber, raBufWithPI(llAddr3, lifetimeSeconds, prefix1, true, true, lifetimeSeconds, lifetimeSeconds)) + if ok, _ := expectRouterEvent(); !ok { + t.Errorf("expected router event for %s on NIC(%d)", llAddr3, nicID2) + } + if ok, _ := expectPrefixEvent(); !ok { + t.Errorf("expected prefix event for %s on NIC(%d)", prefix1, nicID2) + } + if ok, _ := expectAutoGenAddrEvent(); !ok { + t.Errorf("expected auto-gen addr event for %s on NIC(%d)", e1Addr2, nicID2) + } - // Collect invalidation events after becoming a router - gotRouterEvents := make(map[ndpRouterEvent]int) - for i := 0; i < maxEvents; i++ { - ok, e := expectRouterEvent() - if !ok { - t.Errorf("expected %d router events after becoming a router; got = %d", maxEvents, i) - break - } - gotRouterEvents[e]++ - } - gotPrefixEvents := make(map[ndpPrefixEvent]int) - for i := 0; i < maxEvents; i++ { - ok, e := expectPrefixEvent() - if !ok { - t.Errorf("expected %d prefix events after becoming a router; got = %d", maxEvents, i) - break - } - gotPrefixEvents[e]++ - } - gotAutoGenAddrEvents := make(map[ndpAutoGenAddrEvent]int) - for i := 0; i < maxEvents; i++ { - ok, e := expectAutoGenAddrEvent() - if !ok { - t.Errorf("expected %d auto-generated address events after becoming a router; got = %d", maxEvents, i) - break - } - gotAutoGenAddrEvents[e]++ - } + e2.InjectInbound(header.IPv6ProtocolNumber, raBufWithPI(llAddr4, lifetimeSeconds, prefix2, true, true, lifetimeSeconds, lifetimeSeconds)) + if ok, _ := expectRouterEvent(); !ok { + t.Errorf("expected router event for %s on NIC(%d)", llAddr4, nicID2) + } + if ok, _ := expectPrefixEvent(); !ok { + t.Errorf("expected prefix event for %s on NIC(%d)", prefix2, nicID2) + } + if ok, _ := expectAutoGenAddrEvent(); !ok { + t.Errorf("expected auto-gen addr event for %s on NIC(%d)", e2Addr2, nicID2) + } - // No need to proceed any further if we already failed the test (missing - // some invalidation events). - if t.Failed() { - t.FailNow() - } + // We should have the auto-generated addresses added. + nicinfo := s.NICInfo() + nic1Addrs := nicinfo[nicID1].ProtocolAddresses + nic2Addrs := nicinfo[nicID2].ProtocolAddresses + if !containsV6Addr(nic1Addrs, llAddrWithPrefix1) { + t.Errorf("missing %s from the list of addresses for NIC(%d): %+v", llAddrWithPrefix1, nicID1, nic1Addrs) + } + if !containsV6Addr(nic1Addrs, e1Addr1) { + t.Errorf("missing %s from the list of addresses for NIC(%d): %+v", e1Addr1, nicID1, nic1Addrs) + } + if !containsV6Addr(nic1Addrs, e1Addr2) { + t.Errorf("missing %s from the list of addresses for NIC(%d): %+v", e1Addr2, nicID1, nic1Addrs) + } + if !containsV6Addr(nic2Addrs, llAddrWithPrefix2) { + t.Errorf("missing %s from the list of addresses for NIC(%d): %+v", llAddrWithPrefix2, nicID2, nic2Addrs) + } + if !containsV6Addr(nic2Addrs, e2Addr1) { + t.Errorf("missing %s from the list of addresses for NIC(%d): %+v", e2Addr1, nicID2, nic2Addrs) + } + if !containsV6Addr(nic2Addrs, e2Addr2) { + t.Errorf("missing %s from the list of addresses for NIC(%d): %+v", e2Addr2, nicID2, nic2Addrs) + } - expectedRouterEvents := map[ndpRouterEvent]int{ - {nicID: nicID1, addr: llAddr3, discovered: false}: 1, - {nicID: nicID1, addr: llAddr4, discovered: false}: 1, - {nicID: nicID2, addr: llAddr3, discovered: false}: 1, - {nicID: nicID2, addr: llAddr4, discovered: false}: 1, - } - if diff := cmp.Diff(expectedRouterEvents, gotRouterEvents); diff != "" { - t.Errorf("router events mismatch (-want +got):\n%s", diff) - } - expectedPrefixEvents := map[ndpPrefixEvent]int{ - {nicID: nicID1, prefix: subnet1, discovered: false}: 1, - {nicID: nicID1, prefix: subnet2, discovered: false}: 1, - {nicID: nicID2, prefix: subnet1, discovered: false}: 1, - {nicID: nicID2, prefix: subnet2, discovered: false}: 1, - } - if diff := cmp.Diff(expectedPrefixEvents, gotPrefixEvents); diff != "" { - t.Errorf("prefix events mismatch (-want +got):\n%s", diff) - } - expectedAutoGenAddrEvents := map[ndpAutoGenAddrEvent]int{ - {nicID: nicID1, addr: e1Addr1, eventType: invalidatedAddr}: 1, - {nicID: nicID1, addr: e1Addr2, eventType: invalidatedAddr}: 1, - {nicID: nicID2, addr: e2Addr1, eventType: invalidatedAddr}: 1, - {nicID: nicID2, addr: e2Addr2, eventType: invalidatedAddr}: 1, - } - if diff := cmp.Diff(expectedAutoGenAddrEvents, gotAutoGenAddrEvents); diff != "" { - t.Errorf("auto-generated address events mismatch (-want +got):\n%s", diff) - } + // We can't proceed any further if we already failed the test (missing + // some discovery/auto-generated address events or addresses). + if t.Failed() { + t.FailNow() + } - // Make sure the auto-generated addresses got removed. - nicinfo = s.NICInfo() - nic1Addrs = nicinfo[nicID1].ProtocolAddresses - nic2Addrs = nicinfo[nicID2].ProtocolAddresses - if !containsV6Addr(nic1Addrs, llAddrWithPrefix1) { - t.Errorf("missing %s from the list of addresses for NIC(%d): %+v", llAddrWithPrefix1, nicID1, nic1Addrs) - } - if containsV6Addr(nic1Addrs, e1Addr1) { - t.Errorf("still have %s in the list of addresses for NIC(%d): %+v", e1Addr1, nicID1, nic1Addrs) - } - if containsV6Addr(nic1Addrs, e1Addr2) { - t.Errorf("still have %s in the list of addresses for NIC(%d): %+v", e1Addr2, nicID1, nic1Addrs) - } - if !containsV6Addr(nic2Addrs, llAddrWithPrefix2) { - t.Errorf("missing %s from the list of addresses for NIC(%d): %+v", llAddrWithPrefix2, nicID2, nic2Addrs) - } - if containsV6Addr(nic2Addrs, e2Addr1) { - t.Errorf("still have %s in the list of addresses for NIC(%d): %+v", e2Addr1, nicID2, nic2Addrs) - } - if containsV6Addr(nic2Addrs, e2Addr2) { - t.Errorf("still have %s in the list of addresses for NIC(%d): %+v", e2Addr2, nicID2, nic2Addrs) - } + test.cleanupFn(t, s) - // Should not get any more events (invalidation timers should have been - // cancelled when we transitioned into a router). - time.Sleep(lifetimeSeconds*time.Second + defaultTimeout) - select { - case <-ndpDisp.routerC: - t.Error("unexpected router event") - default: - } - select { - case <-ndpDisp.prefixC: - t.Error("unexpected prefix event") - default: - } - select { - case <-ndpDisp.autoGenAddrC: - t.Error("unexpected auto-generated address event") - default: + // Collect invalidation events after having NDP state cleaned up. + gotRouterEvents := make(map[ndpRouterEvent]int) + for i := 0; i < maxRouterAndPrefixEvents; i++ { + ok, e := expectRouterEvent() + if !ok { + t.Errorf("expected %d router events after becoming a router; got = %d", maxRouterAndPrefixEvents, i) + break + } + gotRouterEvents[e]++ + } + gotPrefixEvents := make(map[ndpPrefixEvent]int) + for i := 0; i < maxRouterAndPrefixEvents; i++ { + ok, e := expectPrefixEvent() + if !ok { + t.Errorf("expected %d prefix events after becoming a router; got = %d", maxRouterAndPrefixEvents, i) + break + } + gotPrefixEvents[e]++ + } + gotAutoGenAddrEvents := make(map[ndpAutoGenAddrEvent]int) + for i := 0; i < test.maxAutoGenAddrEvents; i++ { + ok, e := expectAutoGenAddrEvent() + if !ok { + t.Errorf("expected %d auto-generated address events after becoming a router; got = %d", test.maxAutoGenAddrEvents, i) + break + } + gotAutoGenAddrEvents[e]++ + } + + // No need to proceed any further if we already failed the test (missing + // some invalidation events). + if t.Failed() { + t.FailNow() + } + + expectedRouterEvents := map[ndpRouterEvent]int{ + {nicID: nicID1, addr: llAddr3, discovered: false}: 1, + {nicID: nicID1, addr: llAddr4, discovered: false}: 1, + {nicID: nicID2, addr: llAddr3, discovered: false}: 1, + {nicID: nicID2, addr: llAddr4, discovered: false}: 1, + } + if diff := cmp.Diff(expectedRouterEvents, gotRouterEvents); diff != "" { + t.Errorf("router events mismatch (-want +got):\n%s", diff) + } + expectedPrefixEvents := map[ndpPrefixEvent]int{ + {nicID: nicID1, prefix: subnet1, discovered: false}: 1, + {nicID: nicID1, prefix: subnet2, discovered: false}: 1, + {nicID: nicID2, prefix: subnet1, discovered: false}: 1, + {nicID: nicID2, prefix: subnet2, discovered: false}: 1, + } + if diff := cmp.Diff(expectedPrefixEvents, gotPrefixEvents); diff != "" { + t.Errorf("prefix events mismatch (-want +got):\n%s", diff) + } + expectedAutoGenAddrEvents := map[ndpAutoGenAddrEvent]int{ + {nicID: nicID1, addr: e1Addr1, eventType: invalidatedAddr}: 1, + {nicID: nicID1, addr: e1Addr2, eventType: invalidatedAddr}: 1, + {nicID: nicID2, addr: e2Addr1, eventType: invalidatedAddr}: 1, + {nicID: nicID2, addr: e2Addr2, eventType: invalidatedAddr}: 1, + } + + if !test.keepAutoGenLinkLocal { + expectedAutoGenAddrEvents[ndpAutoGenAddrEvent{nicID: nicID1, addr: llAddrWithPrefix1, eventType: invalidatedAddr}] = 1 + expectedAutoGenAddrEvents[ndpAutoGenAddrEvent{nicID: nicID2, addr: llAddrWithPrefix2, eventType: invalidatedAddr}] = 1 + } + + if diff := cmp.Diff(expectedAutoGenAddrEvents, gotAutoGenAddrEvents); diff != "" { + t.Errorf("auto-generated address events mismatch (-want +got):\n%s", diff) + } + + if !test.skipFinalAddrCheck { + // Make sure the auto-generated addresses got removed. + nicinfo = s.NICInfo() + nic1Addrs = nicinfo[nicID1].ProtocolAddresses + nic2Addrs = nicinfo[nicID2].ProtocolAddresses + if containsV6Addr(nic1Addrs, llAddrWithPrefix1) != test.keepAutoGenLinkLocal { + if test.keepAutoGenLinkLocal { + t.Errorf("missing %s from the list of addresses for NIC(%d): %+v", llAddrWithPrefix1, nicID1, nic1Addrs) + } else { + t.Errorf("still have %s in the list of addresses for NIC(%d): %+v", llAddrWithPrefix1, nicID1, nic1Addrs) + } + } + if containsV6Addr(nic1Addrs, e1Addr1) { + t.Errorf("still have %s in the list of addresses for NIC(%d): %+v", e1Addr1, nicID1, nic1Addrs) + } + if containsV6Addr(nic1Addrs, e1Addr2) { + t.Errorf("still have %s in the list of addresses for NIC(%d): %+v", e1Addr2, nicID1, nic1Addrs) + } + if containsV6Addr(nic2Addrs, llAddrWithPrefix2) != test.keepAutoGenLinkLocal { + if test.keepAutoGenLinkLocal { + t.Errorf("missing %s from the list of addresses for NIC(%d): %+v", llAddrWithPrefix2, nicID2, nic2Addrs) + } else { + t.Errorf("still have %s in the list of addresses for NIC(%d): %+v", llAddrWithPrefix2, nicID2, nic2Addrs) + } + } + if containsV6Addr(nic2Addrs, e2Addr1) { + t.Errorf("still have %s in the list of addresses for NIC(%d): %+v", e2Addr1, nicID2, nic2Addrs) + } + if containsV6Addr(nic2Addrs, e2Addr2) { + t.Errorf("still have %s in the list of addresses for NIC(%d): %+v", e2Addr2, nicID2, nic2Addrs) + } + } + + // Should not get any more events (invalidation timers should have been + // cancelled when the NDP state was cleaned up). + time.Sleep(lifetimeSeconds*time.Second + defaultTimeout) + select { + case <-ndpDisp.routerC: + t.Error("unexpected router event") + default: + } + select { + case <-ndpDisp.prefixC: + t.Error("unexpected prefix event") + default: + } + select { + case <-ndpDisp.autoGenAddrC: + t.Error("unexpected auto-generated address event") + default: + } + }) } } @@ -3257,10 +3956,15 @@ func TestDHCPv6ConfigurationFromNDPDA(t *testing.T) { // TestRouterSolicitation tests the initial Router Solicitations that are sent // when a NIC newly becomes enabled. func TestRouterSolicitation(t *testing.T) { - t.Parallel() + const nicID = 1 tests := []struct { name string + linkHeaderLen uint16 + linkAddr tcpip.LinkAddress + nicAddr tcpip.Address + expectedSrcAddr tcpip.Address + expectedNDPOpts []header.NDPOption maxRtrSolicit uint8 rtrSolicitInt time.Duration effectiveRtrSolicitInt time.Duration @@ -3268,31 +3972,54 @@ func TestRouterSolicitation(t *testing.T) { effectiveMaxRtrSolicitDelay time.Duration }{ { - name: "Single RS with delay", + name: "Single RS with 2s delay and interval", + expectedSrcAddr: header.IPv6Any, maxRtrSolicit: 1, - rtrSolicitInt: time.Second, - effectiveRtrSolicitInt: time.Second, - maxRtrSolicitDelay: time.Second, - effectiveMaxRtrSolicitDelay: time.Second, + rtrSolicitInt: 2 * time.Second, + effectiveRtrSolicitInt: 2 * time.Second, + maxRtrSolicitDelay: 2 * time.Second, + effectiveMaxRtrSolicitDelay: 2 * time.Second, + }, + { + name: "Single RS with 4s delay and interval", + expectedSrcAddr: header.IPv6Any, + maxRtrSolicit: 1, + rtrSolicitInt: 4 * time.Second, + effectiveRtrSolicitInt: 4 * time.Second, + maxRtrSolicitDelay: 4 * time.Second, + effectiveMaxRtrSolicitDelay: 4 * time.Second, }, { name: "Two RS with delay", + linkHeaderLen: 1, + nicAddr: llAddr1, + expectedSrcAddr: llAddr1, maxRtrSolicit: 2, - rtrSolicitInt: time.Second, - effectiveRtrSolicitInt: time.Second, + rtrSolicitInt: 2 * time.Second, + effectiveRtrSolicitInt: 2 * time.Second, maxRtrSolicitDelay: 500 * time.Millisecond, effectiveMaxRtrSolicitDelay: 500 * time.Millisecond, }, { - name: "Single RS without delay", + name: "Single RS without delay", + linkHeaderLen: 2, + linkAddr: linkAddr1, + nicAddr: llAddr1, + expectedSrcAddr: llAddr1, + expectedNDPOpts: []header.NDPOption{ + header.NDPSourceLinkLayerAddressOption(linkAddr1), + }, maxRtrSolicit: 1, - rtrSolicitInt: time.Second, - effectiveRtrSolicitInt: time.Second, + rtrSolicitInt: 2 * time.Second, + effectiveRtrSolicitInt: 2 * time.Second, maxRtrSolicitDelay: 0, effectiveMaxRtrSolicitDelay: 0, }, { name: "Two RS without delay and invalid zero interval", + linkHeaderLen: 3, + linkAddr: linkAddr1, + expectedSrcAddr: header.IPv6Any, maxRtrSolicit: 2, rtrSolicitInt: 0, effectiveRtrSolicitInt: 4 * time.Second, @@ -3301,6 +4028,8 @@ func TestRouterSolicitation(t *testing.T) { }, { name: "Three RS without delay", + linkAddr: linkAddr1, + expectedSrcAddr: header.IPv6Any, maxRtrSolicit: 3, rtrSolicitInt: 500 * time.Millisecond, effectiveRtrSolicitInt: 500 * time.Millisecond, @@ -3309,6 +4038,8 @@ func TestRouterSolicitation(t *testing.T) { }, { name: "Two RS with invalid negative delay", + linkAddr: linkAddr1, + expectedSrcAddr: header.IPv6Any, maxRtrSolicit: 2, rtrSolicitInt: time.Second, effectiveRtrSolicitInt: time.Second, @@ -3330,11 +4061,16 @@ func TestRouterSolicitation(t *testing.T) { t.Run(test.name, func(t *testing.T) { t.Parallel() - e := channel.New(int(test.maxRtrSolicit), 1280, linkAddr1) - e.LinkEPCapabilities |= stack.CapabilityResolutionRequired + + e := channelLinkWithHeaderLength{ + Endpoint: channel.New(int(test.maxRtrSolicit), 1280, test.linkAddr), + headerLength: test.linkHeaderLen, + } + e.Endpoint.LinkEPCapabilities |= stack.CapabilityResolutionRequired waitForPkt := func(timeout time.Duration) { t.Helper() - ctx, _ := context.WithTimeout(context.Background(), timeout) + ctx, cancel := context.WithTimeout(context.Background(), timeout) + defer cancel() p, ok := e.ReadContext(ctx) if !ok { t.Fatal("timed out waiting for packet") @@ -3352,15 +4088,20 @@ func TestRouterSolicitation(t *testing.T) { checker.IPv6(t, p.Pkt.Header.View(), - checker.SrcAddr(header.IPv6Any), + checker.SrcAddr(test.expectedSrcAddr), checker.DstAddr(header.IPv6AllRoutersMulticastAddress), checker.TTL(header.NDPHopLimit), - checker.NDPRS(), + checker.NDPRS(checker.NDPRSOptions(test.expectedNDPOpts)), ) + + if l, want := p.Pkt.Header.AvailableLength(), int(test.linkHeaderLen); l != want { + t.Errorf("got p.Pkt.Header.AvailableLength() = %d; want = %d", l, want) + } } waitForNothing := func(timeout time.Duration) { t.Helper() - ctx, _ := context.WithTimeout(context.Background(), timeout) + ctx, cancel := context.WithTimeout(context.Background(), timeout) + defer cancel() if _, ok := e.ReadContext(ctx); ok { t.Fatal("unexpectedly got a packet") } @@ -3373,27 +4114,37 @@ func TestRouterSolicitation(t *testing.T) { MaxRtrSolicitationDelay: test.maxRtrSolicitDelay, }, }) - if err := s.CreateNIC(1, e); err != nil { - t.Fatalf("CreateNIC(1) = %s", err) + if err := s.CreateNIC(nicID, &e); err != nil { + t.Fatalf("CreateNIC(%d, _) = %s", nicID, err) + } + + if addr := test.nicAddr; addr != "" { + if err := s.AddAddress(nicID, header.IPv6ProtocolNumber, addr); err != nil { + t.Fatalf("AddAddress(%d, %d, %s) = %s", nicID, header.IPv6ProtocolNumber, addr, err) + } } - // Make sure each RS got sent at the right - // times. + // Make sure each RS is sent at the right time. remaining := test.maxRtrSolicit if remaining > 0 { waitForPkt(test.effectiveMaxRtrSolicitDelay + defaultAsyncEventTimeout) remaining-- } + for ; remaining > 0; remaining-- { - waitForNothing(test.effectiveRtrSolicitInt - defaultTimeout) - waitForPkt(defaultAsyncEventTimeout) + if test.effectiveRtrSolicitInt > defaultAsyncEventTimeout { + waitForNothing(test.effectiveRtrSolicitInt - defaultAsyncEventTimeout) + waitForPkt(2 * defaultAsyncEventTimeout) + } else { + waitForPkt(test.effectiveRtrSolicitInt * defaultAsyncEventTimeout) + } } // Make sure no more RS. if test.effectiveRtrSolicitInt > test.effectiveMaxRtrSolicitDelay { - waitForNothing(test.effectiveRtrSolicitInt + defaultTimeout) + waitForNothing(test.effectiveRtrSolicitInt + defaultAsyncEventTimeout) } else { - waitForNothing(test.effectiveMaxRtrSolicitDelay + defaultTimeout) + waitForNothing(test.effectiveMaxRtrSolicitDelay + defaultAsyncEventTimeout) } // Make sure the counter got properly @@ -3406,77 +4157,153 @@ func TestRouterSolicitation(t *testing.T) { }) } -// TestStopStartSolicitingRouters tests that when forwarding is enabled or -// disabled, router solicitations are stopped or started, respecitively. func TestStopStartSolicitingRouters(t *testing.T) { - t.Parallel() - + const nicID = 1 + const delay = 0 const interval = 500 * time.Millisecond - const delay = time.Second const maxRtrSolicitations = 3 - e := channel.New(maxRtrSolicitations, 1280, linkAddr1) - waitForPkt := func(timeout time.Duration) { - t.Helper() - ctx, _ := context.WithTimeout(context.Background(), timeout) - p, ok := e.ReadContext(ctx) - if !ok { - t.Fatal("timed out waiting for packet") - return - } - if p.Proto != header.IPv6ProtocolNumber { - t.Fatalf("got Proto = %d, want = %d", p.Proto, header.IPv6ProtocolNumber) - } - checker.IPv6(t, p.Pkt.Header.View(), - checker.SrcAddr(header.IPv6Any), - checker.DstAddr(header.IPv6AllRoutersMulticastAddress), - checker.TTL(header.NDPHopLimit), - checker.NDPRS()) - } - s := stack.New(stack.Options{ - NetworkProtocols: []stack.NetworkProtocol{ipv6.NewProtocol()}, - NDPConfigs: stack.NDPConfigurations{ - MaxRtrSolicitations: maxRtrSolicitations, - RtrSolicitationInterval: interval, - MaxRtrSolicitationDelay: delay, + tests := []struct { + name string + startFn func(t *testing.T, s *stack.Stack) + // first is used to tell stopFn that it is being called for the first time + // after router solicitations were last enabled. + stopFn func(t *testing.T, s *stack.Stack, first bool) + }{ + // Tests that when forwarding is enabled or disabled, router solicitations + // are stopped or started, respectively. + { + name: "Enable and disable forwarding", + startFn: func(t *testing.T, s *stack.Stack) { + t.Helper() + s.SetForwarding(false) + }, + stopFn: func(t *testing.T, s *stack.Stack, _ bool) { + t.Helper() + s.SetForwarding(true) + }, }, - }) - if err := s.CreateNIC(1, e); err != nil { - t.Fatalf("CreateNIC(1) = %s", err) - } - // Enable forwarding which should stop router solicitations. - s.SetForwarding(true) - ctx, _ := context.WithTimeout(context.Background(), delay+defaultTimeout) - if _, ok := e.ReadContext(ctx); ok { - // A single RS may have been sent before forwarding was enabled. - ctx, _ = context.WithTimeout(context.Background(), interval+defaultTimeout) - if _, ok = e.ReadContext(ctx); ok { - t.Fatal("Should not have sent more than one RS message") - } - } + // Tests that when a NIC is enabled or disabled, router solicitations + // are started or stopped, respectively. + { + name: "Enable and disable NIC", + startFn: func(t *testing.T, s *stack.Stack) { + t.Helper() - // Enabling forwarding again should do nothing. - s.SetForwarding(true) - ctx, _ = context.WithTimeout(context.Background(), delay+defaultTimeout) - if _, ok := e.ReadContext(ctx); ok { - t.Fatal("unexpectedly got a packet after becoming a router") - } + if err := s.EnableNIC(nicID); err != nil { + t.Fatalf("s.EnableNIC(%d): %s", nicID, err) + } + }, + stopFn: func(t *testing.T, s *stack.Stack, _ bool) { + t.Helper() - // Disable forwarding which should start router solicitations. - s.SetForwarding(false) - waitForPkt(delay + defaultAsyncEventTimeout) - waitForPkt(interval + defaultAsyncEventTimeout) - waitForPkt(interval + defaultAsyncEventTimeout) - ctx, _ = context.WithTimeout(context.Background(), interval+defaultTimeout) - if _, ok := e.ReadContext(ctx); ok { - t.Fatal("unexpectedly got an extra packet after sending out the expected RSs") + if err := s.DisableNIC(nicID); err != nil { + t.Fatalf("s.DisableNIC(%d): %s", nicID, err) + } + }, + }, + + // Tests that when a NIC is removed, router solicitations are stopped. We + // cannot start router solications on a removed NIC. + { + name: "Remove NIC", + stopFn: func(t *testing.T, s *stack.Stack, first bool) { + t.Helper() + + // Only try to remove the NIC the first time stopFn is called since it's + // impossible to remove an already removed NIC. + if !first { + return + } + + if err := s.RemoveNIC(nicID); err != nil { + t.Fatalf("s.RemoveNIC(%d): %s", nicID, err) + } + }, + }, } - // Disabling forwarding again should do nothing. - s.SetForwarding(false) - ctx, _ = context.WithTimeout(context.Background(), delay+defaultTimeout) - if _, ok := e.ReadContext(ctx); ok { - t.Fatal("unexpectedly got a packet after becoming a router") + for _, test := range tests { + t.Run(test.name, func(t *testing.T) { + e := channel.New(maxRtrSolicitations, 1280, linkAddr1) + waitForPkt := func(timeout time.Duration) { + t.Helper() + + ctx, cancel := context.WithTimeout(context.Background(), timeout) + defer cancel() + p, ok := e.ReadContext(ctx) + if !ok { + t.Fatal("timed out waiting for packet") + } + + if p.Proto != header.IPv6ProtocolNumber { + t.Fatalf("got Proto = %d, want = %d", p.Proto, header.IPv6ProtocolNumber) + } + checker.IPv6(t, p.Pkt.Header.View(), + checker.SrcAddr(header.IPv6Any), + checker.DstAddr(header.IPv6AllRoutersMulticastAddress), + checker.TTL(header.NDPHopLimit), + checker.NDPRS()) + } + s := stack.New(stack.Options{ + NetworkProtocols: []stack.NetworkProtocol{ipv6.NewProtocol()}, + NDPConfigs: stack.NDPConfigurations{ + MaxRtrSolicitations: maxRtrSolicitations, + RtrSolicitationInterval: interval, + MaxRtrSolicitationDelay: delay, + }, + }) + if err := s.CreateNIC(nicID, e); err != nil { + t.Fatalf("CreateNIC(%d, _) = %s", nicID, err) + } + + // Stop soliciting routers. + test.stopFn(t, s, true /* first */) + ctx, cancel := context.WithTimeout(context.Background(), delay+defaultAsyncEventTimeout) + defer cancel() + if _, ok := e.ReadContext(ctx); ok { + // A single RS may have been sent before solicitations were stopped. + ctx, cancel := context.WithTimeout(context.Background(), interval+defaultAsyncEventTimeout) + defer cancel() + if _, ok = e.ReadContext(ctx); ok { + t.Fatal("should not have sent more than one RS message") + } + } + + // Stopping router solicitations after it has already been stopped should + // do nothing. + test.stopFn(t, s, false /* first */) + ctx, cancel = context.WithTimeout(context.Background(), delay+defaultAsyncEventTimeout) + defer cancel() + if _, ok := e.ReadContext(ctx); ok { + t.Fatal("unexpectedly got a packet after router solicitation has been stopepd") + } + + // If test.startFn is nil, there is no way to restart router solications. + if test.startFn == nil { + return + } + + // Start soliciting routers. + test.startFn(t, s) + waitForPkt(delay + defaultAsyncEventTimeout) + waitForPkt(interval + defaultAsyncEventTimeout) + waitForPkt(interval + defaultAsyncEventTimeout) + ctx, cancel = context.WithTimeout(context.Background(), interval+defaultAsyncEventTimeout) + defer cancel() + if _, ok := e.ReadContext(ctx); ok { + t.Fatal("unexpectedly got an extra packet after sending out the expected RSs") + } + + // Starting router solicitations after it has already completed should do + // nothing. + test.startFn(t, s) + ctx, cancel = context.WithTimeout(context.Background(), delay+defaultAsyncEventTimeout) + defer cancel() + if _, ok := e.ReadContext(ctx); ok { + t.Fatal("unexpectedly got a packet after finishing router solicitations") + } + }) } } diff --git a/pkg/tcpip/stack/nic.go b/pkg/tcpip/stack/nic.go index 78d451cca..016dbe15e 100644 --- a/pkg/tcpip/stack/nic.go +++ b/pkg/tcpip/stack/nic.go @@ -15,7 +15,7 @@ package stack import ( - "log" + "fmt" "reflect" "sort" "strings" @@ -27,6 +27,14 @@ import ( "gvisor.dev/gvisor/pkg/tcpip/header" ) +var ipv4BroadcastAddr = tcpip.ProtocolAddress{ + Protocol: header.IPv4ProtocolNumber, + AddressWithPrefix: tcpip.AddressWithPrefix{ + Address: header.IPv4Broadcast, + PrefixLen: 8 * header.IPv4AddressSize, + }, +} + // NIC represents a "network interface card" to which the networking stack is // attached. type NIC struct { @@ -46,7 +54,7 @@ type NIC struct { primary map[tcpip.NetworkProtocolNumber][]*referencedNetworkEndpoint endpoints map[NetworkEndpointID]*referencedNetworkEndpoint addressRanges []tcpip.Subnet - mcastJoins map[NetworkEndpointID]int32 + mcastJoins map[NetworkEndpointID]uint32 // packetEPs is protected by mu, but the contained PacketEndpoint // values are not. packetEPs map[tcpip.NetworkProtocolNumber][]PacketEndpoint @@ -113,15 +121,15 @@ func newNIC(stack *Stack, id tcpip.NICID, name string, ep LinkEndpoint, ctx NICC } nic.mu.primary = make(map[tcpip.NetworkProtocolNumber][]*referencedNetworkEndpoint) nic.mu.endpoints = make(map[NetworkEndpointID]*referencedNetworkEndpoint) - nic.mu.mcastJoins = make(map[NetworkEndpointID]int32) + nic.mu.mcastJoins = make(map[NetworkEndpointID]uint32) nic.mu.packetEPs = make(map[tcpip.NetworkProtocolNumber][]PacketEndpoint) nic.mu.ndp = ndpState{ - nic: nic, - configs: stack.ndpConfigs, - dad: make(map[tcpip.Address]dadState), - defaultRouters: make(map[tcpip.Address]defaultRouterState), - onLinkPrefixes: make(map[tcpip.Subnet]onLinkPrefixState), - autoGenAddresses: make(map[tcpip.Address]autoGenAddressState), + nic: nic, + configs: stack.ndpConfigs, + dad: make(map[tcpip.Address]dadState), + defaultRouters: make(map[tcpip.Address]defaultRouterState), + onLinkPrefixes: make(map[tcpip.Subnet]onLinkPrefixState), + slaacPrefixes: make(map[tcpip.Subnet]slaacPrefixState), } // Register supported packet endpoint protocols. @@ -132,11 +140,86 @@ func newNIC(stack *Stack, id tcpip.NICID, name string, ep LinkEndpoint, ctx NICC nic.mu.packetEPs[netProto.Number()] = []PacketEndpoint{} } + nic.linkEP.Attach(nic) + return nic } -// enable enables the NIC. enable will attach the link to its LinkEndpoint and -// join the IPv6 All-Nodes Multicast address (ff02::1). +// enabled returns true if n is enabled. +func (n *NIC) enabled() bool { + n.mu.RLock() + enabled := n.mu.enabled + n.mu.RUnlock() + return enabled +} + +// disable disables n. +// +// It undoes the work done by enable. +func (n *NIC) disable() *tcpip.Error { + n.mu.RLock() + enabled := n.mu.enabled + n.mu.RUnlock() + if !enabled { + return nil + } + + n.mu.Lock() + err := n.disableLocked() + n.mu.Unlock() + return err +} + +// disableLocked disables n. +// +// It undoes the work done by enable. +// +// n MUST be locked. +func (n *NIC) disableLocked() *tcpip.Error { + if !n.mu.enabled { + return nil + } + + // TODO(b/147015577): Should Routes that are currently bound to n be + // invalidated? Currently, Routes will continue to work when a NIC is enabled + // again, and applications may not know that the underlying NIC was ever + // disabled. + + if _, ok := n.stack.networkProtocols[header.IPv6ProtocolNumber]; ok { + n.mu.ndp.stopSolicitingRouters() + n.mu.ndp.cleanupState(false /* hostOnly */) + + // Stop DAD for all the unicast IPv6 endpoints that are in the + // permanentTentative state. + for _, r := range n.mu.endpoints { + if addr := r.ep.ID().LocalAddress; r.getKind() == permanentTentative && header.IsV6UnicastAddress(addr) { + n.mu.ndp.stopDuplicateAddressDetection(addr) + } + } + + // The NIC may have already left the multicast group. + if err := n.leaveGroupLocked(header.IPv6AllNodesMulticastAddress, false /* force */); err != nil && err != tcpip.ErrBadLocalAddress { + return err + } + } + + if _, ok := n.stack.networkProtocols[header.IPv4ProtocolNumber]; ok { + // The address may have already been removed. + if err := n.removePermanentAddressLocked(ipv4BroadcastAddr.AddressWithPrefix.Address); err != nil && err != tcpip.ErrBadLocalAddress { + return err + } + } + + n.mu.enabled = false + return nil +} + +// enable enables n. +// +// If the stack has IPv6 enabled, enable will join the IPv6 All-Nodes Multicast +// address (ff02::1), start DAD for permanent addresses, and start soliciting +// routers if the stack is not operating as a router. If the stack is also +// configured to auto-generate a link-local address, one will be generated. func (n *NIC) enable() *tcpip.Error { n.mu.RLock() enabled := n.mu.enabled @@ -154,14 +237,9 @@ func (n *NIC) enable() *tcpip.Error { n.mu.enabled = true - n.attachLinkEndpoint() - // Create an endpoint to receive broadcast packets on this interface. if _, ok := n.stack.networkProtocols[header.IPv4ProtocolNumber]; ok { - if _, err := n.addAddressLocked(tcpip.ProtocolAddress{ - Protocol: header.IPv4ProtocolNumber, - AddressWithPrefix: tcpip.AddressWithPrefix{header.IPv4Broadcast, 8 * header.IPv4AddressSize}, - }, NeverPrimaryEndpoint, permanent, static, false /* deprecated */); err != nil { + if _, err := n.addAddressLocked(ipv4BroadcastAddr, NeverPrimaryEndpoint, permanent, static, false /* deprecated */); err != nil { return err } } @@ -183,6 +261,14 @@ func (n *NIC) enable() *tcpip.Error { return nil } + // Join the All-Nodes multicast group before starting DAD as responses to DAD + // (NDP NS) messages may be sent to the All-Nodes multicast group if the + // source address of the NDP NS is the unspecified address, as per RFC 4861 + // section 7.2.4. + if err := n.joinGroupLocked(header.IPv6ProtocolNumber, header.IPv6AllNodesMulticastAddress); err != nil { + return err + } + // Perform DAD on the all the unicast IPv6 endpoints that are in the permanent // state. // @@ -200,10 +286,6 @@ func (n *NIC) enable() *tcpip.Error { } } - if err := n.joinGroupLocked(header.IPv6ProtocolNumber, header.IPv6AllNodesMulticastAddress); err != nil { - return err - } - // Do not auto-generate an IPv6 link-local address for loopback devices. if n.stack.autoGenIPv6LinkLocal && !n.isLoopback() { // The valid and preferred lifetime is infinite for the auto-generated @@ -225,6 +307,42 @@ func (n *NIC) enable() *tcpip.Error { return nil } +// remove detaches NIC from the link endpoint, and marks existing referenced +// network endpoints expired. This guarantees no packets between this NIC and +// the network stack. +func (n *NIC) remove() *tcpip.Error { + n.mu.Lock() + defer n.mu.Unlock() + + n.disableLocked() + + // TODO(b/151378115): come up with a better way to pick an error than the + // first one. + var err *tcpip.Error + + // Forcefully leave multicast groups. + for nid := range n.mu.mcastJoins { + if tempErr := n.leaveGroupLocked(nid.LocalAddress, true /* force */); tempErr != nil && err == nil { + err = tempErr + } + } + + // Remove permanent and permanentTentative addresses, so no packet goes out. + for nid, ref := range n.mu.endpoints { + switch ref.getKind() { + case permanentTentative, permanent: + if tempErr := n.removePermanentAddressLocked(nid.LocalAddress); tempErr != nil && err == nil { + err = tempErr + } + } + } + + // Detach from link endpoint, so no packet comes in. + n.linkEP.Attach(nil) + + return err +} + // becomeIPv6Router transitions n into an IPv6 router. // // When transitioning into an IPv6 router, host-only state (NDP discovered @@ -234,7 +352,7 @@ func (n *NIC) becomeIPv6Router() { n.mu.Lock() defer n.mu.Unlock() - n.mu.ndp.cleanupHostOnlyState() + n.mu.ndp.cleanupState(true /* hostOnly */) n.mu.ndp.stopSolicitingRouters() } @@ -249,12 +367,6 @@ func (n *NIC) becomeIPv6Host() { n.mu.ndp.startSolicitingRouters() } -// attachLinkEndpoint attaches the NIC to the endpoint, which will enable it -// to start delivering packets. -func (n *NIC) attachLinkEndpoint() { - n.linkEP.Attach(n) -} - // setPromiscuousMode enables or disables promiscuous mode. func (n *NIC) setPromiscuousMode(enable bool) { n.mu.Lock() @@ -357,7 +469,7 @@ func (n *NIC) primaryIPv6Endpoint(remoteAddr tcpip.Address) *referencedNetworkEn cs := make([]ipv6AddrCandidate, 0, len(primaryAddrs)) for _, r := range primaryAddrs { // If r is not valid for outgoing connections, it is not a valid endpoint. - if !r.isValidForOutgoing() { + if !r.isValidForOutgoingRLocked() { continue } @@ -367,7 +479,7 @@ func (n *NIC) primaryIPv6Endpoint(remoteAddr tcpip.Address) *referencedNetworkEn // Should never happen as we got r from the primary IPv6 endpoint list and // ScopeForIPv6Address only returns an error if addr is not an IPv6 // address. - log.Fatalf("header.ScopeForIPv6Address(%s): %s", addr, err) + panic(fmt.Sprintf("header.ScopeForIPv6Address(%s): %s", addr, err)) } cs = append(cs, ipv6AddrCandidate{ @@ -379,7 +491,7 @@ func (n *NIC) primaryIPv6Endpoint(remoteAddr tcpip.Address) *referencedNetworkEn remoteScope, err := header.ScopeForIPv6Address(remoteAddr) if err != nil { // primaryIPv6Endpoint should never be called with an invalid IPv6 address. - log.Fatalf("header.ScopeForIPv6Address(%s): %s", remoteAddr, err) + panic(fmt.Sprintf("header.ScopeForIPv6Address(%s): %s", remoteAddr, err)) } // Sort the addresses as per RFC 6724 section 5 rules 1-3. @@ -712,6 +824,7 @@ func (n *NIC) AllAddresses() []tcpip.ProtocolAddress { case permanentExpired, temporary: continue } + addrs = append(addrs, tcpip.ProtocolAddress{ Protocol: ref.protocol, AddressWithPrefix: tcpip.AddressWithPrefix{ @@ -874,6 +987,7 @@ func (n *NIC) removeEndpointLocked(r *referencedNetworkEndpoint) { for i, ref := range refs { if ref == r { n.mu.primary[r.protocol] = append(refs[:i], refs[i+1:]...) + refs[len(refs)-1] = nil break } } @@ -898,35 +1012,42 @@ func (n *NIC) removePermanentAddressLocked(addr tcpip.Address) *tcpip.Error { return tcpip.ErrBadLocalAddress } - isIPv6Unicast := r.protocol == header.IPv6ProtocolNumber && header.IsV6UnicastAddress(addr) + switch r.protocol { + case header.IPv6ProtocolNumber: + return n.removePermanentIPv6EndpointLocked(r, true /* allowSLAAPrefixInvalidation */) + default: + r.expireLocked() + return nil + } +} + +func (n *NIC) removePermanentIPv6EndpointLocked(r *referencedNetworkEndpoint, allowSLAACPrefixInvalidation bool) *tcpip.Error { + addr := r.addrWithPrefix() + + isIPv6Unicast := header.IsV6UnicastAddress(addr.Address) if isIPv6Unicast { - // If we are removing a tentative IPv6 unicast address, stop - // DAD. - if kind == permanentTentative { - n.mu.ndp.stopDuplicateAddressDetection(addr) - } + n.mu.ndp.stopDuplicateAddressDetection(addr.Address) // If we are removing an address generated via SLAAC, cleanup // its SLAAC resources and notify the integrator. if r.configType == slaac { - n.mu.ndp.cleanupAutoGenAddrResourcesAndNotify(addr) + n.mu.ndp.cleanupSLAACAddrResourcesAndNotify(addr, allowSLAACPrefixInvalidation) } } - r.setKind(permanentExpired) - if !r.decRefLocked() { - // The endpoint still has references to it. - return nil - } + r.expireLocked() // At this point the endpoint is deleted. // If we are removing an IPv6 unicast address, leave the solicited-node // multicast address. + // + // We ignore the tcpip.ErrBadLocalAddress error because the solicited-node + // multicast group may be left by user action. if isIPv6Unicast { - snmc := header.SolicitedNodeAddr(addr) - if err := n.leaveGroupLocked(snmc); err != nil { + snmc := header.SolicitedNodeAddr(addr.Address) + if err := n.leaveGroupLocked(snmc, false /* force */); err != nil && err != tcpip.ErrBadLocalAddress { return err } } @@ -986,32 +1107,47 @@ func (n *NIC) leaveGroup(addr tcpip.Address) *tcpip.Error { n.mu.Lock() defer n.mu.Unlock() - return n.leaveGroupLocked(addr) + return n.leaveGroupLocked(addr, false /* force */) } // leaveGroupLocked decrements the count for the given multicast address, and // when it reaches zero removes the endpoint for this address. n MUST be locked // before leaveGroupLocked is called. -func (n *NIC) leaveGroupLocked(addr tcpip.Address) *tcpip.Error { +// +// If force is true, then the count for the multicast addres is ignored and the +// endpoint will be removed immediately. +func (n *NIC) leaveGroupLocked(addr tcpip.Address, force bool) *tcpip.Error { id := NetworkEndpointID{addr} - joins := n.mu.mcastJoins[id] - switch joins { - case 0: + joins, ok := n.mu.mcastJoins[id] + if !ok { // There are no joins with this address on this NIC. return tcpip.ErrBadLocalAddress - case 1: - // This is the last one, clean up. - if err := n.removePermanentAddressLocked(addr); err != nil { - return err - } } - n.mu.mcastJoins[id] = joins - 1 + + joins-- + if force || joins == 0 { + // There are no outstanding joins or we are forced to leave, clean up. + delete(n.mu.mcastJoins, id) + return n.removePermanentAddressLocked(addr) + } + + n.mu.mcastJoins[id] = joins return nil } -func handlePacket(protocol tcpip.NetworkProtocolNumber, dst, src tcpip.Address, localLinkAddr, remotelinkAddr tcpip.LinkAddress, ref *referencedNetworkEndpoint, pkt tcpip.PacketBuffer) { +// isInGroup returns true if n has joined the multicast group addr. +func (n *NIC) isInGroup(addr tcpip.Address) bool { + n.mu.RLock() + joins := n.mu.mcastJoins[NetworkEndpointID{addr}] + n.mu.RUnlock() + + return joins != 0 +} + +func handlePacket(protocol tcpip.NetworkProtocolNumber, dst, src tcpip.Address, localLinkAddr, remotelinkAddr tcpip.LinkAddress, ref *referencedNetworkEndpoint, pkt PacketBuffer) { r := makeRoute(protocol, dst, src, localLinkAddr, ref, false /* handleLocal */, false /* multicastLoop */) r.RemoteLinkAddress = remotelinkAddr + ref.ep.HandlePacket(&r, pkt) ref.decRef() } @@ -1022,7 +1158,7 @@ func handlePacket(protocol tcpip.NetworkProtocolNumber, dst, src tcpip.Address, // Note that the ownership of the slice backing vv is retained by the caller. // This rule applies only to the slice itself, not to the items of the slice; // the ownership of the items is not retained by the caller. -func (n *NIC) DeliverNetworkPacket(linkEP LinkEndpoint, remote, local tcpip.LinkAddress, protocol tcpip.NetworkProtocolNumber, pkt tcpip.PacketBuffer) { +func (n *NIC) DeliverNetworkPacket(linkEP LinkEndpoint, remote, local tcpip.LinkAddress, protocol tcpip.NetworkProtocolNumber, pkt PacketBuffer) { n.mu.RLock() enabled := n.mu.enabled // If the NIC is not yet enabled, don't receive any packets. @@ -1082,6 +1218,16 @@ func (n *NIC) DeliverNetworkPacket(linkEP LinkEndpoint, remote, local tcpip.Link n.stack.stats.IP.InvalidSourceAddressesReceived.Increment() return } + + // TODO(gvisor.dev/issue/170): Not supporting iptables for IPv6 yet. + if protocol == header.IPv4ProtocolNumber { + ipt := n.stack.IPTables() + if ok := ipt.Check(Prerouting, pkt); !ok { + // iptables is telling us to drop the packet. + return + } + } + if ref := n.getRef(protocol, dst); ref != nil { handlePacket(protocol, dst, src, linkEP.LinkAddress(), remote, ref, pkt) return @@ -1097,10 +1243,6 @@ func (n *NIC) DeliverNetworkPacket(linkEP LinkEndpoint, remote, local tcpip.Link n.stack.stats.IP.InvalidDestinationAddressesReceived.Increment() return } - defer r.Release() - - r.LocalLinkAddress = n.linkEP.LinkAddress() - r.RemoteLinkAddress = remote // Found a NIC. n := r.ref.nic @@ -1109,24 +1251,33 @@ func (n *NIC) DeliverNetworkPacket(linkEP LinkEndpoint, remote, local tcpip.Link ok = ok && ref.isValidForOutgoingRLocked() && ref.tryIncRef() n.mu.RUnlock() if ok { + r.LocalLinkAddress = n.linkEP.LinkAddress() + r.RemoteLinkAddress = remote r.RemoteAddress = src // TODO(b/123449044): Update the source NIC as well. ref.ep.HandlePacket(&r, pkt) ref.decRef() - } else { - // n doesn't have a destination endpoint. - // Send the packet out of n. - pkt.Header = buffer.NewPrependableFromView(pkt.Data.First()) - pkt.Data.RemoveFirst() - - // TODO(b/128629022): use route.WritePacket. - if err := n.linkEP.WritePacket(&r, nil /* gso */, protocol, pkt); err != nil { - r.Stats().IP.OutgoingPacketErrors.Increment() - } else { - n.stats.Tx.Packets.Increment() - n.stats.Tx.Bytes.IncrementBy(uint64(pkt.Header.UsedLength() + pkt.Data.Size())) + r.Release() + return + } + + // n doesn't have a destination endpoint. + // Send the packet out of n. + // TODO(b/128629022): move this logic to route.WritePacket. + if ch, err := r.Resolve(nil); err != nil { + if err == tcpip.ErrWouldBlock { + n.stack.forwarder.enqueue(ch, n, &r, protocol, pkt) + // forwarder will release route. + return } + n.stack.stats.IP.InvalidDestinationAddressesReceived.Increment() + r.Release() + return } + + // The link-address resolution finished immediately. + n.forwardPacket(&r, protocol, pkt) + r.Release() return } @@ -1136,9 +1287,38 @@ func (n *NIC) DeliverNetworkPacket(linkEP LinkEndpoint, remote, local tcpip.Link } } +func (n *NIC) forwardPacket(r *Route, protocol tcpip.NetworkProtocolNumber, pkt PacketBuffer) { + // TODO(b/143425874) Decrease the TTL field in forwarded packets. + + firstData := pkt.Data.First() + pkt.Data.RemoveFirst() + + if linkHeaderLen := int(n.linkEP.MaxHeaderLength()); linkHeaderLen == 0 { + pkt.Header = buffer.NewPrependableFromView(firstData) + } else { + firstDataLen := len(firstData) + + // pkt.Header should have enough capacity to hold n.linkEP's headers. + pkt.Header = buffer.NewPrependable(firstDataLen + linkHeaderLen) + + // TODO(b/151227689): avoid copying the packet when forwarding + if n := copy(pkt.Header.Prepend(firstDataLen), firstData); n != firstDataLen { + panic(fmt.Sprintf("copied %d bytes, expected %d", n, firstDataLen)) + } + } + + if err := n.linkEP.WritePacket(r, nil /* gso */, protocol, pkt); err != nil { + r.Stats().IP.OutgoingPacketErrors.Increment() + return + } + + n.stats.Tx.Packets.Increment() + n.stats.Tx.Bytes.IncrementBy(uint64(pkt.Header.UsedLength() + pkt.Data.Size())) +} + // DeliverTransportPacket delivers the packets to the appropriate transport // protocol endpoint. -func (n *NIC) DeliverTransportPacket(r *Route, protocol tcpip.TransportProtocolNumber, pkt tcpip.PacketBuffer) { +func (n *NIC) DeliverTransportPacket(r *Route, protocol tcpip.TransportProtocolNumber, pkt PacketBuffer) { state, ok := n.stack.transportProtocols[protocol] if !ok { n.stack.stats.UnknownProtocolRcvdPackets.Increment() @@ -1184,7 +1364,7 @@ func (n *NIC) DeliverTransportPacket(r *Route, protocol tcpip.TransportProtocolN // DeliverTransportControlPacket delivers control packets to the appropriate // transport protocol endpoint. -func (n *NIC) DeliverTransportControlPacket(local, remote tcpip.Address, net tcpip.NetworkProtocolNumber, trans tcpip.TransportProtocolNumber, typ ControlType, extra uint32, pkt tcpip.PacketBuffer) { +func (n *NIC) DeliverTransportControlPacket(local, remote tcpip.Address, net tcpip.NetworkProtocolNumber, trans tcpip.TransportProtocolNumber, typ ControlType, extra uint32, pkt PacketBuffer) { state, ok := n.stack.transportProtocols[trans] if !ok { return @@ -1215,11 +1395,21 @@ func (n *NIC) ID() tcpip.NICID { return n.id } +// Name returns the name of n. +func (n *NIC) Name() string { + return n.name +} + // Stack returns the instance of the Stack that owns this NIC. func (n *NIC) Stack() *Stack { return n.stack } +// LinkEndpoint returns the link endpoint of n. +func (n *NIC) LinkEndpoint() LinkEndpoint { + return n.linkEP +} + // isAddrTentative returns true if addr is tentative on n. // // Note that if addr is not associated with n, then this function will return @@ -1237,10 +1427,12 @@ func (n *NIC) isAddrTentative(addr tcpip.Address) bool { return ref.getKind() == permanentTentative } -// dupTentativeAddrDetected attempts to inform n that a tentative addr -// is a duplicate on a link. +// dupTentativeAddrDetected attempts to inform n that a tentative addr is a +// duplicate on a link. // -// dupTentativeAddrDetected will delete the tentative address if it exists. +// dupTentativeAddrDetected will remove the tentative address if it exists. If +// the address was generated via SLAAC, an attempt will be made to generate a +// new address. func (n *NIC) dupTentativeAddrDetected(addr tcpip.Address) *tcpip.Error { n.mu.Lock() defer n.mu.Unlock() @@ -1254,7 +1446,17 @@ func (n *NIC) dupTentativeAddrDetected(addr tcpip.Address) *tcpip.Error { return tcpip.ErrInvalidEndpointState } - return n.removePermanentAddressLocked(addr) + // If the address is a SLAAC address, do not invalidate its SLAAC prefix as a + // new address will be generated for it. + if err := n.removePermanentIPv6EndpointLocked(ref, false /* allowSLAACPrefixInvalidation */); err != nil { + return err + } + + if ref.configType == slaac { + n.mu.ndp.regenerateSLAACAddr(ref.addrWithPrefix().Subnet()) + } + + return nil } // setNDPConfigs sets the NDP configurations for n. @@ -1382,6 +1584,13 @@ type referencedNetworkEndpoint struct { deprecated bool } +func (r *referencedNetworkEndpoint) addrWithPrefix() tcpip.AddressWithPrefix { + return tcpip.AddressWithPrefix{ + Address: r.ep.ID().LocalAddress, + PrefixLen: r.ep.PrefixLen(), + } +} + func (r *referencedNetworkEndpoint) getKind() networkEndpointKind { return networkEndpointKind(atomic.LoadInt32((*int32)(&r.kind))) } @@ -1406,7 +1615,14 @@ func (r *referencedNetworkEndpoint) isValidForOutgoing() bool { // // r's NIC must be read locked. func (r *referencedNetworkEndpoint) isValidForOutgoingRLocked() bool { - return r.getKind() != permanentExpired || r.nic.mu.spoofing + return r.nic.mu.enabled && (r.getKind() != permanentExpired || r.nic.mu.spoofing) +} + +// expireLocked decrements the reference count and marks the permanent endpoint +// as expired. +func (r *referencedNetworkEndpoint) expireLocked() { + r.setKind(permanentExpired) + r.decRefLocked() } // decRef decrements the ref count and cleans up the endpoint once it reaches @@ -1418,14 +1634,11 @@ func (r *referencedNetworkEndpoint) decRef() { } // decRefLocked is the same as decRef but assumes that the NIC.mu mutex is -// locked. Returns true if the endpoint was removed. -func (r *referencedNetworkEndpoint) decRefLocked() bool { +// locked. +func (r *referencedNetworkEndpoint) decRefLocked() { if atomic.AddInt32(&r.refs, -1) == 0 { r.nic.removeEndpointLocked(r) - return true } - - return false } // incRef increments the ref count. It must only be called when the caller is diff --git a/pkg/tcpip/stack/nic_test.go b/pkg/tcpip/stack/nic_test.go index edaee3b86..d672fc157 100644 --- a/pkg/tcpip/stack/nic_test.go +++ b/pkg/tcpip/stack/nic_test.go @@ -17,7 +17,6 @@ package stack import ( "testing" - "gvisor.dev/gvisor/pkg/tcpip" "gvisor.dev/gvisor/pkg/tcpip/buffer" ) @@ -45,7 +44,7 @@ func TestDisabledRxStatsWhenNICDisabled(t *testing.T) { t.FailNow() } - nic.DeliverNetworkPacket(nil, "", "", 0, tcpip.PacketBuffer{Data: buffer.View([]byte{1, 2, 3, 4}).ToVectorisedView()}) + nic.DeliverNetworkPacket(nil, "", "", 0, PacketBuffer{Data: buffer.View([]byte{1, 2, 3, 4}).ToVectorisedView()}) if got := nic.stats.DisabledRx.Packets.Value(); got != 1 { t.Errorf("got DisabledRx.Packets = %d, want = 1", got) diff --git a/pkg/tcpip/packet_buffer.go b/pkg/tcpip/stack/packet_buffer.go index ab24372e7..dc125f25e 100644 --- a/pkg/tcpip/packet_buffer.go +++ b/pkg/tcpip/stack/packet_buffer.go @@ -11,18 +11,23 @@ // See the License for the specific language governing permissions and // limitations under the License. -package tcpip +package stack -import "gvisor.dev/gvisor/pkg/tcpip/buffer" +import ( + "gvisor.dev/gvisor/pkg/tcpip" + "gvisor.dev/gvisor/pkg/tcpip/buffer" +) // A PacketBuffer contains all the data of a network packet. // // As a PacketBuffer traverses up the stack, it may be necessary to pass it to // multiple endpoints. Clone() should be called in such cases so that // modifications to the Data field do not affect other copies. -// -// +stateify savable type PacketBuffer struct { + // PacketBufferEntry is used to build an intrusive list of + // PacketBuffers. + PacketBufferEntry + // Data holds the payload of the packet. For inbound packets, it also // holds the headers, which are consumed as the packet moves up the // stack. Headers are guaranteed not to be split across views. @@ -31,14 +36,6 @@ type PacketBuffer struct { // or otherwise modified. Data buffer.VectorisedView - // DataOffset is used for GSO output. It is the offset into the Data - // field where the payload of this packet starts. - DataOffset int - - // DataSize is used for GSO output. It is the size of this packet's - // payload. - DataSize int - // Header holds the headers of outbound packets. As a packet is passed // down the stack, each layer adds to Header. Header buffer.Prependable @@ -55,6 +52,14 @@ type PacketBuffer struct { LinkHeader buffer.View NetworkHeader buffer.View TransportHeader buffer.View + + // Hash is the transport layer hash of this packet. A value of zero + // indicates no valid hash has been set. + Hash uint32 + + // Owner is implemented by task to get the uid and gid. + // Only set for locally generated packets. + Owner tcpip.PacketOwner } // Clone makes a copy of pk. It clones the Data field, which creates a new diff --git a/pkg/tcpip/stack/rand.go b/pkg/tcpip/stack/rand.go new file mode 100644 index 000000000..421fb5c15 --- /dev/null +++ b/pkg/tcpip/stack/rand.go @@ -0,0 +1,40 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package stack + +import ( + mathrand "math/rand" + + "gvisor.dev/gvisor/pkg/sync" +) + +// lockedRandomSource provides a threadsafe rand.Source. +type lockedRandomSource struct { + mu sync.Mutex + src mathrand.Source +} + +func (r *lockedRandomSource) Int63() (n int64) { + r.mu.Lock() + n = r.src.Int63() + r.mu.Unlock() + return n +} + +func (r *lockedRandomSource) Seed(seed int64) { + r.mu.Lock() + r.src.Seed(seed) + r.mu.Unlock() +} diff --git a/pkg/tcpip/stack/registration.go b/pkg/tcpip/stack/registration.go index ec91f60dd..23ca9ee03 100644 --- a/pkg/tcpip/stack/registration.go +++ b/pkg/tcpip/stack/registration.go @@ -67,17 +67,18 @@ type TransportEndpoint interface { // this transport endpoint. It sets pkt.TransportHeader. // // HandlePacket takes ownership of pkt. - HandlePacket(r *Route, id TransportEndpointID, pkt tcpip.PacketBuffer) + HandlePacket(r *Route, id TransportEndpointID, pkt PacketBuffer) // HandleControlPacket is called by the stack when new control (e.g. // ICMP) packets arrive to this transport endpoint. // HandleControlPacket takes ownership of pkt. - HandleControlPacket(id TransportEndpointID, typ ControlType, extra uint32, pkt tcpip.PacketBuffer) + HandleControlPacket(id TransportEndpointID, typ ControlType, extra uint32, pkt PacketBuffer) - // Close puts the endpoint in a closed state and frees all resources - // associated with it. This cleanup may happen asynchronously. Wait can - // be used to block on this asynchronous cleanup. - Close() + // Abort initiates an expedited endpoint teardown. It puts the endpoint + // in a closed state and frees all resources associated with it. This + // cleanup may happen asynchronously. Wait can be used to block on this + // asynchronous cleanup. + Abort() // Wait waits for any worker goroutines owned by the endpoint to stop. // @@ -99,7 +100,7 @@ type RawTransportEndpoint interface { // layer up. // // HandlePacket takes ownership of pkt. - HandlePacket(r *Route, pkt tcpip.PacketBuffer) + HandlePacket(r *Route, pkt PacketBuffer) } // PacketEndpoint is the interface that needs to be implemented by packet @@ -117,7 +118,7 @@ type PacketEndpoint interface { // should construct its own ethernet header for applications. // // HandlePacket takes ownership of pkt. - HandlePacket(nicID tcpip.NICID, addr tcpip.LinkAddress, netProto tcpip.NetworkProtocolNumber, pkt tcpip.PacketBuffer) + HandlePacket(nicID tcpip.NICID, addr tcpip.LinkAddress, netProto tcpip.NetworkProtocolNumber, pkt PacketBuffer) } // TransportProtocol is the interface that needs to be implemented by transport @@ -149,7 +150,7 @@ type TransportProtocol interface { // stats purposes only). // // HandleUnknownDestinationPacket takes ownership of pkt. - HandleUnknownDestinationPacket(r *Route, id TransportEndpointID, pkt tcpip.PacketBuffer) bool + HandleUnknownDestinationPacket(r *Route, id TransportEndpointID, pkt PacketBuffer) bool // SetOption allows enabling/disabling protocol specific features. // SetOption returns an error if the option is not supported or the @@ -160,6 +161,13 @@ type TransportProtocol interface { // Option returns an error if the option is not supported or the // provided option value is invalid. Option(option interface{}) *tcpip.Error + + // Close requests that any worker goroutines owned by the protocol + // stop. + Close() + + // Wait waits for any worker goroutines owned by the protocol to stop. + Wait() } // TransportDispatcher contains the methods used by the network stack to deliver @@ -172,7 +180,7 @@ type TransportDispatcher interface { // pkt.NetworkHeader must be set before calling DeliverTransportPacket. // // DeliverTransportPacket takes ownership of pkt. - DeliverTransportPacket(r *Route, protocol tcpip.TransportProtocolNumber, pkt tcpip.PacketBuffer) + DeliverTransportPacket(r *Route, protocol tcpip.TransportProtocolNumber, pkt PacketBuffer) // DeliverTransportControlPacket delivers control packets to the // appropriate transport protocol endpoint. @@ -181,7 +189,7 @@ type TransportDispatcher interface { // DeliverTransportControlPacket. // // DeliverTransportControlPacket takes ownership of pkt. - DeliverTransportControlPacket(local, remote tcpip.Address, net tcpip.NetworkProtocolNumber, trans tcpip.TransportProtocolNumber, typ ControlType, extra uint32, pkt tcpip.PacketBuffer) + DeliverTransportControlPacket(local, remote tcpip.Address, net tcpip.NetworkProtocolNumber, trans tcpip.TransportProtocolNumber, typ ControlType, extra uint32, pkt PacketBuffer) } // PacketLooping specifies where an outbound packet should be sent. @@ -234,15 +242,15 @@ type NetworkEndpoint interface { // WritePacket writes a packet to the given destination address and // protocol. It sets pkt.NetworkHeader. pkt.TransportHeader must have // already been set. - WritePacket(r *Route, gso *GSO, params NetworkHeaderParams, pkt tcpip.PacketBuffer) *tcpip.Error + WritePacket(r *Route, gso *GSO, params NetworkHeaderParams, pkt PacketBuffer) *tcpip.Error // WritePackets writes packets to the given destination address and // protocol. pkts must not be zero length. - WritePackets(r *Route, gso *GSO, pkts []tcpip.PacketBuffer, params NetworkHeaderParams) (int, *tcpip.Error) + WritePackets(r *Route, gso *GSO, pkts PacketBufferList, params NetworkHeaderParams) (int, *tcpip.Error) // WriteHeaderIncludedPacket writes a packet that includes a network // header to the given destination address. - WriteHeaderIncludedPacket(r *Route, pkt tcpip.PacketBuffer) *tcpip.Error + WriteHeaderIncludedPacket(r *Route, pkt PacketBuffer) *tcpip.Error // ID returns the network protocol endpoint ID. ID() *NetworkEndpointID @@ -257,7 +265,7 @@ type NetworkEndpoint interface { // this network endpoint. It sets pkt.NetworkHeader. // // HandlePacket takes ownership of pkt. - HandlePacket(r *Route, pkt tcpip.PacketBuffer) + HandlePacket(r *Route, pkt PacketBuffer) // Close is called when the endpoint is reomved from a stack. Close() @@ -277,7 +285,7 @@ type NetworkProtocol interface { // DefaultPrefixLen returns the protocol's default prefix length. DefaultPrefixLen() int - // ParsePorts returns the source and destination addresses stored in a + // ParseAddresses returns the source and destination addresses stored in a // packet of this protocol. ParseAddresses(v buffer.View) (src, dst tcpip.Address) @@ -293,6 +301,13 @@ type NetworkProtocol interface { // Option returns an error if the option is not supported or the // provided option value is invalid. Option(option interface{}) *tcpip.Error + + // Close requests that any worker goroutines owned by the protocol + // stop. + Close() + + // Wait waits for any worker goroutines owned by the protocol to stop. + Wait() } // NetworkDispatcher contains the methods used by the network stack to deliver @@ -307,7 +322,7 @@ type NetworkDispatcher interface { // packets sent via loopback), and won't have the field set. // // DeliverNetworkPacket takes ownership of pkt. - DeliverNetworkPacket(linkEP LinkEndpoint, remote, local tcpip.LinkAddress, protocol tcpip.NetworkProtocolNumber, pkt tcpip.PacketBuffer) + DeliverNetworkPacket(linkEP LinkEndpoint, remote, local tcpip.LinkAddress, protocol tcpip.NetworkProtocolNumber, pkt PacketBuffer) } // LinkEndpointCapabilities is the type associated with the capabilities @@ -339,7 +354,7 @@ const ( // LinkEndpoint is the interface implemented by data link layer protocols (e.g., // ethernet, loopback, raw) and used by network layer protocols to send packets // out through the implementer's data link endpoint. When a link header exists, -// it sets each tcpip.PacketBuffer's LinkHeader field before passing it up the +// it sets each PacketBuffer's LinkHeader field before passing it up the // stack. type LinkEndpoint interface { // MTU is the maximum transmission unit for this endpoint. This is @@ -370,7 +385,7 @@ type LinkEndpoint interface { // To participate in transparent bridging, a LinkEndpoint implementation // should call eth.Encode with header.EthernetFields.SrcAddr set to // r.LocalLinkAddress if it is provided. - WritePacket(r *Route, gso *GSO, protocol tcpip.NetworkProtocolNumber, pkt tcpip.PacketBuffer) *tcpip.Error + WritePacket(r *Route, gso *GSO, protocol tcpip.NetworkProtocolNumber, pkt PacketBuffer) *tcpip.Error // WritePackets writes packets with the given protocol through the // given route. pkts must not be zero length. @@ -378,7 +393,7 @@ type LinkEndpoint interface { // Right now, WritePackets is used only when the software segmentation // offload is enabled. If it will be used for something else, it may // require to change syscall filters. - WritePackets(r *Route, gso *GSO, pkts []tcpip.PacketBuffer, protocol tcpip.NetworkProtocolNumber) (int, *tcpip.Error) + WritePackets(r *Route, gso *GSO, pkts PacketBufferList, protocol tcpip.NetworkProtocolNumber) (int, *tcpip.Error) // WriteRawPacket writes a packet directly to the link. The packet // should already have an ethernet header. @@ -386,6 +401,9 @@ type LinkEndpoint interface { // Attach attaches the data link layer endpoint to the network-layer // dispatcher of the stack. + // + // Attach will be called with a nil dispatcher if the receiver's associated + // NIC is being removed. Attach(dispatcher NetworkDispatcher) // IsAttached returns whether a NetworkDispatcher is attached to the @@ -408,7 +426,7 @@ type InjectableLinkEndpoint interface { LinkEndpoint // InjectInbound injects an inbound packet. - InjectInbound(protocol tcpip.NetworkProtocolNumber, pkt tcpip.PacketBuffer) + InjectInbound(protocol tcpip.NetworkProtocolNumber, pkt PacketBuffer) // InjectOutbound writes a fully formed outbound packet directly to the // link. diff --git a/pkg/tcpip/stack/route.go b/pkg/tcpip/stack/route.go index f565aafb2..a0e5e0300 100644 --- a/pkg/tcpip/stack/route.go +++ b/pkg/tcpip/stack/route.go @@ -153,7 +153,7 @@ func (r *Route) IsResolutionRequired() bool { } // WritePacket writes the packet through the given route. -func (r *Route) WritePacket(gso *GSO, params NetworkHeaderParams, pkt tcpip.PacketBuffer) *tcpip.Error { +func (r *Route) WritePacket(gso *GSO, params NetworkHeaderParams, pkt PacketBuffer) *tcpip.Error { if !r.ref.isValidForOutgoing() { return tcpip.ErrInvalidEndpointState } @@ -168,29 +168,32 @@ func (r *Route) WritePacket(gso *GSO, params NetworkHeaderParams, pkt tcpip.Pack return err } -// WritePackets writes the set of packets through the given route. -func (r *Route) WritePackets(gso *GSO, pkts []tcpip.PacketBuffer, params NetworkHeaderParams) (int, *tcpip.Error) { +// WritePackets writes a list of n packets through the given route and returns +// the number of packets written. +func (r *Route) WritePackets(gso *GSO, pkts PacketBufferList, params NetworkHeaderParams) (int, *tcpip.Error) { if !r.ref.isValidForOutgoing() { return 0, tcpip.ErrInvalidEndpointState } n, err := r.ref.ep.WritePackets(r, gso, pkts, params) if err != nil { - r.Stats().IP.OutgoingPacketErrors.IncrementBy(uint64(len(pkts) - n)) + r.Stats().IP.OutgoingPacketErrors.IncrementBy(uint64(pkts.Len() - n)) } r.ref.nic.stats.Tx.Packets.IncrementBy(uint64(n)) - payloadSize := 0 - for i := 0; i < n; i++ { - r.ref.nic.stats.Tx.Bytes.IncrementBy(uint64(pkts[i].Header.UsedLength())) - payloadSize += pkts[i].DataSize + + writtenBytes := 0 + for i, pb := 0, pkts.Front(); i < n && pb != nil; i, pb = i+1, pb.Next() { + writtenBytes += pb.Header.UsedLength() + writtenBytes += pb.Data.Size() } - r.ref.nic.stats.Tx.Bytes.IncrementBy(uint64(payloadSize)) + + r.ref.nic.stats.Tx.Bytes.IncrementBy(uint64(writtenBytes)) return n, err } // WriteHeaderIncludedPacket writes a packet already containing a network // header through the given route. -func (r *Route) WriteHeaderIncludedPacket(pkt tcpip.PacketBuffer) *tcpip.Error { +func (r *Route) WriteHeaderIncludedPacket(pkt PacketBuffer) *tcpip.Error { if !r.ref.isValidForOutgoing() { return tcpip.ErrInvalidEndpointState } diff --git a/pkg/tcpip/stack/stack.go b/pkg/tcpip/stack/stack.go index b793f1d74..41398a1b6 100644 --- a/pkg/tcpip/stack/stack.go +++ b/pkg/tcpip/stack/stack.go @@ -20,7 +20,9 @@ package stack import ( + "bytes" "encoding/binary" + mathrand "math/rand" "sync/atomic" "time" @@ -31,7 +33,6 @@ import ( "gvisor.dev/gvisor/pkg/tcpip" "gvisor.dev/gvisor/pkg/tcpip/buffer" "gvisor.dev/gvisor/pkg/tcpip/header" - "gvisor.dev/gvisor/pkg/tcpip/iptables" "gvisor.dev/gvisor/pkg/tcpip/ports" "gvisor.dev/gvisor/pkg/tcpip/seqnum" "gvisor.dev/gvisor/pkg/waiter" @@ -51,7 +52,7 @@ const ( type transportProtocolState struct { proto TransportProtocol - defaultHandler func(r *Route, id TransportEndpointID, pkt tcpip.PacketBuffer) bool + defaultHandler func(r *Route, id TransportEndpointID, pkt PacketBuffer) bool } // TCPProbeFunc is the expected function type for a TCP probe function to be @@ -428,7 +429,7 @@ type Stack struct { // tables are the iptables packet filtering and manipulation rules. The are // protected by tablesMu.` - tables iptables.IPTables + tables IPTables // resumableEndpoints is a list of endpoints that need to be resumed if the // stack is being restored. @@ -462,6 +463,14 @@ type Stack struct { // opaqueIIDOpts hold the options for generating opaque interface identifiers // (IIDs) as outlined by RFC 7217. opaqueIIDOpts OpaqueInterfaceIdentifierOptions + + // forwarder holds the packets that wait for their link-address resolutions + // to complete, and forwards them when each resolution is done. + forwarder *forwardQueue + + // randomGenerator is an injectable pseudo random generator that can be + // used when a random number is required. + randomGenerator *mathrand.Rand } // UniqueID is an abstract generator of unique identifiers. @@ -522,9 +531,16 @@ type Options struct { // this is non-nil. RawFactory RawFactory - // OpaqueIIDOpts hold the options for generating opaque interface identifiers - // (IIDs) as outlined by RFC 7217. + // OpaqueIIDOpts hold the options for generating opaque interface + // identifiers (IIDs) as outlined by RFC 7217. OpaqueIIDOpts OpaqueInterfaceIdentifierOptions + + // RandSource is an optional source to use to generate random + // numbers. If omitted it defaults to a Source seeded by the data + // returned by rand.Read(). + // + // RandSource must be thread-safe. + RandSource mathrand.Source } // TransportEndpointInfo holds useful information about a transport endpoint @@ -551,11 +567,13 @@ type TransportEndpointInfo struct { RegisterNICID tcpip.NICID } -// AddrNetProto unwraps the specified address if it is a V4-mapped V6 address -// and returns the network protocol number to be used to communicate with the -// specified address. It returns an error if the passed address is incompatible -// with the receiver. -func (e *TransportEndpointInfo) AddrNetProto(addr tcpip.FullAddress, v6only bool) (tcpip.FullAddress, tcpip.NetworkProtocolNumber, *tcpip.Error) { +// AddrNetProtoLocked unwraps the specified address if it is a V4-mapped V6 +// address and returns the network protocol number to be used to communicate +// with the specified address. It returns an error if the passed address is +// incompatible with the receiver. +// +// Preconditon: the parent endpoint mu must be held while calling this method. +func (e *TransportEndpointInfo) AddrNetProtoLocked(addr tcpip.FullAddress, v6only bool) (tcpip.FullAddress, tcpip.NetworkProtocolNumber, *tcpip.Error) { netProto := e.NetProto switch len(addr.Addr) { case header.IPv4AddressSize: @@ -618,6 +636,13 @@ func New(opts Options) *Stack { opts.UniqueID = new(uniqueIDGenerator) } + randSrc := opts.RandSource + if randSrc == nil { + // Source provided by mathrand.NewSource is not thread-safe so + // we wrap it in a simple thread-safe version. + randSrc = &lockedRandomSource{src: mathrand.NewSource(generateRandInt64())} + } + // Make sure opts.NDPConfigs contains valid values only. opts.NDPConfigs.validate() @@ -639,6 +664,8 @@ func New(opts Options) *Stack { uniqueIDGenerator: opts.UniqueID, ndpDisp: opts.NDPDisp, opaqueIIDOpts: opts.OpaqueIIDOpts, + forwarder: newForwardQueue(), + randomGenerator: mathrand.New(randSrc), } // Add specified network protocols. @@ -731,7 +758,7 @@ func (s *Stack) TransportProtocolOption(transport tcpip.TransportProtocolNumber, // // It must be called only during initialization of the stack. Changing it as the // stack is operating is not supported. -func (s *Stack) SetTransportProtocolHandler(p tcpip.TransportProtocolNumber, h func(*Route, TransportEndpointID, tcpip.PacketBuffer) bool) { +func (s *Stack) SetTransportProtocolHandler(p tcpip.TransportProtocolNumber, h func(*Route, TransportEndpointID, PacketBuffer) bool) { state := s.transportProtocols[p] if state != nil { state.defaultHandler = h @@ -881,6 +908,8 @@ type NICOptions struct { // CreateNICWithOptions creates a NIC with the provided id, LinkEndpoint, and // NICOptions. See the documentation on type NICOptions for details on how // NICs can be configured. +// +// LinkEndpoint.Attach will be called to bind ep with a NetworkDispatcher. func (s *Stack) CreateNICWithOptions(id tcpip.NICID, ep LinkEndpoint, opts NICOptions) *tcpip.Error { s.mu.Lock() defer s.mu.Unlock() @@ -890,8 +919,16 @@ func (s *Stack) CreateNICWithOptions(id tcpip.NICID, ep LinkEndpoint, opts NICOp return tcpip.ErrDuplicateNICID } - n := newNIC(s, id, opts.Name, ep, opts.Context) + // Make sure name is unique, unless unnamed. + if opts.Name != "" { + for _, n := range s.nics { + if n.Name() == opts.Name { + return tcpip.ErrDuplicateNICID + } + } + } + n := newNIC(s, id, opts.Name, ep, opts.Context) s.nics[id] = n if !opts.Disabled { return n.enable() @@ -901,34 +938,88 @@ func (s *Stack) CreateNICWithOptions(id tcpip.NICID, ep LinkEndpoint, opts NICOp } // CreateNIC creates a NIC with the provided id and LinkEndpoint and calls -// `LinkEndpoint.Attach` to start delivering packets to it. +// LinkEndpoint.Attach to bind ep with a NetworkDispatcher. func (s *Stack) CreateNIC(id tcpip.NICID, ep LinkEndpoint) *tcpip.Error { return s.CreateNICWithOptions(id, ep, NICOptions{}) } +// GetNICByName gets the NIC specified by name. +func (s *Stack) GetNICByName(name string) (*NIC, bool) { + s.mu.RLock() + defer s.mu.RUnlock() + for _, nic := range s.nics { + if nic.Name() == name { + return nic, true + } + } + return nil, false +} + // EnableNIC enables the given NIC so that the link-layer endpoint can start // delivering packets to it. func (s *Stack) EnableNIC(id tcpip.NICID) *tcpip.Error { s.mu.RLock() defer s.mu.RUnlock() - nic := s.nics[id] - if nic == nil { + nic, ok := s.nics[id] + if !ok { return tcpip.ErrUnknownNICID } return nic.enable() } +// DisableNIC disables the given NIC. +func (s *Stack) DisableNIC(id tcpip.NICID) *tcpip.Error { + s.mu.RLock() + defer s.mu.RUnlock() + + nic, ok := s.nics[id] + if !ok { + return tcpip.ErrUnknownNICID + } + + return nic.disable() +} + // CheckNIC checks if a NIC is usable. func (s *Stack) CheckNIC(id tcpip.NICID) bool { s.mu.RLock() + defer s.mu.RUnlock() + nic, ok := s.nics[id] - s.mu.RUnlock() - if ok { - return nic.linkEP.IsAttached() + if !ok { + return false } - return false + + return nic.enabled() +} + +// RemoveNIC removes NIC and all related routes from the network stack. +func (s *Stack) RemoveNIC(id tcpip.NICID) *tcpip.Error { + s.mu.Lock() + defer s.mu.Unlock() + + nic, ok := s.nics[id] + if !ok { + return tcpip.ErrUnknownNICID + } + delete(s.nics, id) + + // Remove routes in-place. n tracks the number of routes written. + n := 0 + for i, r := range s.routeTable { + if r.NIC != id { + // Keep this route. + if i > n { + s.routeTable[n] = r + } + n++ + } + } + s.routeTable = s.routeTable[:n] + + return nic.remove() } // NICAddressRanges returns a map of NICIDs to their associated subnets. @@ -980,7 +1071,7 @@ func (s *Stack) NICInfo() map[tcpip.NICID]NICInfo { for id, nic := range s.nics { flags := NICStateFlags{ Up: true, // Netstack interfaces are always up. - Running: nic.linkEP.IsAttached(), + Running: nic.enabled(), Promiscuous: nic.isPromiscuousMode(), Loopback: nic.isLoopback(), } @@ -1142,7 +1233,7 @@ func (s *Stack) FindRoute(id tcpip.NICID, localAddr, remoteAddr tcpip.Address, n isMulticast := header.IsV4MulticastAddress(remoteAddr) || header.IsV6MulticastAddress(remoteAddr) needRoute := !(isBroadcast || isMulticast || header.IsV6LinkLocalAddress(remoteAddr)) if id != 0 && !needRoute { - if nic, ok := s.nics[id]; ok { + if nic, ok := s.nics[id]; ok && nic.enabled() { if ref := s.getRefEP(nic, localAddr, remoteAddr, netProto); ref != nil { return makeRoute(netProto, ref.ep.ID().LocalAddress, remoteAddr, nic.linkEP.LinkAddress(), ref, s.handleLocal && !nic.isLoopback(), multicastLoop && !nic.isLoopback()), nil } @@ -1152,7 +1243,7 @@ func (s *Stack) FindRoute(id tcpip.NICID, localAddr, remoteAddr tcpip.Address, n if (id != 0 && id != route.NIC) || (len(remoteAddr) != 0 && !route.Destination.Contains(remoteAddr)) { continue } - if nic, ok := s.nics[route.NIC]; ok { + if nic, ok := s.nics[route.NIC]; ok && nic.enabled() { if ref := s.getRefEP(nic, localAddr, remoteAddr, netProto); ref != nil { if len(remoteAddr) == 0 { // If no remote address was provided, then the route @@ -1382,7 +1473,13 @@ func (s *Stack) RestoreCleanupEndpoints(es []TransportEndpoint) { // Endpoints created or modified during this call may not get closed. func (s *Stack) Close() { for _, e := range s.RegisteredEndpoints() { - e.Close() + e.Abort() + } + for _, p := range s.transportProtocols { + p.proto.Close() + } + for _, p := range s.networkProtocols { + p.Close() } } @@ -1400,6 +1497,12 @@ func (s *Stack) Wait() { for _, e := range s.CleanupEndpoints() { e.Wait() } + for _, p := range s.transportProtocols { + p.proto.Wait() + } + for _, p := range s.networkProtocols { + p.Wait() + } s.mu.RLock() defer s.mu.RUnlock() @@ -1605,8 +1708,20 @@ func (s *Stack) LeaveGroup(protocol tcpip.NetworkProtocolNumber, nicID tcpip.NIC return tcpip.ErrUnknownNICID } +// IsInGroup returns true if the NIC with ID nicID has joined the multicast +// group multicastAddr. +func (s *Stack) IsInGroup(nicID tcpip.NICID, multicastAddr tcpip.Address) (bool, *tcpip.Error) { + s.mu.RLock() + defer s.mu.RUnlock() + + if nic, ok := s.nics[nicID]; ok { + return nic.isInGroup(multicastAddr), nil + } + return false, tcpip.ErrUnknownNICID +} + // IPTables returns the stack's iptables. -func (s *Stack) IPTables() iptables.IPTables { +func (s *Stack) IPTables() IPTables { s.tablesMu.RLock() t := s.tables s.tablesMu.RUnlock() @@ -1614,7 +1729,7 @@ func (s *Stack) IPTables() iptables.IPTables { } // SetIPTables sets the stack's iptables. -func (s *Stack) SetIPTables(ipt iptables.IPTables) { +func (s *Stack) SetIPTables(ipt IPTables) { s.tablesMu.Lock() s.tables = ipt s.tablesMu.Unlock() @@ -1724,6 +1839,12 @@ func (s *Stack) Seed() uint32 { return s.seed } +// Rand returns a reference to a pseudo random generator that can be used +// to generate random numbers as required. +func (s *Stack) Rand() *mathrand.Rand { + return s.randomGenerator +} + func generateRandUint32() uint32 { b := make([]byte, 4) if _, err := rand.Read(b); err != nil { @@ -1731,3 +1852,16 @@ func generateRandUint32() uint32 { } return binary.LittleEndian.Uint32(b) } + +func generateRandInt64() int64 { + b := make([]byte, 8) + if _, err := rand.Read(b); err != nil { + panic(err) + } + buf := bytes.NewReader(b) + var v int64 + if err := binary.Read(buf, binary.LittleEndian, &v); err != nil { + panic(err) + } + return v +} diff --git a/pkg/tcpip/stack/stack_test.go b/pkg/tcpip/stack/stack_test.go index 24133e6f2..c7634ceb1 100644 --- a/pkg/tcpip/stack/stack_test.go +++ b/pkg/tcpip/stack/stack_test.go @@ -33,6 +33,7 @@ import ( "gvisor.dev/gvisor/pkg/tcpip/header" "gvisor.dev/gvisor/pkg/tcpip/link/channel" "gvisor.dev/gvisor/pkg/tcpip/link/loopback" + "gvisor.dev/gvisor/pkg/tcpip/network/ipv4" "gvisor.dev/gvisor/pkg/tcpip/network/ipv6" "gvisor.dev/gvisor/pkg/tcpip/stack" "gvisor.dev/gvisor/pkg/tcpip/transport/udp" @@ -89,7 +90,7 @@ func (f *fakeNetworkEndpoint) ID() *stack.NetworkEndpointID { return &f.id } -func (f *fakeNetworkEndpoint) HandlePacket(r *stack.Route, pkt tcpip.PacketBuffer) { +func (f *fakeNetworkEndpoint) HandlePacket(r *stack.Route, pkt stack.PacketBuffer) { // Increment the received packet count in the protocol descriptor. f.proto.packetCount[int(f.id.LocalAddress[0])%len(f.proto.packetCount)]++ @@ -125,7 +126,7 @@ func (f *fakeNetworkEndpoint) Capabilities() stack.LinkEndpointCapabilities { return f.ep.Capabilities() } -func (f *fakeNetworkEndpoint) WritePacket(r *stack.Route, gso *stack.GSO, params stack.NetworkHeaderParams, pkt tcpip.PacketBuffer) *tcpip.Error { +func (f *fakeNetworkEndpoint) WritePacket(r *stack.Route, gso *stack.GSO, params stack.NetworkHeaderParams, pkt stack.PacketBuffer) *tcpip.Error { // Increment the sent packet count in the protocol descriptor. f.proto.sendPacketCount[int(r.RemoteAddress[0])%len(f.proto.sendPacketCount)]++ @@ -140,7 +141,7 @@ func (f *fakeNetworkEndpoint) WritePacket(r *stack.Route, gso *stack.GSO, params views := make([]buffer.View, 1, 1+len(pkt.Data.Views())) views[0] = pkt.Header.View() views = append(views, pkt.Data.Views()...) - f.HandlePacket(r, tcpip.PacketBuffer{ + f.HandlePacket(r, stack.PacketBuffer{ Data: buffer.NewVectorisedView(len(views[0])+pkt.Data.Size(), views), }) } @@ -152,11 +153,11 @@ func (f *fakeNetworkEndpoint) WritePacket(r *stack.Route, gso *stack.GSO, params } // WritePackets implements stack.LinkEndpoint.WritePackets. -func (f *fakeNetworkEndpoint) WritePackets(r *stack.Route, gso *stack.GSO, pkts []tcpip.PacketBuffer, params stack.NetworkHeaderParams) (int, *tcpip.Error) { +func (f *fakeNetworkEndpoint) WritePackets(r *stack.Route, gso *stack.GSO, pkts stack.PacketBufferList, params stack.NetworkHeaderParams) (int, *tcpip.Error) { panic("not implemented") } -func (*fakeNetworkEndpoint) WriteHeaderIncludedPacket(r *stack.Route, pkt tcpip.PacketBuffer) *tcpip.Error { +func (*fakeNetworkEndpoint) WriteHeaderIncludedPacket(r *stack.Route, pkt stack.PacketBuffer) *tcpip.Error { return tcpip.ErrNotSupported } @@ -234,10 +235,33 @@ func (f *fakeNetworkProtocol) Option(option interface{}) *tcpip.Error { } } +// Close implements TransportProtocol.Close. +func (*fakeNetworkProtocol) Close() {} + +// Wait implements TransportProtocol.Wait. +func (*fakeNetworkProtocol) Wait() {} + func fakeNetFactory() stack.NetworkProtocol { return &fakeNetworkProtocol{} } +// linkEPWithMockedAttach is a stack.LinkEndpoint that tests can use to verify +// that LinkEndpoint.Attach was called. +type linkEPWithMockedAttach struct { + stack.LinkEndpoint + attached bool +} + +// Attach implements stack.LinkEndpoint.Attach. +func (l *linkEPWithMockedAttach) Attach(d stack.NetworkDispatcher) { + l.LinkEndpoint.Attach(d) + l.attached = d != nil +} + +func (l *linkEPWithMockedAttach) isAttached() bool { + return l.attached +} + func TestNetworkReceive(t *testing.T) { // Create a stack with the fake network protocol, one nic, and two // addresses attached to it: 1 & 2. @@ -263,7 +287,7 @@ func TestNetworkReceive(t *testing.T) { // Make sure packet with wrong address is not delivered. buf[0] = 3 - ep.InjectInbound(fakeNetNumber, tcpip.PacketBuffer{ + ep.InjectInbound(fakeNetNumber, stack.PacketBuffer{ Data: buf.ToVectorisedView(), }) if fakeNet.packetCount[1] != 0 { @@ -275,7 +299,7 @@ func TestNetworkReceive(t *testing.T) { // Make sure packet is delivered to first endpoint. buf[0] = 1 - ep.InjectInbound(fakeNetNumber, tcpip.PacketBuffer{ + ep.InjectInbound(fakeNetNumber, stack.PacketBuffer{ Data: buf.ToVectorisedView(), }) if fakeNet.packetCount[1] != 1 { @@ -287,7 +311,7 @@ func TestNetworkReceive(t *testing.T) { // Make sure packet is delivered to second endpoint. buf[0] = 2 - ep.InjectInbound(fakeNetNumber, tcpip.PacketBuffer{ + ep.InjectInbound(fakeNetNumber, stack.PacketBuffer{ Data: buf.ToVectorisedView(), }) if fakeNet.packetCount[1] != 1 { @@ -298,7 +322,7 @@ func TestNetworkReceive(t *testing.T) { } // Make sure packet is not delivered if protocol number is wrong. - ep.InjectInbound(fakeNetNumber-1, tcpip.PacketBuffer{ + ep.InjectInbound(fakeNetNumber-1, stack.PacketBuffer{ Data: buf.ToVectorisedView(), }) if fakeNet.packetCount[1] != 1 { @@ -310,7 +334,7 @@ func TestNetworkReceive(t *testing.T) { // Make sure packet that is too small is dropped. buf.CapLength(2) - ep.InjectInbound(fakeNetNumber, tcpip.PacketBuffer{ + ep.InjectInbound(fakeNetNumber, stack.PacketBuffer{ Data: buf.ToVectorisedView(), }) if fakeNet.packetCount[1] != 1 { @@ -332,7 +356,7 @@ func sendTo(s *stack.Stack, addr tcpip.Address, payload buffer.View) *tcpip.Erro func send(r stack.Route, payload buffer.View) *tcpip.Error { hdr := buffer.NewPrependable(int(r.MaxHeaderLength())) - return r.WritePacket(nil /* gso */, stack.NetworkHeaderParams{Protocol: fakeTransNumber, TTL: 123, TOS: stack.DefaultTOS}, tcpip.PacketBuffer{ + return r.WritePacket(nil /* gso */, stack.NetworkHeaderParams{Protocol: fakeTransNumber, TTL: 123, TOS: stack.DefaultTOS}, stack.PacketBuffer{ Header: hdr, Data: payload.ToVectorisedView(), }) @@ -390,7 +414,7 @@ func testFailingRecv(t *testing.T, fakeNet *fakeNetworkProtocol, localAddrByte b func testRecvInternal(t *testing.T, fakeNet *fakeNetworkProtocol, localAddrByte byte, ep *channel.Endpoint, buf buffer.View, want int) { t.Helper() - ep.InjectInbound(fakeNetNumber, tcpip.PacketBuffer{ + ep.InjectInbound(fakeNetNumber, stack.PacketBuffer{ Data: buf.ToVectorisedView(), }) if got := fakeNet.PacketCount(localAddrByte); got != want { @@ -509,6 +533,340 @@ func testNoRoute(t *testing.T, s *stack.Stack, nic tcpip.NICID, srcAddr, dstAddr } } +// TestAttachToLinkEndpointImmediately tests that a LinkEndpoint is attached to +// a NetworkDispatcher when the NIC is created. +func TestAttachToLinkEndpointImmediately(t *testing.T) { + const nicID = 1 + + tests := []struct { + name string + nicOpts stack.NICOptions + }{ + { + name: "Create enabled NIC", + nicOpts: stack.NICOptions{Disabled: false}, + }, + { + name: "Create disabled NIC", + nicOpts: stack.NICOptions{Disabled: true}, + }, + } + + for _, test := range tests { + t.Run(test.name, func(t *testing.T) { + s := stack.New(stack.Options{ + NetworkProtocols: []stack.NetworkProtocol{fakeNetFactory()}, + }) + + e := linkEPWithMockedAttach{ + LinkEndpoint: loopback.New(), + } + + if err := s.CreateNICWithOptions(nicID, &e, test.nicOpts); err != nil { + t.Fatalf("CreateNICWithOptions(%d, _, %+v) = %s", nicID, test.nicOpts, err) + } + if !e.isAttached() { + t.Fatal("link endpoint not attached to a network dispatcher") + } + }) + } +} + +func TestDisableUnknownNIC(t *testing.T) { + s := stack.New(stack.Options{ + NetworkProtocols: []stack.NetworkProtocol{fakeNetFactory()}, + }) + + if err := s.DisableNIC(1); err != tcpip.ErrUnknownNICID { + t.Fatalf("got s.DisableNIC(1) = %v, want = %s", err, tcpip.ErrUnknownNICID) + } +} + +func TestDisabledNICsNICInfoAndCheckNIC(t *testing.T) { + const nicID = 1 + + s := stack.New(stack.Options{ + NetworkProtocols: []stack.NetworkProtocol{fakeNetFactory()}, + }) + + e := loopback.New() + nicOpts := stack.NICOptions{Disabled: true} + if err := s.CreateNICWithOptions(nicID, e, nicOpts); err != nil { + t.Fatalf("CreateNICWithOptions(%d, _, %+v) = %s", nicID, nicOpts, err) + } + + checkNIC := func(enabled bool) { + t.Helper() + + allNICInfo := s.NICInfo() + nicInfo, ok := allNICInfo[nicID] + if !ok { + t.Errorf("entry for %d missing from allNICInfo = %+v", nicID, allNICInfo) + } else if nicInfo.Flags.Running != enabled { + t.Errorf("got nicInfo.Flags.Running = %t, want = %t", nicInfo.Flags.Running, enabled) + } + + if got := s.CheckNIC(nicID); got != enabled { + t.Errorf("got s.CheckNIC(%d) = %t, want = %t", nicID, got, enabled) + } + } + + // NIC should initially report itself as disabled. + checkNIC(false) + + if err := s.EnableNIC(nicID); err != nil { + t.Fatalf("s.EnableNIC(%d): %s", nicID, err) + } + checkNIC(true) + + // If the NIC is not reporting a correct enabled status, we cannot trust the + // next check so end the test here. + if t.Failed() { + t.FailNow() + } + + if err := s.DisableNIC(nicID); err != nil { + t.Fatalf("s.DisableNIC(%d): %s", nicID, err) + } + checkNIC(false) +} + +func TestRemoveUnknownNIC(t *testing.T) { + s := stack.New(stack.Options{ + NetworkProtocols: []stack.NetworkProtocol{fakeNetFactory()}, + }) + + if err := s.RemoveNIC(1); err != tcpip.ErrUnknownNICID { + t.Fatalf("got s.RemoveNIC(1) = %v, want = %s", err, tcpip.ErrUnknownNICID) + } +} + +func TestRemoveNIC(t *testing.T) { + const nicID = 1 + + s := stack.New(stack.Options{ + NetworkProtocols: []stack.NetworkProtocol{fakeNetFactory()}, + }) + + e := linkEPWithMockedAttach{ + LinkEndpoint: loopback.New(), + } + if err := s.CreateNIC(nicID, &e); err != nil { + t.Fatalf("CreateNIC(%d, _) = %s", nicID, err) + } + + // NIC should be present in NICInfo and attached to a NetworkDispatcher. + allNICInfo := s.NICInfo() + if _, ok := allNICInfo[nicID]; !ok { + t.Errorf("entry for %d missing from allNICInfo = %+v", nicID, allNICInfo) + } + if !e.isAttached() { + t.Fatal("link endpoint not attached to a network dispatcher") + } + + // Removing a NIC should remove it from NICInfo and e should be detached from + // the NetworkDispatcher. + if err := s.RemoveNIC(nicID); err != nil { + t.Fatalf("s.RemoveNIC(%d): %s", nicID, err) + } + if nicInfo, ok := s.NICInfo()[nicID]; ok { + t.Errorf("got unexpected NICInfo entry for deleted NIC %d = %+v", nicID, nicInfo) + } + if e.isAttached() { + t.Error("link endpoint for removed NIC still attached to a network dispatcher") + } +} + +func TestRouteWithDownNIC(t *testing.T) { + tests := []struct { + name string + downFn func(s *stack.Stack, nicID tcpip.NICID) *tcpip.Error + upFn func(s *stack.Stack, nicID tcpip.NICID) *tcpip.Error + }{ + { + name: "Disabled NIC", + downFn: (*stack.Stack).DisableNIC, + upFn: (*stack.Stack).EnableNIC, + }, + + // Once a NIC is removed, it cannot be brought up. + { + name: "Removed NIC", + downFn: (*stack.Stack).RemoveNIC, + }, + } + + const unspecifiedNIC = 0 + const nicID1 = 1 + const nicID2 = 2 + const addr1 = tcpip.Address("\x01") + const addr2 = tcpip.Address("\x02") + const nic1Dst = tcpip.Address("\x05") + const nic2Dst = tcpip.Address("\x06") + + setup := func(t *testing.T) (*stack.Stack, *channel.Endpoint, *channel.Endpoint) { + s := stack.New(stack.Options{ + NetworkProtocols: []stack.NetworkProtocol{fakeNetFactory()}, + }) + + ep1 := channel.New(1, defaultMTU, "") + if err := s.CreateNIC(nicID1, ep1); err != nil { + t.Fatalf("CreateNIC(%d, _): %s", nicID1, err) + } + + if err := s.AddAddress(nicID1, fakeNetNumber, addr1); err != nil { + t.Fatalf("AddAddress(%d, %d, %s): %s", nicID1, fakeNetNumber, addr1, err) + } + + ep2 := channel.New(1, defaultMTU, "") + if err := s.CreateNIC(nicID2, ep2); err != nil { + t.Fatalf("CreateNIC(%d, _): %s", nicID2, err) + } + + if err := s.AddAddress(nicID2, fakeNetNumber, addr2); err != nil { + t.Fatalf("AddAddress(%d, %d, %s): %s", nicID2, fakeNetNumber, addr2, err) + } + + // Set a route table that sends all packets with odd destination + // addresses through the first NIC, and all even destination address + // through the second one. + { + subnet0, err := tcpip.NewSubnet("\x00", "\x01") + if err != nil { + t.Fatal(err) + } + subnet1, err := tcpip.NewSubnet("\x01", "\x01") + if err != nil { + t.Fatal(err) + } + s.SetRouteTable([]tcpip.Route{ + {Destination: subnet1, Gateway: "\x00", NIC: nicID1}, + {Destination: subnet0, Gateway: "\x00", NIC: nicID2}, + }) + } + + return s, ep1, ep2 + } + + // Tests that routes through a down NIC are not used when looking up a route + // for a destination. + t.Run("Find", func(t *testing.T) { + for _, test := range tests { + t.Run(test.name, func(t *testing.T) { + s, _, _ := setup(t) + + // Test routes to odd address. + testRoute(t, s, unspecifiedNIC, "", "\x05", addr1) + testRoute(t, s, unspecifiedNIC, addr1, "\x05", addr1) + testRoute(t, s, nicID1, addr1, "\x05", addr1) + + // Test routes to even address. + testRoute(t, s, unspecifiedNIC, "", "\x06", addr2) + testRoute(t, s, unspecifiedNIC, addr2, "\x06", addr2) + testRoute(t, s, nicID2, addr2, "\x06", addr2) + + // Bringing NIC1 down should result in no routes to odd addresses. Routes to + // even addresses should continue to be available as NIC2 is still up. + if err := test.downFn(s, nicID1); err != nil { + t.Fatalf("test.downFn(_, %d): %s", nicID1, err) + } + testNoRoute(t, s, unspecifiedNIC, "", nic1Dst) + testNoRoute(t, s, unspecifiedNIC, addr1, nic1Dst) + testNoRoute(t, s, nicID1, addr1, nic1Dst) + testRoute(t, s, unspecifiedNIC, "", nic2Dst, addr2) + testRoute(t, s, unspecifiedNIC, addr2, nic2Dst, addr2) + testRoute(t, s, nicID2, addr2, nic2Dst, addr2) + + // Bringing NIC2 down should result in no routes to even addresses. No + // route should be available to any address as routes to odd addresses + // were made unavailable by bringing NIC1 down above. + if err := test.downFn(s, nicID2); err != nil { + t.Fatalf("test.downFn(_, %d): %s", nicID2, err) + } + testNoRoute(t, s, unspecifiedNIC, "", nic1Dst) + testNoRoute(t, s, unspecifiedNIC, addr1, nic1Dst) + testNoRoute(t, s, nicID1, addr1, nic1Dst) + testNoRoute(t, s, unspecifiedNIC, "", nic2Dst) + testNoRoute(t, s, unspecifiedNIC, addr2, nic2Dst) + testNoRoute(t, s, nicID2, addr2, nic2Dst) + + if upFn := test.upFn; upFn != nil { + // Bringing NIC1 up should make routes to odd addresses available + // again. Routes to even addresses should continue to be unavailable + // as NIC2 is still down. + if err := upFn(s, nicID1); err != nil { + t.Fatalf("test.upFn(_, %d): %s", nicID1, err) + } + testRoute(t, s, unspecifiedNIC, "", nic1Dst, addr1) + testRoute(t, s, unspecifiedNIC, addr1, nic1Dst, addr1) + testRoute(t, s, nicID1, addr1, nic1Dst, addr1) + testNoRoute(t, s, unspecifiedNIC, "", nic2Dst) + testNoRoute(t, s, unspecifiedNIC, addr2, nic2Dst) + testNoRoute(t, s, nicID2, addr2, nic2Dst) + } + }) + } + }) + + // Tests that writing a packet using a Route through a down NIC fails. + t.Run("WritePacket", func(t *testing.T) { + for _, test := range tests { + t.Run(test.name, func(t *testing.T) { + s, ep1, ep2 := setup(t) + + r1, err := s.FindRoute(nicID1, addr1, nic1Dst, fakeNetNumber, false /* multicastLoop */) + if err != nil { + t.Errorf("FindRoute(%d, %s, %s, %d, false): %s", nicID1, addr1, nic1Dst, fakeNetNumber, err) + } + defer r1.Release() + + r2, err := s.FindRoute(nicID2, addr2, nic2Dst, fakeNetNumber, false /* multicastLoop */) + if err != nil { + t.Errorf("FindRoute(%d, %s, %s, %d, false): %s", nicID2, addr2, nic2Dst, fakeNetNumber, err) + } + defer r2.Release() + + // If we failed to get routes r1 or r2, we cannot proceed with the test. + if t.Failed() { + t.FailNow() + } + + buf := buffer.View([]byte{1}) + testSend(t, r1, ep1, buf) + testSend(t, r2, ep2, buf) + + // Writes with Routes that use NIC1 after being brought down should fail. + if err := test.downFn(s, nicID1); err != nil { + t.Fatalf("test.downFn(_, %d): %s", nicID1, err) + } + testFailingSend(t, r1, ep1, buf, tcpip.ErrInvalidEndpointState) + testSend(t, r2, ep2, buf) + + // Writes with Routes that use NIC2 after being brought down should fail. + if err := test.downFn(s, nicID2); err != nil { + t.Fatalf("test.downFn(_, %d): %s", nicID2, err) + } + testFailingSend(t, r1, ep1, buf, tcpip.ErrInvalidEndpointState) + testFailingSend(t, r2, ep2, buf, tcpip.ErrInvalidEndpointState) + + if upFn := test.upFn; upFn != nil { + // Writes with Routes that use NIC1 after being brought up should + // succeed. + // + // TODO(b/147015577): Should we instead completely invalidate all + // Routes that were bound to a NIC that was brought down at some + // point? + if err := upFn(s, nicID1); err != nil { + t.Fatalf("test.upFn(_, %d): %s", nicID1, err) + } + testSend(t, r1, ep1, buf) + testFailingSend(t, r2, ep2, buf, tcpip.ErrInvalidEndpointState) + } + }) + } + }) +} + func TestRoutes(t *testing.T) { // Create a stack with the fake network protocol, two nics, and two // addresses per nic, the first nic has odd address, the second one has @@ -1087,19 +1445,19 @@ func TestOutgoingBroadcastWithEmptyRouteTable(t *testing.T) { protoAddr := tcpip.ProtocolAddress{Protocol: fakeNetNumber, AddressWithPrefix: tcpip.AddressWithPrefix{header.IPv4Any, 0}} if err := s.AddProtocolAddress(1, protoAddr); err != nil { - t.Fatalf("AddProtocolAddress(1, %s) failed: %s", protoAddr, err) + t.Fatalf("AddProtocolAddress(1, %v) failed: %v", protoAddr, err) } r, err := s.FindRoute(1, header.IPv4Any, header.IPv4Broadcast, fakeNetNumber, false /* multicastLoop */) if err != nil { - t.Fatalf("FindRoute(1, %s, %s, %d) failed: %s", header.IPv4Any, header.IPv4Broadcast, fakeNetNumber, err) + t.Fatalf("FindRoute(1, %v, %v, %d) failed: %v", header.IPv4Any, header.IPv4Broadcast, fakeNetNumber, err) } if err := verifyRoute(r, stack.Route{LocalAddress: header.IPv4Any, RemoteAddress: header.IPv4Broadcast}); err != nil { - t.Errorf("FindRoute(1, %s, %s, %d) returned unexpected Route: %s)", header.IPv4Any, header.IPv4Broadcast, fakeNetNumber, err) + t.Errorf("FindRoute(1, %v, %v, %d) returned unexpected Route: %v", header.IPv4Any, header.IPv4Broadcast, fakeNetNumber, err) } // If the NIC doesn't exist, it won't work. if _, err := s.FindRoute(2, header.IPv4Any, header.IPv4Broadcast, fakeNetNumber, false /* multicastLoop */); err != tcpip.ErrNetworkUnreachable { - t.Fatalf("got FindRoute(2, %s, %s, %d) = %s want = %s", header.IPv4Any, header.IPv4Broadcast, fakeNetNumber, err, tcpip.ErrNetworkUnreachable) + t.Fatalf("got FindRoute(2, %v, %v, %d) = %v want = %v", header.IPv4Any, header.IPv4Broadcast, fakeNetNumber, err, tcpip.ErrNetworkUnreachable) } } @@ -1125,12 +1483,12 @@ func TestOutgoingBroadcastWithRouteTable(t *testing.T) { } nic1ProtoAddr := tcpip.ProtocolAddress{fakeNetNumber, nic1Addr} if err := s.AddProtocolAddress(1, nic1ProtoAddr); err != nil { - t.Fatalf("AddProtocolAddress(1, %s) failed: %s", nic1ProtoAddr, err) + t.Fatalf("AddProtocolAddress(1, %v) failed: %v", nic1ProtoAddr, err) } nic2ProtoAddr := tcpip.ProtocolAddress{fakeNetNumber, nic2Addr} if err := s.AddProtocolAddress(2, nic2ProtoAddr); err != nil { - t.Fatalf("AddAddress(2, %s) failed: %s", nic2ProtoAddr, err) + t.Fatalf("AddAddress(2, %v) failed: %v", nic2ProtoAddr, err) } // Set the initial route table. @@ -1145,10 +1503,10 @@ func TestOutgoingBroadcastWithRouteTable(t *testing.T) { // When an interface is given, the route for a broadcast goes through it. r, err := s.FindRoute(1, nic1Addr.Address, header.IPv4Broadcast, fakeNetNumber, false /* multicastLoop */) if err != nil { - t.Fatalf("FindRoute(1, %s, %s, %d) failed: %s", nic1Addr.Address, header.IPv4Broadcast, fakeNetNumber, err) + t.Fatalf("FindRoute(1, %v, %v, %d) failed: %v", nic1Addr.Address, header.IPv4Broadcast, fakeNetNumber, err) } if err := verifyRoute(r, stack.Route{LocalAddress: nic1Addr.Address, RemoteAddress: header.IPv4Broadcast}); err != nil { - t.Errorf("FindRoute(1, %s, %s, %d) returned unexpected Route: %s)", nic1Addr.Address, header.IPv4Broadcast, fakeNetNumber, err) + t.Errorf("FindRoute(1, %v, %v, %d) returned unexpected Route: %v", nic1Addr.Address, header.IPv4Broadcast, fakeNetNumber, err) } // When an interface is not given, it consults the route table. @@ -1792,6 +2150,91 @@ func TestAddProtocolAddressWithOptions(t *testing.T) { verifyAddresses(t, expectedAddresses, gotAddresses) } +func TestCreateNICWithOptions(t *testing.T) { + type callArgsAndExpect struct { + nicID tcpip.NICID + opts stack.NICOptions + err *tcpip.Error + } + + tests := []struct { + desc string + calls []callArgsAndExpect + }{ + { + desc: "DuplicateNICID", + calls: []callArgsAndExpect{ + { + nicID: tcpip.NICID(1), + opts: stack.NICOptions{Name: "eth1"}, + err: nil, + }, + { + nicID: tcpip.NICID(1), + opts: stack.NICOptions{Name: "eth2"}, + err: tcpip.ErrDuplicateNICID, + }, + }, + }, + { + desc: "DuplicateName", + calls: []callArgsAndExpect{ + { + nicID: tcpip.NICID(1), + opts: stack.NICOptions{Name: "lo"}, + err: nil, + }, + { + nicID: tcpip.NICID(2), + opts: stack.NICOptions{Name: "lo"}, + err: tcpip.ErrDuplicateNICID, + }, + }, + }, + { + desc: "Unnamed", + calls: []callArgsAndExpect{ + { + nicID: tcpip.NICID(1), + opts: stack.NICOptions{}, + err: nil, + }, + { + nicID: tcpip.NICID(2), + opts: stack.NICOptions{}, + err: nil, + }, + }, + }, + { + desc: "UnnamedDuplicateNICID", + calls: []callArgsAndExpect{ + { + nicID: tcpip.NICID(1), + opts: stack.NICOptions{}, + err: nil, + }, + { + nicID: tcpip.NICID(1), + opts: stack.NICOptions{}, + err: tcpip.ErrDuplicateNICID, + }, + }, + }, + } + for _, test := range tests { + t.Run(test.desc, func(t *testing.T) { + s := stack.New(stack.Options{}) + ep := channel.New(0, 0, tcpip.LinkAddress("\x00\x00\x00\x00\x00\x00")) + for _, call := range test.calls { + if got, want := s.CreateNICWithOptions(call.nicID, ep, call.opts), call.err; got != want { + t.Fatalf("CreateNICWithOptions(%v, _, %+v) = %v, want %v", call.nicID, call.opts, got, want) + } + } + }) + } +} + func TestNICStats(t *testing.T) { s := stack.New(stack.Options{ NetworkProtocols: []stack.NetworkProtocol{fakeNetFactory()}, @@ -1814,7 +2257,7 @@ func TestNICStats(t *testing.T) { // Send a packet to address 1. buf := buffer.NewView(30) - ep1.InjectInbound(fakeNetNumber, tcpip.PacketBuffer{ + ep1.InjectInbound(fakeNetNumber, stack.PacketBuffer{ Data: buf.ToVectorisedView(), }) if got, want := s.NICInfo()[1].Stats.Rx.Packets.Value(), uint64(1); got != want { @@ -1841,56 +2284,84 @@ func TestNICStats(t *testing.T) { } func TestNICForwarding(t *testing.T) { - // Create a stack with the fake network protocol, two NICs, each with - // an address. - s := stack.New(stack.Options{ - NetworkProtocols: []stack.NetworkProtocol{fakeNetFactory()}, - }) - s.SetForwarding(true) + const nicID1 = 1 + const nicID2 = 2 + const dstAddr = tcpip.Address("\x03") - ep1 := channel.New(10, defaultMTU, "") - if err := s.CreateNIC(1, ep1); err != nil { - t.Fatal("CreateNIC #1 failed:", err) - } - if err := s.AddAddress(1, fakeNetNumber, "\x01"); err != nil { - t.Fatal("AddAddress #1 failed:", err) + tests := []struct { + name string + headerLen uint16 + }{ + { + name: "Zero header length", + }, + { + name: "Non-zero header length", + headerLen: 16, + }, } - ep2 := channel.New(10, defaultMTU, "") - if err := s.CreateNIC(2, ep2); err != nil { - t.Fatal("CreateNIC #2 failed:", err) - } - if err := s.AddAddress(2, fakeNetNumber, "\x02"); err != nil { - t.Fatal("AddAddress #2 failed:", err) - } + for _, test := range tests { + t.Run(test.name, func(t *testing.T) { + s := stack.New(stack.Options{ + NetworkProtocols: []stack.NetworkProtocol{fakeNetFactory()}, + }) + s.SetForwarding(true) - // Route all packets to address 3 to NIC 2. - { - subnet, err := tcpip.NewSubnet("\x03", "\xff") - if err != nil { - t.Fatal(err) - } - s.SetRouteTable([]tcpip.Route{{Destination: subnet, Gateway: "\x00", NIC: 2}}) - } + ep1 := channel.New(10, defaultMTU, "") + if err := s.CreateNIC(nicID1, ep1); err != nil { + t.Fatalf("CreateNIC(%d, _): %s", nicID1, err) + } + if err := s.AddAddress(nicID1, fakeNetNumber, "\x01"); err != nil { + t.Fatalf("AddAddress(%d, %d, 0x01): %s", nicID1, fakeNetNumber, err) + } - // Send a packet to address 3. - buf := buffer.NewView(30) - buf[0] = 3 - ep1.InjectInbound(fakeNetNumber, tcpip.PacketBuffer{ - Data: buf.ToVectorisedView(), - }) + ep2 := channelLinkWithHeaderLength{ + Endpoint: channel.New(10, defaultMTU, ""), + headerLength: test.headerLen, + } + if err := s.CreateNIC(nicID2, &ep2); err != nil { + t.Fatalf("CreateNIC(%d, _): %s", nicID2, err) + } + if err := s.AddAddress(nicID2, fakeNetNumber, "\x02"); err != nil { + t.Fatalf("AddAddress(%d, %d, 0x02): %s", nicID2, fakeNetNumber, err) + } - if _, ok := ep2.Read(); !ok { - t.Fatal("Packet not forwarded") - } + // Route all packets to dstAddr to NIC 2. + { + subnet, err := tcpip.NewSubnet(dstAddr, "\xff") + if err != nil { + t.Fatal(err) + } + s.SetRouteTable([]tcpip.Route{{Destination: subnet, Gateway: "\x00", NIC: nicID2}}) + } - // Test that forwarding increments Tx stats correctly. - if got, want := s.NICInfo()[2].Stats.Tx.Packets.Value(), uint64(1); got != want { - t.Errorf("got Tx.Packets.Value() = %d, want = %d", got, want) - } + // Send a packet to dstAddr. + buf := buffer.NewView(30) + buf[0] = dstAddr[0] + ep1.InjectInbound(fakeNetNumber, stack.PacketBuffer{ + Data: buf.ToVectorisedView(), + }) - if got, want := s.NICInfo()[2].Stats.Tx.Bytes.Value(), uint64(len(buf)); got != want { - t.Errorf("got Tx.Bytes.Value() = %d, want = %d", got, want) + pkt, ok := ep2.Read() + if !ok { + t.Fatal("packet not forwarded") + } + + // Test that the link's MaxHeaderLength is honoured. + if capacity, want := pkt.Pkt.Header.AvailableLength(), int(test.headerLen); capacity != want { + t.Errorf("got Header.AvailableLength() = %d, want = %d", capacity, want) + } + + // Test that forwarding increments Tx stats correctly. + if got, want := s.NICInfo()[nicID2].Stats.Tx.Packets.Value(), uint64(1); got != want { + t.Errorf("got Tx.Packets.Value() = %d, want = %d", got, want) + } + + if got, want := s.NICInfo()[nicID2].Stats.Tx.Bytes.Value(), uint64(len(buf)); got != want { + t.Errorf("got Tx.Bytes.Value() = %d, want = %d", got, want) + } + }) } } @@ -1928,7 +2399,7 @@ func TestNICContextPreservation(t *testing.T) { t.Fatalf("got nicinfos[%d] = _, %t, want _, true; nicinfos = %+v", id, ok, nicinfos) } if got, want := nicinfo.Context == test.want, true; got != want { - t.Fatal("got nicinfo.Context == ctx = %t, want %t; nicinfo.Context = %p, ctx = %p", got, want, nicinfo.Context, test.want) + t.Fatalf("got nicinfo.Context == ctx = %t, want %t; nicinfo.Context = %p, ctx = %p", got, want, nicinfo.Context, test.want) } }) } @@ -2088,13 +2559,29 @@ func TestNICAutoGenLinkLocalAddr(t *testing.T) { e := channel.New(0, 1280, test.linkAddr) s := stack.New(opts) - nicOpts := stack.NICOptions{Name: test.nicName} + nicOpts := stack.NICOptions{Name: test.nicName, Disabled: true} if err := s.CreateNICWithOptions(nicID, e, nicOpts); err != nil { t.Fatalf("CreateNICWithOptions(%d, _, %+v) = %s", nicID, opts, err) } - var expectedMainAddr tcpip.AddressWithPrefix + // A new disabled NIC should not have any address, even if auto generation + // was enabled. + allStackAddrs := s.AllAddresses() + allNICAddrs, ok := allStackAddrs[nicID] + if !ok { + t.Fatalf("entry for %d missing from allStackAddrs = %+v", nicID, allStackAddrs) + } + if l := len(allNICAddrs); l != 0 { + t.Fatalf("got len(allNICAddrs) = %d, want = 0", l) + } + // Enabling the NIC should attempt auto-generation of a link-local + // address. + if err := s.EnableNIC(nicID); err != nil { + t.Fatalf("s.EnableNIC(%d): %s", nicID, err) + } + + var expectedMainAddr tcpip.AddressWithPrefix if test.shouldGen { expectedMainAddr = tcpip.AddressWithPrefix{ Address: test.expectedAddr, @@ -2281,7 +2768,7 @@ func TestNewPEBOnPromotionToPermanent(t *testing.T) { { subnet, err := tcpip.NewSubnet("\x00", "\x00") if err != nil { - t.Fatalf("NewSubnet failed:", err) + t.Fatalf("NewSubnet failed: %v", err) } s.SetRouteTable([]tcpip.Route{{Destination: subnet, Gateway: "\x00", NIC: 1}}) } @@ -2295,11 +2782,11 @@ func TestNewPEBOnPromotionToPermanent(t *testing.T) { // permanentExpired kind. r, err := s.FindRoute(1, "\x01", "\x02", fakeNetNumber, false) if err != nil { - t.Fatal("FindRoute failed:", err) + t.Fatalf("FindRoute failed: %v", err) } defer r.Release() if err := s.RemoveAddress(1, "\x01"); err != nil { - t.Fatalf("RemoveAddress failed:", err) + t.Fatalf("RemoveAddress failed: %v", err) } // @@ -2311,7 +2798,7 @@ func TestNewPEBOnPromotionToPermanent(t *testing.T) { // Add some other address with peb set to // FirstPrimaryEndpoint. if err := s.AddAddressWithOptions(1, fakeNetNumber, "\x03", stack.FirstPrimaryEndpoint); err != nil { - t.Fatal("AddAddressWithOptions failed:", err) + t.Fatalf("AddAddressWithOptions failed: %v", err) } @@ -2319,7 +2806,7 @@ func TestNewPEBOnPromotionToPermanent(t *testing.T) { // make sure the new peb was respected. // (The address should just be promoted now). if err := s.AddAddressWithOptions(1, fakeNetNumber, "\x01", ps); err != nil { - t.Fatal("AddAddressWithOptions failed:", err) + t.Fatalf("AddAddressWithOptions failed: %v", err) } var primaryAddrs []tcpip.Address for _, pa := range s.NICInfo()[1].ProtocolAddresses { @@ -2352,11 +2839,11 @@ func TestNewPEBOnPromotionToPermanent(t *testing.T) { // GetMainNICAddress; else, our original address // should be returned. if err := s.RemoveAddress(1, "\x03"); err != nil { - t.Fatalf("RemoveAddress failed:", err) + t.Fatalf("RemoveAddress failed: %v", err) } addr, err = s.GetMainNICAddress(1, fakeNetNumber) if err != nil { - t.Fatal("s.GetMainNICAddress failed:", err) + t.Fatalf("s.GetMainNICAddress failed: %v", err) } if ps == stack.NeverPrimaryEndpoint { if want := (tcpip.AddressWithPrefix{}); addr != want { @@ -2375,13 +2862,14 @@ func TestNewPEBOnPromotionToPermanent(t *testing.T) { func TestIPv6SourceAddressSelectionScopeAndSameAddress(t *testing.T) { const ( - linkLocalAddr1 = tcpip.Address("\xfe\x80\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x01") - linkLocalAddr2 = tcpip.Address("\xfe\x80\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x02") - uniqueLocalAddr1 = tcpip.Address("\xfc\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x01") - uniqueLocalAddr2 = tcpip.Address("\xfd\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x02") - globalAddr1 = tcpip.Address("\xa0\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x01") - globalAddr2 = tcpip.Address("\xa0\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x02") - nicID = 1 + linkLocalAddr1 = tcpip.Address("\xfe\x80\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x01") + linkLocalAddr2 = tcpip.Address("\xfe\x80\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x02") + linkLocalMulticastAddr = tcpip.Address("\xff\x02\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x01") + uniqueLocalAddr1 = tcpip.Address("\xfc\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x01") + uniqueLocalAddr2 = tcpip.Address("\xfd\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x02") + globalAddr1 = tcpip.Address("\xa0\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x01") + globalAddr2 = tcpip.Address("\xa0\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x02") + nicID = 1 ) // Rule 3 is not tested here, and is instead tested by NDP's AutoGenAddr test. @@ -2455,6 +2943,18 @@ func TestIPv6SourceAddressSelectionScopeAndSameAddress(t *testing.T) { expectedLocalAddr: linkLocalAddr1, }, { + name: "Link Local most preferred for link local multicast (last address)", + nicAddrs: []tcpip.Address{globalAddr1, uniqueLocalAddr1, linkLocalAddr1}, + connectAddr: linkLocalMulticastAddr, + expectedLocalAddr: linkLocalAddr1, + }, + { + name: "Link Local most preferred for link local multicast (first address)", + nicAddrs: []tcpip.Address{linkLocalAddr1, uniqueLocalAddr1, globalAddr1}, + connectAddr: linkLocalMulticastAddr, + expectedLocalAddr: linkLocalAddr1, + }, + { name: "Unique Local most preferred (last address)", nicAddrs: []tcpip.Address{uniqueLocalAddr1, globalAddr1, linkLocalAddr1}, connectAddr: uniqueLocalAddr2, @@ -2524,11 +3024,158 @@ func TestIPv6SourceAddressSelectionScopeAndSameAddress(t *testing.T) { } } +func TestAddRemoveIPv4BroadcastAddressOnNICEnableDisable(t *testing.T) { + const nicID = 1 + + e := loopback.New() + s := stack.New(stack.Options{ + NetworkProtocols: []stack.NetworkProtocol{ipv4.NewProtocol()}, + }) + nicOpts := stack.NICOptions{Disabled: true} + if err := s.CreateNICWithOptions(nicID, e, nicOpts); err != nil { + t.Fatalf("CreateNIC(%d, _, %+v) = %s", nicID, nicOpts, err) + } + + allStackAddrs := s.AllAddresses() + allNICAddrs, ok := allStackAddrs[nicID] + if !ok { + t.Fatalf("entry for %d missing from allStackAddrs = %+v", nicID, allStackAddrs) + } + if l := len(allNICAddrs); l != 0 { + t.Fatalf("got len(allNICAddrs) = %d, want = 0", l) + } + + // Enabling the NIC should add the IPv4 broadcast address. + if err := s.EnableNIC(nicID); err != nil { + t.Fatalf("s.EnableNIC(%d): %s", nicID, err) + } + allStackAddrs = s.AllAddresses() + allNICAddrs, ok = allStackAddrs[nicID] + if !ok { + t.Fatalf("entry for %d missing from allStackAddrs = %+v", nicID, allStackAddrs) + } + if l := len(allNICAddrs); l != 1 { + t.Fatalf("got len(allNICAddrs) = %d, want = 1", l) + } + want := tcpip.ProtocolAddress{ + Protocol: header.IPv4ProtocolNumber, + AddressWithPrefix: tcpip.AddressWithPrefix{ + Address: header.IPv4Broadcast, + PrefixLen: 32, + }, + } + if allNICAddrs[0] != want { + t.Fatalf("got allNICAddrs[0] = %+v, want = %+v", allNICAddrs[0], want) + } + + // Disabling the NIC should remove the IPv4 broadcast address. + if err := s.DisableNIC(nicID); err != nil { + t.Fatalf("s.DisableNIC(%d): %s", nicID, err) + } + allStackAddrs = s.AllAddresses() + allNICAddrs, ok = allStackAddrs[nicID] + if !ok { + t.Fatalf("entry for %d missing from allStackAddrs = %+v", nicID, allStackAddrs) + } + if l := len(allNICAddrs); l != 0 { + t.Fatalf("got len(allNICAddrs) = %d, want = 0", l) + } +} + +// TestLeaveIPv6SolicitedNodeAddrBeforeAddrRemoval tests that removing an IPv6 +// address after leaving its solicited node multicast address does not result in +// an error. +func TestLeaveIPv6SolicitedNodeAddrBeforeAddrRemoval(t *testing.T) { + const nicID = 1 + + s := stack.New(stack.Options{ + NetworkProtocols: []stack.NetworkProtocol{ipv6.NewProtocol()}, + }) + e := channel.New(10, 1280, linkAddr1) + if err := s.CreateNIC(1, e); err != nil { + t.Fatalf("CreateNIC(%d, _): %s", nicID, err) + } + + if err := s.AddAddress(nicID, ipv6.ProtocolNumber, addr1); err != nil { + t.Fatalf("AddAddress(%d, %d, %s): %s", nicID, ipv6.ProtocolNumber, addr1, err) + } + + // The NIC should have joined addr1's solicited node multicast address. + snmc := header.SolicitedNodeAddr(addr1) + in, err := s.IsInGroup(nicID, snmc) + if err != nil { + t.Fatalf("IsInGroup(%d, %s): %s", nicID, snmc, err) + } + if !in { + t.Fatalf("got IsInGroup(%d, %s) = false, want = true", nicID, snmc) + } + + if err := s.LeaveGroup(ipv6.ProtocolNumber, nicID, snmc); err != nil { + t.Fatalf("LeaveGroup(%d, %d, %s): %s", ipv6.ProtocolNumber, nicID, snmc, err) + } + in, err = s.IsInGroup(nicID, snmc) + if err != nil { + t.Fatalf("IsInGroup(%d, %s): %s", nicID, snmc, err) + } + if in { + t.Fatalf("got IsInGroup(%d, %s) = true, want = false", nicID, snmc) + } + + if err := s.RemoveAddress(nicID, addr1); err != nil { + t.Fatalf("RemoveAddress(%d, %s) = %s", nicID, addr1, err) + } +} + +func TestJoinLeaveAllNodesMulticastOnNICEnableDisable(t *testing.T) { + const nicID = 1 + + e := loopback.New() + s := stack.New(stack.Options{ + NetworkProtocols: []stack.NetworkProtocol{ipv6.NewProtocol()}, + }) + nicOpts := stack.NICOptions{Disabled: true} + if err := s.CreateNICWithOptions(nicID, e, nicOpts); err != nil { + t.Fatalf("CreateNIC(%d, _, %+v) = %s", nicID, nicOpts, err) + } + + // Should not be in the IPv6 all-nodes multicast group yet because the NIC has + // not been enabled yet. + isInGroup, err := s.IsInGroup(nicID, header.IPv6AllNodesMulticastAddress) + if err != nil { + t.Fatalf("IsInGroup(%d, %s): %s", nicID, header.IPv6AllNodesMulticastAddress, err) + } + if isInGroup { + t.Fatalf("got IsInGroup(%d, %s) = true, want = false", nicID, header.IPv6AllNodesMulticastAddress) + } + + // The all-nodes multicast group should be joined when the NIC is enabled. + if err := s.EnableNIC(nicID); err != nil { + t.Fatalf("s.EnableNIC(%d): %s", nicID, err) + } + isInGroup, err = s.IsInGroup(nicID, header.IPv6AllNodesMulticastAddress) + if err != nil { + t.Fatalf("IsInGroup(%d, %s): %s", nicID, header.IPv6AllNodesMulticastAddress, err) + } + if !isInGroup { + t.Fatalf("got IsInGroup(%d, %s) = false, want = true", nicID, header.IPv6AllNodesMulticastAddress) + } + + // The all-nodes multicast group should be left when the NIC is disabled. + if err := s.DisableNIC(nicID); err != nil { + t.Fatalf("s.DisableNIC(%d): %s", nicID, err) + } + isInGroup, err = s.IsInGroup(nicID, header.IPv6AllNodesMulticastAddress) + if err != nil { + t.Fatalf("IsInGroup(%d, %s): %s", nicID, header.IPv6AllNodesMulticastAddress, err) + } + if isInGroup { + t.Fatalf("got IsInGroup(%d, %s) = true, want = false", nicID, header.IPv6AllNodesMulticastAddress) + } +} + // TestDoDADWhenNICEnabled tests that IPv6 endpoints that were added while a NIC // was disabled have DAD performed on them when the NIC is enabled. func TestDoDADWhenNICEnabled(t *testing.T) { - t.Parallel() - const dadTransmits = 1 const retransmitTimer = time.Second const nicID = 1 diff --git a/pkg/tcpip/stack/transport_demuxer.go b/pkg/tcpip/stack/transport_demuxer.go index d686e6eb8..9a33ed375 100644 --- a/pkg/tcpip/stack/transport_demuxer.go +++ b/pkg/tcpip/stack/transport_demuxer.go @@ -15,9 +15,9 @@ package stack import ( + "container/heap" "fmt" "math/rand" - "sort" "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/tcpip" @@ -35,7 +35,7 @@ type protocolIDs struct { type transportEndpoints struct { // mu protects all fields of the transportEndpoints. mu sync.RWMutex - endpoints map[TransportEndpointID]*endpointsByNic + endpoints map[TransportEndpointID]*endpointsByNIC // rawEndpoints contains endpoints for raw sockets, which receive all // traffic of a given protocol regardless of port. rawEndpoints []RawTransportEndpoint @@ -46,11 +46,11 @@ type transportEndpoints struct { func (eps *transportEndpoints) unregisterEndpoint(id TransportEndpointID, ep TransportEndpoint, bindToDevice tcpip.NICID) { eps.mu.Lock() defer eps.mu.Unlock() - epsByNic, ok := eps.endpoints[id] + epsByNIC, ok := eps.endpoints[id] if !ok { return } - if !epsByNic.unregisterEndpoint(bindToDevice, ep) { + if !epsByNIC.unregisterEndpoint(bindToDevice, ep) { return } delete(eps.endpoints, id) @@ -66,18 +66,85 @@ func (eps *transportEndpoints) transportEndpoints() []TransportEndpoint { return es } -type endpointsByNic struct { +// iterEndpointsLocked yields all endpointsByNIC in eps that match id, in +// descending order of match quality. If a call to yield returns false, +// iterEndpointsLocked stops iteration and returns immediately. +// +// Preconditions: eps.mu must be locked. +func (eps *transportEndpoints) iterEndpointsLocked(id TransportEndpointID, yield func(*endpointsByNIC) bool) { + // Try to find a match with the id as provided. + if ep, ok := eps.endpoints[id]; ok { + if !yield(ep) { + return + } + } + + // Try to find a match with the id minus the local address. + nid := id + + nid.LocalAddress = "" + if ep, ok := eps.endpoints[nid]; ok { + if !yield(ep) { + return + } + } + + // Try to find a match with the id minus the remote part. + nid.LocalAddress = id.LocalAddress + nid.RemoteAddress = "" + nid.RemotePort = 0 + if ep, ok := eps.endpoints[nid]; ok { + if !yield(ep) { + return + } + } + + // Try to find a match with only the local port. + nid.LocalAddress = "" + if ep, ok := eps.endpoints[nid]; ok { + if !yield(ep) { + return + } + } +} + +// findAllEndpointsLocked returns all endpointsByNIC in eps that match id, in +// descending order of match quality. +// +// Preconditions: eps.mu must be locked. +func (eps *transportEndpoints) findAllEndpointsLocked(id TransportEndpointID) []*endpointsByNIC { + var matchedEPs []*endpointsByNIC + eps.iterEndpointsLocked(id, func(ep *endpointsByNIC) bool { + matchedEPs = append(matchedEPs, ep) + return true + }) + return matchedEPs +} + +// findEndpointLocked returns the endpoint that most closely matches the given id. +// +// Preconditions: eps.mu must be locked. +func (eps *transportEndpoints) findEndpointLocked(id TransportEndpointID) *endpointsByNIC { + var matchedEP *endpointsByNIC + eps.iterEndpointsLocked(id, func(ep *endpointsByNIC) bool { + matchedEP = ep + return false + }) + return matchedEP +} + +type endpointsByNIC struct { mu sync.RWMutex endpoints map[tcpip.NICID]*multiPortEndpoint // seed is a random secret for a jenkins hash. seed uint32 } -func (epsByNic *endpointsByNic) transportEndpoints() []TransportEndpoint { - epsByNic.mu.RLock() - defer epsByNic.mu.RUnlock() +func (epsByNIC *endpointsByNIC) transportEndpoints() []TransportEndpoint { + epsByNIC.mu.RLock() + defer epsByNIC.mu.RUnlock() var eps []TransportEndpoint - for _, ep := range epsByNic.endpoints { + for _, ep := range epsByNIC.endpoints { eps = append(eps, ep.transportEndpoints()...) } return eps @@ -85,13 +152,13 @@ func (epsByNic *endpointsByNic) transportEndpoints() []TransportEndpoint { // HandlePacket is called by the stack when new packets arrive to this transport // endpoint. -func (epsByNic *endpointsByNic) handlePacket(r *Route, id TransportEndpointID, pkt tcpip.PacketBuffer) { - epsByNic.mu.RLock() +func (epsByNIC *endpointsByNIC) handlePacket(r *Route, id TransportEndpointID, pkt PacketBuffer) { + epsByNIC.mu.RLock() - mpep, ok := epsByNic.endpoints[r.ref.nic.ID()] + mpep, ok := epsByNIC.endpoints[r.ref.nic.ID()] if !ok { - if mpep, ok = epsByNic.endpoints[0]; !ok { - epsByNic.mu.RUnlock() // Don't use defer for performance reasons. + if mpep, ok = epsByNIC.endpoints[0]; !ok { + epsByNIC.mu.RUnlock() // Don't use defer for performance reasons. return } } @@ -100,29 +167,29 @@ func (epsByNic *endpointsByNic) handlePacket(r *Route, id TransportEndpointID, p // endpoints bound to the right device. if isMulticastOrBroadcast(id.LocalAddress) { mpep.handlePacketAll(r, id, pkt) - epsByNic.mu.RUnlock() // Don't use defer for performance reasons. + epsByNIC.mu.RUnlock() // Don't use defer for performance reasons. return } // multiPortEndpoints are guaranteed to have at least one element. - transEP := selectEndpoint(id, mpep, epsByNic.seed) + transEP := selectEndpoint(id, mpep, epsByNIC.seed) if queuedProtocol, mustQueue := mpep.demux.queuedProtocols[protocolIDs{mpep.netProto, mpep.transProto}]; mustQueue { queuedProtocol.QueuePacket(r, transEP, id, pkt) - epsByNic.mu.RUnlock() + epsByNIC.mu.RUnlock() return } transEP.HandlePacket(r, id, pkt) - epsByNic.mu.RUnlock() // Don't use defer for performance reasons. + epsByNIC.mu.RUnlock() // Don't use defer for performance reasons. } // HandleControlPacket implements stack.TransportEndpoint.HandleControlPacket. -func (epsByNic *endpointsByNic) handleControlPacket(n *NIC, id TransportEndpointID, typ ControlType, extra uint32, pkt tcpip.PacketBuffer) { - epsByNic.mu.RLock() - defer epsByNic.mu.RUnlock() +func (epsByNIC *endpointsByNIC) handleControlPacket(n *NIC, id TransportEndpointID, typ ControlType, extra uint32, pkt PacketBuffer) { + epsByNIC.mu.RLock() + defer epsByNIC.mu.RUnlock() - mpep, ok := epsByNic.endpoints[n.ID()] + mpep, ok := epsByNIC.endpoints[n.ID()] if !ok { - mpep, ok = epsByNic.endpoints[0] + mpep, ok = epsByNIC.endpoints[0] } if !ok { return @@ -132,40 +199,41 @@ func (epsByNic *endpointsByNic) handleControlPacket(n *NIC, id TransportEndpoint // broadcast like we are doing with handlePacket above? // multiPortEndpoints are guaranteed to have at least one element. - selectEndpoint(id, mpep, epsByNic.seed).HandleControlPacket(id, typ, extra, pkt) + selectEndpoint(id, mpep, epsByNIC.seed).HandleControlPacket(id, typ, extra, pkt) } // registerEndpoint returns true if it succeeds. It fails and returns // false if ep already has an element with the same key. -func (epsByNic *endpointsByNic) registerEndpoint(d *transportDemuxer, netProto tcpip.NetworkProtocolNumber, transProto tcpip.TransportProtocolNumber, t TransportEndpoint, reusePort bool, bindToDevice tcpip.NICID) *tcpip.Error { - epsByNic.mu.Lock() - defer epsByNic.mu.Unlock() +func (epsByNIC *endpointsByNIC) registerEndpoint(d *transportDemuxer, netProto tcpip.NetworkProtocolNumber, transProto tcpip.TransportProtocolNumber, t TransportEndpoint, reusePort bool, bindToDevice tcpip.NICID) *tcpip.Error { + epsByNIC.mu.Lock() + defer epsByNIC.mu.Unlock() - if multiPortEp, ok := epsByNic.endpoints[bindToDevice]; ok { - // There was already a bind. - return multiPortEp.singleRegisterEndpoint(t, reusePort) + multiPortEp, ok := epsByNIC.endpoints[bindToDevice] + if !ok { + multiPortEp = &multiPortEndpoint{ + demux: d, + netProto: netProto, + transProto: transProto, + reuse: reusePort, + } + epsByNIC.endpoints[bindToDevice] = multiPortEp } - // This is a new binding. - multiPortEp := &multiPortEndpoint{demux: d, netProto: netProto, transProto: transProto} - multiPortEp.endpointsMap = make(map[TransportEndpoint]int) - multiPortEp.reuse = reusePort - epsByNic.endpoints[bindToDevice] = multiPortEp return multiPortEp.singleRegisterEndpoint(t, reusePort) } -// unregisterEndpoint returns true if endpointsByNic has to be unregistered. -func (epsByNic *endpointsByNic) unregisterEndpoint(bindToDevice tcpip.NICID, t TransportEndpoint) bool { - epsByNic.mu.Lock() - defer epsByNic.mu.Unlock() - multiPortEp, ok := epsByNic.endpoints[bindToDevice] +// unregisterEndpoint returns true if endpointsByNIC has to be unregistered. +func (epsByNIC *endpointsByNIC) unregisterEndpoint(bindToDevice tcpip.NICID, t TransportEndpoint) bool { + epsByNIC.mu.Lock() + defer epsByNIC.mu.Unlock() + multiPortEp, ok := epsByNIC.endpoints[bindToDevice] if !ok { return false } if multiPortEp.unregisterEndpoint(t) { - delete(epsByNic.endpoints, bindToDevice) + delete(epsByNIC.endpoints, bindToDevice) } - return len(epsByNic.endpoints) == 0 + return len(epsByNIC.endpoints) == 0 } // transportDemuxer demultiplexes packets targeted at a transport endpoint @@ -183,7 +251,7 @@ type transportDemuxer struct { // the dispatcher to delivery packets to the QueuePacket method instead of // calling HandlePacket directly on the endpoint. type queuedTransportProtocol interface { - QueuePacket(r *Route, ep TransportEndpoint, id TransportEndpointID, pkt tcpip.PacketBuffer) + QueuePacket(r *Route, ep TransportEndpoint, id TransportEndpointID, pkt PacketBuffer) } func newTransportDemuxer(stack *Stack) *transportDemuxer { @@ -197,7 +265,7 @@ func newTransportDemuxer(stack *Stack) *transportDemuxer { for proto := range stack.transportProtocols { protoIDs := protocolIDs{netProto, proto} d.protocol[protoIDs] = &transportEndpoints{ - endpoints: make(map[TransportEndpointID]*endpointsByNic), + endpoints: make(map[TransportEndpointID]*endpointsByNIC), } qTransProto, isQueued := (stack.transportProtocols[proto].proto).(queuedTransportProtocol) if isQueued { @@ -222,6 +290,35 @@ func (d *transportDemuxer) registerEndpoint(netProtos []tcpip.NetworkProtocolNum return nil } +type transportEndpointHeap []TransportEndpoint + +var _ heap.Interface = (*transportEndpointHeap)(nil) + +func (h *transportEndpointHeap) Len() int { + return len(*h) +} + +func (h *transportEndpointHeap) Less(i, j int) bool { + return (*h)[i].UniqueID() < (*h)[j].UniqueID() +} + +func (h *transportEndpointHeap) Swap(i, j int) { + (*h)[i], (*h)[j] = (*h)[j], (*h)[i] +} + +func (h *transportEndpointHeap) Push(x interface{}) { + *h = append(*h, x.(TransportEndpoint)) +} + +func (h *transportEndpointHeap) Pop() interface{} { + old := *h + n := len(old) + x := old[n-1] + old[n-1] = nil + *h = old[:n-1] + return x +} + // multiPortEndpoint is a container for TransportEndpoints which are bound to // the same pair of address and port. endpointsArr always has at least one // element. @@ -237,15 +334,14 @@ type multiPortEndpoint struct { netProto tcpip.NetworkProtocolNumber transProto tcpip.TransportProtocolNumber - endpointsArr []TransportEndpoint - endpointsMap map[TransportEndpoint]int + endpoints transportEndpointHeap // reuse indicates if more than one endpoint is allowed. reuse bool } func (ep *multiPortEndpoint) transportEndpoints() []TransportEndpoint { ep.mu.RLock() - eps := append([]TransportEndpoint(nil), ep.endpointsArr...) + eps := append([]TransportEndpoint(nil), ep.endpoints...) ep.mu.RUnlock() return eps } @@ -262,8 +358,8 @@ func reciprocalScale(val, n uint32) uint32 { // ports then uses it to select a socket. In this case, all packets from one // address will be sent to same endpoint. func selectEndpoint(id TransportEndpointID, mpep *multiPortEndpoint, seed uint32) TransportEndpoint { - if len(mpep.endpointsArr) == 1 { - return mpep.endpointsArr[0] + if len(mpep.endpoints) == 1 { + return mpep.endpoints[0] } payload := []byte{ @@ -279,51 +375,28 @@ func selectEndpoint(id TransportEndpointID, mpep *multiPortEndpoint, seed uint32 h.Write([]byte(id.RemoteAddress)) hash := h.Sum32() - idx := reciprocalScale(hash, uint32(len(mpep.endpointsArr))) - return mpep.endpointsArr[idx] + idx := reciprocalScale(hash, uint32(len(mpep.endpoints))) + return mpep.endpoints[idx] } -func (ep *multiPortEndpoint) handlePacketAll(r *Route, id TransportEndpointID, pkt tcpip.PacketBuffer) { +func (ep *multiPortEndpoint) handlePacketAll(r *Route, id TransportEndpointID, pkt PacketBuffer) { ep.mu.RLock() queuedProtocol, mustQueue := ep.demux.queuedProtocols[protocolIDs{ep.netProto, ep.transProto}] - for i, endpoint := range ep.endpointsArr { - // HandlePacket takes ownership of pkt, so each endpoint needs - // its own copy except for the final one. - if i == len(ep.endpointsArr)-1 { - if mustQueue { - queuedProtocol.QueuePacket(r, endpoint, id, pkt) - break - } - endpoint.HandlePacket(r, id, pkt) - break - } + // HandlePacket takes ownership of pkt, so each endpoint needs + // its own copy except for the final one. + for _, endpoint := range ep.endpoints[:len(ep.endpoints)-1] { if mustQueue { queuedProtocol.QueuePacket(r, endpoint, id, pkt.Clone()) - continue + } else { + endpoint.HandlePacket(r, id, pkt.Clone()) } - endpoint.HandlePacket(r, id, pkt.Clone()) - } - ep.mu.RUnlock() // Don't use defer for performance reasons. -} - -// Close implements stack.TransportEndpoint.Close. -func (ep *multiPortEndpoint) Close() { - ep.mu.RLock() - eps := append([]TransportEndpoint(nil), ep.endpointsArr...) - ep.mu.RUnlock() - for _, e := range eps { - e.Close() } -} - -// Wait implements stack.TransportEndpoint.Wait. -func (ep *multiPortEndpoint) Wait() { - ep.mu.RLock() - eps := append([]TransportEndpoint(nil), ep.endpointsArr...) - ep.mu.RUnlock() - for _, e := range eps { - e.Wait() + if endpoint := ep.endpoints[len(ep.endpoints)-1]; mustQueue { + queuedProtocol.QueuePacket(r, endpoint, id, pkt) + } else { + endpoint.HandlePacket(r, id, pkt) } + ep.mu.RUnlock() // Don't use defer for performance reasons. } // singleRegisterEndpoint tries to add an endpoint to the multiPortEndpoint @@ -332,26 +405,15 @@ func (ep *multiPortEndpoint) singleRegisterEndpoint(t TransportEndpoint, reusePo ep.mu.Lock() defer ep.mu.Unlock() - if len(ep.endpointsArr) > 0 { + if len(ep.endpoints) != 0 { // If it was previously bound, we need to check if we can bind again. if !ep.reuse || !reusePort { return tcpip.ErrPortInUse } } - // A new endpoint is added into endpointsArr and its index there is saved in - // endpointsMap. This will allow us to remove endpoint from the array fast. - ep.endpointsMap[t] = len(ep.endpointsArr) - ep.endpointsArr = append(ep.endpointsArr, t) + heap.Push(&ep.endpoints, t) - // ep.endpointsArr is sorted by endpoint unique IDs, so that endpoints - // can be restored in the same order. - sort.Slice(ep.endpointsArr, func(i, j int) bool { - return ep.endpointsArr[i].UniqueID() < ep.endpointsArr[j].UniqueID() - }) - for i, e := range ep.endpointsArr { - ep.endpointsMap[e] = i - } return nil } @@ -360,21 +422,13 @@ func (ep *multiPortEndpoint) unregisterEndpoint(t TransportEndpoint) bool { ep.mu.Lock() defer ep.mu.Unlock() - idx, ok := ep.endpointsMap[t] - if !ok { - return false - } - delete(ep.endpointsMap, t) - l := len(ep.endpointsArr) - if l > 1 { - // The last endpoint in endpointsArr is moved instead of the deleted one. - lastEp := ep.endpointsArr[l-1] - ep.endpointsArr[idx] = lastEp - ep.endpointsMap[lastEp] = idx - ep.endpointsArr = ep.endpointsArr[0 : l-1] - return false + for i, endpoint := range ep.endpoints { + if endpoint == t { + heap.Remove(&ep.endpoints, i) + break + } } - return true + return len(ep.endpoints) == 0 } func (d *transportDemuxer) singleRegisterEndpoint(netProto tcpip.NetworkProtocolNumber, protocol tcpip.TransportProtocolNumber, id TransportEndpointID, ep TransportEndpoint, reusePort bool, bindToDevice tcpip.NICID) *tcpip.Error { @@ -391,19 +445,16 @@ func (d *transportDemuxer) singleRegisterEndpoint(netProto tcpip.NetworkProtocol eps.mu.Lock() defer eps.mu.Unlock() - if epsByNic, ok := eps.endpoints[id]; ok { - // There was already a binding. - return epsByNic.registerEndpoint(d, netProto, protocol, ep, reusePort, bindToDevice) - } - - // This is a new binding. - epsByNic := &endpointsByNic{ - endpoints: make(map[tcpip.NICID]*multiPortEndpoint), - seed: rand.Uint32(), + epsByNIC, ok := eps.endpoints[id] + if !ok { + epsByNIC = &endpointsByNIC{ + endpoints: make(map[tcpip.NICID]*multiPortEndpoint), + seed: rand.Uint32(), + } + eps.endpoints[id] = epsByNIC } - eps.endpoints[id] = epsByNic - return epsByNic.registerEndpoint(d, netProto, protocol, ep, reusePort, bindToDevice) + return epsByNIC.registerEndpoint(d, netProto, protocol, ep, reusePort, bindToDevice) } // unregisterEndpoint unregisters the endpoint with the given id such that it @@ -416,84 +467,60 @@ func (d *transportDemuxer) unregisterEndpoint(netProtos []tcpip.NetworkProtocolN } } -var loopbackSubnet = func() tcpip.Subnet { - sn, err := tcpip.NewSubnet("\x7f\x00\x00\x00", "\xff\x00\x00\x00") - if err != nil { - panic(err) - } - return sn -}() - // deliverPacket attempts to find one or more matching transport endpoints, and // then, if matches are found, delivers the packet to them. Returns true if // the packet no longer needs to be handled. -func (d *transportDemuxer) deliverPacket(r *Route, protocol tcpip.TransportProtocolNumber, pkt tcpip.PacketBuffer, id TransportEndpointID) bool { +func (d *transportDemuxer) deliverPacket(r *Route, protocol tcpip.TransportProtocolNumber, pkt PacketBuffer, id TransportEndpointID) bool { eps, ok := d.protocol[protocolIDs{r.NetProto, protocol}] if !ok { return false } - eps.mu.RLock() - - // Determine which transport endpoint or endpoints to deliver this packet to. // If the packet is a UDP broadcast or multicast, then find all matching - // transport endpoints. If the packet is a TCP packet with a non-unicast - // source or destination address, then do nothing further and instruct - // the caller to do the same. - var destEps []*endpointsByNic - switch protocol { - case header.UDPProtocolNumber: - if isMulticastOrBroadcast(id.LocalAddress) { - destEps = d.findAllEndpointsLocked(eps, id) - break - } - - if ep := d.findEndpointLocked(eps, id); ep != nil { - destEps = append(destEps, ep) + // transport endpoints. + if protocol == header.UDPProtocolNumber && isMulticastOrBroadcast(id.LocalAddress) { + eps.mu.RLock() + destEPs := eps.findAllEndpointsLocked(id) + eps.mu.RUnlock() + // Fail if we didn't find at least one matching transport endpoint. + if len(destEPs) == 0 { + r.Stats().UDP.UnknownPortErrors.Increment() + return false } - - case header.TCPProtocolNumber: - if !(isUnicast(r.LocalAddress) && isUnicast(r.RemoteAddress)) { - // TCP can only be used to communicate between a single - // source and a single destination; the addresses must - // be unicast. - eps.mu.RUnlock() - r.Stats().TCP.InvalidSegmentsReceived.Increment() - return true + // handlePacket takes ownership of pkt, so each endpoint needs its own + // copy except for the final one. + for _, ep := range destEPs[:len(destEPs)-1] { + ep.handlePacket(r, id, pkt.Clone()) } + destEPs[len(destEPs)-1].handlePacket(r, id, pkt) + return true + } - fallthrough - - default: - if ep := d.findEndpointLocked(eps, id); ep != nil { - destEps = append(destEps, ep) - } + // If the packet is a TCP packet with a non-unicast source or destination + // address, then do nothing further and instruct the caller to do the same. + if protocol == header.TCPProtocolNumber && (!isUnicast(r.LocalAddress) || !isUnicast(r.RemoteAddress)) { + // TCP can only be used to communicate between a single source and a + // single destination; the addresses must be unicast. + r.Stats().TCP.InvalidSegmentsReceived.Increment() + return true } + eps.mu.RLock() + ep := eps.findEndpointLocked(id) eps.mu.RUnlock() - - // Fail if we didn't find at least one matching transport endpoint. - if len(destEps) == 0 { - // UDP packet could not be delivered to an unknown destination port. + if ep == nil { if protocol == header.UDPProtocolNumber { r.Stats().UDP.UnknownPortErrors.Increment() } return false } - - // HandlePacket takes ownership of pkt, so each endpoint needs its own - // copy except for the final one. - for _, ep := range destEps[:len(destEps)-1] { - ep.handlePacket(r, id, pkt.Clone()) - } - destEps[len(destEps)-1].handlePacket(r, id, pkt) - + ep.handlePacket(r, id, pkt) return true } // deliverRawPacket attempts to deliver the given packet and returns whether it // was delivered successfully. -func (d *transportDemuxer) deliverRawPacket(r *Route, protocol tcpip.TransportProtocolNumber, pkt tcpip.PacketBuffer) bool { +func (d *transportDemuxer) deliverRawPacket(r *Route, protocol tcpip.TransportProtocolNumber, pkt PacketBuffer) bool { eps, ok := d.protocol[protocolIDs{r.NetProto, protocol}] if !ok { return false @@ -517,99 +544,53 @@ func (d *transportDemuxer) deliverRawPacket(r *Route, protocol tcpip.TransportPr // deliverControlPacket attempts to deliver the given control packet. Returns // true if it found an endpoint, false otherwise. -func (d *transportDemuxer) deliverControlPacket(n *NIC, net tcpip.NetworkProtocolNumber, trans tcpip.TransportProtocolNumber, typ ControlType, extra uint32, pkt tcpip.PacketBuffer, id TransportEndpointID) bool { +func (d *transportDemuxer) deliverControlPacket(n *NIC, net tcpip.NetworkProtocolNumber, trans tcpip.TransportProtocolNumber, typ ControlType, extra uint32, pkt PacketBuffer, id TransportEndpointID) bool { eps, ok := d.protocol[protocolIDs{net, trans}] if !ok { return false } - // Try to find the endpoint. eps.mu.RLock() - ep := d.findEndpointLocked(eps, id) + ep := eps.findEndpointLocked(id) eps.mu.RUnlock() - - // Fail if we didn't find one. if ep == nil { return false } - // Deliver the packet. ep.handleControlPacket(n, id, typ, extra, pkt) - return true } -func (d *transportDemuxer) findAllEndpointsLocked(eps *transportEndpoints, id TransportEndpointID) []*endpointsByNic { - var matchedEPs []*endpointsByNic - // Try to find a match with the id as provided. - if ep, ok := eps.endpoints[id]; ok { - matchedEPs = append(matchedEPs, ep) - } - - // Try to find a match with the id minus the local address. - nid := id - - nid.LocalAddress = "" - if ep, ok := eps.endpoints[nid]; ok { - matchedEPs = append(matchedEPs, ep) - } - - // Try to find a match with the id minus the remote part. - nid.LocalAddress = id.LocalAddress - nid.RemoteAddress = "" - nid.RemotePort = 0 - if ep, ok := eps.endpoints[nid]; ok { - matchedEPs = append(matchedEPs, ep) - } - - // Try to find a match with only the local port. - nid.LocalAddress = "" - if ep, ok := eps.endpoints[nid]; ok { - matchedEPs = append(matchedEPs, ep) - } - return matchedEPs -} - // findTransportEndpoint find a single endpoint that most closely matches the provided id. func (d *transportDemuxer) findTransportEndpoint(netProto tcpip.NetworkProtocolNumber, transProto tcpip.TransportProtocolNumber, id TransportEndpointID, r *Route) TransportEndpoint { eps, ok := d.protocol[protocolIDs{netProto, transProto}] if !ok { return nil } - // Try to find the endpoint. + eps.mu.RLock() - epsByNic := d.findEndpointLocked(eps, id) - // Fail if we didn't find one. - if epsByNic == nil { + epsByNIC := eps.findEndpointLocked(id) + if epsByNIC == nil { eps.mu.RUnlock() return nil } - epsByNic.mu.RLock() + epsByNIC.mu.RLock() eps.mu.RUnlock() - mpep, ok := epsByNic.endpoints[r.ref.nic.ID()] + mpep, ok := epsByNIC.endpoints[r.ref.nic.ID()] if !ok { - if mpep, ok = epsByNic.endpoints[0]; !ok { - epsByNic.mu.RUnlock() // Don't use defer for performance reasons. + if mpep, ok = epsByNIC.endpoints[0]; !ok { + epsByNIC.mu.RUnlock() // Don't use defer for performance reasons. return nil } } - ep := selectEndpoint(id, mpep, epsByNic.seed) - epsByNic.mu.RUnlock() + ep := selectEndpoint(id, mpep, epsByNIC.seed) + epsByNIC.mu.RUnlock() return ep } -// findEndpointLocked returns the endpoint that most closely matches the given -// id. -func (d *transportDemuxer) findEndpointLocked(eps *transportEndpoints, id TransportEndpointID) *endpointsByNic { - if matchedEPs := d.findAllEndpointsLocked(eps, id); len(matchedEPs) > 0 { - return matchedEPs[0] - } - return nil -} - // registerRawEndpoint registers the given endpoint with the dispatcher such // that packets of the appropriate protocol are delivered to it. A single // packet can be sent to one or more raw endpoints along with a non-raw @@ -621,8 +602,8 @@ func (d *transportDemuxer) registerRawEndpoint(netProto tcpip.NetworkProtocolNum } eps.mu.Lock() - defer eps.mu.Unlock() eps.rawEndpoints = append(eps.rawEndpoints, ep) + eps.mu.Unlock() return nil } @@ -636,13 +617,16 @@ func (d *transportDemuxer) unregisterRawEndpoint(netProto tcpip.NetworkProtocolN } eps.mu.Lock() - defer eps.mu.Unlock() for i, rawEP := range eps.rawEndpoints { if rawEP == ep { - eps.rawEndpoints = append(eps.rawEndpoints[:i], eps.rawEndpoints[i+1:]...) - return + lastIdx := len(eps.rawEndpoints) - 1 + eps.rawEndpoints[i] = eps.rawEndpoints[lastIdx] + eps.rawEndpoints[lastIdx] = nil + eps.rawEndpoints = eps.rawEndpoints[:lastIdx] + break } } + eps.mu.Unlock() } func isMulticastOrBroadcast(addr tcpip.Address) bool { diff --git a/pkg/tcpip/stack/transport_demuxer_test.go b/pkg/tcpip/stack/transport_demuxer_test.go index 5e9237de9..2474a7db3 100644 --- a/pkg/tcpip/stack/transport_demuxer_test.go +++ b/pkg/tcpip/stack/transport_demuxer_test.go @@ -31,84 +31,58 @@ import ( ) const ( - stackV6Addr = "\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x01" - testV6Addr = "\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x02" + testSrcAddrV6 = "\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x01" + testDstAddrV6 = "\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x02" - stackAddr = "\x0a\x00\x00\x01" - stackPort = 1234 - testPort = 4096 + testSrcAddrV4 = "\x0a\x00\x00\x01" + testDstAddrV4 = "\x0a\x00\x00\x02" + + testDstPort = 1234 + testSrcPort = 4096 ) type testContext struct { - t *testing.T linkEps map[tcpip.NICID]*channel.Endpoint s *stack.Stack - - ep tcpip.Endpoint - wq waiter.Queue -} - -func (c *testContext) cleanup() { - if c.ep != nil { - c.ep.Close() - } -} - -func (c *testContext) createV6Endpoint(v6only bool) { - var err *tcpip.Error - c.ep, err = c.s.NewEndpoint(udp.ProtocolNumber, ipv6.ProtocolNumber, &c.wq) - if err != nil { - c.t.Fatalf("NewEndpoint failed: %v", err) - } - - if err := c.ep.SetSockOptBool(tcpip.V6OnlyOption, v6only); err != nil { - c.t.Fatalf("SetSockOpt failed: %v", err) - } + wq waiter.Queue } // newDualTestContextMultiNIC creates the testing context and also linkEpIDs NICs. func newDualTestContextMultiNIC(t *testing.T, mtu uint32, linkEpIDs []tcpip.NICID) *testContext { s := stack.New(stack.Options{ NetworkProtocols: []stack.NetworkProtocol{ipv4.NewProtocol(), ipv6.NewProtocol()}, - TransportProtocols: []stack.TransportProtocol{udp.NewProtocol()}}) + TransportProtocols: []stack.TransportProtocol{udp.NewProtocol()}, + }) linkEps := make(map[tcpip.NICID]*channel.Endpoint) for _, linkEpID := range linkEpIDs { channelEp := channel.New(256, mtu, "") if err := s.CreateNIC(linkEpID, channelEp); err != nil { - t.Fatalf("CreateNIC failed: %v", err) + t.Fatalf("CreateNIC failed: %s", err) } linkEps[linkEpID] = channelEp - if err := s.AddAddress(linkEpID, ipv4.ProtocolNumber, stackAddr); err != nil { - t.Fatalf("AddAddress IPv4 failed: %v", err) + if err := s.AddAddress(linkEpID, ipv4.ProtocolNumber, testDstAddrV4); err != nil { + t.Fatalf("AddAddress IPv4 failed: %s", err) } - if err := s.AddAddress(linkEpID, ipv6.ProtocolNumber, stackV6Addr); err != nil { - t.Fatalf("AddAddress IPv6 failed: %v", err) + if err := s.AddAddress(linkEpID, ipv6.ProtocolNumber, testDstAddrV6); err != nil { + t.Fatalf("AddAddress IPv6 failed: %s", err) } } s.SetRouteTable([]tcpip.Route{ - { - Destination: header.IPv4EmptySubnet, - NIC: 1, - }, - { - Destination: header.IPv6EmptySubnet, - NIC: 1, - }, + {Destination: header.IPv4EmptySubnet, NIC: 1}, + {Destination: header.IPv6EmptySubnet, NIC: 1}, }) return &testContext{ - t: t, s: s, linkEps: linkEps, } } type headers struct { - srcPort uint16 - dstPort uint16 + srcPort, dstPort uint16 } func newPayload() []byte { @@ -119,6 +93,47 @@ func newPayload() []byte { return b } +func (c *testContext) sendV4Packet(payload []byte, h *headers, linkEpID tcpip.NICID) { + buf := buffer.NewView(header.UDPMinimumSize + header.IPv4MinimumSize + len(payload)) + payloadStart := len(buf) - len(payload) + copy(buf[payloadStart:], payload) + + // Initialize the IP header. + ip := header.IPv4(buf) + ip.Encode(&header.IPv4Fields{ + IHL: header.IPv4MinimumSize, + TOS: 0x80, + TotalLength: uint16(len(buf)), + TTL: 65, + Protocol: uint8(udp.ProtocolNumber), + SrcAddr: testSrcAddrV4, + DstAddr: testDstAddrV4, + }) + ip.SetChecksum(^ip.CalculateChecksum()) + + // Initialize the UDP header. + u := header.UDP(buf[header.IPv4MinimumSize:]) + u.Encode(&header.UDPFields{ + SrcPort: h.srcPort, + DstPort: h.dstPort, + Length: uint16(header.UDPMinimumSize + len(payload)), + }) + + // Calculate the UDP pseudo-header checksum. + xsum := header.PseudoHeaderChecksum(udp.ProtocolNumber, testSrcAddrV4, testDstAddrV4, uint16(len(u))) + + // Calculate the UDP checksum and set it. + xsum = header.Checksum(payload, xsum) + u.SetChecksum(^u.CalculateChecksum(xsum)) + + // Inject packet. + c.linkEps[linkEpID].InjectInbound(ipv4.ProtocolNumber, stack.PacketBuffer{ + Data: buf.ToVectorisedView(), + NetworkHeader: buffer.View(ip), + TransportHeader: buffer.View(u), + }) +} + func (c *testContext) sendV6Packet(payload []byte, h *headers, linkEpID tcpip.NICID) { // Allocate a buffer for data and headers. buf := buffer.NewView(header.UDPMinimumSize + header.IPv6MinimumSize + len(payload)) @@ -130,8 +145,8 @@ func (c *testContext) sendV6Packet(payload []byte, h *headers, linkEpID tcpip.NI PayloadLength: uint16(header.UDPMinimumSize + len(payload)), NextHeader: uint8(udp.ProtocolNumber), HopLimit: 65, - SrcAddr: testV6Addr, - DstAddr: stackV6Addr, + SrcAddr: testSrcAddrV6, + DstAddr: testDstAddrV6, }) // Initialize the UDP header. @@ -143,15 +158,17 @@ func (c *testContext) sendV6Packet(payload []byte, h *headers, linkEpID tcpip.NI }) // Calculate the UDP pseudo-header checksum. - xsum := header.PseudoHeaderChecksum(udp.ProtocolNumber, testV6Addr, stackV6Addr, uint16(len(u))) + xsum := header.PseudoHeaderChecksum(udp.ProtocolNumber, testSrcAddrV6, testDstAddrV6, uint16(len(u))) // Calculate the UDP checksum and set it. xsum = header.Checksum(payload, xsum) u.SetChecksum(^u.CalculateChecksum(xsum)) // Inject packet. - c.linkEps[linkEpID].InjectInbound(ipv6.ProtocolNumber, tcpip.PacketBuffer{ - Data: buf.ToVectorisedView(), + c.linkEps[linkEpID].InjectInbound(ipv6.ProtocolNumber, stack.PacketBuffer{ + Data: buf.ToVectorisedView(), + NetworkHeader: buffer.View(ip), + TransportHeader: buffer.View(u), }) } @@ -167,38 +184,48 @@ func TestTransportDemuxerRegister(t *testing.T) { t.Run(test.name, func(t *testing.T) { s := stack.New(stack.Options{ NetworkProtocols: []stack.NetworkProtocol{ipv4.NewProtocol()}, - TransportProtocols: []stack.TransportProtocol{udp.NewProtocol()}}) - if got, want := s.RegisterTransportEndpoint(0, []tcpip.NetworkProtocolNumber{test.proto}, udp.ProtocolNumber, stack.TransportEndpointID{}, nil, false, 0), test.want; got != want { - t.Fatalf("s.RegisterTransportEndpoint(...) = %v, want %v", got, want) + TransportProtocols: []stack.TransportProtocol{udp.NewProtocol()}, + }) + var wq waiter.Queue + ep, err := s.NewEndpoint(udp.ProtocolNumber, ipv4.ProtocolNumber, &wq) + if err != nil { + t.Fatal(err) + } + tEP, ok := ep.(stack.TransportEndpoint) + if !ok { + t.Fatalf("%T does not implement stack.TransportEndpoint", ep) + } + if got, want := s.RegisterTransportEndpoint(0, []tcpip.NetworkProtocolNumber{test.proto}, udp.ProtocolNumber, stack.TransportEndpointID{}, tEP, false, 0), test.want; got != want { + t.Fatalf("s.RegisterTransportEndpoint(...) = %s, want %s", got, want) } }) } } -// TestReuseBindToDevice injects varied packets on input devices and checks that +// TestBindToDeviceDistribution injects varied packets on input devices and checks that // the distribution of packets received matches expectations. -func TestDistribution(t *testing.T) { +func TestBindToDeviceDistribution(t *testing.T) { type endpointSockopts struct { - reuse int + reuse bool bindToDevice tcpip.NICID } for _, test := range []struct { name string // endpoints will received the inject packets. endpoints []endpointSockopts - // wantedDistribution is the wanted ratio of packets received on each + // wantDistributions is the want ratio of packets received on each // endpoint for each NIC on which packets are injected. - wantedDistributions map[tcpip.NICID][]float64 + wantDistributions map[tcpip.NICID][]float64 }{ { "BindPortReuse", // 5 endpoints that all have reuse set. []endpointSockopts{ - {1, 0}, - {1, 0}, - {1, 0}, - {1, 0}, - {1, 0}, + {reuse: true, bindToDevice: 0}, + {reuse: true, bindToDevice: 0}, + {reuse: true, bindToDevice: 0}, + {reuse: true, bindToDevice: 0}, + {reuse: true, bindToDevice: 0}, }, map[tcpip.NICID][]float64{ // Injected packets on dev0 get distributed evenly. @@ -209,9 +236,9 @@ func TestDistribution(t *testing.T) { "BindToDevice", // 3 endpoints with various bindings. []endpointSockopts{ - {0, 1}, - {0, 2}, - {0, 3}, + {reuse: false, bindToDevice: 1}, + {reuse: false, bindToDevice: 2}, + {reuse: false, bindToDevice: 3}, }, map[tcpip.NICID][]float64{ // Injected packets on dev0 go only to the endpoint bound to dev0. @@ -226,12 +253,12 @@ func TestDistribution(t *testing.T) { "ReuseAndBindToDevice", // 6 endpoints with various bindings. []endpointSockopts{ - {1, 1}, - {1, 1}, - {1, 2}, - {1, 2}, - {1, 2}, - {1, 0}, + {reuse: true, bindToDevice: 1}, + {reuse: true, bindToDevice: 1}, + {reuse: true, bindToDevice: 2}, + {reuse: true, bindToDevice: 2}, + {reuse: true, bindToDevice: 2}, + {reuse: true, bindToDevice: 0}, }, map[tcpip.NICID][]float64{ // Injected packets on dev0 get distributed among endpoints bound to @@ -245,17 +272,17 @@ func TestDistribution(t *testing.T) { }, }, } { - t.Run(test.name, func(t *testing.T) { - for device, wantedDistribution := range test.wantedDistributions { - t.Run(string(device), func(t *testing.T) { + for protoName, netProtoNum := range map[string]tcpip.NetworkProtocolNumber{ + "IPv4": ipv4.ProtocolNumber, + "IPv6": ipv6.ProtocolNumber, + } { + for device, wantDistribution := range test.wantDistributions { + t.Run(test.name+protoName+string(device), func(t *testing.T) { var devices []tcpip.NICID - for d := range test.wantedDistributions { + for d := range test.wantDistributions { devices = append(devices, d) } c := newDualTestContextMultiNIC(t, defaultMTU, devices) - defer c.cleanup() - - c.createV6Endpoint(false) eps := make(map[tcpip.Endpoint]int) @@ -269,9 +296,9 @@ func TestDistribution(t *testing.T) { defer close(ch) var err *tcpip.Error - ep, err := c.s.NewEndpoint(udp.ProtocolNumber, ipv6.ProtocolNumber, &wq) + ep, err := c.s.NewEndpoint(udp.ProtocolNumber, netProtoNum, &wq) if err != nil { - c.t.Fatalf("NewEndpoint failed: %v", err) + t.Fatalf("NewEndpoint failed: %s", err) } eps[ep] = i @@ -282,22 +309,31 @@ func TestDistribution(t *testing.T) { }(ep) defer ep.Close() - reusePortOption := tcpip.ReusePortOption(endpoint.reuse) - if err := ep.SetSockOpt(reusePortOption); err != nil { - c.t.Fatalf("SetSockOpt(%#v) on endpoint %d failed: %v", reusePortOption, i, err) + if err := ep.SetSockOptBool(tcpip.ReusePortOption, endpoint.reuse); err != nil { + t.Fatalf("SetSockOptBool(ReusePortOption, %t) on endpoint %d failed: %s", endpoint.reuse, i, err) } bindToDeviceOption := tcpip.BindToDeviceOption(endpoint.bindToDevice) if err := ep.SetSockOpt(bindToDeviceOption); err != nil { - c.t.Fatalf("SetSockOpt(%#v) on endpoint %d failed: %v", bindToDeviceOption, i, err) + t.Fatalf("SetSockOpt(%#v) on endpoint %d failed: %s", bindToDeviceOption, i, err) + } + + var dstAddr tcpip.Address + switch netProtoNum { + case ipv4.ProtocolNumber: + dstAddr = testDstAddrV4 + case ipv6.ProtocolNumber: + dstAddr = testDstAddrV6 + default: + t.Fatalf("unexpected protocol number: %d", netProtoNum) } - if err := ep.Bind(tcpip.FullAddress{Addr: stackV6Addr, Port: stackPort}); err != nil { - t.Fatalf("ep.Bind(...) on endpoint %d failed: %v", i, err) + if err := ep.Bind(tcpip.FullAddress{Addr: dstAddr, Port: testDstPort}); err != nil { + t.Fatalf("ep.Bind(...) on endpoint %d failed: %s", i, err) } } npackets := 100000 nports := 10000 - if got, want := len(test.endpoints), len(wantedDistribution); got != want { + if got, want := len(test.endpoints), len(wantDistribution); got != want { t.Fatalf("got len(test.endpoints) = %d, want %d", got, want) } ports := make(map[uint16]tcpip.Endpoint) @@ -306,17 +342,22 @@ func TestDistribution(t *testing.T) { // Send a packet. port := uint16(i % nports) payload := newPayload() - c.sendV6Packet(payload, - &headers{ - srcPort: testPort + port, - dstPort: stackPort}, - device) + hdrs := &headers{ + srcPort: testSrcPort + port, + dstPort: testDstPort, + } + switch netProtoNum { + case ipv4.ProtocolNumber: + c.sendV4Packet(payload, hdrs, device) + case ipv6.ProtocolNumber: + c.sendV6Packet(payload, hdrs, device) + default: + t.Fatalf("unexpected protocol number: %d", netProtoNum) + } - var addr tcpip.FullAddress ep := <-pollChannel - _, _, err := ep.Read(&addr) - if err != nil { - c.t.Fatalf("Read on endpoint %d failed: %v", eps[ep], err) + if _, _, err := ep.Read(nil); err != nil { + t.Fatalf("Read on endpoint %d failed: %s", eps[ep], err) } stats[ep]++ if i < nports { @@ -332,17 +373,17 @@ func TestDistribution(t *testing.T) { // Check that a packet distribution is as expected. for ep, i := range eps { - wantedRatio := wantedDistribution[i] - wantedRecv := wantedRatio * float64(npackets) + wantRatio := wantDistribution[i] + wantRecv := wantRatio * float64(npackets) actualRecv := stats[ep] actualRatio := float64(stats[ep]) / float64(npackets) // The deviation is less than 10%. - if math.Abs(actualRatio-wantedRatio) > 0.05 { - t.Errorf("wanted about %.0f%% (%.0f of %d) packets to arrive on endpoint %d, got %.0f%% (%d of %d)", wantedRatio*100, wantedRecv, npackets, i, actualRatio*100, actualRecv, npackets) + if math.Abs(actualRatio-wantRatio) > 0.05 { + t.Errorf("want about %.0f%% (%.0f of %d) packets to arrive on endpoint %d, got %.0f%% (%d of %d)", wantRatio*100, wantRecv, npackets, i, actualRatio*100, actualRecv, npackets) } } }) } - }) + } } } diff --git a/pkg/tcpip/stack/transport_test.go b/pkg/tcpip/stack/transport_test.go index 869c69a6d..3084e6593 100644 --- a/pkg/tcpip/stack/transport_test.go +++ b/pkg/tcpip/stack/transport_test.go @@ -19,7 +19,6 @@ import ( "gvisor.dev/gvisor/pkg/tcpip" "gvisor.dev/gvisor/pkg/tcpip/buffer" - "gvisor.dev/gvisor/pkg/tcpip/iptables" "gvisor.dev/gvisor/pkg/tcpip/link/channel" "gvisor.dev/gvisor/pkg/tcpip/link/loopback" "gvisor.dev/gvisor/pkg/tcpip/stack" @@ -57,10 +56,16 @@ func (f *fakeTransportEndpoint) Stats() tcpip.EndpointStats { return nil } +func (f *fakeTransportEndpoint) SetOwner(owner tcpip.PacketOwner) {} + func newFakeTransportEndpoint(s *stack.Stack, proto *fakeTransportProtocol, netProto tcpip.NetworkProtocolNumber, uniqueID uint64) tcpip.Endpoint { return &fakeTransportEndpoint{stack: s, TransportEndpointInfo: stack.TransportEndpointInfo{NetProto: netProto}, proto: proto, uniqueID: uniqueID} } +func (f *fakeTransportEndpoint) Abort() { + f.Close() +} + func (f *fakeTransportEndpoint) Close() { f.route.Release() } @@ -83,7 +88,7 @@ func (f *fakeTransportEndpoint) Write(p tcpip.Payloader, opts tcpip.WriteOptions if err != nil { return 0, nil, err } - if err := f.route.WritePacket(nil /* gso */, stack.NetworkHeaderParams{Protocol: fakeTransNumber, TTL: 123, TOS: stack.DefaultTOS}, tcpip.PacketBuffer{ + if err := f.route.WritePacket(nil /* gso */, stack.NetworkHeaderParams{Protocol: fakeTransNumber, TTL: 123, TOS: stack.DefaultTOS}, stack.PacketBuffer{ Header: hdr, Data: buffer.View(v).ToVectorisedView(), }); err != nil { @@ -210,7 +215,7 @@ func (*fakeTransportEndpoint) GetRemoteAddress() (tcpip.FullAddress, *tcpip.Erro return tcpip.FullAddress{}, nil } -func (f *fakeTransportEndpoint) HandlePacket(r *stack.Route, id stack.TransportEndpointID, _ tcpip.PacketBuffer) { +func (f *fakeTransportEndpoint) HandlePacket(r *stack.Route, id stack.TransportEndpointID, _ stack.PacketBuffer) { // Increment the number of received packets. f.proto.packetCount++ if f.acceptQueue != nil { @@ -227,7 +232,7 @@ func (f *fakeTransportEndpoint) HandlePacket(r *stack.Route, id stack.TransportE } } -func (f *fakeTransportEndpoint) HandleControlPacket(stack.TransportEndpointID, stack.ControlType, uint32, tcpip.PacketBuffer) { +func (f *fakeTransportEndpoint) HandleControlPacket(stack.TransportEndpointID, stack.ControlType, uint32, stack.PacketBuffer) { // Increment the number of received control packets. f.proto.controlCount++ } @@ -238,8 +243,8 @@ func (f *fakeTransportEndpoint) State() uint32 { func (f *fakeTransportEndpoint) ModerateRecvBuf(copied int) {} -func (f *fakeTransportEndpoint) IPTables() (iptables.IPTables, error) { - return iptables.IPTables{}, nil +func (f *fakeTransportEndpoint) IPTables() (stack.IPTables, error) { + return stack.IPTables{}, nil } func (f *fakeTransportEndpoint) Resume(*stack.Stack) {} @@ -272,7 +277,7 @@ func (f *fakeTransportProtocol) NewEndpoint(stack *stack.Stack, netProto tcpip.N return newFakeTransportEndpoint(stack, f, netProto, stack.UniqueID()), nil } -func (f *fakeTransportProtocol) NewRawEndpoint(stack *stack.Stack, netProto tcpip.NetworkProtocolNumber, _ *waiter.Queue) (tcpip.Endpoint, *tcpip.Error) { +func (*fakeTransportProtocol) NewRawEndpoint(stack *stack.Stack, netProto tcpip.NetworkProtocolNumber, _ *waiter.Queue) (tcpip.Endpoint, *tcpip.Error) { return nil, tcpip.ErrUnknownProtocol } @@ -284,7 +289,7 @@ func (*fakeTransportProtocol) ParsePorts(buffer.View) (src, dst uint16, err *tcp return 0, 0, nil } -func (*fakeTransportProtocol) HandleUnknownDestinationPacket(*stack.Route, stack.TransportEndpointID, tcpip.PacketBuffer) bool { +func (*fakeTransportProtocol) HandleUnknownDestinationPacket(*stack.Route, stack.TransportEndpointID, stack.PacketBuffer) bool { return true } @@ -310,6 +315,15 @@ func (f *fakeTransportProtocol) Option(option interface{}) *tcpip.Error { } } +// Abort implements TransportProtocol.Abort. +func (*fakeTransportProtocol) Abort() {} + +// Close implements tcpip.Endpoint.Close. +func (*fakeTransportProtocol) Close() {} + +// Wait implements TransportProtocol.Wait. +func (*fakeTransportProtocol) Wait() {} + func fakeTransFactory() stack.TransportProtocol { return &fakeTransportProtocol{} } @@ -355,7 +369,7 @@ func TestTransportReceive(t *testing.T) { // Make sure packet with wrong protocol is not delivered. buf[0] = 1 buf[2] = 0 - linkEP.InjectInbound(fakeNetNumber, tcpip.PacketBuffer{ + linkEP.InjectInbound(fakeNetNumber, stack.PacketBuffer{ Data: buf.ToVectorisedView(), }) if fakeTrans.packetCount != 0 { @@ -366,7 +380,7 @@ func TestTransportReceive(t *testing.T) { buf[0] = 1 buf[1] = 3 buf[2] = byte(fakeTransNumber) - linkEP.InjectInbound(fakeNetNumber, tcpip.PacketBuffer{ + linkEP.InjectInbound(fakeNetNumber, stack.PacketBuffer{ Data: buf.ToVectorisedView(), }) if fakeTrans.packetCount != 0 { @@ -377,7 +391,7 @@ func TestTransportReceive(t *testing.T) { buf[0] = 1 buf[1] = 2 buf[2] = byte(fakeTransNumber) - linkEP.InjectInbound(fakeNetNumber, tcpip.PacketBuffer{ + linkEP.InjectInbound(fakeNetNumber, stack.PacketBuffer{ Data: buf.ToVectorisedView(), }) if fakeTrans.packetCount != 1 { @@ -432,7 +446,7 @@ func TestTransportControlReceive(t *testing.T) { buf[fakeNetHeaderLen+0] = 0 buf[fakeNetHeaderLen+1] = 1 buf[fakeNetHeaderLen+2] = 0 - linkEP.InjectInbound(fakeNetNumber, tcpip.PacketBuffer{ + linkEP.InjectInbound(fakeNetNumber, stack.PacketBuffer{ Data: buf.ToVectorisedView(), }) if fakeTrans.controlCount != 0 { @@ -443,7 +457,7 @@ func TestTransportControlReceive(t *testing.T) { buf[fakeNetHeaderLen+0] = 3 buf[fakeNetHeaderLen+1] = 1 buf[fakeNetHeaderLen+2] = byte(fakeTransNumber) - linkEP.InjectInbound(fakeNetNumber, tcpip.PacketBuffer{ + linkEP.InjectInbound(fakeNetNumber, stack.PacketBuffer{ Data: buf.ToVectorisedView(), }) if fakeTrans.controlCount != 0 { @@ -454,7 +468,7 @@ func TestTransportControlReceive(t *testing.T) { buf[fakeNetHeaderLen+0] = 2 buf[fakeNetHeaderLen+1] = 1 buf[fakeNetHeaderLen+2] = byte(fakeTransNumber) - linkEP.InjectInbound(fakeNetNumber, tcpip.PacketBuffer{ + linkEP.InjectInbound(fakeNetNumber, stack.PacketBuffer{ Data: buf.ToVectorisedView(), }) if fakeTrans.controlCount != 1 { @@ -609,7 +623,7 @@ func TestTransportForwarding(t *testing.T) { req[0] = 1 req[1] = 3 req[2] = byte(fakeTransNumber) - ep2.InjectInbound(fakeNetNumber, tcpip.PacketBuffer{ + ep2.InjectInbound(fakeNetNumber, stack.PacketBuffer{ Data: req.ToVectorisedView(), }) diff --git a/pkg/tcpip/tcpip.go b/pkg/tcpip/tcpip.go index 0e944712f..1ca4088c9 100644 --- a/pkg/tcpip/tcpip.go +++ b/pkg/tcpip/tcpip.go @@ -323,11 +323,26 @@ type ControlMessages struct { // TOS is the IPv4 type of service of the associated packet. TOS uint8 - // HasTClass indicates whether Tclass is valid/set. + // HasTClass indicates whether TClass is valid/set. HasTClass bool - // Tclass is the IPv6 traffic class of the associated packet. - TClass int32 + // TClass is the IPv6 traffic class of the associated packet. + TClass uint32 + + // HasIPPacketInfo indicates whether PacketInfo is set. + HasIPPacketInfo bool + + // PacketInfo holds interface and address data on an incoming packet. + PacketInfo IPPacketInfo +} + +// PacketOwner is used to get UID and GID of the packet. +type PacketOwner interface { + // UID returns UID of the packet. + UID() uint32 + + // GID returns GID of the packet. + GID() uint32 } // Endpoint is the interface implemented by transport protocols (e.g., tcp, udp) @@ -335,9 +350,15 @@ type ControlMessages struct { // networking stack. type Endpoint interface { // Close puts the endpoint in a closed state and frees all resources - // associated with it. + // associated with it. Close initiates the teardown process, the + // Endpoint may not be fully closed when Close returns. Close() + // Abort initiates an expedited endpoint teardown. As compared to + // Close, Abort prioritizes closing the Endpoint quickly over cleanly. + // Abort is best effort; implementing Abort with Close is acceptable. + Abort() + // Read reads data from the endpoint and optionally returns the sender. // // This method does not block if there is no data pending. It will also @@ -458,6 +479,9 @@ type Endpoint interface { // Stats returns a reference to the endpoint stats. Stats() EndpointStats + + // SetOwner sets the task owner to the endpoint owner. + SetOwner(owner PacketOwner) } // EndpointInfo is the interface implemented by each endpoint info struct. @@ -496,9 +520,56 @@ type WriteOptions struct { type SockOptBool int const ( + // BroadcastOption is used by SetSockOpt/GetSockOpt to specify whether + // datagram sockets are allowed to send packets to a broadcast address. + BroadcastOption SockOptBool = iota + + // CorkOption is used by SetSockOpt/GetSockOpt to specify if data should be + // held until segments are full by the TCP transport protocol. + CorkOption + + // DelayOption is used by SetSockOpt/GetSockOpt to specify if data + // should be sent out immediately by the transport protocol. For TCP, + // it determines if the Nagle algorithm is on or off. + DelayOption + + // KeepaliveEnabledOption is used by SetSockOpt/GetSockOpt to specify whether + // TCP keepalive is enabled for this socket. + KeepaliveEnabledOption + + // MulticastLoopOption is used by SetSockOpt/GetSockOpt to specify whether + // multicast packets sent over a non-loopback interface will be looped back. + MulticastLoopOption + + // PasscredOption is used by SetSockOpt/GetSockOpt to specify whether + // SCM_CREDENTIALS socket control messages are enabled. + // + // Only supported on Unix sockets. + PasscredOption + + // QuickAckOption is stubbed out in SetSockOpt/GetSockOpt. + QuickAckOption + + // ReceiveTClassOption is used by SetSockOpt/GetSockOpt to specify if the + // IPV6_TCLASS ancillary message is passed with incoming packets. + ReceiveTClassOption + // ReceiveTOSOption is used by SetSockOpt/GetSockOpt to specify if the TOS // ancillary message is passed with incoming packets. - ReceiveTOSOption SockOptBool = iota + ReceiveTOSOption + + // ReceiveIPPacketInfoOption is used by {G,S}etSockOptBool to specify + // if more inforamtion is provided with incoming packets such + // as interface index and address. + ReceiveIPPacketInfoOption + + // ReuseAddressOption is used by SetSockOpt/GetSockOpt to specify whether Bind() + // should allow reuse of local address. + ReuseAddressOption + + // ReusePortOption is used by SetSockOpt/GetSockOpt to permit multiple sockets + // to be bound to an identical socket address. + ReusePortOption // V6OnlyOption is used by {G,S}etSockOptBool to specify whether an IPv6 // socket is to be restricted to sending and receiving IPv6 packets only. @@ -509,9 +580,30 @@ const ( type SockOptInt int const ( + // KeepaliveCountOption is used by SetSockOpt/GetSockOpt to specify the number + // of un-ACKed TCP keepalives that will be sent before the connection is + // closed. + KeepaliveCountOption SockOptInt = iota + + // IPv4TOSOption is used by SetSockOpt/GetSockOpt to specify TOS + // for all subsequent outgoing IPv4 packets from the endpoint. + IPv4TOSOption + + // IPv6TrafficClassOption is used by SetSockOpt/GetSockOpt to specify TOS + // for all subsequent outgoing IPv6 packets from the endpoint. + IPv6TrafficClassOption + + // MaxSegOption is used by SetSockOpt/GetSockOpt to set/get the current + // Maximum Segment Size(MSS) value as specified using the TCP_MAXSEG option. + MaxSegOption + + // MulticastTTLOption is used by SetSockOpt/GetSockOpt to control the default + // TTL value for multicast messages. The default is 1. + MulticastTTLOption + // ReceiveQueueSizeOption is used in GetSockOptInt to specify that the // number of unread bytes in the input buffer should be returned. - ReceiveQueueSizeOption SockOptInt = iota + ReceiveQueueSizeOption // SendBufferSizeOption is used by SetSockOptInt/GetSockOptInt to // specify the send buffer size option. @@ -525,44 +617,21 @@ const ( // number of unread bytes in the output buffer should be returned. SendQueueSizeOption - // DelayOption is used by SetSockOpt/GetSockOpt to specify if data - // should be sent out immediately by the transport protocol. For TCP, - // it determines if the Nagle algorithm is on or off. - DelayOption - - // TODO(b/137664753): convert all int socket options to be handled via - // GetSockOptInt. + // TTLOption is used by SetSockOpt/GetSockOpt to control the default TTL/hop + // limit value for unicast messages. The default is protocol specific. + // + // A zero value indicates the default. + TTLOption ) // ErrorOption is used in GetSockOpt to specify that the last error reported by // the endpoint should be cleared and returned. type ErrorOption struct{} -// CorkOption is used by SetSockOpt/GetSockOpt to specify if data should be -// held until segments are full by the TCP transport protocol. -type CorkOption int - -// ReuseAddressOption is used by SetSockOpt/GetSockOpt to specify whether Bind() -// should allow reuse of local address. -type ReuseAddressOption int - -// ReusePortOption is used by SetSockOpt/GetSockOpt to permit multiple sockets -// to be bound to an identical socket address. -type ReusePortOption int - // BindToDeviceOption is used by SetSockOpt/GetSockOpt to specify that sockets // should bind only on a specific NIC. type BindToDeviceOption NICID -// QuickAckOption is stubbed out in SetSockOpt/GetSockOpt. -type QuickAckOption int - -// PasscredOption is used by SetSockOpt/GetSockOpt to specify whether -// SCM_CREDENTIALS socket control messages are enabled. -// -// Only supported on Unix sockets. -type PasscredOption int - // TCPInfoOption is used by GetSockOpt to expose TCP statistics. // // TODO(b/64800844): Add and populate stat fields. @@ -571,10 +640,6 @@ type TCPInfoOption struct { RTTVar time.Duration } -// KeepaliveEnabledOption is used by SetSockOpt/GetSockOpt to specify whether -// TCP keepalive is enabled for this socket. -type KeepaliveEnabledOption int - // KeepaliveIdleOption is used by SetSockOpt/GetSockOpt to specify the time a // connection must remain idle before the first TCP keepalive packet is sent. // Once this time is reached, KeepaliveIntervalOption is used instead. @@ -584,11 +649,6 @@ type KeepaliveIdleOption time.Duration // interval between sending TCP keepalive packets. type KeepaliveIntervalOption time.Duration -// KeepaliveCountOption is used by SetSockOpt/GetSockOpt to specify the number -// of un-ACKed TCP keepalives that will be sent before the connection is -// closed. -type KeepaliveCountOption int - // TCPUserTimeoutOption is used by SetSockOpt/GetSockOpt to specify a user // specified timeout for a given TCP connection. // See: RFC5482 for details. @@ -602,20 +662,9 @@ type CongestionControlOption string // control algorithms. type AvailableCongestionControlOption string -// ModerateReceiveBufferOption allows the caller to enable/disable TCP receive // buffer moderation. type ModerateReceiveBufferOption bool -// MaxSegOption is used by SetSockOpt/GetSockOpt to set/get the current -// Maximum Segment Size(MSS) value as specified using the TCP_MAXSEG option. -type MaxSegOption int - -// TTLOption is used by SetSockOpt/GetSockOpt to control the default TTL/hop -// limit value for unicast messages. The default is protocol specific. -// -// A zero value indicates the default. -type TTLOption uint8 - // TCPLingerTimeoutOption is used by SetSockOpt/GetSockOpt to set/get the // maximum duration for which a socket lingers in the TCP_FIN_WAIT_2 state // before being marked closed. @@ -632,9 +681,14 @@ type TCPTimeWaitTimeoutOption time.Duration // for a handshake till the specified timeout until a segment with data arrives. type TCPDeferAcceptOption time.Duration -// MulticastTTLOption is used by SetSockOpt/GetSockOpt to control the default -// TTL value for multicast messages. The default is 1. -type MulticastTTLOption uint8 +// TCPMinRTOOption is use by SetSockOpt/GetSockOpt to allow overriding +// default MinRTO used by the Stack. +type TCPMinRTOOption time.Duration + +// TCPSynRcvdCountThresholdOption is used by SetSockOpt/GetSockOpt to specify +// the number of endpoints that can be in SYN-RCVD state before the stack +// switches to using SYN cookies. +type TCPSynRcvdCountThresholdOption uint64 // MulticastInterfaceOption is used by SetSockOpt/GetSockOpt to specify a // default interface for multicast. @@ -643,10 +697,6 @@ type MulticastInterfaceOption struct { InterfaceAddr Address } -// MulticastLoopOption is used by SetSockOpt/GetSockOpt to specify whether -// multicast packets sent over a non-loopback interface will be looped back. -type MulticastLoopOption bool - // MembershipOption is used by SetSockOpt/GetSockOpt as an argument to // AddMembershipOption and RemoveMembershipOption. type MembershipOption struct { @@ -669,21 +719,23 @@ type RemoveMembershipOption MembershipOption // TCP out-of-band data is delivered along with the normal in-band data. type OutOfBandInlineOption int -// BroadcastOption is used by SetSockOpt/GetSockOpt to specify whether -// datagram sockets are allowed to send packets to a broadcast address. -type BroadcastOption int - // DefaultTTLOption is used by stack.(*Stack).NetworkProtocolOption to specify // a default TTL. type DefaultTTLOption uint8 -// IPv4TOSOption is used by SetSockOpt/GetSockOpt to specify TOS -// for all subsequent outgoing IPv4 packets from the endpoint. -type IPv4TOSOption uint8 +// IPPacketInfo is the message struture for IP_PKTINFO. +// +// +stateify savable +type IPPacketInfo struct { + // NIC is the ID of the NIC to be used. + NIC NICID + + // LocalAddr is the local address. + LocalAddr Address -// IPv6TrafficClassOption is used by SetSockOpt/GetSockOpt to specify TOS -// for all subsequent outgoing IPv6 packets from the endpoint. -type IPv6TrafficClassOption uint8 + // DestinationAddr is the destination address. + DestinationAddr Address +} // Route is a row in the routing table. It specifies through which NIC (and // gateway) sets of packets should be routed. A row is considered viable if the diff --git a/pkg/tcpip/tcpip_test.go b/pkg/tcpip/tcpip_test.go index 8c0aacffa..1c8e2bc34 100644 --- a/pkg/tcpip/tcpip_test.go +++ b/pkg/tcpip/tcpip_test.go @@ -218,7 +218,7 @@ func TestAddressWithPrefixSubnet(t *testing.T) { gotSubnet := ap.Subnet() wantSubnet, err := NewSubnet(tt.subnetAddr, tt.subnetMask) if err != nil { - t.Error("NewSubnet(%q, %q) failed: %s", tt.subnetAddr, tt.subnetMask, err) + t.Errorf("NewSubnet(%q, %q) failed: %s", tt.subnetAddr, tt.subnetMask, err) continue } if gotSubnet != wantSubnet { diff --git a/pkg/tcpip/time_unsafe.go b/pkg/tcpip/time_unsafe.go index 48764b978..2f98a996f 100644 --- a/pkg/tcpip/time_unsafe.go +++ b/pkg/tcpip/time_unsafe.go @@ -25,6 +25,8 @@ import ( ) // StdClock implements Clock with the time package. +// +// +stateify savable type StdClock struct{} var _ Clock = (*StdClock)(nil) diff --git a/pkg/tcpip/timer.go b/pkg/tcpip/timer.go index f5f01f32f..67f66fc72 100644 --- a/pkg/tcpip/timer.go +++ b/pkg/tcpip/timer.go @@ -131,10 +131,14 @@ func (t *CancellableTimer) StopLocked() { func (t *CancellableTimer) Reset(d time.Duration) { // Create a new instance. earlyReturn := false + + // Capture the locker so that updating the timer does not cause a data race + // when a timer fires and tries to obtain the lock (read the timer's locker). + locker := t.locker t.instance = cancellableTimerInstance{ timer: time.AfterFunc(d, func() { - t.locker.Lock() - defer t.locker.Unlock() + locker.Lock() + defer locker.Unlock() if earlyReturn { // If we reach this point, it means that the timer fired while another diff --git a/pkg/tcpip/timer_test.go b/pkg/tcpip/timer_test.go index 2d20f7ef3..730134906 100644 --- a/pkg/tcpip/timer_test.go +++ b/pkg/tcpip/timer_test.go @@ -28,6 +28,31 @@ const ( longDuration = 1 * time.Second ) +func TestCancellableTimerReassignment(t *testing.T) { + var timer tcpip.CancellableTimer + var wg sync.WaitGroup + var lock sync.Mutex + + for i := 0; i < 2; i++ { + wg.Add(1) + + go func() { + lock.Lock() + // Assigning a new timer value updates the timer's locker and function. + // This test makes sure there is no data race when reassigning a timer + // that has an active timer (even if it has been stopped as a stopped + // timer may be blocked on a lock before it can check if it has been + // stopped while another goroutine holds the same lock). + timer = tcpip.MakeCancellableTimer(&lock, func() { + wg.Done() + }) + timer.Reset(shortDuration) + lock.Unlock() + }() + } + wg.Wait() +} + func TestCancellableTimerFire(t *testing.T) { t.Parallel() diff --git a/pkg/tcpip/transport/icmp/BUILD b/pkg/tcpip/transport/icmp/BUILD index ac18ec5b1..9ce625c17 100644 --- a/pkg/tcpip/transport/icmp/BUILD +++ b/pkg/tcpip/transport/icmp/BUILD @@ -31,7 +31,6 @@ go_library( "//pkg/tcpip", "//pkg/tcpip/buffer", "//pkg/tcpip/header", - "//pkg/tcpip/iptables", "//pkg/tcpip/stack", "//pkg/tcpip/transport/raw", "//pkg/tcpip/transport/tcp", diff --git a/pkg/tcpip/transport/icmp/endpoint.go b/pkg/tcpip/transport/icmp/endpoint.go index 42afb3f5b..feef8dca0 100644 --- a/pkg/tcpip/transport/icmp/endpoint.go +++ b/pkg/tcpip/transport/icmp/endpoint.go @@ -19,7 +19,6 @@ import ( "gvisor.dev/gvisor/pkg/tcpip" "gvisor.dev/gvisor/pkg/tcpip/buffer" "gvisor.dev/gvisor/pkg/tcpip/header" - "gvisor.dev/gvisor/pkg/tcpip/iptables" "gvisor.dev/gvisor/pkg/tcpip/stack" "gvisor.dev/gvisor/pkg/waiter" ) @@ -74,6 +73,9 @@ type endpoint struct { route stack.Route `state:"manual"` ttl uint8 stats tcpip.TransportEndpointStats `state:"nosave"` + + // owner is used to get uid and gid of the packet. + owner tcpip.PacketOwner } func newEndpoint(s *stack.Stack, netProto tcpip.NetworkProtocolNumber, transProto tcpip.TransportProtocolNumber, waiterQueue *waiter.Queue) (tcpip.Endpoint, *tcpip.Error) { @@ -96,6 +98,11 @@ func (e *endpoint) UniqueID() uint64 { return e.uniqueID } +// Abort implements stack.TransportEndpoint.Abort. +func (e *endpoint) Abort() { + e.Close() +} + // Close puts the endpoint in a closed state and frees all resources // associated with it. func (e *endpoint) Close() { @@ -129,8 +136,12 @@ func (e *endpoint) Close() { // ModerateRecvBuf implements tcpip.Endpoint.ModerateRecvBuf. func (e *endpoint) ModerateRecvBuf(copied int) {} +func (e *endpoint) SetOwner(owner tcpip.PacketOwner) { + e.owner = owner +} + // IPTables implements tcpip.Endpoint.IPTables. -func (e *endpoint) IPTables() (iptables.IPTables, error) { +func (e *endpoint) IPTables() (stack.IPTables, error) { return e.stack.IPTables(), nil } @@ -286,15 +297,13 @@ func (e *endpoint) write(p tcpip.Payloader, opts tcpip.WriteOptions) (int64, <-c nicID = e.BindNICID } - toCopy := *to - to = &toCopy - netProto, err := e.checkV4Mapped(to) + dst, netProto, err := e.checkV4MappedLocked(*to) if err != nil { return 0, nil, err } - // Find the enpoint. - r, err := e.stack.FindRoute(nicID, e.BindAddr, to.Addr, netProto, false /* multicastLoop */) + // Find the endpoint. + r, err := e.stack.FindRoute(nicID, e.BindAddr, dst.Addr, netProto, false /* multicastLoop */) if err != nil { return 0, nil, err } @@ -319,7 +328,7 @@ func (e *endpoint) write(p tcpip.Payloader, opts tcpip.WriteOptions) (int64, <-c switch e.NetProto { case header.IPv4ProtocolNumber: - err = send4(route, e.ID.LocalPort, v, e.ttl) + err = send4(route, e.ID.LocalPort, v, e.ttl, e.owner) case header.IPv6ProtocolNumber: err = send6(route, e.ID.LocalPort, v, e.ttl) @@ -339,13 +348,6 @@ func (e *endpoint) Peek([][]byte) (int64, tcpip.ControlMessages, *tcpip.Error) { // SetSockOpt sets a socket option. func (e *endpoint) SetSockOpt(opt interface{}) *tcpip.Error { - switch o := opt.(type) { - case tcpip.TTLOption: - e.mu.Lock() - e.ttl = uint8(o) - e.mu.Unlock() - } - return nil } @@ -356,12 +358,25 @@ func (e *endpoint) SetSockOptBool(opt tcpip.SockOptBool, v bool) *tcpip.Error { // SetSockOptInt sets a socket option. Currently not supported. func (e *endpoint) SetSockOptInt(opt tcpip.SockOptInt, v int) *tcpip.Error { + switch opt { + case tcpip.TTLOption: + e.mu.Lock() + e.ttl = uint8(v) + e.mu.Unlock() + + } return nil } // GetSockOptBool implements tcpip.Endpoint.GetSockOptBool. func (e *endpoint) GetSockOptBool(opt tcpip.SockOptBool) (bool, *tcpip.Error) { - return false, tcpip.ErrUnknownProtocolOption + switch opt { + case tcpip.KeepaliveEnabledOption: + return false, nil + + default: + return false, tcpip.ErrUnknownProtocolOption + } } // GetSockOptInt implements tcpip.Endpoint.GetSockOptInt. @@ -388,32 +403,29 @@ func (e *endpoint) GetSockOptInt(opt tcpip.SockOptInt) (int, *tcpip.Error) { e.rcvMu.Unlock() return v, nil + case tcpip.TTLOption: + e.rcvMu.Lock() + v := int(e.ttl) + e.rcvMu.Unlock() + return v, nil + + default: + return -1, tcpip.ErrUnknownProtocolOption } - return -1, tcpip.ErrUnknownProtocolOption } // GetSockOpt implements tcpip.Endpoint.GetSockOpt. func (e *endpoint) GetSockOpt(opt interface{}) *tcpip.Error { - switch o := opt.(type) { + switch opt.(type) { case tcpip.ErrorOption: return nil - case *tcpip.KeepaliveEnabledOption: - *o = 0 - return nil - - case *tcpip.TTLOption: - e.rcvMu.Lock() - *o = tcpip.TTLOption(e.ttl) - e.rcvMu.Unlock() - return nil - default: return tcpip.ErrUnknownProtocolOption } } -func send4(r *stack.Route, ident uint16, data buffer.View, ttl uint8) *tcpip.Error { +func send4(r *stack.Route, ident uint16, data buffer.View, ttl uint8, owner tcpip.PacketOwner) *tcpip.Error { if len(data) < header.ICMPv4MinimumSize { return tcpip.ErrInvalidEndpointState } @@ -438,10 +450,11 @@ func send4(r *stack.Route, ident uint16, data buffer.View, ttl uint8) *tcpip.Err if ttl == 0 { ttl = r.DefaultTTL() } - return r.WritePacket(nil /* gso */, stack.NetworkHeaderParams{Protocol: header.ICMPv4ProtocolNumber, TTL: ttl, TOS: stack.DefaultTOS}, tcpip.PacketBuffer{ + return r.WritePacket(nil /* gso */, stack.NetworkHeaderParams{Protocol: header.ICMPv4ProtocolNumber, TTL: ttl, TOS: stack.DefaultTOS}, stack.PacketBuffer{ Header: hdr, Data: data.ToVectorisedView(), TransportHeader: buffer.View(icmpv4), + Owner: owner, }) } @@ -468,20 +481,21 @@ func send6(r *stack.Route, ident uint16, data buffer.View, ttl uint8) *tcpip.Err if ttl == 0 { ttl = r.DefaultTTL() } - return r.WritePacket(nil /* gso */, stack.NetworkHeaderParams{Protocol: header.ICMPv6ProtocolNumber, TTL: ttl, TOS: stack.DefaultTOS}, tcpip.PacketBuffer{ + return r.WritePacket(nil /* gso */, stack.NetworkHeaderParams{Protocol: header.ICMPv6ProtocolNumber, TTL: ttl, TOS: stack.DefaultTOS}, stack.PacketBuffer{ Header: hdr, Data: dataVV, TransportHeader: buffer.View(icmpv6), }) } -func (e *endpoint) checkV4Mapped(addr *tcpip.FullAddress) (tcpip.NetworkProtocolNumber, *tcpip.Error) { - unwrapped, netProto, err := e.TransportEndpointInfo.AddrNetProto(*addr, false /* v6only */) +// checkV4MappedLocked determines the effective network protocol and converts +// addr to its canonical form. +func (e *endpoint) checkV4MappedLocked(addr tcpip.FullAddress) (tcpip.FullAddress, tcpip.NetworkProtocolNumber, *tcpip.Error) { + unwrapped, netProto, err := e.TransportEndpointInfo.AddrNetProtoLocked(addr, false /* v6only */) if err != nil { - return 0, err + return tcpip.FullAddress{}, 0, err } - *addr = unwrapped - return netProto, nil + return unwrapped, netProto, nil } // Disconnect implements tcpip.Endpoint.Disconnect. @@ -512,7 +526,7 @@ func (e *endpoint) Connect(addr tcpip.FullAddress) *tcpip.Error { return tcpip.ErrInvalidEndpointState } - netProto, err := e.checkV4Mapped(&addr) + addr, netProto, err := e.checkV4MappedLocked(addr) if err != nil { return err } @@ -625,7 +639,7 @@ func (e *endpoint) bindLocked(addr tcpip.FullAddress) *tcpip.Error { return tcpip.ErrInvalidEndpointState } - netProto, err := e.checkV4Mapped(&addr) + addr, netProto, err := e.checkV4MappedLocked(addr) if err != nil { return err } @@ -729,7 +743,7 @@ func (e *endpoint) Readiness(mask waiter.EventMask) waiter.EventMask { // HandlePacket is called by the stack when new packets arrive to this transport // endpoint. -func (e *endpoint) HandlePacket(r *stack.Route, id stack.TransportEndpointID, pkt tcpip.PacketBuffer) { +func (e *endpoint) HandlePacket(r *stack.Route, id stack.TransportEndpointID, pkt stack.PacketBuffer) { // Only accept echo replies. switch e.NetProto { case header.IPv4ProtocolNumber: @@ -791,7 +805,7 @@ func (e *endpoint) HandlePacket(r *stack.Route, id stack.TransportEndpointID, pk } // HandleControlPacket implements stack.TransportEndpoint.HandleControlPacket. -func (e *endpoint) HandleControlPacket(id stack.TransportEndpointID, typ stack.ControlType, extra uint32, pkt tcpip.PacketBuffer) { +func (e *endpoint) HandleControlPacket(id stack.TransportEndpointID, typ stack.ControlType, extra uint32, pkt stack.PacketBuffer) { } // State implements tcpip.Endpoint.State. The ICMP endpoint currently doesn't diff --git a/pkg/tcpip/transport/icmp/protocol.go b/pkg/tcpip/transport/icmp/protocol.go index 9ce500e80..3c47692b2 100644 --- a/pkg/tcpip/transport/icmp/protocol.go +++ b/pkg/tcpip/transport/icmp/protocol.go @@ -104,20 +104,26 @@ func (p *protocol) ParsePorts(v buffer.View) (src, dst uint16, err *tcpip.Error) // HandleUnknownDestinationPacket handles packets targeted at this protocol but // that don't match any existing endpoint. -func (p *protocol) HandleUnknownDestinationPacket(*stack.Route, stack.TransportEndpointID, tcpip.PacketBuffer) bool { +func (*protocol) HandleUnknownDestinationPacket(*stack.Route, stack.TransportEndpointID, stack.PacketBuffer) bool { return true } -// SetOption implements TransportProtocol.SetOption. -func (p *protocol) SetOption(option interface{}) *tcpip.Error { +// SetOption implements stack.TransportProtocol.SetOption. +func (*protocol) SetOption(option interface{}) *tcpip.Error { return tcpip.ErrUnknownProtocolOption } -// Option implements TransportProtocol.Option. -func (p *protocol) Option(option interface{}) *tcpip.Error { +// Option implements stack.TransportProtocol.Option. +func (*protocol) Option(option interface{}) *tcpip.Error { return tcpip.ErrUnknownProtocolOption } +// Close implements stack.TransportProtocol.Close. +func (*protocol) Close() {} + +// Wait implements stack.TransportProtocol.Wait. +func (*protocol) Wait() {} + // NewProtocol4 returns an ICMPv4 transport protocol. func NewProtocol4() stack.TransportProtocol { return &protocol{ProtocolNumber4} diff --git a/pkg/tcpip/transport/packet/BUILD b/pkg/tcpip/transport/packet/BUILD index d22de6b26..b989b1209 100644 --- a/pkg/tcpip/transport/packet/BUILD +++ b/pkg/tcpip/transport/packet/BUILD @@ -31,7 +31,6 @@ go_library( "//pkg/tcpip", "//pkg/tcpip/buffer", "//pkg/tcpip/header", - "//pkg/tcpip/iptables", "//pkg/tcpip/stack", "//pkg/waiter", ], diff --git a/pkg/tcpip/transport/packet/endpoint.go b/pkg/tcpip/transport/packet/endpoint.go index fc5bc69fa..23158173d 100644 --- a/pkg/tcpip/transport/packet/endpoint.go +++ b/pkg/tcpip/transport/packet/endpoint.go @@ -29,7 +29,6 @@ import ( "gvisor.dev/gvisor/pkg/tcpip" "gvisor.dev/gvisor/pkg/tcpip/buffer" "gvisor.dev/gvisor/pkg/tcpip/header" - "gvisor.dev/gvisor/pkg/tcpip/iptables" "gvisor.dev/gvisor/pkg/tcpip/stack" "gvisor.dev/gvisor/pkg/waiter" ) @@ -76,6 +75,7 @@ type endpoint struct { sndBufSize int closed bool stats tcpip.TransportEndpointStats `state:"nosave"` + bound bool } // NewEndpoint returns a new packet endpoint. @@ -98,6 +98,11 @@ func NewEndpoint(s *stack.Stack, cooked bool, netProto tcpip.NetworkProtocolNumb return ep, nil } +// Abort implements stack.TransportEndpoint.Abort. +func (ep *endpoint) Abort() { + ep.Close() +} + // Close implements tcpip.Endpoint.Close. func (ep *endpoint) Close() { ep.mu.Lock() @@ -120,6 +125,7 @@ func (ep *endpoint) Close() { } ep.closed = true + ep.bound = false ep.waiterQueue.Notify(waiter.EventHUp | waiter.EventErr | waiter.EventIn | waiter.EventOut) } @@ -127,7 +133,7 @@ func (ep *endpoint) Close() { func (ep *endpoint) ModerateRecvBuf(copied int) {} // IPTables implements tcpip.Endpoint.IPTables. -func (ep *endpoint) IPTables() (iptables.IPTables, error) { +func (ep *endpoint) IPTables() (stack.IPTables, error) { return ep.stack.IPTables(), nil } @@ -211,7 +217,24 @@ func (ep *endpoint) Bind(addr tcpip.FullAddress) *tcpip.Error { // sll_family (should be AF_PACKET), sll_protocol, and sll_ifindex." // - packet(7). - return tcpip.ErrNotSupported + ep.mu.Lock() + defer ep.mu.Unlock() + + if ep.bound { + return tcpip.ErrAlreadyBound + } + + // Unregister endpoint with all the nics. + ep.stack.UnregisterPacketEndpoint(0, ep.netProto, ep) + + // Bind endpoint to receive packets from specific interface. + if err := ep.stack.RegisterPacketEndpoint(addr.NIC, ep.netProto, ep); err != nil { + return err + } + + ep.bound = true + + return nil } // GetLocalAddress implements tcpip.Endpoint.GetLocalAddress. @@ -275,7 +298,7 @@ func (ep *endpoint) GetSockOptInt(opt tcpip.SockOptInt) (int, *tcpip.Error) { } // HandlePacket implements stack.PacketEndpoint.HandlePacket. -func (ep *endpoint) HandlePacket(nicID tcpip.NICID, localAddr tcpip.LinkAddress, netProto tcpip.NetworkProtocolNumber, pkt tcpip.PacketBuffer) { +func (ep *endpoint) HandlePacket(nicID tcpip.NICID, localAddr tcpip.LinkAddress, netProto tcpip.NetworkProtocolNumber, pkt stack.PacketBuffer) { ep.rcvMu.Lock() // Drop the packet if our buffer is currently full. @@ -369,3 +392,5 @@ func (ep *endpoint) Info() tcpip.EndpointInfo { func (ep *endpoint) Stats() tcpip.EndpointStats { return &ep.stats } + +func (ep *endpoint) SetOwner(owner tcpip.PacketOwner) {} diff --git a/pkg/tcpip/transport/raw/BUILD b/pkg/tcpip/transport/raw/BUILD index c9baf4600..2eab09088 100644 --- a/pkg/tcpip/transport/raw/BUILD +++ b/pkg/tcpip/transport/raw/BUILD @@ -32,7 +32,6 @@ go_library( "//pkg/tcpip", "//pkg/tcpip/buffer", "//pkg/tcpip/header", - "//pkg/tcpip/iptables", "//pkg/tcpip/stack", "//pkg/tcpip/transport/packet", "//pkg/waiter", diff --git a/pkg/tcpip/transport/raw/endpoint.go b/pkg/tcpip/transport/raw/endpoint.go index ee9c4c58b..eee754a5a 100644 --- a/pkg/tcpip/transport/raw/endpoint.go +++ b/pkg/tcpip/transport/raw/endpoint.go @@ -30,7 +30,6 @@ import ( "gvisor.dev/gvisor/pkg/tcpip" "gvisor.dev/gvisor/pkg/tcpip/buffer" "gvisor.dev/gvisor/pkg/tcpip/header" - "gvisor.dev/gvisor/pkg/tcpip/iptables" "gvisor.dev/gvisor/pkg/tcpip/stack" "gvisor.dev/gvisor/pkg/waiter" ) @@ -81,6 +80,9 @@ type endpoint struct { // Connect(), and is valid only when conneted is true. route stack.Route `state:"manual"` stats tcpip.TransportEndpointStats `state:"nosave"` + + // owner is used to get uid and gid of the packet. + owner tcpip.PacketOwner } // NewEndpoint returns a raw endpoint for the given protocols. @@ -121,6 +123,11 @@ func newEndpoint(s *stack.Stack, netProto tcpip.NetworkProtocolNumber, transProt return e, nil } +// Abort implements stack.TransportEndpoint.Abort. +func (e *endpoint) Abort() { + e.Close() +} + // Close implements tcpip.Endpoint.Close. func (e *endpoint) Close() { e.mu.Lock() @@ -155,8 +162,12 @@ func (e *endpoint) Close() { // ModerateRecvBuf implements tcpip.Endpoint.ModerateRecvBuf. func (e *endpoint) ModerateRecvBuf(copied int) {} +func (e *endpoint) SetOwner(owner tcpip.PacketOwner) { + e.owner = owner +} + // IPTables implements tcpip.Endpoint.IPTables. -func (e *endpoint) IPTables() (iptables.IPTables, error) { +func (e *endpoint) IPTables() (stack.IPTables, error) { return e.stack.IPTables(), nil } @@ -337,17 +348,19 @@ func (e *endpoint) finishWrite(payloadBytes []byte, route *stack.Route) (int64, switch e.NetProto { case header.IPv4ProtocolNumber: if !e.associated { - if err := route.WriteHeaderIncludedPacket(tcpip.PacketBuffer{ + if err := route.WriteHeaderIncludedPacket(stack.PacketBuffer{ Data: buffer.View(payloadBytes).ToVectorisedView(), }); err != nil { return 0, nil, err } break } + hdr := buffer.NewPrependable(len(payloadBytes) + int(route.MaxHeaderLength())) - if err := route.WritePacket(nil /* gso */, stack.NetworkHeaderParams{Protocol: e.TransProto, TTL: route.DefaultTTL(), TOS: stack.DefaultTOS}, tcpip.PacketBuffer{ + if err := route.WritePacket(nil /* gso */, stack.NetworkHeaderParams{Protocol: e.TransProto, TTL: route.DefaultTTL(), TOS: stack.DefaultTOS}, stack.PacketBuffer{ Header: hdr, Data: buffer.View(payloadBytes).ToVectorisedView(), + Owner: e.owner, }); err != nil { return 0, nil, err } @@ -520,14 +533,10 @@ func (e *endpoint) SetSockOptInt(opt tcpip.SockOptInt, v int) *tcpip.Error { // GetSockOpt implements tcpip.Endpoint.GetSockOpt. func (e *endpoint) GetSockOpt(opt interface{}) *tcpip.Error { - switch o := opt.(type) { + switch opt.(type) { case tcpip.ErrorOption: return nil - case *tcpip.KeepaliveEnabledOption: - *o = 0 - return nil - default: return tcpip.ErrUnknownProtocolOption } @@ -535,7 +544,13 @@ func (e *endpoint) GetSockOpt(opt interface{}) *tcpip.Error { // GetSockOptBool implements tcpip.Endpoint.GetSockOptBool. func (e *endpoint) GetSockOptBool(opt tcpip.SockOptBool) (bool, *tcpip.Error) { - return false, tcpip.ErrUnknownProtocolOption + switch opt { + case tcpip.KeepaliveEnabledOption: + return false, nil + + default: + return false, tcpip.ErrUnknownProtocolOption + } } // GetSockOptInt implements tcpip.Endpoint.GetSockOptInt. @@ -563,13 +578,13 @@ func (e *endpoint) GetSockOptInt(opt tcpip.SockOptInt) (int, *tcpip.Error) { e.rcvMu.Unlock() return v, nil + default: + return -1, tcpip.ErrUnknownProtocolOption } - - return -1, tcpip.ErrUnknownProtocolOption } // HandlePacket implements stack.RawTransportEndpoint.HandlePacket. -func (e *endpoint) HandlePacket(route *stack.Route, pkt tcpip.PacketBuffer) { +func (e *endpoint) HandlePacket(route *stack.Route, pkt stack.PacketBuffer) { e.rcvMu.Lock() // Drop the packet if our buffer is currently full. diff --git a/pkg/tcpip/transport/tcp/BUILD b/pkg/tcpip/transport/tcp/BUILD index 272e8f570..61426623c 100644 --- a/pkg/tcpip/transport/tcp/BUILD +++ b/pkg/tcpip/transport/tcp/BUILD @@ -32,6 +32,7 @@ go_library( srcs = [ "accept.go", "connect.go", + "connect_unsafe.go", "cubic.go", "cubic_state.go", "dispatcher.go", @@ -65,12 +66,10 @@ go_library( "//pkg/tcpip/buffer", "//pkg/tcpip/hash/jenkins", "//pkg/tcpip/header", - "//pkg/tcpip/iptables", "//pkg/tcpip/ports", "//pkg/tcpip/seqnum", "//pkg/tcpip/stack", "//pkg/tcpip/transport/raw", - "//pkg/tmutex", "//pkg/waiter", "@com_github_google_btree//:go_default_library", ], @@ -88,7 +87,9 @@ go_test( "tcp_timestamp_test.go", ], # FIXME(b/68809571) - tags = ["flaky"], + tags = [ + "flaky", + ], deps = [ ":tcp", "//pkg/sync", @@ -105,5 +106,16 @@ go_test( "//pkg/tcpip/stack", "//pkg/tcpip/transport/tcp/testing/context", "//pkg/waiter", + "//runsc/testutil", + ], +) + +go_test( + name = "rcv_test", + size = "small", + srcs = ["rcv_test.go"], + deps = [ + ":tcp", + "//pkg/tcpip/seqnum", ], ) diff --git a/pkg/tcpip/transport/tcp/accept.go b/pkg/tcpip/transport/tcp/accept.go index 08afb7c17..e6a23c978 100644 --- a/pkg/tcpip/transport/tcp/accept.go +++ b/pkg/tcpip/transport/tcp/accept.go @@ -17,6 +17,7 @@ package tcp import ( "crypto/sha1" "encoding/binary" + "fmt" "hash" "io" "time" @@ -25,7 +26,6 @@ import ( "gvisor.dev/gvisor/pkg/sleep" "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/tcpip" - "gvisor.dev/gvisor/pkg/tcpip/buffer" "gvisor.dev/gvisor/pkg/tcpip/header" "gvisor.dev/gvisor/pkg/tcpip/seqnum" "gvisor.dev/gvisor/pkg/tcpip/stack" @@ -49,17 +49,14 @@ const ( // timestamp and the current timestamp. If the difference is greater // than maxTSDiff, the cookie is expired. maxTSDiff = 2 -) -var ( - // SynRcvdCountThreshold is the global maximum number of connections - // that are allowed to be in SYN-RCVD state before TCP starts using SYN - // cookies to accept connections. - // - // It is an exported variable only for testing, and should not otherwise - // be used by importers of this package. + // SynRcvdCountThreshold is the default global maximum number of + // connections that are allowed to be in SYN-RCVD state before TCP + // starts using SYN cookies to accept connections. SynRcvdCountThreshold uint64 = 1000 +) +var ( // mssTable is a slice containing the possible MSS values that we // encode in the SYN cookie with two bits. mssTable = []uint16{536, 1300, 1440, 1460} @@ -74,29 +71,42 @@ func encodeMSS(mss uint16) uint32 { return 0 } -// syncRcvdCount is the number of endpoints in the SYN-RCVD state. The value is -// protected by a mutex so that we can increment only when it's guaranteed not -// to go above a threshold. -var synRcvdCount struct { - sync.Mutex - value uint64 - pending sync.WaitGroup -} - // listenContext is used by a listening endpoint to store state used while // listening for connections. This struct is allocated by the listen goroutine // and must not be accessed or have its methods called concurrently as they // may mutate the stored objects. type listenContext struct { - stack *stack.Stack - rcvWnd seqnum.Size - nonce [2][sha1.BlockSize]byte + stack *stack.Stack + + // synRcvdCount is a reference to the stack level synRcvdCount. + synRcvdCount *synRcvdCounter + + // rcvWnd is the receive window that is sent by this listening context + // in the initial SYN-ACK. + rcvWnd seqnum.Size + + // nonce are random bytes that are initialized once when the context + // is created and used to seed the hash function when generating + // the SYN cookie. + nonce [2][sha1.BlockSize]byte + + // listenEP is a reference to the listening endpoint associated with + // this context. Can be nil if the context is created by the forwarder. listenEP *endpoint + // hasherMu protects hasher. hasherMu sync.Mutex - hasher hash.Hash - v6only bool + // hasher is the hash function used to generate a SYN cookie. + hasher hash.Hash + + // v6Only is true if listenEP is a dual stack socket and has the + // IPV6_V6ONLY option set. + v6Only bool + + // netProto indicates the network protocol(IPv4/v6) for the listening + // endpoint. netProto tcpip.NetworkProtocolNumber + // pendingMu protects pendingEndpoints. This should only be accessed // by the listening endpoint's worker goroutine. // @@ -115,55 +125,22 @@ func timeStamp() uint32 { return uint32(time.Now().Unix()>>6) & tsMask } -// incSynRcvdCount tries to increment the global number of endpoints in SYN-RCVD -// state. It succeeds if the increment doesn't make the count go beyond the -// threshold, and fails otherwise. -func incSynRcvdCount() bool { - synRcvdCount.Lock() - - if synRcvdCount.value >= SynRcvdCountThreshold { - synRcvdCount.Unlock() - return false - } - - synRcvdCount.pending.Add(1) - synRcvdCount.value++ - - synRcvdCount.Unlock() - return true -} - -// decSynRcvdCount atomically decrements the global number of endpoints in -// SYN-RCVD state. It must only be called if a previous call to incSynRcvdCount -// succeeded. -func decSynRcvdCount() { - synRcvdCount.Lock() - - synRcvdCount.value-- - synRcvdCount.pending.Done() - synRcvdCount.Unlock() -} - -// synCookiesInUse() returns true if the synRcvdCount is greater than -// SynRcvdCountThreshold. -func synCookiesInUse() bool { - synRcvdCount.Lock() - v := synRcvdCount.value - synRcvdCount.Unlock() - return v >= SynRcvdCountThreshold -} - // newListenContext creates a new listen context. -func newListenContext(stk *stack.Stack, listenEP *endpoint, rcvWnd seqnum.Size, v6only bool, netProto tcpip.NetworkProtocolNumber) *listenContext { +func newListenContext(stk *stack.Stack, listenEP *endpoint, rcvWnd seqnum.Size, v6Only bool, netProto tcpip.NetworkProtocolNumber) *listenContext { l := &listenContext{ stack: stk, rcvWnd: rcvWnd, hasher: sha1.New(), - v6only: v6only, + v6Only: v6Only, netProto: netProto, listenEP: listenEP, pendingEndpoints: make(map[stack.TransportEndpointID]*endpoint), } + p, ok := stk.TransportProtocolInstance(ProtocolNumber).(*protocol) + if !ok { + panic(fmt.Sprintf("unable to get TCP protocol instance from stack: %+v", stk)) + } + l.synRcvdCount = p.SynRcvdCounter() rand.Read(l.nonce[0][:]) rand.Read(l.nonce[1][:]) @@ -221,7 +198,8 @@ func (l *listenContext) isCookieValid(id stack.TransportEndpointID, cookie seqnu } // createConnectingEndpoint creates a new endpoint in a connecting state, with -// the connection parameters given by the arguments. +// the connection parameters given by the arguments. The endpoint is returned +// with n.mu held. func (l *listenContext) createConnectingEndpoint(s *segment, iss seqnum.Value, irs seqnum.Value, rcvdSynOpts *header.TCPSynOptions, queue *waiter.Queue) (*endpoint, *tcpip.Error) { // Create a new endpoint. netProto := l.netProto @@ -229,34 +207,20 @@ func (l *listenContext) createConnectingEndpoint(s *segment, iss seqnum.Value, i netProto = s.route.NetProto } n := newEndpoint(l.stack, netProto, queue) - n.v6only = l.v6only + n.v6only = l.v6Only n.ID = s.id n.boundNICID = s.route.NICID() n.route = s.route.Clone() n.effectiveNetProtos = []tcpip.NetworkProtocolNumber{s.route.NetProto} n.rcvBufSize = int(l.rcvWnd) n.amss = mssForRoute(&n.route) + n.setEndpointState(StateConnecting) n.maybeEnableTimestamp(rcvdSynOpts) n.maybeEnableSACKPermitted(rcvdSynOpts) n.initGSO() - // Now inherit any socket options that should be inherited from the - // listening endpoint. - // In case of Forwarder listenEP will be nil and hence this check. - if l.listenEP != nil { - l.listenEP.propagateInheritableOptions(n) - } - - // Register new endpoint so that packets are routed to it. - if err := n.stack.RegisterTransportEndpoint(n.boundNICID, n.effectiveNetProtos, ProtocolNumber, n.ID, n, n.reusePort, n.boundBindToDevice); err != nil { - n.Close() - return nil, err - } - - n.isRegistered = true - // Create sender and receiver. // // The receiver at least temporarily has a zero receive window scale, @@ -268,12 +232,28 @@ func (l *listenContext) createConnectingEndpoint(s *segment, iss seqnum.Value, i // window to grow to a really large value. n.rcvAutoParams.prevCopied = n.initialReceiveWindow() + // Lock the endpoint before registering to ensure that no out of + // band changes are possible due to incoming packets etc till + // the endpoint is done initializing. + n.mu.Lock() + + // Register new endpoint so that packets are routed to it. + if err := n.stack.RegisterTransportEndpoint(n.boundNICID, n.effectiveNetProtos, ProtocolNumber, n.ID, n, n.reusePort, n.boundBindToDevice); err != nil { + n.mu.Unlock() + n.Close() + return nil, err + } + + n.isRegistered = true + return n, nil } // createEndpointAndPerformHandshake creates a new endpoint in connected state // and then performs the TCP 3-way handshake. -func (l *listenContext) createEndpointAndPerformHandshake(s *segment, opts *header.TCPSynOptions, queue *waiter.Queue) (*endpoint, *tcpip.Error) { +// +// The new endpoint is returned with e.mu held. +func (l *listenContext) createEndpointAndPerformHandshake(s *segment, opts *header.TCPSynOptions, queue *waiter.Queue, owner tcpip.PacketOwner) (*endpoint, *tcpip.Error) { // Create new endpoint. irs := s.sequenceNumber isn := generateSecureISN(s.id, l.stack.Seed()) @@ -281,6 +261,7 @@ func (l *listenContext) createEndpointAndPerformHandshake(s *segment, opts *head if err != nil { return nil, err } + ep.owner = owner // listenEP is nil when listenContext is used by tcp.Forwarder. deferAccept := time.Duration(0) @@ -288,25 +269,50 @@ func (l *listenContext) createEndpointAndPerformHandshake(s *segment, opts *head l.listenEP.mu.Lock() if l.listenEP.EndpointState() != StateListen { l.listenEP.mu.Unlock() + // Ensure we release any registrations done by the newly + // created endpoint. + ep.mu.Unlock() + ep.Close() + + // Wake up any waiters. This is strictly not required normally + // as a socket that was never accepted can't really have any + // registered waiters except when stack.Wait() is called which + // waits for all registered endpoints to stop and expects an + // EventHUp. + ep.waiterQueue.Notify(waiter.EventHUp | waiter.EventErr | waiter.EventIn | waiter.EventOut) return nil, tcpip.ErrConnectionAborted } l.addPendingEndpoint(ep) + + // Propagate any inheritable options from the listening endpoint + // to the newly created endpoint. + l.listenEP.propagateInheritableOptionsLocked(ep) + deferAccept = l.listenEP.deferAccept l.listenEP.mu.Unlock() } // Perform the 3-way handshake. - h := newPassiveHandshake(ep, seqnum.Size(ep.initialReceiveWindow()), isn, irs, opts, deferAccept) + h := newPassiveHandshake(ep, ep.rcv.rcvWnd, isn, irs, opts, deferAccept) if err := h.execute(); err != nil { + ep.mu.Unlock() ep.Close() + // Wake up any waiters. This is strictly not required normally + // as a socket that was never accepted can't really have any + // registered waiters except when stack.Wait() is called which + // waits for all registered endpoints to stop and expects an + // EventHUp. + ep.waiterQueue.Notify(waiter.EventHUp | waiter.EventErr | waiter.EventIn | waiter.EventOut) + if l.listenEP != nil { l.removePendingEndpoint(ep) } + + ep.drainClosingSegmentQueue() + return nil, err } - ep.mu.Lock() ep.isConnectNotified = true - ep.mu.Unlock() // Update the receive window scaling. We can't do it before the // handshake because it's possible that the peer doesn't support window @@ -340,30 +346,38 @@ func (l *listenContext) closeAllPendingEndpoints() { } // deliverAccepted delivers the newly-accepted endpoint to the listener. If the -// endpoint has transitioned out of the listen state, the new endpoint is closed -// instead. +// endpoint has transitioned out of the listen state (acceptedChan is nil), +// the new endpoint is closed instead. func (e *endpoint) deliverAccepted(n *endpoint) { e.mu.Lock() - state := e.EndpointState() e.pendingAccepted.Add(1) - defer e.pendingAccepted.Done() - acceptedChan := e.acceptedChan e.mu.Unlock() + defer e.pendingAccepted.Done() - if state == StateListen { - acceptedChan <- n - e.waiterQueue.Notify(waiter.EventIn) - } else { - n.Close() + e.acceptMu.Lock() + for { + if e.acceptedChan == nil { + e.acceptMu.Unlock() + n.notifyProtocolGoroutine(notifyReset) + return + } + select { + case e.acceptedChan <- n: + e.acceptMu.Unlock() + e.waiterQueue.Notify(waiter.EventIn) + return + default: + e.acceptCond.Wait() + } } } -// propagateInheritableOptions propagates any options set on the listening +// propagateInheritableOptionsLocked propagates any options set on the listening // endpoint to the newly created endpoint. -func (e *endpoint) propagateInheritableOptions(n *endpoint) { - e.mu.Lock() +// +// Precondition: e.mu and n.mu must be held. +func (e *endpoint) propagateInheritableOptionsLocked(n *endpoint) { n.userTimeout = e.userTimeout - e.mu.Unlock() } // handleSynSegment is called in its own goroutine once the listening endpoint @@ -373,11 +387,15 @@ func (e *endpoint) propagateInheritableOptions(n *endpoint) { // A limited number of these goroutines are allowed before TCP starts using SYN // cookies to accept connections. func (e *endpoint) handleSynSegment(ctx *listenContext, s *segment, opts *header.TCPSynOptions) { - defer decSynRcvdCount() - defer e.decSynRcvdCount() + defer ctx.synRcvdCount.dec() + defer func() { + e.mu.Lock() + e.decSynRcvdCount() + e.mu.Unlock() + }() defer s.decRef() - n, err := ctx.createEndpointAndPerformHandshake(s, opts, &waiter.Queue{}) + n, err := ctx.createEndpointAndPerformHandshake(s, opts, &waiter.Queue{}, e.owner) if err != nil { e.stack.Stats().TCP.FailedConnectionAttempts.Increment() e.stats.FailedConnectionAttempts.Increment() @@ -391,40 +409,39 @@ func (e *endpoint) handleSynSegment(ctx *listenContext, s *segment, opts *header } func (e *endpoint) incSynRcvdCount() bool { - e.mu.Lock() - if e.synRcvdCount >= cap(e.acceptedChan) { - e.mu.Unlock() - return false + e.acceptMu.Lock() + canInc := e.synRcvdCount < cap(e.acceptedChan) + e.acceptMu.Unlock() + if canInc { + e.synRcvdCount++ } - e.synRcvdCount++ - e.mu.Unlock() - return true + return canInc } func (e *endpoint) decSynRcvdCount() { - e.mu.Lock() e.synRcvdCount-- - e.mu.Unlock() } func (e *endpoint) acceptQueueIsFull() bool { - e.mu.Lock() - if l, c := len(e.acceptedChan)+e.synRcvdCount, cap(e.acceptedChan); l >= c { - e.mu.Unlock() - return true - } - e.mu.Unlock() - return false + e.acceptMu.Lock() + full := len(e.acceptedChan)+e.synRcvdCount >= cap(e.acceptedChan) + e.acceptMu.Unlock() + return full } // handleListenSegment is called when a listening endpoint receives a segment // and needs to handle it. func (e *endpoint) handleListenSegment(ctx *listenContext, s *segment) { - if s.flagsAreSet(header.TCPFlagSyn | header.TCPFlagAck) { + e.rcvListMu.Lock() + rcvClosed := e.rcvClosed + e.rcvListMu.Unlock() + if rcvClosed || s.flagsAreSet(header.TCPFlagSyn|header.TCPFlagAck) { + // If the endpoint is shutdown, reply with reset. + // // RFC 793 section 3.4 page 35 (figure 12) outlines that a RST // must be sent in response to a SYN-ACK while in the listen // state to prevent completing a handshake from an old SYN. - e.sendTCP(&s.route, s.id, buffer.VectorisedView{}, e.ttl, e.sendTOS, header.TCPFlagRst, s.ackNumber, 0, 0, nil, nil) + replyWithReset(s, e.sendTOS, e.ttl) return } @@ -434,7 +451,7 @@ func (e *endpoint) handleListenSegment(ctx *listenContext, s *segment) { switch { case s.flags == header.TCPFlagSyn: opts := parseSynSegmentOptions(s) - if incSynRcvdCount() { + if ctx.synRcvdCount.inc() { // Only handle the syn if the following conditions hold // - accept queue is not full. // - number of connections in synRcvd state is less than the @@ -444,7 +461,7 @@ func (e *endpoint) handleListenSegment(ctx *listenContext, s *segment) { go e.handleSynSegment(ctx, s, &opts) // S/R-SAFE: synRcvdCount is the barrier. return } - decSynRcvdCount() + ctx.synRcvdCount.dec() e.stack.Stats().TCP.ListenOverflowSynDrop.Increment() e.stats.ReceiveErrors.ListenOverflowSynDrop.Increment() e.stack.Stats().DroppedPackets.Increment() @@ -472,7 +489,15 @@ func (e *endpoint) handleListenSegment(ctx *listenContext, s *segment) { TSEcr: opts.TSVal, MSS: mssForRoute(&s.route), } - e.sendSynTCP(&s.route, s.id, e.ttl, e.sendTOS, header.TCPFlagSyn|header.TCPFlagAck, cookie, s.sequenceNumber+1, ctx.rcvWnd, synOpts) + e.sendSynTCP(&s.route, tcpFields{ + id: s.id, + ttl: e.ttl, + tos: e.sendTOS, + flags: header.TCPFlagSyn | header.TCPFlagAck, + seq: cookie, + ack: s.sequenceNumber + 1, + rcvWnd: ctx.rcvWnd, + }, synOpts) e.stack.Stats().TCP.ListenOverflowSynCookieSent.Increment() } @@ -489,7 +514,7 @@ func (e *endpoint) handleListenSegment(ctx *listenContext, s *segment) { return } - if !synCookiesInUse() { + if !ctx.synRcvdCount.synCookiesInUse() { // When not using SYN cookies, as per RFC 793, section 3.9, page 64: // Any acknowledgment is bad if it arrives on a connection still in // the LISTEN state. An acceptable reset segment should be formed @@ -505,7 +530,7 @@ func (e *endpoint) handleListenSegment(ctx *listenContext, s *segment) { // The only time we should reach here when a connection // was opened and closed really quickly and a delayed // ACK was received from the sender. - replyWithReset(s) + replyWithReset(s, e.sendTOS, e.ttl) return } @@ -551,6 +576,10 @@ func (e *endpoint) handleListenSegment(ctx *listenContext, s *segment) { return } + // Propagate any inheritable options from the listening endpoint + // to the newly created endpoint. + e.propagateInheritableOptionsLocked(n) + // clear the tsOffset for the newly created // endpoint as the Timestamp was already // randomly offset when the original SYN-ACK was @@ -584,15 +613,13 @@ func (e *endpoint) handleListenSegment(ctx *listenContext, s *segment) { // its own goroutine and is responsible for handling connection requests. func (e *endpoint) protocolListenLoop(rcvWnd seqnum.Size) *tcpip.Error { e.mu.Lock() - v6only := e.v6only - e.mu.Unlock() - ctx := newListenContext(e.stack, e, rcvWnd, v6only, e.NetProto) + v6Only := e.v6only + ctx := newListenContext(e.stack, e, rcvWnd, v6Only, e.NetProto) defer func() { // Mark endpoint as closed. This will prevent goroutines running // handleSynSegment() from attempting to queue new connections // to the endpoint. - e.mu.Lock() e.setEndpointState(StateClose) // close any endpoints in SYN-RCVD state. @@ -606,15 +633,20 @@ func (e *endpoint) protocolListenLoop(rcvWnd seqnum.Size) *tcpip.Error { } e.mu.Unlock() + e.drainClosingSegmentQueue() + // Notify waiters that the endpoint is shutdown. - e.waiterQueue.Notify(waiter.EventIn | waiter.EventOut) + e.waiterQueue.Notify(waiter.EventIn | waiter.EventOut | waiter.EventHUp | waiter.EventErr) }() s := sleep.Sleeper{} s.AddWaker(&e.notificationWaker, wakerForNotification) s.AddWaker(&e.newSegmentWaker, wakerForNewSegment) for { - switch index, _ := s.Fetch(true); index { + e.mu.Unlock() + index, _ := s.Fetch(true) + e.mu.Lock() + switch index { case wakerForNotification: n := e.fetchNotifications() if n¬ifyClose != 0 { @@ -627,7 +659,9 @@ func (e *endpoint) protocolListenLoop(rcvWnd seqnum.Size) *tcpip.Error { s.decRef() } close(e.drainDone) + e.mu.Unlock() <-e.undrain + e.mu.Lock() } case wakerForNewSegment: diff --git a/pkg/tcpip/transport/tcp/connect.go b/pkg/tcpip/transport/tcp/connect.go index 5c5397823..76e27bf26 100644 --- a/pkg/tcpip/transport/tcp/connect.go +++ b/pkg/tcpip/transport/tcp/connect.go @@ -61,6 +61,9 @@ const ( ) // handshake holds the state used during a TCP 3-way handshake. +// +// NOTE: handshake.ep.mu is held during handshake processing. It is released if +// we are going to block and reacquired when we start processing an event. type handshake struct { ep *endpoint state handshakeState @@ -102,24 +105,11 @@ type handshake struct { } func newHandshake(ep *endpoint, rcvWnd seqnum.Size) handshake { - rcvWndScale := ep.rcvWndScaleForHandshake() - - // Round-down the rcvWnd to a multiple of wndScale. This ensures that the - // window offered in SYN won't be reduced due to the loss of precision if - // window scaling is enabled after the handshake. - rcvWnd = (rcvWnd >> uint8(rcvWndScale)) << uint8(rcvWndScale) - - // Ensure we can always accept at least 1 byte if the scale specified - // was too high for the provided rcvWnd. - if rcvWnd == 0 { - rcvWnd = 1 - } - h := handshake{ ep: ep, active: true, rcvWnd: rcvWnd, - rcvWndScale: int(rcvWndScale), + rcvWndScale: ep.rcvWndScaleForHandshake(), } h.resetState() return h @@ -209,9 +199,7 @@ func (h *handshake) resetToSynRcvd(iss seqnum.Value, irs seqnum.Value, opts *hea h.mss = opts.MSS h.sndWndScale = opts.WS h.deferAccept = deferAccept - h.ep.mu.Lock() h.ep.setEndpointState(StateSynRecv) - h.ep.mu.Unlock() } // checkAck checks if the ACK number, if present, of a segment received during @@ -241,9 +229,7 @@ func (h *handshake) synSentState(s *segment) *tcpip.Error { // RFC 793, page 67, states that "If the RST bit is set [and] If the ACK // was acceptable then signal the user "error: connection reset", drop // the segment, enter CLOSED state, delete TCB, and return." - h.ep.mu.Lock() h.ep.workerCleanup = true - h.ep.mu.Unlock() // Although the RFC above calls out ECONNRESET, Linux actually returns // ECONNREFUSED here so we do as well. return tcpip.ErrConnectionRefused @@ -281,9 +267,7 @@ func (h *handshake) synSentState(s *segment) *tcpip.Error { if s.flagIsSet(header.TCPFlagAck) { h.state = handshakeCompleted - h.ep.mu.Lock() h.ep.transitionToStateEstablishedLocked(h) - h.ep.mu.Unlock() h.ep.sendRaw(buffer.VectorisedView{}, header.TCPFlagAck, h.iss+1, h.ackNum, h.rcvWnd>>h.effectiveRcvWndScale()) return nil @@ -293,10 +277,9 @@ func (h *handshake) synSentState(s *segment) *tcpip.Error { // but resend our own SYN and wait for it to be acknowledged in the // SYN-RCVD state. h.state = handshakeSynRcvd - h.ep.mu.Lock() ttl := h.ep.ttl + amss := h.ep.amss h.ep.setEndpointState(StateSynRecv) - h.ep.mu.Unlock() synOpts := header.TCPSynOptions{ WS: int(h.effectiveRcvWndScale()), TS: rcvSynOpts.TS, @@ -307,12 +290,20 @@ func (h *handshake) synSentState(s *segment) *tcpip.Error { // permits SACK. This is not explicitly defined in the RFC but // this is the behaviour implemented by Linux. SACKPermitted: rcvSynOpts.SACKPermitted, - MSS: h.ep.amss, + MSS: amss, } if ttl == 0 { ttl = s.route.DefaultTTL() } - h.ep.sendSynTCP(&s.route, h.ep.ID, ttl, h.ep.sendTOS, h.flags, h.iss, h.ackNum, h.rcvWnd, synOpts) + h.ep.sendSynTCP(&s.route, tcpFields{ + id: h.ep.ID, + ttl: ttl, + tos: h.ep.sendTOS, + flags: h.flags, + seq: h.iss, + ack: h.ackNum, + rcvWnd: h.rcvWnd, + }, synOpts) return nil } @@ -365,7 +356,15 @@ func (h *handshake) synRcvdState(s *segment) *tcpip.Error { SACKPermitted: h.ep.sackPermitted, MSS: h.ep.amss, } - h.ep.sendSynTCP(&s.route, h.ep.ID, h.ep.ttl, h.ep.sendTOS, h.flags, h.iss, h.ackNum, h.rcvWnd, synOpts) + h.ep.sendSynTCP(&s.route, tcpFields{ + id: h.ep.ID, + ttl: h.ep.ttl, + tos: h.ep.sendTOS, + flags: h.flags, + seq: h.iss, + ack: h.ackNum, + rcvWnd: h.rcvWnd, + }, synOpts) return nil } @@ -394,15 +393,14 @@ func (h *handshake) synRcvdState(s *segment) *tcpip.Error { } h.state = handshakeCompleted - h.ep.mu.Lock() h.ep.transitionToStateEstablishedLocked(h) + // If the segment has data then requeue it for the receiver // to process it again once main loop is started. if s.data.Size() > 0 { s.incRef() h.ep.enqueueSegment(s) } - h.ep.mu.Unlock() return nil } @@ -488,7 +486,9 @@ func (h *handshake) resolveRoute() *tcpip.Error { } if n¬ifyDrain != 0 { close(h.ep.drainDone) + h.ep.mu.Unlock() <-h.ep.undrain + h.ep.mu.Lock() } } @@ -553,10 +553,23 @@ func (h *handshake) execute() *tcpip.Error { synOpts.WS = -1 } } - h.ep.sendSynTCP(&h.ep.route, h.ep.ID, h.ep.ttl, h.ep.sendTOS, h.flags, h.iss, h.ackNum, h.rcvWnd, synOpts) + + h.ep.sendSynTCP(&h.ep.route, tcpFields{ + id: h.ep.ID, + ttl: h.ep.ttl, + tos: h.ep.sendTOS, + flags: h.flags, + seq: h.iss, + ack: h.ackNum, + rcvWnd: h.rcvWnd, + }, synOpts) for h.state != handshakeCompleted { - switch index, _ := s.Fetch(true); index { + h.ep.mu.Unlock() + index, _ := s.Fetch(true) + h.ep.mu.Lock() + switch index { + case wakerForResend: timeOut *= 2 if timeOut > MaxRTO { @@ -572,12 +585,20 @@ func (h *handshake) execute() *tcpip.Error { // the connection with another ACK or data (as ACKs are never // retransmitted on their own). if h.active || !h.acked || h.deferAccept != 0 && time.Since(h.startTime) > h.deferAccept { - h.ep.sendSynTCP(&h.ep.route, h.ep.ID, h.ep.ttl, h.ep.sendTOS, h.flags, h.iss, h.ackNum, h.rcvWnd, synOpts) + h.ep.sendSynTCP(&h.ep.route, tcpFields{ + id: h.ep.ID, + ttl: h.ep.ttl, + tos: h.ep.sendTOS, + flags: h.flags, + seq: h.iss, + ack: h.ackNum, + rcvWnd: h.rcvWnd, + }, synOpts) } case wakerForNotification: n := h.ep.fetchNotifications() - if n¬ifyClose != 0 { + if (n¬ifyClose)|(n¬ifyAbort) != 0 { return tcpip.ErrAborted } if n¬ifyDrain != 0 { @@ -593,7 +614,9 @@ func (h *handshake) execute() *tcpip.Error { } } close(h.ep.drainDone) + h.ep.mu.Unlock() <-h.ep.undrain + h.ep.mu.Lock() } case wakerForNewSegment: @@ -617,17 +640,17 @@ func parseSynSegmentOptions(s *segment) header.TCPSynOptions { var optionPool = sync.Pool{ New: func() interface{} { - return make([]byte, maxOptionSize) + return &[maxOptionSize]byte{} }, } func getOptions() []byte { - return optionPool.Get().([]byte) + return (*optionPool.Get().(*[maxOptionSize]byte))[:] } func putOptions(options []byte) { // Reslice to full capacity. - optionPool.Put(options[0:cap(options)]) + optionPool.Put(optionsToArray(options)) } func makeSynOptions(opts header.TCPSynOptions) []byte { @@ -683,18 +706,33 @@ func makeSynOptions(opts header.TCPSynOptions) []byte { return options[:offset] } -func (e *endpoint) sendSynTCP(r *stack.Route, id stack.TransportEndpointID, ttl, tos uint8, flags byte, seq, ack seqnum.Value, rcvWnd seqnum.Size, opts header.TCPSynOptions) *tcpip.Error { - options := makeSynOptions(opts) +// tcpFields is a struct to carry different parameters required by the +// send*TCP variant functions below. +type tcpFields struct { + id stack.TransportEndpointID + ttl uint8 + tos uint8 + flags byte + seq seqnum.Value + ack seqnum.Value + rcvWnd seqnum.Size + opts []byte + txHash uint32 +} + +func (e *endpoint) sendSynTCP(r *stack.Route, tf tcpFields, opts header.TCPSynOptions) *tcpip.Error { + tf.opts = makeSynOptions(opts) // We ignore SYN send errors and let the callers re-attempt send. - if err := e.sendTCP(r, id, buffer.VectorisedView{}, ttl, tos, flags, seq, ack, rcvWnd, options, nil); err != nil { + if err := e.sendTCP(r, tf, buffer.VectorisedView{}, nil); err != nil { e.stats.SendErrors.SynSendToNetworkFailed.Increment() } - putOptions(options) + putOptions(tf.opts) return nil } -func (e *endpoint) sendTCP(r *stack.Route, id stack.TransportEndpointID, data buffer.VectorisedView, ttl, tos uint8, flags byte, seq, ack seqnum.Value, rcvWnd seqnum.Size, opts []byte, gso *stack.GSO) *tcpip.Error { - if err := sendTCP(r, id, data, ttl, tos, flags, seq, ack, rcvWnd, opts, gso); err != nil { +func (e *endpoint) sendTCP(r *stack.Route, tf tcpFields, data buffer.VectorisedView, gso *stack.GSO) *tcpip.Error { + tf.txHash = e.txHash + if err := sendTCP(r, tf, data, gso, e.owner); err != nil { e.stats.SendErrors.SegmentSendToNetworkFailed.Increment() return err } @@ -702,24 +740,23 @@ func (e *endpoint) sendTCP(r *stack.Route, id stack.TransportEndpointID, data bu return nil } -func buildTCPHdr(r *stack.Route, id stack.TransportEndpointID, pkt *tcpip.PacketBuffer, flags byte, seq, ack seqnum.Value, rcvWnd seqnum.Size, opts []byte, gso *stack.GSO) { - optLen := len(opts) +func buildTCPHdr(r *stack.Route, tf tcpFields, pkt *stack.PacketBuffer, gso *stack.GSO) { + optLen := len(tf.opts) hdr := &pkt.Header - packetSize := pkt.DataSize - off := pkt.DataOffset + packetSize := pkt.Data.Size() // Initialize the header. tcp := header.TCP(hdr.Prepend(header.TCPMinimumSize + optLen)) pkt.TransportHeader = buffer.View(tcp) tcp.Encode(&header.TCPFields{ - SrcPort: id.LocalPort, - DstPort: id.RemotePort, - SeqNum: uint32(seq), - AckNum: uint32(ack), + SrcPort: tf.id.LocalPort, + DstPort: tf.id.RemotePort, + SeqNum: uint32(tf.seq), + AckNum: uint32(tf.ack), DataOffset: uint8(header.TCPMinimumSize + optLen), - Flags: flags, - WindowSize: uint16(rcvWnd), + Flags: tf.flags, + WindowSize: uint16(tf.rcvWnd), }) - copy(tcp[header.TCPMinimumSize:], opts) + copy(tcp[header.TCPMinimumSize:], tf.opts) length := uint16(hdr.UsedLength() + packetSize) xsum := r.PseudoHeaderChecksum(ProtocolNumber, length) @@ -731,48 +768,49 @@ func buildTCPHdr(r *stack.Route, id stack.TransportEndpointID, pkt *tcpip.Packet // header and data and get the right sum of the TCP packet. tcp.SetChecksum(xsum) } else if r.Capabilities()&stack.CapabilityTXChecksumOffload == 0 { - xsum = header.ChecksumVVWithOffset(pkt.Data, xsum, off, packetSize) + xsum = header.ChecksumVV(pkt.Data, xsum) tcp.SetChecksum(^tcp.CalculateChecksum(xsum)) } - } -func sendTCPBatch(r *stack.Route, id stack.TransportEndpointID, data buffer.VectorisedView, ttl, tos uint8, flags byte, seq, ack seqnum.Value, rcvWnd seqnum.Size, opts []byte, gso *stack.GSO) *tcpip.Error { - optLen := len(opts) - if rcvWnd > 0xffff { - rcvWnd = 0xffff +func sendTCPBatch(r *stack.Route, tf tcpFields, data buffer.VectorisedView, gso *stack.GSO, owner tcpip.PacketOwner) *tcpip.Error { + // We need to shallow clone the VectorisedView here as ReadToView will + // split the VectorisedView and Trim underlying views as it splits. Not + // doing the clone here will cause the underlying views of data itself + // to be altered. + data = data.Clone(nil) + + optLen := len(tf.opts) + if tf.rcvWnd > 0xffff { + tf.rcvWnd = 0xffff } mss := int(gso.MSS) n := (data.Size() + mss - 1) / mss - // Allocate one big slice for all the headers. - hdrSize := header.TCPMinimumSize + int(r.MaxHeaderLength()) + optLen - buf := make([]byte, n*hdrSize) - pkts := make([]tcpip.PacketBuffer, n) - for i := range pkts { - pkts[i].Header = buffer.NewEmptyPrependableFromView(buf[i*hdrSize:][:hdrSize]) - } - size := data.Size() - off := 0 + hdrSize := header.TCPMinimumSize + int(r.MaxHeaderLength()) + optLen + var pkts stack.PacketBufferList for i := 0; i < n; i++ { packetSize := mss if packetSize > size { packetSize = size } size -= packetSize - pkts[i].DataOffset = off - pkts[i].DataSize = packetSize - pkts[i].Data = data - buildTCPHdr(r, id, &pkts[i], flags, seq, ack, rcvWnd, opts, gso) - off += packetSize - seq = seq.Add(seqnum.Size(packetSize)) + var pkt stack.PacketBuffer + pkt.Header = buffer.NewPrependable(hdrSize) + pkt.Hash = tf.txHash + pkt.Owner = owner + data.ReadToVV(&pkt.Data, packetSize) + buildTCPHdr(r, tf, &pkt, gso) + tf.seq = tf.seq.Add(seqnum.Size(packetSize)) + pkts.PushBack(&pkt) } - if ttl == 0 { - ttl = r.DefaultTTL() + + if tf.ttl == 0 { + tf.ttl = r.DefaultTTL() } - sent, err := r.WritePackets(gso, pkts, stack.NetworkHeaderParams{Protocol: ProtocolNumber, TTL: ttl, TOS: tos}) + sent, err := r.WritePackets(gso, pkts, stack.NetworkHeaderParams{Protocol: ProtocolNumber, TTL: tf.ttl, TOS: tf.tos}) if err != nil { r.Stats().TCP.SegmentSendErrors.IncrementBy(uint64(n - sent)) } @@ -782,33 +820,33 @@ func sendTCPBatch(r *stack.Route, id stack.TransportEndpointID, data buffer.Vect // sendTCP sends a TCP segment with the provided options via the provided // network endpoint and under the provided identity. -func sendTCP(r *stack.Route, id stack.TransportEndpointID, data buffer.VectorisedView, ttl, tos uint8, flags byte, seq, ack seqnum.Value, rcvWnd seqnum.Size, opts []byte, gso *stack.GSO) *tcpip.Error { - optLen := len(opts) - if rcvWnd > 0xffff { - rcvWnd = 0xffff +func sendTCP(r *stack.Route, tf tcpFields, data buffer.VectorisedView, gso *stack.GSO, owner tcpip.PacketOwner) *tcpip.Error { + optLen := len(tf.opts) + if tf.rcvWnd > 0xffff { + tf.rcvWnd = 0xffff } if r.Loop&stack.PacketLoop == 0 && gso != nil && gso.Type == stack.GSOSW && int(gso.MSS) < data.Size() { - return sendTCPBatch(r, id, data, ttl, tos, flags, seq, ack, rcvWnd, opts, gso) + return sendTCPBatch(r, tf, data, gso, owner) } - pkt := tcpip.PacketBuffer{ - Header: buffer.NewPrependable(header.TCPMinimumSize + int(r.MaxHeaderLength()) + optLen), - DataOffset: 0, - DataSize: data.Size(), - Data: data, + pkt := stack.PacketBuffer{ + Header: buffer.NewPrependable(header.TCPMinimumSize + int(r.MaxHeaderLength()) + optLen), + Data: data, + Hash: tf.txHash, + Owner: owner, } - buildTCPHdr(r, id, &pkt, flags, seq, ack, rcvWnd, opts, gso) + buildTCPHdr(r, tf, &pkt, gso) - if ttl == 0 { - ttl = r.DefaultTTL() + if tf.ttl == 0 { + tf.ttl = r.DefaultTTL() } - if err := r.WritePacket(gso, stack.NetworkHeaderParams{Protocol: ProtocolNumber, TTL: ttl, TOS: tos}, pkt); err != nil { + if err := r.WritePacket(gso, stack.NetworkHeaderParams{Protocol: ProtocolNumber, TTL: tf.ttl, TOS: tf.tos}, pkt); err != nil { r.Stats().TCP.SegmentSendErrors.Increment() return err } r.Stats().TCP.SegmentsSent.Increment() - if (flags & header.TCPFlagRst) != 0 { + if (tf.flags & header.TCPFlagRst) != 0 { r.Stats().TCP.ResetsSent.Increment() } return nil @@ -860,7 +898,16 @@ func (e *endpoint) sendRaw(data buffer.VectorisedView, flags byte, seq, ack seqn sackBlocks = e.sack.Blocks[:e.sack.NumBlocks] } options := e.makeOptions(sackBlocks) - err := e.sendTCP(&e.route, e.ID, data, e.ttl, e.sendTOS, flags, seq, ack, rcvWnd, options, e.gso) + err := e.sendTCP(&e.route, tcpFields{ + id: e.ID, + ttl: e.ttl, + tos: e.sendTOS, + flags: flags, + seq: seq, + ack: ack, + rcvWnd: rcvWnd, + opts: options, + }, data, e.gso) putOptions(options) return err } @@ -875,7 +922,6 @@ func (e *endpoint) handleWrite() *tcpip.Error { first := e.sndQueue.Front() if first != nil { e.snd.writeList.PushBackList(&e.sndQueue) - e.snd.sndNxtList.UpdateForward(e.sndBufInQueue) e.sndBufInQueue = 0 } @@ -994,22 +1040,40 @@ func (e *endpoint) tryDeliverSegmentFromClosedEndpoint(s *segment) { ep = e.stack.FindTransportEndpoint(header.IPv4ProtocolNumber, e.TransProto, e.ID, &s.route) } if ep == nil { - replyWithReset(s) + replyWithReset(s, stack.DefaultTOS, s.route.DefaultTTL()) s.decRef() return } + + if e == ep { + panic("current endpoint not removed from demuxer, enqueing segments to itself") + } + if ep.(*endpoint).enqueueSegment(s) { ep.(*endpoint).newSegmentWaker.Assert() } } +// Drain segment queue from the endpoint and try to re-match the segment to a +// different endpoint. This is used when the current endpoint is transitioned to +// StateClose and has been unregistered from the transport demuxer. +func (e *endpoint) drainClosingSegmentQueue() { + for { + s := e.segmentQueue.dequeue() + if s == nil { + break + } + + e.tryDeliverSegmentFromClosedEndpoint(s) + } +} + func (e *endpoint) handleReset(s *segment) (ok bool, err *tcpip.Error) { if e.rcv.acceptable(s.sequenceNumber, 0) { // RFC 793, page 37 states that "in all states // except SYN-SENT, all reset (RST) segments are // validated by checking their SEQ-fields." So // we only process it if it's acceptable. - e.mu.Lock() switch e.EndpointState() { // In case of a RST in CLOSE-WAIT linux moves // the socket to closed state with an error set @@ -1033,11 +1097,9 @@ func (e *endpoint) handleReset(s *segment) (ok bool, err *tcpip.Error) { case StateCloseWait: e.transitionToStateCloseLocked() e.HardError = tcpip.ErrAborted - e.mu.Unlock() e.notifyProtocolGoroutine(notifyTickleWorker) return false, nil default: - e.mu.Unlock() // RFC 793, page 37 states that "in all states // except SYN-SENT, all reset (RST) segments are // validated by checking their SEQ-fields." So @@ -1150,9 +1212,7 @@ func (e *endpoint) handleSegment(s *segment) (cont bool, err *tcpip.Error) { // Now check if the received segment has caused us to transition // to a CLOSED state, if yes then terminate processing and do // not invoke the sender. - e.mu.RLock() state := e.state - e.mu.RUnlock() if state == StateClose { // When we get into StateClose while processing from the queue, // return immediately and let the protocolMainloop handle it. @@ -1175,9 +1235,7 @@ func (e *endpoint) handleSegment(s *segment) (cont bool, err *tcpip.Error) { // keepalive packets periodically when the connection is idle. If we don't hear // from the other side after a number of tries, we terminate the connection. func (e *endpoint) keepaliveTimerExpired() *tcpip.Error { - e.mu.RLock() userTimeout := e.userTimeout - e.mu.RUnlock() e.keepalive.Lock() if !e.keepalive.enabled || !e.keepalive.timer.checkExpiration() { @@ -1241,6 +1299,7 @@ func (e *endpoint) disableKeepaliveTimer() { // goroutine and is responsible for sending segments and handling received // segments. func (e *endpoint) protocolMainLoop(handshake bool, wakerInitDone chan<- struct{}) *tcpip.Error { + e.mu.Lock() var closeTimer *time.Timer var closeWaker sleep.Waker @@ -1262,7 +1321,9 @@ func (e *endpoint) protocolMainLoop(handshake bool, wakerInitDone chan<- struct{ } e.mu.Unlock() - e.workMu.Unlock() + + e.drainClosingSegmentQueue() + // When the protocol loop exits we should wake up our waiters. e.waiterQueue.Notify(waiter.EventHUp | waiter.EventErr | waiter.EventIn | waiter.EventOut) } @@ -1273,16 +1334,13 @@ func (e *endpoint) protocolMainLoop(handshake bool, wakerInitDone chan<- struct{ // completion. initialRcvWnd := e.initialReceiveWindow() h := newHandshake(e, seqnum.Size(initialRcvWnd)) - e.mu.Lock() h.ep.setEndpointState(StateSynSent) - e.mu.Unlock() if err := h.execute(); err != nil { e.lastErrorMu.Lock() e.lastError = err e.lastErrorMu.Unlock() - e.mu.Lock() e.setEndpointState(StateError) e.HardError = err @@ -1295,9 +1353,7 @@ func (e *endpoint) protocolMainLoop(handshake bool, wakerInitDone chan<- struct{ e.keepalive.timer.init(&e.keepalive.waker) defer e.keepalive.timer.cleanup() - e.mu.Lock() drained := e.drainDone != nil - e.mu.Unlock() if drained { close(e.drainDone) <-e.undrain @@ -1323,10 +1379,8 @@ func (e *endpoint) protocolMainLoop(handshake bool, wakerInitDone chan<- struct{ // This means the socket is being closed due // to the TCP-FIN-WAIT2 timeout was hit. Just // mark the socket as closed. - e.mu.Lock() e.transitionToStateCloseLocked() e.workerCleanup = true - e.mu.Unlock() return nil }, }, @@ -1372,7 +1426,7 @@ func (e *endpoint) protocolMainLoop(handshake bool, wakerInitDone chan<- struct{ e.snd.updateMaxPayloadSize(mtu, count) } - if n¬ifyReset != 0 { + if n¬ifyReset != 0 || n¬ifyAbort != 0 { return tcpip.ErrConnectionAborted } @@ -1381,7 +1435,6 @@ func (e *endpoint) protocolMainLoop(handshake bool, wakerInitDone chan<- struct{ } if n¬ifyClose != 0 && closeTimer == nil { - e.mu.Lock() if e.EndpointState() == StateFinWait2 && e.closed { // The socket has been closed and we are in FIN_WAIT2 // so start the FIN_WAIT2 timer. @@ -1390,7 +1443,6 @@ func (e *endpoint) protocolMainLoop(handshake bool, wakerInitDone chan<- struct{ }) e.waiterQueue.Notify(waiter.EventHUp | waiter.EventErr | waiter.EventIn | waiter.EventOut) } - e.mu.Unlock() } if n¬ifyKeepaliveChanged != 0 { @@ -1410,7 +1462,9 @@ func (e *endpoint) protocolMainLoop(handshake bool, wakerInitDone chan<- struct{ // Only block the worker if the endpoint // is not in closed state or error state. close(e.drainDone) + e.mu.Unlock() <-e.undrain + e.mu.Lock() } } @@ -1453,7 +1507,6 @@ func (e *endpoint) protocolMainLoop(handshake bool, wakerInitDone chan<- struct{ } e.rcvListMu.Unlock() - e.mu.Lock() if e.workerCleanup { e.notifyProtocolGoroutine(notifyClose) } @@ -1461,7 +1514,6 @@ func (e *endpoint) protocolMainLoop(handshake bool, wakerInitDone chan<- struct{ // Main loop. Handle segments until both send and receive ends of the // connection have completed. cleanupOnError := func(err *tcpip.Error) { - e.mu.Lock() e.workerCleanup = true if err != nil { e.resetConnectionLocked(err) @@ -1473,16 +1525,11 @@ func (e *endpoint) protocolMainLoop(handshake bool, wakerInitDone chan<- struct{ loop: for e.EndpointState() != StateTimeWait && e.EndpointState() != StateClose && e.EndpointState() != StateError { e.mu.Unlock() - e.workMu.Unlock() v, _ := s.Fetch(true) - e.workMu.Lock() + e.mu.Lock() - // We need to double check here because the notification maybe + // We need to double check here because the notification may be // stale by the time we got around to processing it. - // - // NOTE: since we now hold the workMu the processors cannot - // change the state of the endpoint so it's safe to proceed - // after this check. switch e.EndpointState() { case StateError: // If the endpoint has already transitioned to an ERROR @@ -1495,21 +1542,17 @@ loop: case StateTimeWait: fallthrough case StateClose: - e.mu.Lock() break loop default: if err := funcs[v].f(); err != nil { cleanupOnError(err) return nil } - e.mu.Lock() } } - state := e.EndpointState() - e.mu.Unlock() var reuseTW func() - if state == StateTimeWait { + if e.EndpointState() == StateTimeWait { // Disable close timer as we now entering real TIME_WAIT. if closeTimer != nil { closeTimer.Stop() @@ -1519,14 +1562,11 @@ loop: s.Done() // Wake up any waiters before we enter TIME_WAIT. e.waiterQueue.Notify(waiter.EventHUp | waiter.EventErr | waiter.EventIn | waiter.EventOut) - e.mu.Lock() e.workerCleanup = true - e.mu.Unlock() reuseTW = e.doTimeWait() } // Mark endpoint as closed. - e.mu.Lock() if e.EndpointState() != StateError { e.transitionToStateCloseLocked() } @@ -1534,19 +1574,6 @@ loop: // Lock released below. epilogue() - // epilogue removes the endpoint from the transport-demuxer and - // unlocks e.mu. Now that no new segments can get enqueued to this - // endpoint, try to re-match the segment to a different endpoint - // as the current endpoint is closed. - for { - s := e.segmentQueue.dequeue() - if s == nil { - break - } - - e.tryDeliverSegmentFromClosedEndpoint(s) - } - // A new SYN was received during TIME_WAIT and we need to abort // the timewait and redirect the segment to the listener queue if reuseTW != nil { @@ -1632,6 +1659,7 @@ func (e *endpoint) doTimeWait() (twReuse func()) { const timeWaitDone = 3 s := sleep.Sleeper{} + defer s.Done() s.AddWaker(&e.newSegmentWaker, newSegment) s.AddWaker(&e.notificationWaker, notification) @@ -1641,9 +1669,9 @@ func (e *endpoint) doTimeWait() (twReuse func()) { defer timeWaitTimer.Stop() for { - e.workMu.Unlock() + e.mu.Unlock() v, _ := s.Fetch(true) - e.workMu.Lock() + e.mu.Lock() switch v { case newSegment: extendTimeWait, reuseTW := e.handleTimeWaitSegments() @@ -1655,7 +1683,7 @@ func (e *endpoint) doTimeWait() (twReuse func()) { } case notification: n := e.fetchNotifications() - if n¬ifyClose != 0 { + if n¬ifyClose != 0 || n¬ifyAbort != 0 { return nil } if n¬ifyDrain != 0 { @@ -1666,7 +1694,9 @@ func (e *endpoint) doTimeWait() (twReuse func()) { e.handleTimeWaitSegments() } close(e.drainDone) + e.mu.Unlock() <-e.undrain + e.mu.Lock() return nil } case timeWaitDone: diff --git a/pkg/tcpip/transport/tcp/connect_unsafe.go b/pkg/tcpip/transport/tcp/connect_unsafe.go new file mode 100644 index 000000000..cfc304616 --- /dev/null +++ b/pkg/tcpip/transport/tcp/connect_unsafe.go @@ -0,0 +1,30 @@ +// Copyright 2018 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package tcp + +import ( + "reflect" + "unsafe" +) + +// optionsToArray converts a slice of capacity >-= maxOptionSize to an array. +// +// optionsToArray panics if the capacity of options is smaller than +// maxOptionSize. +func optionsToArray(options []byte) *[maxOptionSize]byte { + // Reslice to full capacity. + options = options[0:maxOptionSize] + return (*[maxOptionSize]byte)(unsafe.Pointer((*reflect.SliceHeader)(unsafe.Pointer(&options)).Data)) +} diff --git a/pkg/tcpip/transport/tcp/dispatcher.go b/pkg/tcpip/transport/tcp/dispatcher.go index e18012ac0..6062ca916 100644 --- a/pkg/tcpip/transport/tcp/dispatcher.go +++ b/pkg/tcpip/transport/tcp/dispatcher.go @@ -18,7 +18,6 @@ import ( "gvisor.dev/gvisor/pkg/rand" "gvisor.dev/gvisor/pkg/sleep" "gvisor.dev/gvisor/pkg/sync" - "gvisor.dev/gvisor/pkg/tcpip" "gvisor.dev/gvisor/pkg/tcpip/hash/jenkins" "gvisor.dev/gvisor/pkg/tcpip/header" "gvisor.dev/gvisor/pkg/tcpip/stack" @@ -68,17 +67,28 @@ func (q *epQueue) empty() bool { type processor struct { epQ epQueue newEndpointWaker sleep.Waker + closeWaker sleep.Waker id int + wg sync.WaitGroup } func newProcessor(id int) *processor { p := &processor{ id: id, } + p.wg.Add(1) go p.handleSegments() return p } +func (p *processor) close() { + p.closeWaker.Assert() +} + +func (p *processor) wait() { + p.wg.Wait() +} + func (p *processor) queueEndpoint(ep *endpoint) { // Queue an endpoint for processing by the processor goroutine. p.epQ.enqueue(ep) @@ -87,11 +97,17 @@ func (p *processor) queueEndpoint(ep *endpoint) { func (p *processor) handleSegments() { const newEndpointWaker = 1 + const closeWaker = 2 s := sleep.Sleeper{} s.AddWaker(&p.newEndpointWaker, newEndpointWaker) + s.AddWaker(&p.closeWaker, closeWaker) defer s.Done() for { - s.Fetch(true) + id, ok := s.Fetch(true) + if ok && id == closeWaker { + p.wg.Done() + return + } for ep := p.epQ.dequeue(); ep != nil; ep = p.epQ.dequeue() { if ep.segmentQueue.empty() { continue @@ -111,7 +127,7 @@ func (p *processor) handleSegments() { continue } - if !ep.workMu.TryLock() { + if !ep.mu.TryLock() { ep.newSegmentWaker.Assert() continue } @@ -121,12 +137,10 @@ func (p *processor) handleSegments() { if err := ep.handleSegments(true /* fastPath */); err != nil || ep.EndpointState() == StateClose { // Send any active resets if required. if err != nil { - ep.mu.Lock() ep.resetConnectionLocked(err) - ep.mu.Unlock() } ep.notifyProtocolGoroutine(notifyTickleWorker) - ep.workMu.Unlock() + ep.mu.Unlock() continue } @@ -134,7 +148,7 @@ func (p *processor) handleSegments() { p.epQ.enqueue(ep) } - ep.workMu.Unlock() + ep.mu.Unlock() } } } @@ -160,7 +174,19 @@ func newDispatcher(nProcessors int) *dispatcher { } } -func (d *dispatcher) queuePacket(r *stack.Route, stackEP stack.TransportEndpoint, id stack.TransportEndpointID, pkt tcpip.PacketBuffer) { +func (d *dispatcher) close() { + for _, p := range d.processors { + p.close() + } +} + +func (d *dispatcher) wait() { + for _, p := range d.processors { + p.wait() + } +} + +func (d *dispatcher) queuePacket(r *stack.Route, stackEP stack.TransportEndpoint, id stack.TransportEndpointID, pkt stack.PacketBuffer) { ep := stackEP.(*endpoint) s := newSegment(r, id, pkt) if !s.parse() { diff --git a/pkg/tcpip/transport/tcp/dual_stack_test.go b/pkg/tcpip/transport/tcp/dual_stack_test.go index 4f361b226..804e95aea 100644 --- a/pkg/tcpip/transport/tcp/dual_stack_test.go +++ b/pkg/tcpip/transport/tcp/dual_stack_test.go @@ -568,11 +568,10 @@ func TestV4AcceptOnV4(t *testing.T) { func testV4ListenClose(t *testing.T, c *context.Context) { // Set the SynRcvd threshold to zero to force a syn cookie based accept // to happen. - saved := tcp.SynRcvdCountThreshold - defer func() { - tcp.SynRcvdCountThreshold = saved - }() - tcp.SynRcvdCountThreshold = 0 + if err := c.Stack().SetTransportProtocolOption(tcp.ProtocolNumber, tcpip.TCPSynRcvdCountThresholdOption(0)); err != nil { + t.Fatalf("setting TCPSynRcvdCountThresholdOption failed: %s", err) + } + const n = uint16(32) // Start listening. diff --git a/pkg/tcpip/transport/tcp/endpoint.go b/pkg/tcpip/transport/tcp/endpoint.go index f2be0e651..45f2aa78b 100644 --- a/pkg/tcpip/transport/tcp/endpoint.go +++ b/pkg/tcpip/transport/tcp/endpoint.go @@ -18,6 +18,7 @@ import ( "encoding/binary" "fmt" "math" + "runtime" "strings" "sync/atomic" "time" @@ -29,11 +30,9 @@ import ( "gvisor.dev/gvisor/pkg/tcpip/buffer" "gvisor.dev/gvisor/pkg/tcpip/hash/jenkins" "gvisor.dev/gvisor/pkg/tcpip/header" - "gvisor.dev/gvisor/pkg/tcpip/iptables" "gvisor.dev/gvisor/pkg/tcpip/ports" "gvisor.dev/gvisor/pkg/tcpip/seqnum" "gvisor.dev/gvisor/pkg/tcpip/stack" - "gvisor.dev/gvisor/pkg/tmutex" "gvisor.dev/gvisor/pkg/waiter" ) @@ -121,6 +120,8 @@ const ( notifyDrain notifyReset notifyResetByPeer + // notifyAbort is a request for an expedited teardown. + notifyAbort notifyKeepaliveChanged notifyMSSChanged // notifyTickleWorker is used to tickle the protocol main loop during a @@ -281,6 +282,38 @@ func (*EndpointInfo) IsEndpointInfo() {} // synchronized. The protocol implementation, however, runs in a single // goroutine. // +// Each endpoint has a few mutexes: +// +// e.mu -> Primary mutex for an endpoint must be held for all operations except +// in e.Readiness where acquiring it will result in a deadlock in epoll +// implementation. +// +// The following three mutexes can be acquired independent of e.mu but if +// acquired with e.mu then e.mu must be acquired first. +// +// e.acceptMu -> protects acceptedChan. +// e.rcvListMu -> Protects the rcvList and associated fields. +// e.sndBufMu -> Protects the sndQueue and associated fields. +// e.lastErrorMu -> Protects the lastError field. +// +// LOCKING/UNLOCKING of the endpoint. The locking of an endpoint is different +// based on the context in which the lock is acquired. In the syscall context +// e.LockUser/e.UnlockUser should be used and when doing background processing +// e.mu.Lock/e.mu.Unlock should be used. The distinction is described below +// in brief. +// +// The reason for this locking behaviour is to avoid wakeups to handle packets. +// In cases where the endpoint is already locked the background processor can +// queue the packet up and go its merry way and the lock owner will eventually +// process the backlog when releasing the lock. Similarly when acquiring the +// lock from say a syscall goroutine we can implement a bit of spinning if we +// know that the lock is not held by another syscall goroutine. Background +// processors should never hold the lock for long and we can avoid an expensive +// sleep/wakeup by spinning for a shortwhile. +// +// For more details please see the detailed documentation on +// e.LockUser/e.UnlockUser methods. +// // +stateify savable type endpoint struct { EndpointInfo @@ -297,12 +330,6 @@ type endpoint struct { // Precondition: epQueue.mu must be held to read/write this field.. pendingProcessing bool `state:"nosave"` - // workMu is used to arbitrate which goroutine may perform protocol - // work. Only the main protocol goroutine is expected to call Lock() on - // it, but other goroutines (e.g., send) may call TryLock() to eagerly - // perform work without having to wait for the main one to wake up. - workMu tmutex.Mutex `state:"nosave"` - // The following fields are initialized at creation time and do not // change throughout the lifetime of the endpoint. stack *stack.Stack `state:"manual"` @@ -328,15 +355,11 @@ type endpoint struct { rcvBufSize int rcvBufUsed int rcvAutoParams rcvBufAutoTuneParams - // zeroWindow indicates that the window was closed due to receive buffer - // space being filled up. This is set by the worker goroutine before - // moving a segment to the rcvList. This setting is cleared by the - // endpoint when a Read() call reads enough data for the new window to - // be non-zero. - zeroWindow bool - // The following fields are protected by the mutex. - mu sync.RWMutex `state:"nosave"` + // mu protects all endpoint fields unless documented otherwise. mu must + // be acquired before interacting with the endpoint fields. + mu sync.Mutex `state:"nosave"` + ownedByUser uint32 // state must be read/set using the EndpointState()/setEndpointState() methods. state EndpointState `state:".(EndpointState)"` @@ -511,6 +534,23 @@ type endpoint struct { // to the acceptedChan below terminate before we close acceptedChan. pendingAccepted sync.WaitGroup `state:"nosave"` + // acceptMu protects acceptedChan. + acceptMu sync.Mutex `state:"nosave"` + + // acceptCond is a condition variable that can be used to block on when + // acceptedChan is full and an endpoint is ready to be delivered. + // + // This condition variable is required because just blocking on sending + // to acceptedChan does not work in cases where endpoint.Listen is + // called twice with different backlog values. In such cases the channel + // is closed and a new one created. Any pending goroutines blocking on + // the write to the channel will panic. + // + // We use this condition variable to block/unblock goroutines which + // tried to deliver an endpoint but couldn't because accept backlog was + // full ( See: endpoint.deliverAccepted ). + acceptCond *sync.Cond `state:"nosave"` + // acceptedChan is used by a listening endpoint protocol goroutine to // send newly accepted connections to the endpoint so that they can be // read by Accept() calls. @@ -559,6 +599,13 @@ type endpoint struct { // endpoint and at this point the endpoint is only around // to complete the TCP shutdown. closed bool + + // txHash is the transport layer hash to be set on outbound packets + // emitted by this endpoint. + txHash uint32 + + // owner is used to get uid and gid of the packet. + owner tcpip.PacketOwner } // UniqueID implements stack.TransportEndpoint.UniqueID. @@ -581,14 +628,93 @@ func calculateAdvertisedMSS(userMSS uint16, r stack.Route) uint16 { return maxMSS } +// LockUser tries to lock e.mu and if it fails it will check if the lock is held +// by another syscall goroutine. If yes, then it will goto sleep waiting for the +// lock to be released, if not then it will spin till it acquires the lock or +// another syscall goroutine acquires it in which case it will goto sleep as +// described above. +// +// The assumption behind spinning here being that background packet processing +// should not be holding the lock for long and spinning reduces latency as we +// avoid an expensive sleep/wakeup of of the syscall goroutine). +func (e *endpoint) LockUser() { + for { + // Try first if the sock is locked then check if it's owned + // by another user goroutine if not then we spin, otherwise + // we just goto sleep on the Lock() and wait. + if !e.mu.TryLock() { + // If socket is owned by the user then just goto sleep + // as the lock could be held for a reasonably long time. + if atomic.LoadUint32(&e.ownedByUser) == 1 { + e.mu.Lock() + atomic.StoreUint32(&e.ownedByUser, 1) + return + } + // Spin but yield the processor since the lower half + // should yield the lock soon. + runtime.Gosched() + continue + } + atomic.StoreUint32(&e.ownedByUser, 1) + return + } +} + +// UnlockUser will check if there are any segments already queued for processing +// and process any such segments before unlocking e.mu. This is required because +// we when packets arrive and endpoint lock is already held then such packets +// are queued up to be processed. If the lock is held by the endpoint goroutine +// then it will process these packets but if the lock is instead held by the +// syscall goroutine then we can have the syscall goroutine process the backlog +// before unlocking. +// +// This avoids an unnecessary wakeup of the endpoint protocol goroutine for the +// endpoint. It's also required eventually when we get rid of the endpoint +// protocol goroutine altogether. +// +// Precondition: e.LockUser() must have been called before calling e.UnlockUser() +func (e *endpoint) UnlockUser() { + // Lock segment queue before checking so that we avoid a race where + // segments can be queued between the time we check if queue is empty + // and actually unlock the endpoint mutex. + for { + e.segmentQueue.mu.Lock() + if e.segmentQueue.emptyLocked() { + if atomic.SwapUint32(&e.ownedByUser, 0) != 1 { + panic("e.UnlockUser() called without calling e.LockUser()") + } + e.mu.Unlock() + e.segmentQueue.mu.Unlock() + return + } + e.segmentQueue.mu.Unlock() + + switch e.EndpointState() { + case StateEstablished: + if err := e.handleSegments(true /* fastPath */); err != nil { + e.notifyProtocolGoroutine(notifyTickleWorker) + } + default: + // Since we are waking the endpoint goroutine here just unlock + // and let it process the queued segments. + e.newSegmentWaker.Assert() + if atomic.SwapUint32(&e.ownedByUser, 0) != 1 { + panic("e.UnlockUser() called without calling e.LockUser()") + } + e.mu.Unlock() + return + } + } +} + // StopWork halts packet processing. Only to be used in tests. func (e *endpoint) StopWork() { - e.workMu.Lock() + e.mu.Lock() } // ResumeWork resumes packet processing. Only to be used in tests. func (e *endpoint) ResumeWork() { - e.workMu.Unlock() + e.mu.Unlock() } // setEndpointState updates the state of the endpoint to state atomically. This @@ -670,6 +796,7 @@ func newEndpoint(s *stack.Stack, netProto tcpip.NetworkProtocolNumber, waiterQue count: 9, }, uniqueID: s.UniqueID(), + txHash: s.Rand().Uint32(), } var ss SendBufferSizeOption @@ -694,7 +821,7 @@ func newEndpoint(s *stack.Stack, netProto tcpip.NetworkProtocolNumber, waiterQue var de DelayEnabled if err := s.TransportProtocolOption(ProtocolNumber, &de); err == nil && de { - e.SetSockOptInt(tcpip.DelayOption, 1) + e.SetSockOptBool(tcpip.DelayOption, true) } var tcpLT tcpip.TCPLingerTimeoutOption @@ -707,9 +834,8 @@ func newEndpoint(s *stack.Stack, netProto tcpip.NetworkProtocolNumber, waiterQue } e.segmentQueue.setLimit(MaxUnprocessedSegments) - e.workMu.Init() - e.workMu.Lock() e.tsOffset = timeStampOffset() + e.acceptCond = sync.NewCond(&e.acceptMu) return e } @@ -719,9 +845,6 @@ func newEndpoint(s *stack.Stack, netProto tcpip.NetworkProtocolNumber, waiterQue func (e *endpoint) Readiness(mask waiter.EventMask) waiter.EventMask { result := waiter.EventMask(0) - e.mu.RLock() - defer e.mu.RUnlock() - switch e.EndpointState() { case StateInitial, StateBound, StateConnecting, StateSynSent, StateSynRecv: // Ready for nothing. @@ -733,9 +856,11 @@ func (e *endpoint) Readiness(mask waiter.EventMask) waiter.EventMask { case StateListen: // Check if there's anything in the accepted channel. if (mask & waiter.EventIn) != 0 { + e.acceptMu.Lock() if len(e.acceptedChan) > 0 { result |= waiter.EventIn } + e.acceptMu.Unlock() } } if e.EndpointState().connected() { @@ -785,29 +910,58 @@ func (e *endpoint) notifyProtocolGoroutine(n uint32) { } } +// Abort implements stack.TransportEndpoint.Abort. +func (e *endpoint) Abort() { + // The abort notification is not processed synchronously, so no + // synchronization is needed. + // + // If the endpoint becomes connected after this check, we still close + // the endpoint. This worst case results in a slower abort. + // + // If the endpoint disconnected after the check, nothing needs to be + // done, so sending a notification which will potentially be ignored is + // fine. + // + // If the endpoint connecting finishes after the check, the endpoint + // is either in a connected state (where we would notifyAbort anyway), + // SYN-RECV (where we would also notifyAbort anyway), or in an error + // state where nothing is required and the notification can be safely + // ignored. + // + // Endpoints where a Close during connecting or SYN-RECV state would be + // problematic are set to state connecting before being registered (and + // thus possible to be Aborted). They are never available in initial + // state. + // + // Endpoints transitioning from initial to connecting state may be + // safely either closed or sent notifyAbort. + if s := e.EndpointState(); s == StateConnecting || s == StateSynRecv || s.connected() { + e.notifyProtocolGoroutine(notifyAbort) + return + } + e.Close() +} + // Close puts the endpoint in a closed state and frees all resources associated // with it. It must be called only once and with no other concurrent calls to // the endpoint. func (e *endpoint) Close() { - e.mu.Lock() - closed := e.closed - e.mu.Unlock() - if closed { + e.LockUser() + defer e.UnlockUser() + if e.closed { return } // Issue a shutdown so that the peer knows we won't send any more data // if we're connected, or stop accepting if we're listening. - e.Shutdown(tcpip.ShutdownWrite | tcpip.ShutdownRead) - e.closeNoShutdown() + e.shutdownLocked(tcpip.ShutdownWrite | tcpip.ShutdownRead) + e.closeNoShutdownLocked() } // closeNoShutdown closes the endpoint without doing a full shutdown. This is // used when a connection needs to be aborted with a RST and we want to skip // a full 4 way TCP shutdown. -func (e *endpoint) closeNoShutdown() { - e.mu.Lock() - +func (e *endpoint) closeNoShutdownLocked() { // For listening sockets, we always release ports inline so that they // are immediately available for reuse after Close() is called. If also // registered, we unregister as well otherwise the next user would fail @@ -826,57 +980,55 @@ func (e *endpoint) closeNoShutdown() { // Mark endpoint as closed. e.closed = true + + switch e.EndpointState() { + case StateClose, StateError: + return + } + // Either perform the local cleanup or kick the worker to make sure it // knows it needs to cleanup. - tcpip.AddDanglingEndpoint(e) - if !e.workerRunning { - e.cleanupLocked() - } else { + if e.workerRunning { e.workerCleanup = true + tcpip.AddDanglingEndpoint(e) + // Worker will remove the dangling endpoint when the endpoint + // goroutine terminates. e.notifyProtocolGoroutine(notifyClose) + } else { + e.transitionToStateCloseLocked() } - - e.mu.Unlock() } // closePendingAcceptableConnections closes all connections that have completed // handshake but not yet been delivered to the application. func (e *endpoint) closePendingAcceptableConnectionsLocked() { - done := make(chan struct{}) - // Spin a goroutine up as ranging on e.acceptedChan will just block when - // there are no more connections in the channel. Using a non-blocking - // select does not work as it can potentially select the default case - // even when there are pending writes but that are not yet written to - // the channel. - go func() { - defer close(done) - for n := range e.acceptedChan { - n.notifyProtocolGoroutine(notifyReset) - // close all connections that have completed but - // not accepted by the application. - n.Close() - } - }() - // pendingAccepted(see endpoint.deliverAccepted) tracks the number of - // endpoints which have completed handshake but are not yet written to - // the e.acceptedChan. We wait here till the goroutine above can drain - // all such connections from e.acceptedChan. - e.pendingAccepted.Wait() + e.acceptMu.Lock() + if e.acceptedChan == nil { + e.acceptMu.Unlock() + return + } close(e.acceptedChan) - <-done + ch := e.acceptedChan e.acceptedChan = nil + e.acceptCond.Broadcast() + e.acceptMu.Unlock() + + // Reset all connections that are waiting to be accepted. + for n := range ch { + n.notifyProtocolGoroutine(notifyReset) + } + // Wait for reset of all endpoints that are still waiting to be delivered to + // the now closed acceptedChan. + e.pendingAccepted.Wait() } // cleanupLocked frees all resources associated with the endpoint. It is called // after Close() is called and the worker goroutine (if any) is done with its // work. func (e *endpoint) cleanupLocked() { - // Close all endpoints that might have been accepted by TCP but not by // the client. - if e.acceptedChan != nil { - e.closePendingAcceptableConnectionsLocked() - } + e.closePendingAcceptableConnectionsLocked() e.workerCleanup = false @@ -910,12 +1062,28 @@ func (e *endpoint) initialReceiveWindow() int { if rcvWnd > routeWnd { rcvWnd = routeWnd } + rcvWndScale := e.rcvWndScaleForHandshake() + + // Round-down the rcvWnd to a multiple of wndScale. This ensures that the + // window offered in SYN won't be reduced due to the loss of precision if + // window scaling is enabled after the handshake. + rcvWnd = (rcvWnd >> uint8(rcvWndScale)) << uint8(rcvWndScale) + + // Ensure we can always accept at least 1 byte if the scale specified + // was too high for the provided rcvWnd. + if rcvWnd == 0 { + rcvWnd = 1 + } + return rcvWnd } // ModerateRecvBuf adjusts the receive buffer and the advertised window // based on the number of bytes copied to user space. func (e *endpoint) ModerateRecvBuf(copied int) { + e.LockUser() + defer e.UnlockUser() + e.rcvListMu.Lock() if e.rcvAutoParams.disabled { e.rcvListMu.Unlock() @@ -965,7 +1133,7 @@ func (e *endpoint) ModerateRecvBuf(copied int) { e.rcvBufSize = rcvWnd availAfter := e.receiveBufferAvailableLocked() mask := uint32(notifyReceiveWindowChanged) - if crossed, above := e.windowCrossedACKThreshold(availAfter - availBefore); crossed && above { + if crossed, above := e.windowCrossedACKThresholdLocked(availAfter - availBefore); crossed && above { mask |= notifyNonZeroReceiveWindow } e.notifyProtocolGoroutine(mask) @@ -982,14 +1150,18 @@ func (e *endpoint) ModerateRecvBuf(copied int) { e.rcvListMu.Unlock() } +func (e *endpoint) SetOwner(owner tcpip.PacketOwner) { + e.owner = owner +} + // IPTables implements tcpip.Endpoint.IPTables. -func (e *endpoint) IPTables() (iptables.IPTables, error) { +func (e *endpoint) IPTables() (stack.IPTables, error) { return e.stack.IPTables(), nil } // Read reads data from the endpoint. func (e *endpoint) Read(*tcpip.FullAddress) (buffer.View, tcpip.ControlMessages, *tcpip.Error) { - e.mu.RLock() + e.LockUser() // The endpoint can be read if it's connected, or if it's already closed // but has some pending unread data. Also note that a RST being received // would cause the state to become StateError so we should allow the @@ -999,7 +1171,7 @@ func (e *endpoint) Read(*tcpip.FullAddress) (buffer.View, tcpip.ControlMessages, if s := e.EndpointState(); !s.connected() && s != StateClose && bufUsed == 0 { e.rcvListMu.Unlock() he := e.HardError - e.mu.RUnlock() + e.UnlockUser() if s == StateError { return buffer.View{}, tcpip.ControlMessages{}, he } @@ -1009,8 +1181,7 @@ func (e *endpoint) Read(*tcpip.FullAddress) (buffer.View, tcpip.ControlMessages, v, err := e.readLocked() e.rcvListMu.Unlock() - - e.mu.RUnlock() + e.UnlockUser() if err == tcpip.ErrClosedForReceive { e.stats.ReadErrors.ReadClosed.Increment() @@ -1042,7 +1213,7 @@ func (e *endpoint) readLocked() (buffer.View, *tcpip.Error) { // enough buffer space, to either fit an aMSS or half a receive buffer // (whichever smaller), then notify the protocol goroutine to send a // window update. - if crossed, above := e.windowCrossedACKThreshold(len(v)); crossed && above { + if crossed, above := e.windowCrossedACKThresholdLocked(len(v)); crossed && above { e.notifyProtocolGoroutine(notifyNonZeroReceiveWindow) } @@ -1083,13 +1254,13 @@ func (e *endpoint) Write(p tcpip.Payloader, opts tcpip.WriteOptions) (int64, <-c // (without the MSG_FASTOPEN flag). Corking is unimplemented, so opts.More // and opts.EndOfRecord are also ignored. - e.mu.RLock() + e.LockUser() e.sndBufMu.Lock() avail, err := e.isEndpointWritableLocked() if err != nil { e.sndBufMu.Unlock() - e.mu.RUnlock() + e.UnlockUser() e.stats.WriteErrors.WriteClosed.Increment() return 0, nil, err } @@ -1101,113 +1272,68 @@ func (e *endpoint) Write(p tcpip.Payloader, opts tcpip.WriteOptions) (int64, <-c // are copying data in. if !opts.Atomic { e.sndBufMu.Unlock() - e.mu.RUnlock() + e.UnlockUser() } // Fetch data. v, perr := p.Payload(avail) if perr != nil || len(v) == 0 { - if opts.Atomic { // See above. + // Note that perr may be nil if len(v) == 0. + if opts.Atomic { e.sndBufMu.Unlock() - e.mu.RUnlock() + e.UnlockUser() } - // Note that perr may be nil if len(v) == 0. return 0, nil, perr } - if opts.Atomic { + queueAndSend := func() (int64, <-chan struct{}, *tcpip.Error) { // Add data to the send queue. s := newSegmentFromView(&e.route, e.ID, v) e.sndBufUsed += len(v) e.sndBufInQueue += seqnum.Size(len(v)) e.sndQueue.PushBack(s) e.sndBufMu.Unlock() - // Release the endpoint lock to prevent deadlocks due to lock - // order inversion when acquiring workMu. - e.mu.RUnlock() - } - - if e.workMu.TryLock() { - // Since we released locks in between it's possible that the - // endpoint transitioned to a CLOSED/ERROR states so make - // sure endpoint is still writable before trying to write. - if !opts.Atomic { // See above. - e.mu.RLock() - e.sndBufMu.Lock() - - // Because we released the lock before copying, check state again - // to make sure the endpoint is still in a valid state for a write. - avail, err = e.isEndpointWritableLocked() - if err != nil { - e.sndBufMu.Unlock() - e.mu.RUnlock() - e.stats.WriteErrors.WriteClosed.Increment() - return 0, nil, err - } - // Discard any excess data copied in due to avail being reduced due - // to a simultaneous write call to the socket. - if avail < len(v) { - v = v[:avail] - } - // Add data to the send queue. - s := newSegmentFromView(&e.route, e.ID, v) - e.sndBufUsed += len(v) - e.sndBufInQueue += seqnum.Size(len(v)) - e.sndQueue.PushBack(s) - e.sndBufMu.Unlock() - // Release the endpoint lock to prevent deadlocks due to lock - // order inversion when acquiring workMu. - e.mu.RUnlock() - - } // Do the work inline. e.handleWrite() - e.workMu.Unlock() - } else { - if !opts.Atomic { // See above. - e.mu.RLock() - e.sndBufMu.Lock() + e.UnlockUser() + return int64(len(v)), nil, nil + } - // Because we released the lock before copying, check state again - // to make sure the endpoint is still in a valid state for a write. - avail, err = e.isEndpointWritableLocked() - if err != nil { - e.sndBufMu.Unlock() - e.mu.RUnlock() - e.stats.WriteErrors.WriteClosed.Increment() - return 0, nil, err - } + if opts.Atomic { + // Locks released in queueAndSend() + return queueAndSend() + } - // Discard any excess data copied in due to avail being reduced due - // to a simultaneous write call to the socket. - if avail < len(v) { - v = v[:avail] - } - // Add data to the send queue. - s := newSegmentFromView(&e.route, e.ID, v) - e.sndBufUsed += len(v) - e.sndBufInQueue += seqnum.Size(len(v)) - e.sndQueue.PushBack(s) - e.sndBufMu.Unlock() - // Release the endpoint lock to prevent deadlocks due to lock - // order inversion when acquiring workMu. - e.mu.RUnlock() + // Since we released locks in between it's possible that the + // endpoint transitioned to a CLOSED/ERROR states so make + // sure endpoint is still writable before trying to write. + e.LockUser() + e.sndBufMu.Lock() + avail, err = e.isEndpointWritableLocked() + if err != nil { + e.sndBufMu.Unlock() + e.UnlockUser() + e.stats.WriteErrors.WriteClosed.Increment() + return 0, nil, err + } - } - // Let the protocol goroutine do the work. - e.sndWaker.Assert() + // Discard any excess data copied in due to avail being reduced due + // to a simultaneous write call to the socket. + if avail < len(v) { + v = v[:avail] } - return int64(len(v)), nil, nil + // Locks released in queueAndSend() + return queueAndSend() } // Peek reads data without consuming it from the endpoint. // // This method does not block if there is no data pending. func (e *endpoint) Peek(vec [][]byte) (int64, tcpip.ControlMessages, *tcpip.Error) { - e.mu.RLock() - defer e.mu.RUnlock() + e.LockUser() + defer e.UnlockUser() // The endpoint can be read if it's connected, or if it's already closed // but has some pending unread data. @@ -1260,9 +1386,9 @@ func (e *endpoint) Peek(vec [][]byte) (int64, tcpip.ControlMessages, *tcpip.Erro return num, tcpip.ControlMessages{}, nil } -// windowCrossedACKThreshold checks if the receive window to be announced now -// would be under aMSS or under half receive buffer, whichever smaller. This is -// useful as a receive side silly window syndrome prevention mechanism. If +// windowCrossedACKThresholdLocked checks if the receive window to be announced +// now would be under aMSS or under half receive buffer, whichever smaller. This +// is useful as a receive side silly window syndrome prevention mechanism. If // window grows to reasonable value, we should send ACK to the sender to inform // the rx space is now large. We also want ensure a series of small read()'s // won't trigger a flood of spurious tiny ACK's. @@ -1273,7 +1399,9 @@ func (e *endpoint) Peek(vec [][]byte) (int64, tcpip.ControlMessages, *tcpip.Erro // crossed will be true if the window size crossed the ACK threshold. // above will be true if the new window is >= ACK threshold and false // otherwise. -func (e *endpoint) windowCrossedACKThreshold(deltaBefore int) (crossed bool, above bool) { +// +// Precondition: e.mu and e.rcvListMu must be held. +func (e *endpoint) windowCrossedACKThresholdLocked(deltaBefore int) (crossed bool, above bool) { newAvail := e.receiveBufferAvailableLocked() oldAvail := newAvail - deltaBefore if oldAvail < 0 { @@ -1297,21 +1425,71 @@ func (e *endpoint) windowCrossedACKThreshold(deltaBefore int) (crossed bool, abo // SetSockOptBool sets a socket option. func (e *endpoint) SetSockOptBool(opt tcpip.SockOptBool, v bool) *tcpip.Error { switch opt { + + case tcpip.BroadcastOption: + e.LockUser() + e.broadcast = v + e.UnlockUser() + + case tcpip.CorkOption: + e.LockUser() + if !v { + atomic.StoreUint32(&e.cork, 0) + + // Handle the corked data. + e.sndWaker.Assert() + } else { + atomic.StoreUint32(&e.cork, 1) + } + e.UnlockUser() + + case tcpip.DelayOption: + if v { + atomic.StoreUint32(&e.delay, 1) + } else { + atomic.StoreUint32(&e.delay, 0) + + // Handle delayed data. + e.sndWaker.Assert() + } + + case tcpip.KeepaliveEnabledOption: + e.keepalive.Lock() + e.keepalive.enabled = v + e.keepalive.Unlock() + e.notifyProtocolGoroutine(notifyKeepaliveChanged) + + case tcpip.QuickAckOption: + o := uint32(1) + if v { + o = 0 + } + atomic.StoreUint32(&e.slowAck, o) + + case tcpip.ReuseAddressOption: + e.LockUser() + e.reuseAddr = v + e.UnlockUser() + + case tcpip.ReusePortOption: + e.LockUser() + e.reusePort = v + e.UnlockUser() + case tcpip.V6OnlyOption: // We only recognize this option on v6 endpoints. if e.NetProto != header.IPv6ProtocolNumber { return tcpip.ErrInvalidEndpointState } - e.mu.Lock() - defer e.mu.Unlock() - // We only allow this to be set when we're in the initial state. if e.EndpointState() != StateInitial { return tcpip.ErrInvalidEndpointState } + e.LockUser() e.v6only = v + e.UnlockUser() } return nil @@ -1319,23 +1497,56 @@ func (e *endpoint) SetSockOptBool(opt tcpip.SockOptBool, v bool) *tcpip.Error { // SetSockOptInt sets a socket option. func (e *endpoint) SetSockOptInt(opt tcpip.SockOptInt, v int) *tcpip.Error { + // Lower 2 bits represents ECN bits. RFC 3168, section 23.1 + const inetECNMask = 3 + switch opt { + case tcpip.KeepaliveCountOption: + e.keepalive.Lock() + e.keepalive.count = v + e.keepalive.Unlock() + e.notifyProtocolGoroutine(notifyKeepaliveChanged) + + case tcpip.IPv4TOSOption: + e.LockUser() + // TODO(gvisor.dev/issue/995): ECN is not currently supported, + // ignore the bits for now. + e.sendTOS = uint8(v) & ^uint8(inetECNMask) + e.UnlockUser() + + case tcpip.IPv6TrafficClassOption: + e.LockUser() + // TODO(gvisor.dev/issue/995): ECN is not currently supported, + // ignore the bits for now. + e.sendTOS = uint8(v) & ^uint8(inetECNMask) + e.UnlockUser() + + case tcpip.MaxSegOption: + userMSS := v + if userMSS < header.TCPMinimumMSS || userMSS > header.TCPMaximumMSS { + return tcpip.ErrInvalidOptionValue + } + e.LockUser() + e.userMSS = uint16(userMSS) + e.UnlockUser() + e.notifyProtocolGoroutine(notifyMSSChanged) + case tcpip.ReceiveBufferSizeOption: // Make sure the receive buffer size is within the min and max // allowed. var rs ReceiveBufferSizeOption - size := int(v) if err := e.stack.TransportProtocolOption(ProtocolNumber, &rs); err == nil { - if size < rs.Min { - size = rs.Min + if v < rs.Min { + v = rs.Min } - if size > rs.Max { - size = rs.Max + if v > rs.Max { + v = rs.Max } } mask := uint32(notifyReceiveWindowChanged) + e.LockUser() e.rcvListMu.Lock() // Make sure the receive buffer size allows us to send a @@ -1344,17 +1555,17 @@ func (e *endpoint) SetSockOptInt(opt tcpip.SockOptInt, v int) *tcpip.Error { if e.rcv != nil { scale = e.rcv.rcvWndScale } - if size>>scale == 0 { - size = 1 << scale + if v>>scale == 0 { + v = 1 << scale } // Make sure 2*size doesn't overflow. - if size > math.MaxInt32/2 { - size = math.MaxInt32 / 2 + if v > math.MaxInt32/2 { + v = math.MaxInt32 / 2 } availBefore := e.receiveBufferAvailableLocked() - e.rcvBufSize = size + e.rcvBufSize = v availAfter := e.receiveBufferAvailableLocked() e.rcvAutoParams.disabled = true @@ -1362,151 +1573,71 @@ func (e *endpoint) SetSockOptInt(opt tcpip.SockOptInt, v int) *tcpip.Error { // Immediately send an ACK to uncork the sender silly window // syndrome prevetion, when our available space grows above aMSS // or half receive buffer, whichever smaller. - if crossed, above := e.windowCrossedACKThreshold(availAfter - availBefore); crossed && above { + if crossed, above := e.windowCrossedACKThresholdLocked(availAfter - availBefore); crossed && above { mask |= notifyNonZeroReceiveWindow } - e.rcvListMu.Unlock() + e.rcvListMu.Unlock() + e.UnlockUser() e.notifyProtocolGoroutine(mask) - return nil case tcpip.SendBufferSizeOption: // Make sure the send buffer size is within the min and max // allowed. - size := int(v) var ss SendBufferSizeOption if err := e.stack.TransportProtocolOption(ProtocolNumber, &ss); err == nil { - if size < ss.Min { - size = ss.Min + if v < ss.Min { + v = ss.Min } - if size > ss.Max { - size = ss.Max + if v > ss.Max { + v = ss.Max } } e.sndBufMu.Lock() - e.sndBufSize = size + e.sndBufSize = v e.sndBufMu.Unlock() - return nil - case tcpip.DelayOption: - if v == 0 { - atomic.StoreUint32(&e.delay, 0) - - // Handle delayed data. - e.sndWaker.Assert() - } else { - atomic.StoreUint32(&e.delay, 1) - } - return nil + case tcpip.TTLOption: + e.LockUser() + e.ttl = uint8(v) + e.UnlockUser() - default: - return nil } + return nil } // SetSockOpt sets a socket option. func (e *endpoint) SetSockOpt(opt interface{}) *tcpip.Error { - // Lower 2 bits represents ECN bits. RFC 3168, section 23.1 - const inetECNMask = 3 switch v := opt.(type) { - case tcpip.CorkOption: - if v == 0 { - atomic.StoreUint32(&e.cork, 0) - - // Handle the corked data. - e.sndWaker.Assert() - } else { - atomic.StoreUint32(&e.cork, 1) - } - return nil - - case tcpip.ReuseAddressOption: - e.mu.Lock() - e.reuseAddr = v != 0 - e.mu.Unlock() - return nil - - case tcpip.ReusePortOption: - e.mu.Lock() - e.reusePort = v != 0 - e.mu.Unlock() - return nil - case tcpip.BindToDeviceOption: id := tcpip.NICID(v) if id != 0 && !e.stack.HasNIC(id) { return tcpip.ErrUnknownDevice } - e.mu.Lock() + e.LockUser() e.bindToDevice = id - e.mu.Unlock() - return nil - - case tcpip.QuickAckOption: - if v == 0 { - atomic.StoreUint32(&e.slowAck, 1) - } else { - atomic.StoreUint32(&e.slowAck, 0) - } - return nil - - case tcpip.MaxSegOption: - userMSS := v - if userMSS < header.TCPMinimumMSS || userMSS > header.TCPMaximumMSS { - return tcpip.ErrInvalidOptionValue - } - e.mu.Lock() - e.userMSS = uint16(userMSS) - e.mu.Unlock() - e.notifyProtocolGoroutine(notifyMSSChanged) - return nil - - case tcpip.TTLOption: - e.mu.Lock() - e.ttl = uint8(v) - e.mu.Unlock() - return nil - - case tcpip.KeepaliveEnabledOption: - e.keepalive.Lock() - e.keepalive.enabled = v != 0 - e.keepalive.Unlock() - e.notifyProtocolGoroutine(notifyKeepaliveChanged) - return nil + e.UnlockUser() case tcpip.KeepaliveIdleOption: e.keepalive.Lock() e.keepalive.idle = time.Duration(v) e.keepalive.Unlock() e.notifyProtocolGoroutine(notifyKeepaliveChanged) - return nil case tcpip.KeepaliveIntervalOption: e.keepalive.Lock() e.keepalive.interval = time.Duration(v) e.keepalive.Unlock() e.notifyProtocolGoroutine(notifyKeepaliveChanged) - return nil - case tcpip.KeepaliveCountOption: - e.keepalive.Lock() - e.keepalive.count = int(v) - e.keepalive.Unlock() - e.notifyProtocolGoroutine(notifyKeepaliveChanged) - return nil + case tcpip.OutOfBandInlineOption: + // We don't currently support disabling this option. case tcpip.TCPUserTimeoutOption: - e.mu.Lock() + e.LockUser() e.userTimeout = time.Duration(v) - e.mu.Unlock() - return nil - - case tcpip.BroadcastOption: - e.mu.Lock() - e.broadcast = v != 0 - e.mu.Unlock() - return nil + e.UnlockUser() case tcpip.CongestionControlOption: // Query the available cc algorithms in the stack and @@ -1519,22 +1650,16 @@ func (e *endpoint) SetSockOpt(opt interface{}) *tcpip.Error { availCC := strings.Split(string(avail), " ") for _, cc := range availCC { if v == tcpip.CongestionControlOption(cc) { - // Acquire the work mutex as we may need to - // reinitialize the congestion control state. - e.mu.Lock() + e.LockUser() state := e.EndpointState() e.cc = v - e.mu.Unlock() switch state { case StateEstablished: - e.workMu.Lock() - e.mu.Lock() if e.EndpointState() == state { e.snd.cc = e.snd.initCongestionControl(e.cc) } - e.mu.Unlock() - e.workMu.Unlock() } + e.UnlockUser() return nil } } @@ -1543,24 +1668,8 @@ func (e *endpoint) SetSockOpt(opt interface{}) *tcpip.Error { // control algorithm is specified. return tcpip.ErrNoSuchFile - case tcpip.IPv4TOSOption: - e.mu.Lock() - // TODO(gvisor.dev/issue/995): ECN is not currently supported, - // ignore the bits for now. - e.sendTOS = uint8(v) & ^uint8(inetECNMask) - e.mu.Unlock() - return nil - - case tcpip.IPv6TrafficClassOption: - e.mu.Lock() - // TODO(gvisor.dev/issue/995): ECN is not currently supported, - // ignore the bits for now. - e.sendTOS = uint8(v) & ^uint8(inetECNMask) - e.mu.Unlock() - return nil - case tcpip.TCPLingerTimeoutOption: - e.mu.Lock() + e.LockUser() if v < 0 { // Same as effectively disabling TCPLinger timeout. v = 0 @@ -1578,27 +1687,26 @@ func (e *endpoint) SetSockOpt(opt interface{}) *tcpip.Error { v = stkTCPLingerTimeout } e.tcpLingerTimeout = time.Duration(v) - e.mu.Unlock() - return nil + e.UnlockUser() case tcpip.TCPDeferAcceptOption: - e.mu.Lock() + e.LockUser() if time.Duration(v) > MaxRTO { v = tcpip.TCPDeferAcceptOption(MaxRTO) } e.deferAccept = time.Duration(v) - e.mu.Unlock() - return nil + e.UnlockUser() default: return nil } + return nil } // readyReceiveSize returns the number of bytes ready to be received. func (e *endpoint) readyReceiveSize() (int, *tcpip.Error) { - e.mu.RLock() - defer e.mu.RUnlock() + e.LockUser() + defer e.UnlockUser() // The endpoint cannot be in listen state. if e.EndpointState() == StateListen { @@ -1614,25 +1722,89 @@ func (e *endpoint) readyReceiveSize() (int, *tcpip.Error) { // GetSockOptBool implements tcpip.Endpoint.GetSockOptBool. func (e *endpoint) GetSockOptBool(opt tcpip.SockOptBool) (bool, *tcpip.Error) { switch opt { + case tcpip.BroadcastOption: + e.LockUser() + v := e.broadcast + e.UnlockUser() + return v, nil + + case tcpip.CorkOption: + return atomic.LoadUint32(&e.cork) != 0, nil + + case tcpip.DelayOption: + return atomic.LoadUint32(&e.delay) != 0, nil + + case tcpip.KeepaliveEnabledOption: + e.keepalive.Lock() + v := e.keepalive.enabled + e.keepalive.Unlock() + + return v, nil + + case tcpip.QuickAckOption: + v := atomic.LoadUint32(&e.slowAck) == 0 + return v, nil + + case tcpip.ReuseAddressOption: + e.LockUser() + v := e.reuseAddr + e.UnlockUser() + + return v, nil + + case tcpip.ReusePortOption: + e.LockUser() + v := e.reusePort + e.UnlockUser() + + return v, nil + case tcpip.V6OnlyOption: // We only recognize this option on v6 endpoints. if e.NetProto != header.IPv6ProtocolNumber { return false, tcpip.ErrUnknownProtocolOption } - e.mu.Lock() + e.LockUser() v := e.v6only - e.mu.Unlock() + e.UnlockUser() return v, nil - } - return false, tcpip.ErrUnknownProtocolOption + default: + return false, tcpip.ErrUnknownProtocolOption + } } // GetSockOptInt implements tcpip.Endpoint.GetSockOptInt. func (e *endpoint) GetSockOptInt(opt tcpip.SockOptInt) (int, *tcpip.Error) { switch opt { + case tcpip.KeepaliveCountOption: + e.keepalive.Lock() + v := e.keepalive.count + e.keepalive.Unlock() + return v, nil + + case tcpip.IPv4TOSOption: + e.LockUser() + v := int(e.sendTOS) + e.UnlockUser() + return v, nil + + case tcpip.IPv6TrafficClassOption: + e.LockUser() + v := int(e.sendTOS) + e.UnlockUser() + return v, nil + + case tcpip.MaxSegOption: + // This is just stubbed out. Linux never returns the user_mss + // value as it either returns the defaultMSS or returns the + // actual current MSS. Netstack just returns the defaultMSS + // always for now. + v := header.TCPDefaultMSS + return v, nil + case tcpip.ReceiveQueueSizeOption: return e.readyReceiveSize() @@ -1648,12 +1820,11 @@ func (e *endpoint) GetSockOptInt(opt tcpip.SockOptInt) (int, *tcpip.Error) { e.rcvListMu.Unlock() return v, nil - case tcpip.DelayOption: - var o int - if v := atomic.LoadUint32(&e.delay); v != 0 { - o = 1 - } - return o, nil + case tcpip.TTLOption: + e.LockUser() + v := int(e.ttl) + e.UnlockUser() + return v, nil default: return -1, tcpip.ErrUnknownProtocolOption @@ -1670,168 +1841,71 @@ func (e *endpoint) GetSockOpt(opt interface{}) *tcpip.Error { e.lastErrorMu.Unlock() return err - case *tcpip.MaxSegOption: - // This is just stubbed out. Linux never returns the user_mss - // value as it either returns the defaultMSS or returns the - // actual current MSS. Netstack just returns the defaultMSS - // always for now. - *o = header.TCPDefaultMSS - return nil - - case *tcpip.CorkOption: - *o = 0 - if v := atomic.LoadUint32(&e.cork); v != 0 { - *o = 1 - } - return nil - - case *tcpip.ReuseAddressOption: - e.mu.RLock() - v := e.reuseAddr - e.mu.RUnlock() - - *o = 0 - if v { - *o = 1 - } - return nil - - case *tcpip.ReusePortOption: - e.mu.RLock() - v := e.reusePort - e.mu.RUnlock() - - *o = 0 - if v { - *o = 1 - } - return nil - case *tcpip.BindToDeviceOption: - e.mu.RLock() + e.LockUser() *o = tcpip.BindToDeviceOption(e.bindToDevice) - e.mu.RUnlock() - return nil - - case *tcpip.QuickAckOption: - *o = 1 - if v := atomic.LoadUint32(&e.slowAck); v != 0 { - *o = 0 - } - return nil - - case *tcpip.TTLOption: - e.mu.Lock() - *o = tcpip.TTLOption(e.ttl) - e.mu.Unlock() - return nil + e.UnlockUser() case *tcpip.TCPInfoOption: *o = tcpip.TCPInfoOption{} - e.mu.RLock() + e.LockUser() snd := e.snd - e.mu.RUnlock() + e.UnlockUser() if snd != nil { snd.rtt.Lock() o.RTT = snd.rtt.srtt o.RTTVar = snd.rtt.rttvar snd.rtt.Unlock() } - return nil - - case *tcpip.KeepaliveEnabledOption: - e.keepalive.Lock() - v := e.keepalive.enabled - e.keepalive.Unlock() - - *o = 0 - if v { - *o = 1 - } - return nil case *tcpip.KeepaliveIdleOption: e.keepalive.Lock() *o = tcpip.KeepaliveIdleOption(e.keepalive.idle) e.keepalive.Unlock() - return nil case *tcpip.KeepaliveIntervalOption: e.keepalive.Lock() *o = tcpip.KeepaliveIntervalOption(e.keepalive.interval) e.keepalive.Unlock() - return nil - - case *tcpip.KeepaliveCountOption: - e.keepalive.Lock() - *o = tcpip.KeepaliveCountOption(e.keepalive.count) - e.keepalive.Unlock() - return nil case *tcpip.TCPUserTimeoutOption: - e.mu.Lock() + e.LockUser() *o = tcpip.TCPUserTimeoutOption(e.userTimeout) - e.mu.Unlock() - return nil + e.UnlockUser() case *tcpip.OutOfBandInlineOption: // We don't currently support disabling this option. *o = 1 - return nil - - case *tcpip.BroadcastOption: - e.mu.Lock() - v := e.broadcast - e.mu.Unlock() - - *o = 0 - if v { - *o = 1 - } - return nil case *tcpip.CongestionControlOption: - e.mu.Lock() + e.LockUser() *o = e.cc - e.mu.Unlock() - return nil - - case *tcpip.IPv4TOSOption: - e.mu.RLock() - *o = tcpip.IPv4TOSOption(e.sendTOS) - e.mu.RUnlock() - return nil - - case *tcpip.IPv6TrafficClassOption: - e.mu.RLock() - *o = tcpip.IPv6TrafficClassOption(e.sendTOS) - e.mu.RUnlock() - return nil + e.UnlockUser() case *tcpip.TCPLingerTimeoutOption: - e.mu.Lock() + e.LockUser() *o = tcpip.TCPLingerTimeoutOption(e.tcpLingerTimeout) - e.mu.Unlock() - return nil + e.UnlockUser() case *tcpip.TCPDeferAcceptOption: - e.mu.Lock() + e.LockUser() *o = tcpip.TCPDeferAcceptOption(e.deferAccept) - e.mu.Unlock() - return nil + e.UnlockUser() default: return tcpip.ErrUnknownProtocolOption } + return nil } -func (e *endpoint) checkV4Mapped(addr *tcpip.FullAddress) (tcpip.NetworkProtocolNumber, *tcpip.Error) { - unwrapped, netProto, err := e.TransportEndpointInfo.AddrNetProto(*addr, e.v6only) +// checkV4MappedLocked determines the effective network protocol and converts +// addr to its canonical form. +func (e *endpoint) checkV4MappedLocked(addr tcpip.FullAddress) (tcpip.FullAddress, tcpip.NetworkProtocolNumber, *tcpip.Error) { + unwrapped, netProto, err := e.TransportEndpointInfo.AddrNetProtoLocked(addr, e.v6only) if err != nil { - return 0, err + return tcpip.FullAddress{}, 0, err } - *addr = unwrapped - return netProto, nil + return unwrapped, netProto, nil } // Disconnect implements tcpip.Endpoint.Disconnect. @@ -1856,12 +1930,12 @@ func (e *endpoint) Connect(addr tcpip.FullAddress) *tcpip.Error { // yet accepted by the app, they are restored without running the main goroutine // here. func (e *endpoint) connect(addr tcpip.FullAddress, handshake bool, run bool) *tcpip.Error { - e.mu.Lock() - defer e.mu.Unlock() + e.LockUser() + defer e.UnlockUser() connectingAddr := addr.Addr - netProto, err := e.checkV4Mapped(&addr) + addr, netProto, err := e.checkV4MappedLocked(addr) if err != nil { return err } @@ -2026,13 +2100,17 @@ func (*endpoint) ConnectEndpoint(tcpip.Endpoint) *tcpip.Error { // Shutdown closes the read and/or write end of the endpoint connection to its // peer. func (e *endpoint) Shutdown(flags tcpip.ShutdownFlags) *tcpip.Error { - e.mu.Lock() + e.LockUser() + defer e.UnlockUser() + return e.shutdownLocked(flags) +} + +func (e *endpoint) shutdownLocked(flags tcpip.ShutdownFlags) *tcpip.Error { e.shutdownFlags |= flags - finQueued := false switch { case e.EndpointState().connected(): // Close for read. - if (e.shutdownFlags & tcpip.ShutdownRead) != 0 { + if e.shutdownFlags&tcpip.ShutdownRead != 0 { // Mark read side as closed. e.rcvListMu.Lock() e.rcvClosed = true @@ -2041,69 +2119,58 @@ func (e *endpoint) Shutdown(flags tcpip.ShutdownFlags) *tcpip.Error { // If we're fully closed and we have unread data we need to abort // the connection with a RST. - if (e.shutdownFlags&tcpip.ShutdownWrite) != 0 && rcvBufUsed > 0 { - e.mu.Unlock() - // Try to send an active reset immediately if the - // work mutex is available. - if e.workMu.TryLock() { - e.mu.Lock() - // We need to double check here to make - // sure worker has not transitioned the - // endpoint out of a connected state - // before trying to send a reset. - if e.EndpointState().connected() { - e.resetConnectionLocked(tcpip.ErrConnectionAborted) - e.notifyProtocolGoroutine(notifyTickleWorker) - } - e.mu.Unlock() - e.workMu.Unlock() - } else { - e.notifyProtocolGoroutine(notifyReset) - } + if e.shutdownFlags&tcpip.ShutdownWrite != 0 && rcvBufUsed > 0 { + e.resetConnectionLocked(tcpip.ErrConnectionAborted) + // Wake up worker to terminate loop. + e.notifyProtocolGoroutine(notifyTickleWorker) return nil } } // Close for write. - if (e.shutdownFlags & tcpip.ShutdownWrite) != 0 { + if e.shutdownFlags&tcpip.ShutdownWrite != 0 { e.sndBufMu.Lock() - if e.sndClosed { // Already closed. e.sndBufMu.Unlock() - break + if e.EndpointState() == StateTimeWait { + return tcpip.ErrNotConnected + } + return nil } // Queue fin segment. s := newSegmentFromView(&e.route, e.ID, nil) e.sndQueue.PushBack(s) e.sndBufInQueue++ - finQueued = true // Mark endpoint as closed. e.sndClosed = true e.sndBufMu.Unlock() + e.handleClose() } + return nil case e.EndpointState() == StateListen: - // Tell protocolListenLoop to stop. - if flags&tcpip.ShutdownRead != 0 { - e.notifyProtocolGoroutine(notifyClose) + if e.shutdownFlags&tcpip.ShutdownRead != 0 { + // Reset all connections from the accept queue and keep the + // worker running so that it can continue handling incoming + // segments by replying with RST. + // + // By not removing this endpoint from the demuxer mapping, we + // ensure that any other bind to the same port fails, as on Linux. + // TODO(gvisor.dev/issue/2468): We need to enable applications to + // start listening on this endpoint again similar to Linux. + e.rcvListMu.Lock() + e.rcvClosed = true + e.rcvListMu.Unlock() + e.closePendingAcceptableConnectionsLocked() + // Notify waiters that the endpoint is shutdown. + e.waiterQueue.Notify(waiter.EventIn | waiter.EventOut | waiter.EventHUp | waiter.EventErr) } + return nil default: - e.mu.Unlock() return tcpip.ErrNotConnected } - e.mu.Unlock() - if finQueued { - if e.workMu.TryLock() { - e.handleClose() - e.workMu.Unlock() - } else { - // Tell protocol goroutine to close. - e.sndCloseWaker.Assert() - } - } - return nil } // Listen puts the endpoint in "listen" mode, which allows it to accept @@ -2118,8 +2185,8 @@ func (e *endpoint) Listen(backlog int) *tcpip.Error { } func (e *endpoint) listen(backlog int) *tcpip.Error { - e.mu.Lock() - defer e.mu.Unlock() + e.LockUser() + defer e.UnlockUser() // Allow the backlog to be adjusted if the endpoint is not shutting down. // When the endpoint shuts down, it sets workerCleanup to true, and from @@ -2128,6 +2195,8 @@ func (e *endpoint) listen(backlog int) *tcpip.Error { if e.EndpointState() == StateListen && !e.workerCleanup { // Adjust the size of the channel iff we can fix existing // pending connections into the new one. + e.acceptMu.Lock() + defer e.acceptMu.Unlock() if len(e.acceptedChan) > backlog { return tcpip.ErrInvalidEndpointState } @@ -2140,6 +2209,11 @@ func (e *endpoint) listen(backlog int) *tcpip.Error { for ep := range origChan { e.acceptedChan <- ep } + + // Notify any blocked goroutines that they can attempt to + // deliver endpoints again. + e.acceptCond.Broadcast() + return nil } @@ -2169,9 +2243,12 @@ func (e *endpoint) listen(backlog int) *tcpip.Error { // The channel may be non-nil when we're restoring the endpoint, and it // may be pre-populated with some previously accepted (but not Accepted) // endpoints. + e.acceptMu.Lock() if e.acceptedChan == nil { e.acceptedChan = make(chan *endpoint, backlog) } + e.acceptMu.Unlock() + e.workerRunning = true go e.protocolListenLoop( // S/R-SAFE: drained on save. seqnum.Size(e.receiveBufferAvailable())) @@ -2181,7 +2258,6 @@ func (e *endpoint) listen(backlog int) *tcpip.Error { // startAcceptedLoop sets up required state and starts a goroutine with the // main loop for accepted connections. func (e *endpoint) startAcceptedLoop() { - e.mu.Lock() e.workerRunning = true e.mu.Unlock() wakerInitDone := make(chan struct{}) @@ -2192,18 +2268,24 @@ func (e *endpoint) startAcceptedLoop() { // Accept returns a new endpoint if a peer has established a connection // to an endpoint previously set to listen mode. func (e *endpoint) Accept() (tcpip.Endpoint, *waiter.Queue, *tcpip.Error) { - e.mu.RLock() - defer e.mu.RUnlock() + e.LockUser() + defer e.UnlockUser() + e.rcvListMu.Lock() + rcvClosed := e.rcvClosed + e.rcvListMu.Unlock() // Endpoint must be in listen state before it can accept connections. - if e.EndpointState() != StateListen { + if rcvClosed || e.EndpointState() != StateListen { return nil, nil, tcpip.ErrInvalidEndpointState } // Get the new accepted endpoint. + e.acceptMu.Lock() + defer e.acceptMu.Unlock() var n *endpoint select { case n = <-e.acceptedChan: + e.acceptCond.Signal() default: return nil, nil, tcpip.ErrWouldBlock } @@ -2212,8 +2294,8 @@ func (e *endpoint) Accept() (tcpip.Endpoint, *waiter.Queue, *tcpip.Error) { // Bind binds the endpoint to a specific local port and optionally address. func (e *endpoint) Bind(addr tcpip.FullAddress) (err *tcpip.Error) { - e.mu.Lock() - defer e.mu.Unlock() + e.LockUser() + defer e.UnlockUser() return e.bindLocked(addr) } @@ -2227,7 +2309,7 @@ func (e *endpoint) bindLocked(addr tcpip.FullAddress) (err *tcpip.Error) { } e.BindAddr = addr.Addr - netProto, err := e.checkV4Mapped(&addr) + addr, netProto, err := e.checkV4MappedLocked(addr) if err != nil { return err } @@ -2291,8 +2373,8 @@ func (e *endpoint) bindLocked(addr tcpip.FullAddress) (err *tcpip.Error) { // GetLocalAddress returns the address to which the endpoint is bound. func (e *endpoint) GetLocalAddress() (tcpip.FullAddress, *tcpip.Error) { - e.mu.RLock() - defer e.mu.RUnlock() + e.LockUser() + defer e.UnlockUser() return tcpip.FullAddress{ Addr: e.ID.LocalAddress, @@ -2303,8 +2385,8 @@ func (e *endpoint) GetLocalAddress() (tcpip.FullAddress, *tcpip.Error) { // GetRemoteAddress returns the address to which the endpoint is connected. func (e *endpoint) GetRemoteAddress() (tcpip.FullAddress, *tcpip.Error) { - e.mu.RLock() - defer e.mu.RUnlock() + e.LockUser() + defer e.UnlockUser() if !e.EndpointState().connected() { return tcpip.FullAddress{}, tcpip.ErrNotConnected @@ -2317,7 +2399,7 @@ func (e *endpoint) GetRemoteAddress() (tcpip.FullAddress, *tcpip.Error) { }, nil } -func (e *endpoint) HandlePacket(r *stack.Route, id stack.TransportEndpointID, pkt tcpip.PacketBuffer) { +func (e *endpoint) HandlePacket(r *stack.Route, id stack.TransportEndpointID, pkt stack.PacketBuffer) { // TCP HandlePacket is not required anymore as inbound packets first // land at the Dispatcher which then can either delivery using the // worker go routine or directly do the invoke the tcp processing inline @@ -2336,7 +2418,7 @@ func (e *endpoint) enqueueSegment(s *segment) bool { } // HandleControlPacket implements stack.TransportEndpoint.HandleControlPacket. -func (e *endpoint) HandleControlPacket(id stack.TransportEndpointID, typ stack.ControlType, extra uint32, pkt tcpip.PacketBuffer) { +func (e *endpoint) HandleControlPacket(id stack.TransportEndpointID, typ stack.ControlType, extra uint32, pkt stack.PacketBuffer) { switch typ { case stack.ControlPacketTooBig: e.sndBufMu.Lock() @@ -2377,7 +2459,7 @@ func (e *endpoint) readyToRead(s *segment) { e.rcvBufUsed += s.data.Size() // Increase counter if the receive window falls down below MSS // or half receive buffer size, whichever smaller. - if crossed, above := e.windowCrossedACKThreshold(-s.data.Size()); crossed && !above { + if crossed, above := e.windowCrossedACKThresholdLocked(-s.data.Size()); crossed && !above { e.stats.ReceiveErrors.ZeroRcvWindowState.Increment() } e.rcvList.PushBack(s) @@ -2385,7 +2467,6 @@ func (e *endpoint) readyToRead(s *segment) { e.rcvClosed = true } e.rcvListMu.Unlock() - e.waiterQueue.Notify(waiter.EventIn) } @@ -2529,9 +2610,7 @@ func (e *endpoint) completeState() stack.TCPEndpointState { s.SegTime = time.Now() // Copy EndpointID. - e.mu.Lock() s.ID = stack.TCPEndpointID(e.ID) - e.mu.Unlock() // Copy endpoint rcv state. e.rcvListMu.Lock() @@ -2661,10 +2740,10 @@ func (e *endpoint) State() uint32 { // Info returns a copy of the endpoint info. func (e *endpoint) Info() tcpip.EndpointInfo { - e.mu.RLock() + e.LockUser() // Make a copy of the endpoint info. ret := e.EndpointInfo - e.mu.RUnlock() + e.UnlockUser() return &ret } @@ -2679,9 +2758,9 @@ func (e *endpoint) Wait() { e.waiterQueue.EventRegister(&waitEntry, waiter.EventHUp) defer e.waiterQueue.EventUnregister(&waitEntry) for { - e.mu.Lock() + e.LockUser() running := e.workerRunning - e.mu.Unlock() + e.UnlockUser() if !running { break } diff --git a/pkg/tcpip/transport/tcp/endpoint_state.go b/pkg/tcpip/transport/tcp/endpoint_state.go index 4a46f0ec5..c3c692555 100644 --- a/pkg/tcpip/transport/tcp/endpoint_state.go +++ b/pkg/tcpip/transport/tcp/endpoint_state.go @@ -162,8 +162,8 @@ func (e *endpoint) loadState(state EndpointState) { connectingLoading.Add(1) } // Directly update the state here rather than using e.setEndpointState - // as the endpoint is still being loaded and the stack reference to increment - // metrics is not yet initialized. + // as the endpoint is still being loaded and the stack reference is not + // yet initialized. atomic.StoreUint32((*uint32)(&e.state), uint32(state)) } @@ -173,6 +173,9 @@ func (e *endpoint) afterLoad() { // Restore the endpoint to InitialState as it will be moved to // its origEndpointState during Resume. e.state = StateInitial + // Condition variables and mutexs are not S/R'ed so reinitialize + // acceptCond with e.acceptMu. + e.acceptCond = sync.NewCond(&e.acceptMu) stack.StackFromEnv.RegisterRestoredEndpoint(e) } @@ -180,7 +183,6 @@ func (e *endpoint) afterLoad() { func (e *endpoint) Resume(s *stack.Stack) { e.stack = s e.segmentQueue.setLimit(MaxUnprocessedSegments) - e.workMu.Init() state := e.origEndpointState switch state { case StateInitial, StateBound, StateListen, StateConnecting, StateEstablished: diff --git a/pkg/tcpip/transport/tcp/forwarder.go b/pkg/tcpip/transport/tcp/forwarder.go index c9ee5bf06..704d01c64 100644 --- a/pkg/tcpip/transport/tcp/forwarder.go +++ b/pkg/tcpip/transport/tcp/forwarder.go @@ -61,7 +61,7 @@ func NewForwarder(s *stack.Stack, rcvWnd, maxInFlight int, handler func(*Forward // // This function is expected to be passed as an argument to the // stack.SetTransportProtocolHandler function. -func (f *Forwarder) HandlePacket(r *stack.Route, id stack.TransportEndpointID, pkt tcpip.PacketBuffer) bool { +func (f *Forwarder) HandlePacket(r *stack.Route, id stack.TransportEndpointID, pkt stack.PacketBuffer) bool { s := newSegment(r, id, pkt) defer s.decRef() @@ -130,7 +130,7 @@ func (r *ForwarderRequest) Complete(sendReset bool) { // If the caller requested, send a reset. if sendReset { - replyWithReset(r.segment) + replyWithReset(r.segment, stack.DefaultTOS, r.segment.route.DefaultTTL()) } // Release all resources. @@ -157,7 +157,7 @@ func (r *ForwarderRequest) CreateEndpoint(queue *waiter.Queue) (tcpip.Endpoint, TSVal: r.synOptions.TSVal, TSEcr: r.synOptions.TSEcr, SACKPermitted: r.synOptions.SACKPermitted, - }, queue) + }, queue, nil) if err != nil { return nil, err } diff --git a/pkg/tcpip/transport/tcp/protocol.go b/pkg/tcpip/transport/tcp/protocol.go index 958c06fa7..cfd9a4e8e 100644 --- a/pkg/tcpip/transport/tcp/protocol.go +++ b/pkg/tcpip/transport/tcp/protocol.go @@ -94,8 +94,65 @@ const ( ccCubic = "cubic" ) +// syncRcvdCounter tracks the number of endpoints in the SYN-RCVD state. The +// value is protected by a mutex so that we can increment only when it's +// guaranteed not to go above a threshold. +type synRcvdCounter struct { + sync.Mutex + value uint64 + pending sync.WaitGroup + threshold uint64 +} + +// inc tries to increment the global number of endpoints in SYN-RCVD state. It +// succeeds if the increment doesn't make the count go beyond the threshold, and +// fails otherwise. +func (s *synRcvdCounter) inc() bool { + s.Lock() + defer s.Unlock() + if s.value >= s.threshold { + return false + } + + s.pending.Add(1) + s.value++ + + return true +} + +// dec atomically decrements the global number of endpoints in SYN-RCVD +// state. It must only be called if a previous call to inc succeeded. +func (s *synRcvdCounter) dec() { + s.Lock() + defer s.Unlock() + s.value-- + s.pending.Done() +} + +// synCookiesInUse returns true if the synRcvdCount is greater than +// SynRcvdCountThreshold. +func (s *synRcvdCounter) synCookiesInUse() bool { + s.Lock() + defer s.Unlock() + return s.value >= s.threshold +} + +// SetThreshold sets synRcvdCounter.Threshold to ths new threshold. +func (s *synRcvdCounter) SetThreshold(threshold uint64) { + s.Lock() + defer s.Unlock() + s.threshold = threshold +} + +// Threshold returns the current value of synRcvdCounter.Threhsold. +func (s *synRcvdCounter) Threshold() uint64 { + s.Lock() + defer s.Unlock() + return s.threshold +} + type protocol struct { - mu sync.Mutex + mu sync.RWMutex sackEnabled bool delayEnabled bool sendBufferSize SendBufferSizeOption @@ -105,6 +162,8 @@ type protocol struct { moderateReceiveBuffer bool tcpLingerTimeout time.Duration tcpTimeWaitTimeout time.Duration + minRTO time.Duration + synRcvdCount synRcvdCounter dispatcher *dispatcher } @@ -140,7 +199,7 @@ func (*protocol) ParsePorts(v buffer.View) (src, dst uint16, err *tcpip.Error) { // to a specific processing queue. Each queue is serviced by its own processor // goroutine which is responsible for dequeuing and doing full TCP dispatch of // the packet. -func (p *protocol) QueuePacket(r *stack.Route, ep stack.TransportEndpoint, id stack.TransportEndpointID, pkt tcpip.PacketBuffer) { +func (p *protocol) QueuePacket(r *stack.Route, ep stack.TransportEndpoint, id stack.TransportEndpointID, pkt stack.PacketBuffer) { p.dispatcher.queuePacket(r, ep, id, pkt) } @@ -151,7 +210,7 @@ func (p *protocol) QueuePacket(r *stack.Route, ep stack.TransportEndpoint, id st // a reset is sent in response to any incoming segment except another reset. In // particular, SYNs addressed to a non-existent connection are rejected by this // means." -func (*protocol) HandleUnknownDestinationPacket(r *stack.Route, id stack.TransportEndpointID, pkt tcpip.PacketBuffer) bool { +func (*protocol) HandleUnknownDestinationPacket(r *stack.Route, id stack.TransportEndpointID, pkt stack.PacketBuffer) bool { s := newSegment(r, id, pkt) defer s.decRef() @@ -164,12 +223,12 @@ func (*protocol) HandleUnknownDestinationPacket(r *stack.Route, id stack.Transpo return true } - replyWithReset(s) + replyWithReset(s, stack.DefaultTOS, s.route.DefaultTTL()) return true } // replyWithReset replies to the given segment with a reset segment. -func replyWithReset(s *segment) { +func replyWithReset(s *segment, tos, ttl uint8) { // Get the seqnum from the packet if the ack flag is set. seq := seqnum.Value(0) ack := seqnum.Value(0) @@ -191,10 +250,18 @@ func replyWithReset(s *segment) { flags |= header.TCPFlagAck ack = s.sequenceNumber.Add(s.logicalLen()) } - sendTCP(&s.route, s.id, buffer.VectorisedView{}, s.route.DefaultTTL(), stack.DefaultTOS, flags, seq, ack, 0 /* rcvWnd */, nil /* options */, nil /* gso */) + sendTCP(&s.route, tcpFields{ + id: s.id, + ttl: ttl, + tos: tos, + flags: flags, + seq: seq, + ack: ack, + rcvWnd: 0, + }, buffer.VectorisedView{}, nil /* gso */, nil /* PacketOwner */) } -// SetOption implements TransportProtocol.SetOption. +// SetOption implements stack.TransportProtocol.SetOption. func (p *protocol) SetOption(option interface{}) *tcpip.Error { switch v := option.(type) { case SACKEnabled: @@ -264,66 +331,93 @@ func (p *protocol) SetOption(option interface{}) *tcpip.Error { p.mu.Unlock() return nil + case tcpip.TCPMinRTOOption: + if v < 0 { + v = tcpip.TCPMinRTOOption(MinRTO) + } + p.mu.Lock() + p.minRTO = time.Duration(v) + p.mu.Unlock() + return nil + + case tcpip.TCPSynRcvdCountThresholdOption: + p.mu.Lock() + p.synRcvdCount.SetThreshold(uint64(v)) + p.mu.Unlock() + return nil + default: return tcpip.ErrUnknownProtocolOption } } -// Option implements TransportProtocol.Option. +// Option implements stack.TransportProtocol.Option. func (p *protocol) Option(option interface{}) *tcpip.Error { switch v := option.(type) { case *SACKEnabled: - p.mu.Lock() + p.mu.RLock() *v = SACKEnabled(p.sackEnabled) - p.mu.Unlock() + p.mu.RUnlock() return nil case *DelayEnabled: - p.mu.Lock() + p.mu.RLock() *v = DelayEnabled(p.delayEnabled) - p.mu.Unlock() + p.mu.RUnlock() return nil case *SendBufferSizeOption: - p.mu.Lock() + p.mu.RLock() *v = p.sendBufferSize - p.mu.Unlock() + p.mu.RUnlock() return nil case *ReceiveBufferSizeOption: - p.mu.Lock() + p.mu.RLock() *v = p.recvBufferSize - p.mu.Unlock() + p.mu.RUnlock() return nil case *tcpip.CongestionControlOption: - p.mu.Lock() + p.mu.RLock() *v = tcpip.CongestionControlOption(p.congestionControl) - p.mu.Unlock() + p.mu.RUnlock() return nil case *tcpip.AvailableCongestionControlOption: - p.mu.Lock() + p.mu.RLock() *v = tcpip.AvailableCongestionControlOption(strings.Join(p.availableCongestionControl, " ")) - p.mu.Unlock() + p.mu.RUnlock() return nil case *tcpip.ModerateReceiveBufferOption: - p.mu.Lock() + p.mu.RLock() *v = tcpip.ModerateReceiveBufferOption(p.moderateReceiveBuffer) - p.mu.Unlock() + p.mu.RUnlock() return nil case *tcpip.TCPLingerTimeoutOption: - p.mu.Lock() + p.mu.RLock() *v = tcpip.TCPLingerTimeoutOption(p.tcpLingerTimeout) - p.mu.Unlock() + p.mu.RUnlock() return nil case *tcpip.TCPTimeWaitTimeoutOption: - p.mu.Lock() + p.mu.RLock() *v = tcpip.TCPTimeWaitTimeoutOption(p.tcpTimeWaitTimeout) - p.mu.Unlock() + p.mu.RUnlock() + return nil + + case *tcpip.TCPMinRTOOption: + p.mu.RLock() + *v = tcpip.TCPMinRTOOption(p.minRTO) + p.mu.RUnlock() + return nil + + case *tcpip.TCPSynRcvdCountThresholdOption: + p.mu.RLock() + *v = tcpip.TCPSynRcvdCountThresholdOption(p.synRcvdCount.Threshold()) + p.mu.RUnlock() return nil default: @@ -331,6 +425,22 @@ func (p *protocol) Option(option interface{}) *tcpip.Error { } } +// Close implements stack.TransportProtocol.Close. +func (p *protocol) Close() { + p.dispatcher.close() +} + +// Wait implements stack.TransportProtocol.Wait. +func (p *protocol) Wait() { + p.dispatcher.wait() +} + +// SynRcvdCounter returns a reference to the synRcvdCount for this protocol +// instance. +func (p *protocol) SynRcvdCounter() *synRcvdCounter { + return &p.synRcvdCount +} + // NewProtocol returns a TCP transport protocol. func NewProtocol() stack.TransportProtocol { return &protocol{ @@ -340,6 +450,8 @@ func NewProtocol() stack.TransportProtocol { availableCongestionControl: []string{ccReno, ccCubic}, tcpLingerTimeout: DefaultTCPLingerTimeout, tcpTimeWaitTimeout: DefaultTCPTimeWaitTimeout, + synRcvdCount: synRcvdCounter{threshold: SynRcvdCountThreshold}, dispatcher: newDispatcher(runtime.GOMAXPROCS(0)), + minRTO: MinRTO, } } diff --git a/pkg/tcpip/transport/tcp/rcv.go b/pkg/tcpip/transport/tcp/rcv.go index 958f03ac1..a4b73b588 100644 --- a/pkg/tcpip/transport/tcp/rcv.go +++ b/pkg/tcpip/transport/tcp/rcv.go @@ -70,13 +70,24 @@ func newReceiver(ep *endpoint, irs seqnum.Value, rcvWnd seqnum.Size, rcvWndScale // acceptable checks if the segment sequence number range is acceptable // according to the table on page 26 of RFC 793. func (r *receiver) acceptable(segSeq seqnum.Value, segLen seqnum.Size) bool { - rcvWnd := r.rcvNxt.Size(r.rcvAcc) - if rcvWnd == 0 { - return segLen == 0 && segSeq == r.rcvNxt - } + return Acceptable(segSeq, segLen, r.rcvNxt, r.rcvAcc) +} - return segSeq.InWindow(r.rcvNxt, rcvWnd) || - seqnum.Overlap(r.rcvNxt, rcvWnd, segSeq, segLen) +// Acceptable checks if a segment that starts at segSeq and has length segLen is +// "acceptable" for arriving in a receive window that starts at rcvNxt and ends +// before rcvAcc, according to the table on page 26 and 69 of RFC 793. +func Acceptable(segSeq seqnum.Value, segLen seqnum.Size, rcvNxt, rcvAcc seqnum.Value) bool { + if rcvNxt == rcvAcc { + return segLen == 0 && segSeq == rcvNxt + } + if segLen == 0 { + // rcvWnd is incremented by 1 because that is Linux's behavior despite the + // RFC. + return segSeq.InRange(rcvNxt, rcvAcc.Add(1)) + } + // Page 70 of RFC 793 allows packets that can be made "acceptable" by trimming + // the payload, so we'll accept any payload that overlaps the receieve window. + return rcvNxt.LessThan(segSeq.Add(segLen)) && segSeq.LessThan(rcvAcc) } // getSendParams returns the parameters needed by the sender when building @@ -168,7 +179,6 @@ func (r *receiver) consumeSegment(s *segment, segSeq seqnum.Value, segLen seqnum // We just received a FIN, our next state depends on whether we sent a // FIN already or not. - r.ep.mu.Lock() switch r.ep.EndpointState() { case StateEstablished: r.ep.setEndpointState(StateCloseWait) @@ -183,7 +193,6 @@ func (r *receiver) consumeSegment(s *segment, segSeq seqnum.Value, segLen seqnum case StateFinWait2: r.ep.setEndpointState(StateTimeWait) } - r.ep.mu.Unlock() // Flush out any pending segments, except the very first one if // it happens to be the one we're handling now because the @@ -195,6 +204,10 @@ func (r *receiver) consumeSegment(s *segment, segSeq seqnum.Value, segLen seqnum for i := first; i < len(r.pendingRcvdSegments); i++ { r.pendingRcvdSegments[i].decRef() + // Note that slice truncation does not allow garbage collection of + // truncated items, thus truncated items must be set to nil to avoid + // memory leaks. + r.pendingRcvdSegments[i] = nil } r.pendingRcvdSegments = r.pendingRcvdSegments[:first] @@ -204,7 +217,6 @@ func (r *receiver) consumeSegment(s *segment, segSeq seqnum.Value, segLen seqnum // Handle ACK (not FIN-ACK, which we handled above) during one of the // shutdown states. if s.flagIsSet(header.TCPFlagAck) && s.ackNumber == r.ep.snd.sndNxt { - r.ep.mu.Lock() switch r.ep.EndpointState() { case StateFinWait1: r.ep.setEndpointState(StateFinWait2) @@ -218,7 +230,6 @@ func (r *receiver) consumeSegment(s *segment, segSeq seqnum.Value, segLen seqnum case StateLastAck: r.ep.transitionToStateCloseLocked() } - r.ep.mu.Unlock() } return true @@ -332,10 +343,8 @@ func (r *receiver) handleRcvdSegmentClosing(s *segment, state EndpointState, clo // handleRcvdSegment handles TCP segments directed at the connection managed by // r as they arrive. It is called by the protocol main loop. func (r *receiver) handleRcvdSegment(s *segment) (drop bool, err *tcpip.Error) { - r.ep.mu.RLock() state := r.ep.EndpointState() closed := r.ep.closed - r.ep.mu.RUnlock() if state != StateEstablished { drop, err := r.handleRcvdSegmentClosing(s, state, closed) diff --git a/pkg/tcpip/transport/tcp/rcv_test.go b/pkg/tcpip/transport/tcp/rcv_test.go new file mode 100644 index 000000000..dc02729ce --- /dev/null +++ b/pkg/tcpip/transport/tcp/rcv_test.go @@ -0,0 +1,74 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package rcv_test + +import ( + "testing" + + "gvisor.dev/gvisor/pkg/tcpip/seqnum" + "gvisor.dev/gvisor/pkg/tcpip/transport/tcp" +) + +func TestAcceptable(t *testing.T) { + for _, tt := range []struct { + segSeq seqnum.Value + segLen seqnum.Size + rcvNxt, rcvAcc seqnum.Value + want bool + }{ + // The segment is smaller than the window. + {105, 2, 100, 104, false}, + {105, 2, 101, 105, false}, + {105, 2, 102, 106, true}, + {105, 2, 103, 107, true}, + {105, 2, 104, 108, true}, + {105, 2, 105, 109, true}, + {105, 2, 106, 110, true}, + {105, 2, 107, 111, false}, + + // The segment is larger than the window. + {105, 4, 103, 105, false}, + {105, 4, 104, 106, true}, + {105, 4, 105, 107, true}, + {105, 4, 106, 108, true}, + {105, 4, 107, 109, true}, + {105, 4, 108, 110, true}, + {105, 4, 109, 111, false}, + {105, 4, 110, 112, false}, + + // The segment has no width. + {105, 0, 100, 102, false}, + {105, 0, 101, 103, false}, + {105, 0, 102, 104, false}, + {105, 0, 103, 105, true}, + {105, 0, 104, 106, true}, + {105, 0, 105, 107, true}, + {105, 0, 106, 108, false}, + {105, 0, 107, 109, false}, + + // The receive window has no width. + {105, 2, 103, 103, false}, + {105, 2, 104, 104, false}, + {105, 2, 105, 105, false}, + {105, 2, 106, 106, false}, + {105, 2, 107, 107, false}, + {105, 2, 108, 108, false}, + {105, 2, 109, 109, false}, + } { + if got := tcp.Acceptable(tt.segSeq, tt.segLen, tt.rcvNxt, tt.rcvAcc); got != tt.want { + t.Errorf("tcp.Acceptable(%d, %d, %d, %d) = %t, want %t", tt.segSeq, tt.segLen, tt.rcvNxt, tt.rcvAcc, got, tt.want) + } + } +} diff --git a/pkg/tcpip/transport/tcp/segment.go b/pkg/tcpip/transport/tcp/segment.go index 1c10da5ca..40461fd31 100644 --- a/pkg/tcpip/transport/tcp/segment.go +++ b/pkg/tcpip/transport/tcp/segment.go @@ -18,7 +18,6 @@ import ( "sync/atomic" "time" - "gvisor.dev/gvisor/pkg/tcpip" "gvisor.dev/gvisor/pkg/tcpip/buffer" "gvisor.dev/gvisor/pkg/tcpip/header" "gvisor.dev/gvisor/pkg/tcpip/seqnum" @@ -56,12 +55,12 @@ type segment struct { options []byte `state:".([]byte)"` hasNewSACKInfo bool rcvdTime time.Time `state:".(unixTime)"` - // xmitTime is the last transmit time of this segment. A zero value - // indicates that the segment has yet to be transmitted. - xmitTime time.Time `state:".(unixTime)"` + // xmitTime is the last transmit time of this segment. + xmitTime time.Time `state:".(unixTime)"` + xmitCount uint32 } -func newSegment(r *stack.Route, id stack.TransportEndpointID, pkt tcpip.PacketBuffer) *segment { +func newSegment(r *stack.Route, id stack.TransportEndpointID, pkt stack.PacketBuffer) *segment { s := &segment{ refCnt: 1, id: id, @@ -78,9 +77,11 @@ func newSegmentFromView(r *stack.Route, id stack.TransportEndpointID, v buffer.V id: id, route: r.Clone(), } - s.views[0] = v - s.data = buffer.NewVectorisedView(len(v), s.views[:1]) s.rcvdTime = time.Now() + if len(v) != 0 { + s.views[0] = v + s.data = buffer.NewVectorisedView(len(v), s.views[:1]) + } return s } diff --git a/pkg/tcpip/transport/tcp/segment_heap.go b/pkg/tcpip/transport/tcp/segment_heap.go index 9fd061d7d..8d3ddce4b 100644 --- a/pkg/tcpip/transport/tcp/segment_heap.go +++ b/pkg/tcpip/transport/tcp/segment_heap.go @@ -14,21 +14,25 @@ package tcp +import "container/heap" + type segmentHeap []*segment +var _ heap.Interface = (*segmentHeap)(nil) + // Len returns the length of h. -func (h segmentHeap) Len() int { - return len(h) +func (h *segmentHeap) Len() int { + return len(*h) } // Less determines whether the i-th element of h is less than the j-th element. -func (h segmentHeap) Less(i, j int) bool { - return h[i].sequenceNumber.LessThan(h[j].sequenceNumber) +func (h *segmentHeap) Less(i, j int) bool { + return (*h)[i].sequenceNumber.LessThan((*h)[j].sequenceNumber) } // Swap swaps the i-th and j-th elements of h. -func (h segmentHeap) Swap(i, j int) { - h[i], h[j] = h[j], h[i] +func (h *segmentHeap) Swap(i, j int) { + (*h)[i], (*h)[j] = (*h)[j], (*h)[i] } // Push adds x as the last element of h. @@ -41,6 +45,7 @@ func (h *segmentHeap) Pop() interface{} { old := *h n := len(old) x := old[n-1] + old[n-1] = nil *h = old[:n-1] return x } diff --git a/pkg/tcpip/transport/tcp/segment_queue.go b/pkg/tcpip/transport/tcp/segment_queue.go index bd20a7ee9..48a257137 100644 --- a/pkg/tcpip/transport/tcp/segment_queue.go +++ b/pkg/tcpip/transport/tcp/segment_queue.go @@ -28,10 +28,16 @@ type segmentQueue struct { used int } +// emptyLocked determines if the queue is empty. +// Preconditions: q.mu must be held. +func (q *segmentQueue) emptyLocked() bool { + return q.used == 0 +} + // empty determines if the queue is empty. func (q *segmentQueue) empty() bool { q.mu.Lock() - r := q.used == 0 + r := q.emptyLocked() q.mu.Unlock() return r diff --git a/pkg/tcpip/transport/tcp/snd.go b/pkg/tcpip/transport/tcp/snd.go index b74b61e7d..d8cfe3115 100644 --- a/pkg/tcpip/transport/tcp/snd.go +++ b/pkg/tcpip/transport/tcp/snd.go @@ -15,6 +15,7 @@ package tcp import ( + "fmt" "math" "sync/atomic" "time" @@ -126,10 +127,6 @@ type sender struct { // sndNxt is the sequence number of the next segment to be sent. sndNxt seqnum.Value - // sndNxtList is the sequence number of the next segment to be added to - // the send list. - sndNxtList seqnum.Value - // rttMeasureSeqNum is the sequence number being used for the latest RTT // measurement. rttMeasureSeqNum seqnum.Value @@ -153,6 +150,9 @@ type sender struct { rtt rtt rto time.Duration + // minRTO is the minimum permitted value for sender.rto. + minRTO time.Duration + // maxPayloadSize is the maximum size of the payload of a given segment. // It is initialized on demand. maxPayloadSize int @@ -229,7 +229,6 @@ func newSender(ep *endpoint, iss, irs seqnum.Value, sndWnd seqnum.Size, mss uint sndWnd: sndWnd, sndUna: iss + 1, sndNxt: iss + 1, - sndNxtList: iss + 1, rto: 1 * time.Second, rttMeasureSeqNum: iss + 1, lastSendTime: time.Now(), @@ -265,6 +264,13 @@ func newSender(ep *endpoint, iss, irs seqnum.Value, sndWnd seqnum.Size, mss uint // etc. s.ep.scoreboard = NewSACKScoreboard(uint16(s.maxPayloadSize), iss) + // Get Stack wide minRTO. + var v tcpip.TCPMinRTOOption + if err := ep.stack.TransportProtocolOption(ProtocolNumber, &v); err != nil { + panic(fmt.Sprintf("unable to get minRTO from stack: %s", err)) + } + s.minRTO = time.Duration(v) + return s } @@ -399,8 +405,8 @@ func (s *sender) updateRTO(rtt time.Duration) { s.rto = s.rtt.srtt + 4*s.rtt.rttvar s.rtt.Unlock() - if s.rto < MinRTO { - s.rto = MinRTO + if s.rto < s.minRTO { + s.rto = s.minRTO } } @@ -455,9 +461,7 @@ func (s *sender) retransmitTimerExpired() bool { // Give up if we've waited more than a minute since the last resend or // if a user time out is set and we have exceeded the user specified // timeout since the first retransmission. - s.ep.mu.RLock() uto := s.ep.userTimeout - s.ep.mu.RUnlock() if s.firstRetransmittedSegXmitTime.IsZero() { // We store the original xmitTime of the segment that we are @@ -713,7 +717,6 @@ func (s *sender) maybeSendSegment(seg *segment, limit int, end seqnum.Value) (se default: s.ep.setEndpointState(StateFinWait1) } - } else { // We're sending a non-FIN segment. if seg.flags&header.TCPFlagFin != 0 { @@ -1229,7 +1232,7 @@ func (s *sender) handleRcvdSegment(seg *segment) { // sendSegment sends the specified segment. func (s *sender) sendSegment(seg *segment) *tcpip.Error { - if !seg.xmitTime.IsZero() { + if seg.xmitCount > 0 { s.ep.stack.Stats().TCP.Retransmits.Increment() s.ep.stats.SendErrors.Retransmits.Increment() if s.sndCwnd < s.sndSsthresh { @@ -1237,6 +1240,7 @@ func (s *sender) sendSegment(seg *segment) *tcpip.Error { } } seg.xmitTime = time.Now() + seg.xmitCount++ return s.sendSegmentFromView(seg.data, seg.flags, seg.sequenceNumber) } diff --git a/pkg/tcpip/transport/tcp/tcp_noracedetector_test.go b/pkg/tcpip/transport/tcp/tcp_noracedetector_test.go index 782d7b42c..359a75e73 100644 --- a/pkg/tcpip/transport/tcp/tcp_noracedetector_test.go +++ b/pkg/tcpip/transport/tcp/tcp_noracedetector_test.go @@ -31,6 +31,7 @@ import ( "gvisor.dev/gvisor/pkg/tcpip/header" "gvisor.dev/gvisor/pkg/tcpip/transport/tcp" "gvisor.dev/gvisor/pkg/tcpip/transport/tcp/testing/context" + "gvisor.dev/gvisor/runsc/testutil" ) func TestFastRecovery(t *testing.T) { @@ -40,7 +41,7 @@ func TestFastRecovery(t *testing.T) { c.CreateConnected(789, 30000, -1 /* epRcvBuf */) - const iterations = 7 + const iterations = 3 data := buffer.NewView(2 * maxPayload * (tcp.InitialCwnd << (iterations + 1))) for i := range data { data[i] = byte(i) @@ -86,16 +87,23 @@ func TestFastRecovery(t *testing.T) { // Receive the retransmitted packet. c.ReceiveAndCheckPacket(data, rtxOffset, maxPayload) - if got, want := c.Stack().Stats().TCP.FastRetransmit.Value(), uint64(1); got != want { - t.Errorf("got stats.TCP.FastRetransmit.Value = %v, want = %v", got, want) - } + // Wait before checking metrics. + metricPollFn := func() error { + if got, want := c.Stack().Stats().TCP.FastRetransmit.Value(), uint64(1); got != want { + return fmt.Errorf("got stats.TCP.FastRetransmit.Value = %v, want = %v", got, want) + } + if got, want := c.Stack().Stats().TCP.Retransmits.Value(), uint64(1); got != want { + return fmt.Errorf("got stats.TCP.Retransmit.Value = %v, want = %v", got, want) + } - if got, want := c.Stack().Stats().TCP.Retransmits.Value(), uint64(1); got != want { - t.Errorf("got stats.TCP.Retransmit.Value = %v, want = %v", got, want) + if got, want := c.Stack().Stats().TCP.FastRecovery.Value(), uint64(1); got != want { + return fmt.Errorf("got stats.TCP.FastRecovery.Value = %v, want = %v", got, want) + } + return nil } - if got, want := c.Stack().Stats().TCP.FastRecovery.Value(), uint64(1); got != want { - t.Errorf("got stats.TCP.FastRecovery.Value = %v, want = %v", got, want) + if err := testutil.Poll(metricPollFn, 1*time.Second); err != nil { + t.Error(err) } // Now send 7 mode duplicate acks. Each of these should cause a window @@ -117,12 +125,18 @@ func TestFastRecovery(t *testing.T) { // Receive the retransmit due to partial ack. c.ReceiveAndCheckPacket(data, rtxOffset, maxPayload) - if got, want := c.Stack().Stats().TCP.FastRetransmit.Value(), uint64(2); got != want { - t.Errorf("got stats.TCP.FastRetransmit.Value = %v, want = %v", got, want) + // Wait before checking metrics. + metricPollFn = func() error { + if got, want := c.Stack().Stats().TCP.FastRetransmit.Value(), uint64(2); got != want { + return fmt.Errorf("got stats.TCP.FastRetransmit.Value = %v, want = %v", got, want) + } + if got, want := c.Stack().Stats().TCP.Retransmits.Value(), uint64(2); got != want { + return fmt.Errorf("got stats.TCP.Retransmit.Value = %v, want = %v", got, want) + } + return nil } - - if got, want := c.Stack().Stats().TCP.Retransmits.Value(), uint64(2); got != want { - t.Errorf("got stats.TCP.Retransmit.Value = %v, want = %v", got, want) + if err := testutil.Poll(metricPollFn, 1*time.Second); err != nil { + t.Error(err) } // Receive the 10 extra packets that should have been released due to @@ -192,7 +206,7 @@ func TestExponentialIncreaseDuringSlowStart(t *testing.T) { c.CreateConnected(789, 30000, -1 /* epRcvBuf */) - const iterations = 7 + const iterations = 3 data := buffer.NewView(maxPayload * (tcp.InitialCwnd << (iterations + 1))) for i := range data { data[i] = byte(i) @@ -234,7 +248,7 @@ func TestCongestionAvoidance(t *testing.T) { c.CreateConnected(789, 30000, -1 /* epRcvBuf */) - const iterations = 7 + const iterations = 3 data := buffer.NewView(2 * maxPayload * (tcp.InitialCwnd << (iterations + 1))) for i := range data { data[i] = byte(i) @@ -338,7 +352,7 @@ func TestCubicCongestionAvoidance(t *testing.T) { c.CreateConnected(789, 30000, -1 /* epRcvBuf */) - const iterations = 7 + const iterations = 3 data := buffer.NewView(2 * maxPayload * (tcp.InitialCwnd << (iterations + 1))) for i := range data { @@ -447,7 +461,7 @@ func TestRetransmit(t *testing.T) { c.CreateConnected(789, 30000, -1 /* epRcvBuf */) - const iterations = 7 + const iterations = 3 data := buffer.NewView(maxPayload * (tcp.InitialCwnd << (iterations + 1))) for i := range data { data[i] = byte(i) @@ -492,24 +506,33 @@ func TestRetransmit(t *testing.T) { rtxOffset := bytesRead - maxPayload*expected c.ReceiveAndCheckPacket(data, rtxOffset, maxPayload) - if got, want := c.Stack().Stats().TCP.Timeouts.Value(), uint64(1); got != want { - t.Errorf("got stats.TCP.Timeouts.Value = %v, want = %v", got, want) - } + metricPollFn := func() error { + if got, want := c.Stack().Stats().TCP.Timeouts.Value(), uint64(1); got != want { + return fmt.Errorf("got stats.TCP.Timeouts.Value = %v, want = %v", got, want) + } - if got, want := c.Stack().Stats().TCP.Retransmits.Value(), uint64(1); got != want { - t.Errorf("got stats.TCP.Retransmits.Value = %v, want = %v", got, want) - } + if got, want := c.Stack().Stats().TCP.Retransmits.Value(), uint64(1); got != want { + return fmt.Errorf("got stats.TCP.Retransmits.Value = %v, want = %v", got, want) + } - if got, want := c.EP.Stats().(*tcp.Stats).SendErrors.Timeouts.Value(), uint64(1); got != want { - t.Errorf("got EP SendErrors.Timeouts.Value = %v, want = %v", got, want) - } + if got, want := c.EP.Stats().(*tcp.Stats).SendErrors.Timeouts.Value(), uint64(1); got != want { + return fmt.Errorf("got EP SendErrors.Timeouts.Value = %v, want = %v", got, want) + } + + if got, want := c.EP.Stats().(*tcp.Stats).SendErrors.Retransmits.Value(), uint64(1); got != want { + return fmt.Errorf("got EP stats SendErrors.Retransmits.Value = %v, want = %v", got, want) + } + + if got, want := c.Stack().Stats().TCP.SlowStartRetransmits.Value(), uint64(1); got != want { + return fmt.Errorf("got stats.TCP.SlowStartRetransmits.Value = %v, want = %v", got, want) + } - if got, want := c.EP.Stats().(*tcp.Stats).SendErrors.Retransmits.Value(), uint64(1); got != want { - t.Errorf("got EP stats SendErrors.Retransmits.Value = %v, want = %v", got, want) + return nil } - if got, want := c.Stack().Stats().TCP.SlowStartRetransmits.Value(), uint64(1); got != want { - t.Errorf("got stats.TCP.SlowStartRetransmits.Value = %v, want = %v", got, want) + // Poll when checking metrics. + if err := testutil.Poll(metricPollFn, 1*time.Second); err != nil { + t.Error(err) } // Acknowledge half of the pending data. diff --git a/pkg/tcpip/transport/tcp/tcp_sack_test.go b/pkg/tcpip/transport/tcp/tcp_sack_test.go index afea124ec..1dd63dd61 100644 --- a/pkg/tcpip/transport/tcp/tcp_sack_test.go +++ b/pkg/tcpip/transport/tcp/tcp_sack_test.go @@ -149,21 +149,22 @@ func TestSackPermittedAccept(t *testing.T) { {true, false, -1, 0xffff}, // When cookie is used window scaling is disabled. {false, true, 5, 0x8000}, // 0x8000 * 2^5 = 1<<20 = 1MB window (the default). } - savedSynCountThreshold := tcp.SynRcvdCountThreshold - defer func() { - tcp.SynRcvdCountThreshold = savedSynCountThreshold - }() + for _, tc := range testCases { t.Run(fmt.Sprintf("test: %#v", tc), func(t *testing.T) { - if tc.cookieEnabled { - tcp.SynRcvdCountThreshold = 0 - } else { - tcp.SynRcvdCountThreshold = savedSynCountThreshold - } for _, sackEnabled := range []bool{false, true} { t.Run(fmt.Sprintf("test stack.sackEnabled: %v", sackEnabled), func(t *testing.T) { c := context.New(t, defaultMTU) defer c.Cleanup() + + if tc.cookieEnabled { + // Set the SynRcvd threshold to + // zero to force a syn cookie + // based accept to happen. + if err := c.Stack().SetTransportProtocolOption(tcp.ProtocolNumber, tcpip.TCPSynRcvdCountThresholdOption(0)); err != nil { + t.Fatalf("setting TCPSynRcvdCountThresholdOption to 0 failed: %s", err) + } + } setStackSACKPermitted(t, c, sackEnabled) rep := c.AcceptWithOptions(tc.wndScale, header.TCPSynOptions{MSS: defaultIPv4MSS, SACKPermitted: tc.sackPermitted}) @@ -222,21 +223,23 @@ func TestSackDisabledAccept(t *testing.T) { {true, -1, 0xffff}, // When cookie is used window scaling is disabled. {false, 5, 0x8000}, // 0x8000 * 2^5 = 1<<20 = 1MB window (the default). } - savedSynCountThreshold := tcp.SynRcvdCountThreshold - defer func() { - tcp.SynRcvdCountThreshold = savedSynCountThreshold - }() + for _, tc := range testCases { t.Run(fmt.Sprintf("test: %#v", tc), func(t *testing.T) { - if tc.cookieEnabled { - tcp.SynRcvdCountThreshold = 0 - } else { - tcp.SynRcvdCountThreshold = savedSynCountThreshold - } for _, sackEnabled := range []bool{false, true} { t.Run(fmt.Sprintf("test: sackEnabled: %v", sackEnabled), func(t *testing.T) { c := context.New(t, defaultMTU) defer c.Cleanup() + + if tc.cookieEnabled { + // Set the SynRcvd threshold to + // zero to force a syn cookie + // based accept to happen. + if err := c.Stack().SetTransportProtocolOption(tcp.ProtocolNumber, tcpip.TCPSynRcvdCountThresholdOption(0)); err != nil { + t.Fatalf("setting TCPSynRcvdCountThresholdOption to 0 failed: %s", err) + } + } + setStackSACKPermitted(t, c, sackEnabled) rep := c.AcceptWithOptions(tc.wndScale, header.TCPSynOptions{MSS: defaultIPv4MSS}) @@ -387,7 +390,7 @@ func TestSACKRecovery(t *testing.T) { setStackSACKPermitted(t, c, true) createConnectedWithSACKAndTS(c) - const iterations = 7 + const iterations = 3 data := buffer.NewView(2 * maxPayload * (tcp.InitialCwnd << (iterations + 1))) for i := range data { data[i] = byte(i) diff --git a/pkg/tcpip/transport/tcp/tcp_test.go b/pkg/tcpip/transport/tcp/tcp_test.go index cc118c993..ab1014c7f 100644 --- a/pkg/tcpip/transport/tcp/tcp_test.go +++ b/pkg/tcpip/transport/tcp/tcp_test.go @@ -284,7 +284,7 @@ func TestTCPResetSentForACKWhenNotUsingSynCookies(t *testing.T) { // are released instantly on Close. tcpTW := tcpip.TCPTimeWaitTimeoutOption(1 * time.Millisecond) if err := c.Stack().SetTransportProtocolOption(tcp.ProtocolNumber, tcpTW); err != nil { - t.Fatalf("e.stack.SetTransportProtocolOption(%d, %s) = %s", tcp.ProtocolNumber, tcpTW, err) + t.Fatalf("e.stack.SetTransportProtocolOption(%d, %v) = %v", tcp.ProtocolNumber, tcpTW, err) } c.EP.Close() @@ -543,8 +543,9 @@ func TestCurrentConnectedIncrement(t *testing.T) { ), ) - // Wait for the TIME-WAIT state to transition to CLOSED. - time.Sleep(1 * time.Second) + // Wait for a little more than the TIME-WAIT duration for the socket to + // transition to CLOSED state. + time.Sleep(1200 * time.Millisecond) if got := c.Stack().Stats().TCP.CurrentEstablished.Value(); got != 0 { t.Errorf("got stats.TCP.CurrentEstablished.Value() = %v, want = 0", got) @@ -589,6 +590,10 @@ func TestClosingWithEnqueuedSegments(t *testing.T) { ), ) + // Give the stack a few ms to transition the endpoint out of ESTABLISHED + // state. + time.Sleep(10 * time.Millisecond) + if got, want := tcp.EndpointState(ep.State()), tcp.StateCloseWait; got != want { t.Errorf("Unexpected endpoint state: want %v, got %v", want, got) } @@ -727,7 +732,7 @@ func TestUserSuppliedMSSOnConnectV4(t *testing.T) { const maxMSS = mtu - header.IPv4MinimumSize - header.TCPMinimumSize tests := []struct { name string - setMSS uint16 + setMSS int expMSS uint16 }{ { @@ -755,15 +760,14 @@ func TestUserSuppliedMSSOnConnectV4(t *testing.T) { c.Create(-1) // Set the MSS socket option. - opt := tcpip.MaxSegOption(test.setMSS) - if err := c.EP.SetSockOpt(opt); err != nil { - t.Fatalf("SetSockOpt(%#v) failed: %s", opt, err) + if err := c.EP.SetSockOptInt(tcpip.MaxSegOption, test.setMSS); err != nil { + t.Fatalf("SetSockOptInt(MaxSegOption, %d) failed: %s", test.setMSS, err) } // Get expected window size. rcvBufSize, err := c.EP.GetSockOptInt(tcpip.ReceiveBufferSizeOption) if err != nil { - t.Fatalf("GetSockOpt(%v) failed: %s", tcpip.ReceiveBufferSizeOption, err) + t.Fatalf("GetSockOptInt(ReceiveBufferSizeOption) failed: %s", err) } ws := tcp.FindWndScale(seqnum.Size(rcvBufSize)) @@ -817,15 +821,14 @@ func TestUserSuppliedMSSOnConnectV6(t *testing.T) { c.CreateV6Endpoint(true) // Set the MSS socket option. - opt := tcpip.MaxSegOption(test.setMSS) - if err := c.EP.SetSockOpt(opt); err != nil { - t.Fatalf("SetSockOpt(%#v) failed: %s", opt, err) + if err := c.EP.SetSockOptInt(tcpip.MaxSegOption, int(test.setMSS)); err != nil { + t.Fatalf("SetSockOptInt(MaxSegOption, %d) failed: %s", test.setMSS, err) } // Get expected window size. rcvBufSize, err := c.EP.GetSockOptInt(tcpip.ReceiveBufferSizeOption) if err != nil { - t.Fatalf("GetSockOpt(%v) failed: %s", tcpip.ReceiveBufferSizeOption, err) + t.Fatalf("GetSockOptInt(ReceiveBufferSizeOption) failed: %s", err) } ws := tcp.FindWndScale(seqnum.Size(rcvBufSize)) @@ -1031,8 +1034,8 @@ func TestSendRstOnListenerRxAckV6(t *testing.T) { checker.SeqNum(200))) } -// TestListenShutdown tests for the listening endpoint not processing -// any receive when it is on read shutdown. +// TestListenShutdown tests for the listening endpoint replying with RST +// on read shutdown. func TestListenShutdown(t *testing.T) { c := context.New(t, defaultMTU) defer c.Cleanup() @@ -1043,7 +1046,7 @@ func TestListenShutdown(t *testing.T) { t.Fatal("Bind failed:", err) } - if err := c.EP.Listen(10 /* backlog */); err != nil { + if err := c.EP.Listen(1 /* backlog */); err != nil { t.Fatal("Listen failed:", err) } @@ -1051,9 +1054,6 @@ func TestListenShutdown(t *testing.T) { t.Fatal("Shutdown failed:", err) } - // Wait for the endpoint state to be propagated. - time.Sleep(10 * time.Millisecond) - c.SendPacket(nil, &context.Headers{ SrcPort: context.TestPort, DstPort: context.StackPort, @@ -1062,7 +1062,49 @@ func TestListenShutdown(t *testing.T) { AckNum: 200, }) - c.CheckNoPacket("Packet received when listening socket was shutdown") + // Expect the listening endpoint to reset the connection. + checker.IPv4(t, c.GetPacket(), + checker.TCP( + checker.DstPort(context.TestPort), + checker.TCPFlags(header.TCPFlagAck|header.TCPFlagRst), + )) +} + +// TestListenCloseWhileConnect tests for the listening endpoint to +// drain the accept-queue when closed. This should reset all of the +// pending connections that are waiting to be accepted. +func TestListenCloseWhileConnect(t *testing.T) { + c := context.New(t, defaultMTU) + defer c.Cleanup() + + c.Create(-1 /* epRcvBuf */) + + if err := c.EP.Bind(tcpip.FullAddress{Port: context.StackPort}); err != nil { + t.Fatal("Bind failed:", err) + } + + if err := c.EP.Listen(1 /* backlog */); err != nil { + t.Fatal("Listen failed:", err) + } + + waitEntry, notifyCh := waiter.NewChannelEntry(nil) + c.WQ.EventRegister(&waitEntry, waiter.EventIn) + defer c.WQ.EventUnregister(&waitEntry) + + executeHandshake(t, c, context.TestPort, false /* synCookiesInUse */) + // Wait for the new endpoint created because of handshake to be delivered + // to the listening endpoint's accept queue. + <-notifyCh + + // Close the listening endpoint. + c.EP.Close() + + // Expect the listening endpoint to reset the connection. + checker.IPv4(t, c.GetPacket(), + checker.TCP( + checker.DstPort(context.TestPort), + checker.TCPFlags(header.TCPFlagAck|header.TCPFlagRst), + )) } func TestTOSV4(t *testing.T) { @@ -1076,17 +1118,17 @@ func TestTOSV4(t *testing.T) { c.EP = ep const tos = 0xC0 - if err := c.EP.SetSockOpt(tcpip.IPv4TOSOption(tos)); err != nil { - t.Errorf("SetSockOpt(%#v) failed: %s", tcpip.IPv4TOSOption(tos), err) + if err := c.EP.SetSockOptInt(tcpip.IPv4TOSOption, tos); err != nil { + t.Errorf("SetSockOptInt(IPv4TOSOption, %d) failed: %s", tos, err) } - var v tcpip.IPv4TOSOption - if err := c.EP.GetSockOpt(&v); err != nil { - t.Errorf("GetSockopt failed: %s", err) + v, err := c.EP.GetSockOptInt(tcpip.IPv4TOSOption) + if err != nil { + t.Errorf("GetSockoptInt(IPv4TOSOption) failed: %s", err) } - if want := tcpip.IPv4TOSOption(tos); v != want { - t.Errorf("got GetSockOpt(...) = %#v, want = %#v", v, want) + if v != tos { + t.Errorf("got GetSockOptInt(IPv4TOSOption) = %d, want = %d", v, tos) } testV4Connect(t, c, checker.TOS(tos, 0)) @@ -1124,17 +1166,17 @@ func TestTrafficClassV6(t *testing.T) { c.CreateV6Endpoint(false) const tos = 0xC0 - if err := c.EP.SetSockOpt(tcpip.IPv6TrafficClassOption(tos)); err != nil { - t.Errorf("SetSockOpt(%#v) failed: %s", tcpip.IPv6TrafficClassOption(tos), err) + if err := c.EP.SetSockOptInt(tcpip.IPv6TrafficClassOption, tos); err != nil { + t.Errorf("SetSockOpInt(IPv6TrafficClassOption, %d) failed: %s", tos, err) } - var v tcpip.IPv6TrafficClassOption - if err := c.EP.GetSockOpt(&v); err != nil { - t.Fatalf("GetSockopt failed: %s", err) + v, err := c.EP.GetSockOptInt(tcpip.IPv6TrafficClassOption) + if err != nil { + t.Fatalf("GetSockoptInt(IPv6TrafficClassOption) failed: %s", err) } - if want := tcpip.IPv6TrafficClassOption(tos); v != want { - t.Errorf("got GetSockOpt(...) = %#v, want = %#v", v, want) + if v != tos { + t.Errorf("got GetSockOptInt(IPv6TrafficClassOption) = %d, want = %d", v, tos) } // Test the connection request. @@ -1710,7 +1752,7 @@ func TestNoWindowShrinking(t *testing.T) { c.CreateConnected(789, 30000, 10) if err := c.EP.SetSockOptInt(tcpip.ReceiveBufferSizeOption, 5); err != nil { - t.Fatalf("SetSockOpt failed: %v", err) + t.Fatalf("SetSockOptInt(ReceiveBufferSizeOption, 5) failed: %v", err) } we, ch := waiter.NewChannelEntry(nil) @@ -1983,7 +2025,7 @@ func TestScaledWindowAccept(t *testing.T) { // Set the window size greater than the maximum non-scaled window. if err := ep.SetSockOptInt(tcpip.ReceiveBufferSizeOption, 65535*3); err != nil { - t.Fatalf("SetSockOpt failed failed: %v", err) + t.Fatalf("SetSockOptInt(ReceiveBufferSizeOption, 65535*3) failed failed: %v", err) } if err := ep.Bind(tcpip.FullAddress{Port: context.StackPort}); err != nil { @@ -2056,7 +2098,7 @@ func TestNonScaledWindowAccept(t *testing.T) { // Set the window size greater than the maximum non-scaled window. if err := ep.SetSockOptInt(tcpip.ReceiveBufferSizeOption, 65535*3); err != nil { - t.Fatalf("SetSockOpt failed failed: %v", err) + t.Fatalf("SetSockOptInt(ReceiveBufferSizeOption, 65535*3) failed failed: %v", err) } if err := ep.Bind(tcpip.FullAddress{Port: context.StackPort}); err != nil { @@ -2220,10 +2262,10 @@ func TestSegmentMerging(t *testing.T) { { "cork", func(ep tcpip.Endpoint) { - ep.SetSockOpt(tcpip.CorkOption(1)) + ep.SetSockOptBool(tcpip.CorkOption, true) }, func(ep tcpip.Endpoint) { - ep.SetSockOpt(tcpip.CorkOption(0)) + ep.SetSockOptBool(tcpip.CorkOption, false) }, }, } @@ -2235,9 +2277,18 @@ func TestSegmentMerging(t *testing.T) { c.CreateConnected(789, 30000, -1 /* epRcvBuf */) - // Prevent the endpoint from processing packets. - test.stop(c.EP) + // Send tcp.InitialCwnd number of segments to fill up + // InitialWindow but don't ACK. That should prevent + // anymore packets from going out. + for i := 0; i < tcp.InitialCwnd; i++ { + view := buffer.NewViewFromBytes([]byte{0}) + if _, _, err := c.EP.Write(tcpip.SlicePayload(view), tcpip.WriteOptions{}); err != nil { + t.Fatalf("Write #%d failed: %s", i+1, err) + } + } + // Now send the segments that should get merged as the congestion + // window is full and we won't be able to send any more packets. var allData []byte for i, data := range [][]byte{{1, 2, 3, 4}, {5, 6, 7}, {8, 9}, {10}, {11}} { allData = append(allData, data...) @@ -2247,8 +2298,29 @@ func TestSegmentMerging(t *testing.T) { } } - // Let the endpoint process the segments that we just sent. - test.resume(c.EP) + // Check that we get tcp.InitialCwnd packets. + for i := 0; i < tcp.InitialCwnd; i++ { + b := c.GetPacket() + checker.IPv4(t, b, + checker.PayloadLen(header.TCPMinimumSize+1), + checker.TCP( + checker.DstPort(context.TestPort), + checker.SeqNum(uint32(c.IRS)+uint32(i)+1), + checker.AckNum(790), + checker.TCPFlagsMatch(header.TCPFlagAck, ^uint8(header.TCPFlagPsh)), + ), + ) + } + + // Acknowledge the data. + c.SendPacket(nil, &context.Headers{ + SrcPort: context.TestPort, + DstPort: c.Port, + Flags: header.TCPFlagAck, + SeqNum: 790, + AckNum: c.IRS.Add(1 + 10), // 10 for the 10 bytes of payload. + RcvWnd: 30000, + }) // Check that data is received. b := c.GetPacket() @@ -2256,7 +2328,7 @@ func TestSegmentMerging(t *testing.T) { checker.PayloadLen(len(allData)+header.TCPMinimumSize), checker.TCP( checker.DstPort(context.TestPort), - checker.SeqNum(uint32(c.IRS)+1), + checker.SeqNum(uint32(c.IRS)+11), checker.AckNum(790), checker.TCPFlagsMatch(header.TCPFlagAck, ^uint8(header.TCPFlagPsh)), ), @@ -2272,7 +2344,7 @@ func TestSegmentMerging(t *testing.T) { DstPort: c.Port, Flags: header.TCPFlagAck, SeqNum: 790, - AckNum: c.IRS.Add(1 + seqnum.Size(len(allData))), + AckNum: c.IRS.Add(11 + seqnum.Size(len(allData))), RcvWnd: 30000, }) }) @@ -2285,7 +2357,7 @@ func TestDelay(t *testing.T) { c.CreateConnected(789, 30000, -1 /* epRcvBuf */) - c.EP.SetSockOptInt(tcpip.DelayOption, 1) + c.EP.SetSockOptBool(tcpip.DelayOption, true) var allData []byte for i, data := range [][]byte{{0}, {1, 2, 3, 4}, {5, 6, 7}, {8, 9}, {10}, {11}} { @@ -2333,7 +2405,7 @@ func TestUndelay(t *testing.T) { c.CreateConnected(789, 30000, -1 /* epRcvBuf */) - c.EP.SetSockOptInt(tcpip.DelayOption, 1) + c.EP.SetSockOptBool(tcpip.DelayOption, true) allData := [][]byte{{0}, {1, 2, 3}} for i, data := range allData { @@ -2366,7 +2438,7 @@ func TestUndelay(t *testing.T) { // Check that we don't get the second packet yet. c.CheckNoPacketTimeout("delayed second packet transmitted", 100*time.Millisecond) - c.EP.SetSockOptInt(tcpip.DelayOption, 0) + c.EP.SetSockOptBool(tcpip.DelayOption, false) // Check that data is received. second := c.GetPacket() @@ -2403,8 +2475,8 @@ func TestMSSNotDelayed(t *testing.T) { fn func(tcpip.Endpoint) }{ {"no-op", func(tcpip.Endpoint) {}}, - {"delay", func(ep tcpip.Endpoint) { ep.SetSockOptInt(tcpip.DelayOption, 1) }}, - {"cork", func(ep tcpip.Endpoint) { ep.SetSockOpt(tcpip.CorkOption(1)) }}, + {"delay", func(ep tcpip.Endpoint) { ep.SetSockOptBool(tcpip.DelayOption, true) }}, + {"cork", func(ep tcpip.Endpoint) { ep.SetSockOptBool(tcpip.CorkOption, true) }}, } for _, test := range tests { @@ -2545,12 +2617,12 @@ func TestSetTTL(t *testing.T) { t.Fatalf("NewEndpoint failed: %v", err) } - if err := c.EP.SetSockOpt(tcpip.TTLOption(wantTTL)); err != nil { - t.Fatalf("SetSockOpt failed: %v", err) + if err := c.EP.SetSockOptInt(tcpip.TTLOption, int(wantTTL)); err != nil { + t.Fatalf("SetSockOptInt(TTLOption, %d) failed: %s", wantTTL, err) } if err := c.EP.Connect(tcpip.FullAddress{Addr: context.TestAddr, Port: context.TestPort}); err != tcpip.ErrConnectStarted { - t.Fatalf("Unexpected return value from Connect: %v", err) + t.Fatalf("Unexpected return value from Connect: %s", err) } // Receive SYN packet. @@ -2590,7 +2662,7 @@ func TestPassiveSendMSSLessThanMTU(t *testing.T) { // window scaling option. const rcvBufferSize = 0x20000 if err := ep.SetSockOptInt(tcpip.ReceiveBufferSizeOption, rcvBufferSize); err != nil { - t.Fatalf("SetSockOpt failed failed: %v", err) + t.Fatalf("SetSockOptInt(ReceiveBufferSizeOption, %d) failed failed: %s", rcvBufferSize, err) } if err := ep.Bind(tcpip.FullAddress{Port: context.StackPort}); err != nil { @@ -2636,26 +2708,24 @@ func TestSynCookiePassiveSendMSSLessThanMTU(t *testing.T) { // Set the SynRcvd threshold to zero to force a syn cookie based accept // to happen. - saved := tcp.SynRcvdCountThreshold - defer func() { - tcp.SynRcvdCountThreshold = saved - }() - tcp.SynRcvdCountThreshold = 0 + if err := c.Stack().SetTransportProtocolOption(tcp.ProtocolNumber, tcpip.TCPSynRcvdCountThresholdOption(0)); err != nil { + t.Fatalf("setting TCPSynRcvdCountThresholdOption to 0 failed: %s", err) + } // Create EP and start listening. wq := &waiter.Queue{} ep, err := c.Stack().NewEndpoint(tcp.ProtocolNumber, ipv4.ProtocolNumber, wq) if err != nil { - t.Fatalf("NewEndpoint failed: %v", err) + t.Fatalf("NewEndpoint failed: %s", err) } defer ep.Close() if err := ep.Bind(tcpip.FullAddress{Port: context.StackPort}); err != nil { - t.Fatalf("Bind failed: %v", err) + t.Fatalf("Bind failed: %s", err) } if err := ep.Listen(10); err != nil { - t.Fatalf("Listen failed: %v", err) + t.Fatalf("Listen failed: %s", err) } // Do 3-way handshake. @@ -2673,7 +2743,7 @@ func TestSynCookiePassiveSendMSSLessThanMTU(t *testing.T) { case <-ch: c.EP, _, err = ep.Accept() if err != nil { - t.Fatalf("Accept failed: %v", err) + t.Fatalf("Accept failed: %s", err) } case <-time.After(1 * time.Second): @@ -2734,7 +2804,7 @@ func TestSynOptionsOnActiveConnect(t *testing.T) { const rcvBufferSize = 0x20000 const wndScale = 2 if err := c.EP.SetSockOptInt(tcpip.ReceiveBufferSizeOption, rcvBufferSize); err != nil { - t.Fatalf("SetSockOpt failed failed: %v", err) + t.Fatalf("SetSockOptInt(ReceiveBufferSizeOption, %d) failed failed: %s", rcvBufferSize, err) } // Start connection attempt. @@ -3851,26 +3921,26 @@ func TestMinMaxBufferSizes(t *testing.T) { // Set values below the min. if err := ep.SetSockOptInt(tcpip.ReceiveBufferSizeOption, 199); err != nil { - t.Fatalf("GetSockOpt failed: %v", err) + t.Fatalf("SetSockOptInt(ReceiveBufferSizeOption, 199) failed: %s", err) } checkRecvBufferSize(t, ep, 200) if err := ep.SetSockOptInt(tcpip.SendBufferSizeOption, 299); err != nil { - t.Fatalf("GetSockOpt failed: %v", err) + t.Fatalf("SetSockOptInt(SendBufferSizeOption, 299) failed: %s", err) } checkSendBufferSize(t, ep, 300) // Set values above the max. if err := ep.SetSockOptInt(tcpip.ReceiveBufferSizeOption, 1+tcp.DefaultReceiveBufferSize*20); err != nil { - t.Fatalf("GetSockOpt failed: %v", err) + t.Fatalf("SetSockOptInt(ReceiveBufferSizeOption) failed: %s", err) } checkRecvBufferSize(t, ep, tcp.DefaultReceiveBufferSize*20) if err := ep.SetSockOptInt(tcpip.SendBufferSizeOption, 1+tcp.DefaultSendBufferSize*30); err != nil { - t.Fatalf("GetSockOpt failed: %v", err) + t.Fatalf("SetSockOptInt(SendBufferSizeOption) failed: %s", err) } checkSendBufferSize(t, ep, tcp.DefaultSendBufferSize*30) @@ -4116,11 +4186,11 @@ func TestConnectAvoidsBoundPorts(t *testing.T) { case "ipv4": case "ipv6": if err := ep.SetSockOptBool(tcpip.V6OnlyOption, true); err != nil { - t.Fatalf("SetSockOpt(V6OnlyOption(true)) failed: %v", err) + t.Fatalf("SetSockOptBool(V6OnlyOption(true)) failed: %s", err) } case "dual": if err := ep.SetSockOptBool(tcpip.V6OnlyOption, false); err != nil { - t.Fatalf("SetSockOpt(V6OnlyOption(false)) failed: %v", err) + t.Fatalf("SetSockOptBool(V6OnlyOption(false)) failed: %s", err) } default: t.Fatalf("unknown network: '%s'", network) @@ -4443,11 +4513,11 @@ func TestKeepalive(t *testing.T) { c.CreateConnected(789, 30000, -1 /* epRcvBuf */) - const keepAliveInterval = 10 * time.Millisecond - c.EP.SetSockOpt(tcpip.KeepaliveIdleOption(10 * time.Millisecond)) + const keepAliveInterval = 3 * time.Second + c.EP.SetSockOpt(tcpip.KeepaliveIdleOption(100 * time.Millisecond)) c.EP.SetSockOpt(tcpip.KeepaliveIntervalOption(keepAliveInterval)) - c.EP.SetSockOpt(tcpip.KeepaliveCountOption(5)) - c.EP.SetSockOpt(tcpip.KeepaliveEnabledOption(1)) + c.EP.SetSockOptInt(tcpip.KeepaliveCountOption, 5) + c.EP.SetSockOptBool(tcpip.KeepaliveEnabledOption, true) // 5 unacked keepalives are sent. ACK each one, and check that the // connection stays alive after 5. @@ -4538,7 +4608,7 @@ func TestKeepalive(t *testing.T) { // Sleep for a litte over the KeepAlive interval to make sure // the timer has time to fire after the last ACK and close the // close the socket. - time.Sleep(keepAliveInterval + 5*time.Millisecond) + time.Sleep(keepAliveInterval + keepAliveInterval/2) // The connection should be terminated after 5 unacked keepalives. // Send an ACK to trigger a RST from the stack as the endpoint should @@ -5073,25 +5143,23 @@ func TestListenSynRcvdQueueFull(t *testing.T) { } func TestListenBacklogFullSynCookieInUse(t *testing.T) { - saved := tcp.SynRcvdCountThreshold - defer func() { - tcp.SynRcvdCountThreshold = saved - }() - tcp.SynRcvdCountThreshold = 1 - c := context.New(t, defaultMTU) defer c.Cleanup() + if err := c.Stack().SetTransportProtocolOption(tcp.ProtocolNumber, tcpip.TCPSynRcvdCountThresholdOption(1)); err != nil { + t.Fatalf("setting TCPSynRcvdCountThresholdOption to 1 failed: %s", err) + } + // Create TCP endpoint. var err *tcpip.Error c.EP, err = c.Stack().NewEndpoint(tcp.ProtocolNumber, ipv4.ProtocolNumber, &c.WQ) if err != nil { - t.Fatalf("NewEndpoint failed: %v", err) + t.Fatalf("NewEndpoint failed: %s", err) } // Bind to wildcard. if err := c.EP.Bind(tcpip.FullAddress{Port: context.StackPort}); err != nil { - t.Fatalf("Bind failed: %v", err) + t.Fatalf("Bind failed: %s", err) } // Test acceptance. @@ -5099,7 +5167,7 @@ func TestListenBacklogFullSynCookieInUse(t *testing.T) { listenBacklog := 1 portOffset := uint16(0) if err := c.EP.Listen(listenBacklog); err != nil { - t.Fatalf("Listen failed: %v", err) + t.Fatalf("Listen failed: %s", err) } executeHandshake(t, c, context.TestPort+portOffset, false) @@ -5578,7 +5646,7 @@ func TestReceiveBufferAutoTuningApplicationLimited(t *testing.T) { return } if w := tcp.WindowSize(); w == 0 || w > uint16(wantRcvWnd) { - t.Errorf("expected a non-zero window: got %d, want <= wantRcvWnd", w, wantRcvWnd) + t.Errorf("expected a non-zero window: got %d, want <= wantRcvWnd", w) } }, )) @@ -5739,14 +5807,14 @@ func TestReceiveBufferAutoTuning(t *testing.T) { func TestDelayEnabled(t *testing.T) { c := context.New(t, defaultMTU) defer c.Cleanup() - checkDelayOption(t, c, false, 0) // Delay is disabled by default. + checkDelayOption(t, c, false, false) // Delay is disabled by default. for _, v := range []struct { delayEnabled tcp.DelayEnabled - wantDelayOption int + wantDelayOption bool }{ - {delayEnabled: false, wantDelayOption: 0}, - {delayEnabled: true, wantDelayOption: 1}, + {delayEnabled: false, wantDelayOption: false}, + {delayEnabled: true, wantDelayOption: true}, } { c := context.New(t, defaultMTU) defer c.Cleanup() @@ -5757,7 +5825,7 @@ func TestDelayEnabled(t *testing.T) { } } -func checkDelayOption(t *testing.T, c *context.Context, wantDelayEnabled tcp.DelayEnabled, wantDelayOption int) { +func checkDelayOption(t *testing.T, c *context.Context, wantDelayEnabled tcp.DelayEnabled, wantDelayOption bool) { t.Helper() var gotDelayEnabled tcp.DelayEnabled @@ -5772,12 +5840,12 @@ func checkDelayOption(t *testing.T, c *context.Context, wantDelayEnabled tcp.Del if err != nil { t.Fatalf("NewEndPoint(tcp, ipv4, new(waiter.Queue)) failed: %v", err) } - gotDelayOption, err := ep.GetSockOptInt(tcpip.DelayOption) + gotDelayOption, err := ep.GetSockOptBool(tcpip.DelayOption) if err != nil { - t.Fatalf("ep.GetSockOptInt(tcpip.DelayOption) failed: %v", err) + t.Fatalf("ep.GetSockOptBool(tcpip.DelayOption) failed: %s", err) } if gotDelayOption != wantDelayOption { - t.Errorf("ep.GetSockOptInt(tcpip.DelayOption) got: %d, want: %d", gotDelayOption, wantDelayOption) + t.Errorf("ep.GetSockOptBool(tcpip.DelayOption) got: %t, want: %t", gotDelayOption, wantDelayOption) } } @@ -6586,14 +6654,17 @@ func TestKeepaliveWithUserTimeout(t *testing.T) { origEstablishedTimedout := c.Stack().Stats().TCP.EstablishedTimedout.Value() - const keepAliveInterval = 10 * time.Millisecond - c.EP.SetSockOpt(tcpip.KeepaliveIdleOption(10 * time.Millisecond)) + const keepAliveInterval = 3 * time.Second + c.EP.SetSockOpt(tcpip.KeepaliveIdleOption(100 * time.Millisecond)) c.EP.SetSockOpt(tcpip.KeepaliveIntervalOption(keepAliveInterval)) - c.EP.SetSockOpt(tcpip.KeepaliveCountOption(10)) - c.EP.SetSockOpt(tcpip.KeepaliveEnabledOption(1)) - - // Set userTimeout to be the duration for 3 keepalive probes. - userTimeout := 30 * time.Millisecond + c.EP.SetSockOptInt(tcpip.KeepaliveCountOption, 10) + c.EP.SetSockOptBool(tcpip.KeepaliveEnabledOption, true) + + // Set userTimeout to be the duration to be 1 keepalive + // probes. Which means that after the first probe is sent + // the second one should cause the connection to be + // closed due to userTimeout being hit. + userTimeout := 1 * keepAliveInterval c.EP.SetSockOpt(tcpip.TCPUserTimeoutOption(userTimeout)) // Check that the connection is still alive. @@ -6601,28 +6672,23 @@ func TestKeepaliveWithUserTimeout(t *testing.T) { t.Fatalf("got c.EP.Read(nil) = %v, want = %v", err, tcpip.ErrWouldBlock) } - // Now receive 2 keepalives, but don't ACK them. The connection should - // be reset when the 3rd one should be sent due to userTimeout being - // 30ms and each keepalive probe should be sent 10ms apart as set above after - // the connection has been idle for 10ms. - for i := 0; i < 2; i++ { - b := c.GetPacket() - checker.IPv4(t, b, - checker.TCP( - checker.DstPort(context.TestPort), - checker.SeqNum(uint32(c.IRS)), - checker.AckNum(uint32(790)), - checker.TCPFlags(header.TCPFlagAck), - ), - ) - } + // Now receive 1 keepalives, but don't ACK it. + b := c.GetPacket() + checker.IPv4(t, b, + checker.TCP( + checker.DstPort(context.TestPort), + checker.SeqNum(uint32(c.IRS)), + checker.AckNum(uint32(790)), + checker.TCPFlags(header.TCPFlagAck), + ), + ) // Sleep for a litte over the KeepAlive interval to make sure // the timer has time to fire after the last ACK and close the // close the socket. - time.Sleep(keepAliveInterval + 5*time.Millisecond) + time.Sleep(keepAliveInterval + keepAliveInterval/2) - // The connection should be terminated after 30ms. + // The connection should be closed with a timeout. // Send an ACK to trigger a RST from the stack as the endpoint should // be dead. c.SendPacket(nil, &context.Headers{ diff --git a/pkg/tcpip/transport/tcp/tcp_timestamp_test.go b/pkg/tcpip/transport/tcp/tcp_timestamp_test.go index a641e953d..8edbff964 100644 --- a/pkg/tcpip/transport/tcp/tcp_timestamp_test.go +++ b/pkg/tcpip/transport/tcp/tcp_timestamp_test.go @@ -127,16 +127,14 @@ func TestTimeStampDisabledConnect(t *testing.T) { } func timeStampEnabledAccept(t *testing.T, cookieEnabled bool, wndScale int, wndSize uint16) { - savedSynCountThreshold := tcp.SynRcvdCountThreshold - defer func() { - tcp.SynRcvdCountThreshold = savedSynCountThreshold - }() + c := context.New(t, defaultMTU) + defer c.Cleanup() if cookieEnabled { - tcp.SynRcvdCountThreshold = 0 + if err := c.Stack().SetTransportProtocolOption(tcp.ProtocolNumber, tcpip.TCPSynRcvdCountThresholdOption(0)); err != nil { + t.Fatalf("setting TCPSynRcvdCountThresholdOption to 0 failed: %s", err) + } } - c := context.New(t, defaultMTU) - defer c.Cleanup() t.Logf("Test w/ CookieEnabled = %v", cookieEnabled) tsVal := rand.Uint32() @@ -148,7 +146,7 @@ func timeStampEnabledAccept(t *testing.T, cookieEnabled bool, wndScale int, wndS copy(view, data) if _, _, err := c.EP.Write(tcpip.SlicePayload(view), tcpip.WriteOptions{}); err != nil { - t.Fatalf("Unexpected error from Write: %v", err) + t.Fatalf("Unexpected error from Write: %s", err) } // Check that data is received and that the timestamp option TSEcr field @@ -190,17 +188,15 @@ func TestTimeStampEnabledAccept(t *testing.T) { } func timeStampDisabledAccept(t *testing.T, cookieEnabled bool, wndScale int, wndSize uint16) { - savedSynCountThreshold := tcp.SynRcvdCountThreshold - defer func() { - tcp.SynRcvdCountThreshold = savedSynCountThreshold - }() - if cookieEnabled { - tcp.SynRcvdCountThreshold = 0 - } - c := context.New(t, defaultMTU) defer c.Cleanup() + if cookieEnabled { + if err := c.Stack().SetTransportProtocolOption(tcp.ProtocolNumber, tcpip.TCPSynRcvdCountThresholdOption(0)); err != nil { + t.Fatalf("setting TCPSynRcvdCountThresholdOption to 0 failed: %s", err) + } + } + t.Logf("Test w/ CookieEnabled = %v", cookieEnabled) c.AcceptWithOptions(wndScale, header.TCPSynOptions{MSS: defaultIPv4MSS}) @@ -211,7 +207,7 @@ func timeStampDisabledAccept(t *testing.T, cookieEnabled bool, wndScale int, wnd copy(view, data) if _, _, err := c.EP.Write(tcpip.SlicePayload(view), tcpip.WriteOptions{}); err != nil { - t.Fatalf("Unexpected error from Write: %v", err) + t.Fatalf("Unexpected error from Write: %s", err) } // Check that data is received and that the timestamp option is disabled diff --git a/pkg/tcpip/transport/tcp/testing/context/context.go b/pkg/tcpip/transport/tcp/testing/context/context.go index 1e9a0dea3..7b1d72cf4 100644 --- a/pkg/tcpip/transport/tcp/testing/context/context.go +++ b/pkg/tcpip/transport/tcp/testing/context/context.go @@ -152,6 +152,13 @@ func New(t *testing.T, mtu uint32) *Context { t.Fatalf("SetTransportProtocolOption failed: %v", err) } + // Increase minimum RTO in tests to avoid test flakes due to early + // retransmit in case the test executors are overloaded and cause timers + // to fire earlier than expected. + if err := s.SetTransportProtocolOption(tcp.ProtocolNumber, tcpip.TCPMinRTOOption(3*time.Second)); err != nil { + t.Fatalf("failed to set stack-wide minRTO: %s", err) + } + // Some of the congestion control tests send up to 640 packets, we so // set the channel size to 1000. ep := channel.New(1000, mtu, "") @@ -204,6 +211,7 @@ func (c *Context) Cleanup() { if c.EP != nil { c.EP.Close() } + c.Stack().Close() } // Stack returns a reference to the stack in the Context. @@ -216,7 +224,8 @@ func (c *Context) Stack() *stack.Stack { func (c *Context) CheckNoPacketTimeout(errMsg string, wait time.Duration) { c.t.Helper() - ctx, _ := context.WithTimeout(context.Background(), wait) + ctx, cancel := context.WithTimeout(context.Background(), wait) + defer cancel() if _, ok := c.linkEP.ReadContext(ctx); ok { c.t.Fatal(errMsg) } @@ -234,7 +243,8 @@ func (c *Context) CheckNoPacket(errMsg string) { func (c *Context) GetPacket() []byte { c.t.Helper() - ctx, _ := context.WithTimeout(context.Background(), 2*time.Second) + ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second) + defer cancel() p, ok := c.linkEP.ReadContext(ctx) if !ok { c.t.Fatalf("Packet wasn't written out") @@ -306,7 +316,7 @@ func (c *Context) SendICMPPacket(typ header.ICMPv4Type, code uint8, p1, p2 []byt copy(icmp[header.ICMPv4PayloadOffset:], p2) // Inject packet. - c.linkEP.InjectInbound(ipv4.ProtocolNumber, tcpip.PacketBuffer{ + c.linkEP.InjectInbound(ipv4.ProtocolNumber, stack.PacketBuffer{ Data: buf.ToVectorisedView(), }) } @@ -362,7 +372,7 @@ func (c *Context) BuildSegmentWithAddrs(payload []byte, h *Headers, src, dst tcp // SendSegment sends a TCP segment that has already been built and written to a // buffer.VectorisedView. func (c *Context) SendSegment(s buffer.VectorisedView) { - c.linkEP.InjectInbound(ipv4.ProtocolNumber, tcpip.PacketBuffer{ + c.linkEP.InjectInbound(ipv4.ProtocolNumber, stack.PacketBuffer{ Data: s, }) } @@ -370,7 +380,7 @@ func (c *Context) SendSegment(s buffer.VectorisedView) { // SendPacket builds and sends a TCP segment(with the provided payload & TCP // headers) in an IPv4 packet via the link layer endpoint. func (c *Context) SendPacket(payload []byte, h *Headers) { - c.linkEP.InjectInbound(ipv4.ProtocolNumber, tcpip.PacketBuffer{ + c.linkEP.InjectInbound(ipv4.ProtocolNumber, stack.PacketBuffer{ Data: c.BuildSegment(payload, h), }) } @@ -379,7 +389,7 @@ func (c *Context) SendPacket(payload []byte, h *Headers) { // & TCPheaders) in an IPv4 packet via the link layer endpoint using the // provided source and destination IPv4 addresses. func (c *Context) SendPacketWithAddrs(payload []byte, h *Headers, src, dst tcpip.Address) { - c.linkEP.InjectInbound(ipv4.ProtocolNumber, tcpip.PacketBuffer{ + c.linkEP.InjectInbound(ipv4.ProtocolNumber, stack.PacketBuffer{ Data: c.BuildSegmentWithAddrs(payload, h, src, dst), }) } @@ -414,6 +424,8 @@ func (c *Context) SendAckWithSACK(seq seqnum.Value, bytesReceived int, sackBlock // verifies that the packet packet payload of packet matches the slice // of data indicated by offset & size. func (c *Context) ReceiveAndCheckPacket(data []byte, offset, size int) { + c.t.Helper() + c.ReceiveAndCheckPacketWithOptions(data, offset, size, 0) } @@ -422,6 +434,8 @@ func (c *Context) ReceiveAndCheckPacket(data []byte, offset, size int) { // data indicated by offset & size and skips optlen bytes in addition to the IP // TCP headers when comparing the data. func (c *Context) ReceiveAndCheckPacketWithOptions(data []byte, offset, size, optlen int) { + c.t.Helper() + b := c.GetPacket() checker.IPv4(c.t, b, checker.PayloadLen(size+header.TCPMinimumSize+optlen), @@ -444,6 +458,8 @@ func (c *Context) ReceiveAndCheckPacketWithOptions(data []byte, offset, size, op // data indicated by offset & size. It returns true if a packet was received and // processed. func (c *Context) ReceiveNonBlockingAndCheckPacket(data []byte, offset, size int) bool { + c.t.Helper() + b := c.GetPacketNonBlocking() if b == nil { return false @@ -485,7 +501,8 @@ func (c *Context) CreateV6Endpoint(v6only bool) { func (c *Context) GetV6Packet() []byte { c.t.Helper() - ctx, _ := context.WithTimeout(context.Background(), 2*time.Second) + ctx, cancel := context.WithTimeout(context.Background(), 2*time.Second) + defer cancel() p, ok := c.linkEP.ReadContext(ctx) if !ok { c.t.Fatalf("Packet wasn't written out") @@ -547,7 +564,7 @@ func (c *Context) SendV6PacketWithAddrs(payload []byte, h *Headers, src, dst tcp t.SetChecksum(^t.CalculateChecksum(xsum)) // Inject packet. - c.linkEP.InjectInbound(ipv6.ProtocolNumber, tcpip.PacketBuffer{ + c.linkEP.InjectInbound(ipv6.ProtocolNumber, stack.PacketBuffer{ Data: buf.ToVectorisedView(), }) } @@ -566,6 +583,8 @@ func (c *Context) CreateConnected(iss seqnum.Value, rcvWnd seqnum.Size, epRcvBuf // // PreCondition: c.EP must already be created. func (c *Context) Connect(iss seqnum.Value, rcvWnd seqnum.Size, options []byte) { + c.t.Helper() + // Start connection attempt. waitEntry, notifyCh := waiter.NewChannelEntry(nil) c.WQ.EventRegister(&waitEntry, waiter.EventOut) diff --git a/pkg/tcpip/transport/tcpconntrack/BUILD b/pkg/tcpip/transport/tcpconntrack/BUILD index 3ad6994a7..2025ff757 100644 --- a/pkg/tcpip/transport/tcpconntrack/BUILD +++ b/pkg/tcpip/transport/tcpconntrack/BUILD @@ -9,6 +9,7 @@ go_library( deps = [ "//pkg/tcpip/header", "//pkg/tcpip/seqnum", + "//pkg/tcpip/transport/tcp", ], ) diff --git a/pkg/tcpip/transport/tcpconntrack/tcp_conntrack.go b/pkg/tcpip/transport/tcpconntrack/tcp_conntrack.go index 93712cd45..30d05200f 100644 --- a/pkg/tcpip/transport/tcpconntrack/tcp_conntrack.go +++ b/pkg/tcpip/transport/tcpconntrack/tcp_conntrack.go @@ -20,6 +20,7 @@ package tcpconntrack import ( "gvisor.dev/gvisor/pkg/tcpip/header" "gvisor.dev/gvisor/pkg/tcpip/seqnum" + "gvisor.dev/gvisor/pkg/tcpip/transport/tcp" ) // Result is returned when the state of a TCB is updated in response to an @@ -311,17 +312,7 @@ type stream struct { // the window is zero, if it's a packet with no payload and sequence number // equal to una. func (s *stream) acceptable(segSeq seqnum.Value, segLen seqnum.Size) bool { - wnd := s.una.Size(s.end) - if wnd == 0 { - return segLen == 0 && segSeq == s.una - } - - // Make sure [segSeq, seqSeq+segLen) is non-empty. - if segLen == 0 { - segLen = 1 - } - - return seqnum.Overlap(s.una, wnd, segSeq, segLen) + return tcp.Acceptable(segSeq, segLen, s.una, s.end) } // closed determines if the stream has already been closed. This happens when diff --git a/pkg/tcpip/transport/udp/BUILD b/pkg/tcpip/transport/udp/BUILD index adc908e24..b5d2d0ba6 100644 --- a/pkg/tcpip/transport/udp/BUILD +++ b/pkg/tcpip/transport/udp/BUILD @@ -32,7 +32,6 @@ go_library( "//pkg/tcpip", "//pkg/tcpip/buffer", "//pkg/tcpip/header", - "//pkg/tcpip/iptables", "//pkg/tcpip/ports", "//pkg/tcpip/stack", "//pkg/tcpip/transport/raw", diff --git a/pkg/tcpip/transport/udp/endpoint.go b/pkg/tcpip/transport/udp/endpoint.go index c9cbed8f4..edb54f0be 100644 --- a/pkg/tcpip/transport/udp/endpoint.go +++ b/pkg/tcpip/transport/udp/endpoint.go @@ -19,7 +19,6 @@ import ( "gvisor.dev/gvisor/pkg/tcpip" "gvisor.dev/gvisor/pkg/tcpip/buffer" "gvisor.dev/gvisor/pkg/tcpip/header" - "gvisor.dev/gvisor/pkg/tcpip/iptables" "gvisor.dev/gvisor/pkg/tcpip/ports" "gvisor.dev/gvisor/pkg/tcpip/stack" "gvisor.dev/gvisor/pkg/waiter" @@ -29,9 +28,11 @@ import ( type udpPacket struct { udpPacketEntry senderAddress tcpip.FullAddress + packetInfo tcpip.IPPacketInfo data buffer.VectorisedView `state:".(buffer.VectorisedView)"` timestamp int64 - tos uint8 + // tos stores either the receiveTOS or receiveTClass value. + tos uint8 } // EndpointState represents the state of a UDP endpoint. @@ -118,6 +119,13 @@ type endpoint struct { // as ancillary data to ControlMessages on Read. receiveTOS bool + // receiveTClass determines if the incoming IPv6 TClass header field is + // passed as ancillary data to ControlMessages on Read. + receiveTClass bool + + // receiveIPPacketInfo determines if the packet info is returned by Read. + receiveIPPacketInfo bool + // shutdownFlags represent the current shutdown state of the endpoint. shutdownFlags tcpip.ShutdownFlags @@ -135,6 +143,9 @@ type endpoint struct { // TODO(b/142022063): Add ability to save and restore per endpoint stats. stats tcpip.TransportEndpointStats `state:"nosave"` + + // owner is used to get uid and gid of the packet. + owner tcpip.PacketOwner } // +stateify savable @@ -177,6 +188,11 @@ func (e *endpoint) UniqueID() uint64 { return e.uniqueID } +// Abort implements stack.TransportEndpoint.Abort. +func (e *endpoint) Abort() { + e.Close() +} + // Close puts the endpoint in a closed state and frees all resources // associated with it. func (e *endpoint) Close() { @@ -220,7 +236,7 @@ func (e *endpoint) Close() { func (e *endpoint) ModerateRecvBuf(copied int) {} // IPTables implements tcpip.Endpoint.IPTables. -func (e *endpoint) IPTables() (iptables.IPTables, error) { +func (e *endpoint) IPTables() (stack.IPTables, error) { return e.stack.IPTables(), nil } @@ -254,11 +270,22 @@ func (e *endpoint) Read(addr *tcpip.FullAddress) (buffer.View, tcpip.ControlMess } e.mu.RLock() receiveTOS := e.receiveTOS + receiveTClass := e.receiveTClass + receiveIPPacketInfo := e.receiveIPPacketInfo e.mu.RUnlock() if receiveTOS { cm.HasTOS = true cm.TOS = p.tos } + if receiveTClass { + cm.HasTClass = true + // Although TClass is an 8-bit value it's read in the CMsg as a uint32. + cm.TClass = uint32(p.tos) + } + if receiveIPPacketInfo { + cm.HasIPPacketInfo = true + cm.PacketInfo = p.packetInfo + } return p.data.ToView(), cm, nil } @@ -418,19 +445,19 @@ func (e *endpoint) write(p tcpip.Payloader, opts tcpip.WriteOptions) (int64, <-c return 0, nil, tcpip.ErrBroadcastDisabled } - netProto, err := e.checkV4Mapped(to) + dst, netProto, err := e.checkV4MappedLocked(*to) if err != nil { return 0, nil, err } - r, _, err := e.connectRoute(nicID, *to, netProto) + r, _, err := e.connectRoute(nicID, dst, netProto) if err != nil { return 0, nil, err } defer r.Release() route = &r - dstPort = to.Port + dstPort = dst.Port } if route.IsResolutionRequired() { @@ -460,7 +487,7 @@ func (e *endpoint) write(p tcpip.Payloader, opts tcpip.WriteOptions) (int64, <-c useDefaultTTL = false } - if err := sendUDP(route, buffer.View(v).ToVectorisedView(), e.ID.LocalPort, dstPort, ttl, useDefaultTTL, e.sendTOS); err != nil { + if err := sendUDP(route, buffer.View(v).ToVectorisedView(), e.ID.LocalPort, dstPort, ttl, useDefaultTTL, e.sendTOS, e.owner); err != nil { return 0, nil, err } return int64(len(v)), nil, nil @@ -474,11 +501,42 @@ func (e *endpoint) Peek([][]byte) (int64, tcpip.ControlMessages, *tcpip.Error) { // SetSockOptBool implements tcpip.Endpoint.SetSockOptBool. func (e *endpoint) SetSockOptBool(opt tcpip.SockOptBool, v bool) *tcpip.Error { switch opt { + case tcpip.BroadcastOption: + e.mu.Lock() + e.broadcast = v + e.mu.Unlock() + + case tcpip.MulticastLoopOption: + e.mu.Lock() + e.multicastLoop = v + e.mu.Unlock() + case tcpip.ReceiveTOSOption: e.mu.Lock() e.receiveTOS = v e.mu.Unlock() - return nil + + case tcpip.ReceiveTClassOption: + // We only support this option on v6 endpoints. + if e.NetProto != header.IPv6ProtocolNumber { + return tcpip.ErrNotSupported + } + + e.mu.Lock() + e.receiveTClass = v + e.mu.Unlock() + + case tcpip.ReceiveIPPacketInfoOption: + e.mu.Lock() + e.receiveIPPacketInfo = v + e.mu.Unlock() + + case tcpip.ReuseAddressOption: + + case tcpip.ReusePortOption: + e.mu.Lock() + e.reusePort = v + e.mu.Unlock() case tcpip.V6OnlyOption: // We only recognize this option on v6 endpoints. @@ -502,28 +560,44 @@ func (e *endpoint) SetSockOptBool(opt tcpip.SockOptBool, v bool) *tcpip.Error { // SetSockOptInt implements tcpip.Endpoint.SetSockOptInt. func (e *endpoint) SetSockOptInt(opt tcpip.SockOptInt, v int) *tcpip.Error { - return nil -} + switch opt { + case tcpip.MulticastTTLOption: + e.mu.Lock() + e.multicastTTL = uint8(v) + e.mu.Unlock() -// SetSockOpt implements tcpip.Endpoint.SetSockOpt. -func (e *endpoint) SetSockOpt(opt interface{}) *tcpip.Error { - switch v := opt.(type) { case tcpip.TTLOption: e.mu.Lock() e.ttl = uint8(v) e.mu.Unlock() - case tcpip.MulticastTTLOption: + case tcpip.IPv4TOSOption: e.mu.Lock() - e.multicastTTL = uint8(v) + e.sendTOS = uint8(v) + e.mu.Unlock() + + case tcpip.IPv6TrafficClassOption: + e.mu.Lock() + e.sendTOS = uint8(v) e.mu.Unlock() + case tcpip.ReceiveBufferSizeOption: + case tcpip.SendBufferSizeOption: + + } + + return nil +} + +// SetSockOpt implements tcpip.Endpoint.SetSockOpt. +func (e *endpoint) SetSockOpt(opt interface{}) *tcpip.Error { + switch v := opt.(type) { case tcpip.MulticastInterfaceOption: e.mu.Lock() defer e.mu.Unlock() fa := tcpip.FullAddress{Addr: v.InterfaceAddr} - netProto, err := e.checkV4Mapped(&fa) + fa, netProto, err := e.checkV4MappedLocked(fa) if err != nil { return err } @@ -641,16 +715,6 @@ func (e *endpoint) SetSockOpt(opt interface{}) *tcpip.Error { e.multicastMemberships[memToRemoveIndex] = e.multicastMemberships[len(e.multicastMemberships)-1] e.multicastMemberships = e.multicastMemberships[:len(e.multicastMemberships)-1] - case tcpip.MulticastLoopOption: - e.mu.Lock() - e.multicastLoop = bool(v) - e.mu.Unlock() - - case tcpip.ReusePortOption: - e.mu.Lock() - e.reusePort = v != 0 - e.mu.Unlock() - case tcpip.BindToDeviceOption: id := tcpip.NICID(v) if id != 0 && !e.stack.HasNIC(id) { @@ -659,26 +723,6 @@ func (e *endpoint) SetSockOpt(opt interface{}) *tcpip.Error { e.mu.Lock() e.bindToDevice = id e.mu.Unlock() - return nil - - case tcpip.BroadcastOption: - e.mu.Lock() - e.broadcast = v != 0 - e.mu.Unlock() - - return nil - - case tcpip.IPv4TOSOption: - e.mu.Lock() - e.sendTOS = uint8(v) - e.mu.Unlock() - return nil - - case tcpip.IPv6TrafficClassOption: - e.mu.Lock() - e.sendTOS = uint8(v) - e.mu.Unlock() - return nil } return nil } @@ -686,12 +730,54 @@ func (e *endpoint) SetSockOpt(opt interface{}) *tcpip.Error { // GetSockOptBool implements tcpip.Endpoint.GetSockOptBool. func (e *endpoint) GetSockOptBool(opt tcpip.SockOptBool) (bool, *tcpip.Error) { switch opt { + case tcpip.BroadcastOption: + e.mu.RLock() + v := e.broadcast + e.mu.RUnlock() + return v, nil + + case tcpip.KeepaliveEnabledOption: + return false, nil + + case tcpip.MulticastLoopOption: + e.mu.RLock() + v := e.multicastLoop + e.mu.RUnlock() + return v, nil + case tcpip.ReceiveTOSOption: e.mu.RLock() v := e.receiveTOS e.mu.RUnlock() return v, nil + case tcpip.ReceiveTClassOption: + // We only support this option on v6 endpoints. + if e.NetProto != header.IPv6ProtocolNumber { + return false, tcpip.ErrNotSupported + } + + e.mu.RLock() + v := e.receiveTClass + e.mu.RUnlock() + return v, nil + + case tcpip.ReceiveIPPacketInfoOption: + e.mu.RLock() + v := e.receiveIPPacketInfo + e.mu.RUnlock() + return v, nil + + case tcpip.ReuseAddressOption: + return false, nil + + case tcpip.ReusePortOption: + e.mu.RLock() + v := e.reusePort + e.mu.RUnlock() + + return v, nil + case tcpip.V6OnlyOption: // We only recognize this option on v6 endpoints. if e.NetProto != header.IPv6ProtocolNumber { @@ -703,14 +789,33 @@ func (e *endpoint) GetSockOptBool(opt tcpip.SockOptBool) (bool, *tcpip.Error) { e.mu.RUnlock() return v, nil - } - return false, tcpip.ErrUnknownProtocolOption + default: + return false, tcpip.ErrUnknownProtocolOption + } } // GetSockOptInt implements tcpip.Endpoint.GetSockOptInt. func (e *endpoint) GetSockOptInt(opt tcpip.SockOptInt) (int, *tcpip.Error) { switch opt { + case tcpip.IPv4TOSOption: + e.mu.RLock() + v := int(e.sendTOS) + e.mu.RUnlock() + return v, nil + + case tcpip.IPv6TrafficClassOption: + e.mu.RLock() + v := int(e.sendTOS) + e.mu.RUnlock() + return v, nil + + case tcpip.MulticastTTLOption: + e.mu.Lock() + v := int(e.multicastTTL) + e.mu.Unlock() + return v, nil + case tcpip.ReceiveQueueSizeOption: v := 0 e.rcvMu.Lock() @@ -732,29 +837,22 @@ func (e *endpoint) GetSockOptInt(opt tcpip.SockOptInt) (int, *tcpip.Error) { v := e.rcvBufSizeMax e.rcvMu.Unlock() return v, nil - } - return -1, tcpip.ErrUnknownProtocolOption + case tcpip.TTLOption: + e.mu.Lock() + v := int(e.ttl) + e.mu.Unlock() + return v, nil + + default: + return -1, tcpip.ErrUnknownProtocolOption + } } // GetSockOpt implements tcpip.Endpoint.GetSockOpt. func (e *endpoint) GetSockOpt(opt interface{}) *tcpip.Error { switch o := opt.(type) { case tcpip.ErrorOption: - return nil - - case *tcpip.TTLOption: - e.mu.Lock() - *o = tcpip.TTLOption(e.ttl) - e.mu.Unlock() - return nil - - case *tcpip.MulticastTTLOption: - e.mu.Lock() - *o = tcpip.MulticastTTLOption(e.multicastTTL) - e.mu.Unlock() - return nil - case *tcpip.MulticastInterfaceOption: e.mu.Lock() *o = tcpip.MulticastInterfaceOption{ @@ -762,72 +860,21 @@ func (e *endpoint) GetSockOpt(opt interface{}) *tcpip.Error { e.multicastAddr, } e.mu.Unlock() - return nil - - case *tcpip.MulticastLoopOption: - e.mu.RLock() - v := e.multicastLoop - e.mu.RUnlock() - - *o = tcpip.MulticastLoopOption(v) - return nil - - case *tcpip.ReuseAddressOption: - *o = 0 - return nil - - case *tcpip.ReusePortOption: - e.mu.RLock() - v := e.reusePort - e.mu.RUnlock() - - *o = 0 - if v { - *o = 1 - } - return nil case *tcpip.BindToDeviceOption: e.mu.RLock() *o = tcpip.BindToDeviceOption(e.bindToDevice) e.mu.RUnlock() - return nil - - case *tcpip.KeepaliveEnabledOption: - *o = 0 - return nil - - case *tcpip.BroadcastOption: - e.mu.RLock() - v := e.broadcast - e.mu.RUnlock() - - *o = 0 - if v { - *o = 1 - } - return nil - - case *tcpip.IPv4TOSOption: - e.mu.RLock() - *o = tcpip.IPv4TOSOption(e.sendTOS) - e.mu.RUnlock() - return nil - - case *tcpip.IPv6TrafficClassOption: - e.mu.RLock() - *o = tcpip.IPv6TrafficClassOption(e.sendTOS) - e.mu.RUnlock() - return nil default: return tcpip.ErrUnknownProtocolOption } + return nil } // sendUDP sends a UDP segment via the provided network endpoint and under the // provided identity. -func sendUDP(r *stack.Route, data buffer.VectorisedView, localPort, remotePort uint16, ttl uint8, useDefaultTTL bool, tos uint8) *tcpip.Error { +func sendUDP(r *stack.Route, data buffer.VectorisedView, localPort, remotePort uint16, ttl uint8, useDefaultTTL bool, tos uint8, owner tcpip.PacketOwner) *tcpip.Error { // Allocate a buffer for the UDP header. hdr := buffer.NewPrependable(header.UDPMinimumSize + int(r.MaxHeaderLength())) @@ -853,10 +900,11 @@ func sendUDP(r *stack.Route, data buffer.VectorisedView, localPort, remotePort u if useDefaultTTL { ttl = r.DefaultTTL() } - if err := r.WritePacket(nil /* gso */, stack.NetworkHeaderParams{Protocol: ProtocolNumber, TTL: ttl, TOS: tos}, tcpip.PacketBuffer{ + if err := r.WritePacket(nil /* gso */, stack.NetworkHeaderParams{Protocol: ProtocolNumber, TTL: ttl, TOS: tos}, stack.PacketBuffer{ Header: hdr, Data: data, TransportHeader: buffer.View(udp), + Owner: owner, }); err != nil { r.Stats().UDP.PacketSendErrors.Increment() return err @@ -867,13 +915,14 @@ func sendUDP(r *stack.Route, data buffer.VectorisedView, localPort, remotePort u return nil } -func (e *endpoint) checkV4Mapped(addr *tcpip.FullAddress) (tcpip.NetworkProtocolNumber, *tcpip.Error) { - unwrapped, netProto, err := e.TransportEndpointInfo.AddrNetProto(*addr, e.v6only) +// checkV4MappedLocked determines the effective network protocol and converts +// addr to its canonical form. +func (e *endpoint) checkV4MappedLocked(addr tcpip.FullAddress) (tcpip.FullAddress, tcpip.NetworkProtocolNumber, *tcpip.Error) { + unwrapped, netProto, err := e.TransportEndpointInfo.AddrNetProtoLocked(addr, e.v6only) if err != nil { - return 0, err + return tcpip.FullAddress{}, 0, err } - *addr = unwrapped - return netProto, nil + return unwrapped, netProto, nil } // Disconnect implements tcpip.Endpoint.Disconnect. @@ -921,10 +970,6 @@ func (e *endpoint) Disconnect() *tcpip.Error { // Connect connects the endpoint to its peer. Specifying a NIC is optional. func (e *endpoint) Connect(addr tcpip.FullAddress) *tcpip.Error { - netProto, err := e.checkV4Mapped(&addr) - if err != nil { - return err - } if addr.Port == 0 { // We don't support connecting to port zero. return tcpip.ErrInvalidEndpointState @@ -952,6 +997,11 @@ func (e *endpoint) Connect(addr tcpip.FullAddress) *tcpip.Error { return tcpip.ErrInvalidEndpointState } + addr, netProto, err := e.checkV4MappedLocked(addr) + if err != nil { + return err + } + r, nicID, err := e.connectRoute(nicID, addr, netProto) if err != nil { return err @@ -1079,7 +1129,7 @@ func (e *endpoint) bindLocked(addr tcpip.FullAddress) *tcpip.Error { return tcpip.ErrInvalidEndpointState } - netProto, err := e.checkV4Mapped(&addr) + addr, netProto, err := e.checkV4MappedLocked(addr) if err != nil { return err } @@ -1198,7 +1248,7 @@ func (e *endpoint) Readiness(mask waiter.EventMask) waiter.EventMask { // HandlePacket is called by the stack when new packets arrive to this transport // endpoint. -func (e *endpoint) HandlePacket(r *stack.Route, id stack.TransportEndpointID, pkt tcpip.PacketBuffer) { +func (e *endpoint) HandlePacket(r *stack.Route, id stack.TransportEndpointID, pkt stack.PacketBuffer) { // Get the header then trim it from the view. hdr := header.UDP(pkt.Data.First()) if int(hdr.Length()) > pkt.Data.Size() { @@ -1247,6 +1297,11 @@ func (e *endpoint) HandlePacket(r *stack.Route, id stack.TransportEndpointID, pk switch r.NetProto { case header.IPv4ProtocolNumber: packet.tos, _ = header.IPv4(pkt.NetworkHeader).TOS() + packet.packetInfo.LocalAddr = r.LocalAddress + packet.packetInfo.DestinationAddr = r.RemoteAddress + packet.packetInfo.NIC = r.NICID() + case header.IPv6ProtocolNumber: + packet.tos, _ = header.IPv6(pkt.NetworkHeader).TOS() } packet.timestamp = e.stack.NowNanoseconds() @@ -1260,7 +1315,7 @@ func (e *endpoint) HandlePacket(r *stack.Route, id stack.TransportEndpointID, pk } // HandleControlPacket implements stack.TransportEndpoint.HandleControlPacket. -func (e *endpoint) HandleControlPacket(id stack.TransportEndpointID, typ stack.ControlType, extra uint32, pkt tcpip.PacketBuffer) { +func (e *endpoint) HandleControlPacket(id stack.TransportEndpointID, typ stack.ControlType, extra uint32, pkt stack.PacketBuffer) { } // State implements tcpip.Endpoint.State. @@ -1290,3 +1345,7 @@ func (*endpoint) Wait() {} func isBroadcastOrMulticast(a tcpip.Address) bool { return a == header.IPv4Broadcast || header.IsV4MulticastAddress(a) || header.IsV6MulticastAddress(a) } + +func (e *endpoint) SetOwner(owner tcpip.PacketOwner) { + e.owner = owner +} diff --git a/pkg/tcpip/transport/udp/endpoint_state.go b/pkg/tcpip/transport/udp/endpoint_state.go index 43fb047ed..466bd9381 100644 --- a/pkg/tcpip/transport/udp/endpoint_state.go +++ b/pkg/tcpip/transport/udp/endpoint_state.go @@ -69,6 +69,9 @@ func (e *endpoint) afterLoad() { // Resume implements tcpip.ResumableEndpoint.Resume. func (e *endpoint) Resume(s *stack.Stack) { + e.mu.Lock() + defer e.mu.Unlock() + e.stack = s for _, m := range e.multicastMemberships { diff --git a/pkg/tcpip/transport/udp/forwarder.go b/pkg/tcpip/transport/udp/forwarder.go index fc706ede2..a674ceb68 100644 --- a/pkg/tcpip/transport/udp/forwarder.go +++ b/pkg/tcpip/transport/udp/forwarder.go @@ -43,7 +43,7 @@ func NewForwarder(s *stack.Stack, handler func(*ForwarderRequest)) *Forwarder { // // This function is expected to be passed as an argument to the // stack.SetTransportProtocolHandler function. -func (f *Forwarder) HandlePacket(r *stack.Route, id stack.TransportEndpointID, pkt tcpip.PacketBuffer) bool { +func (f *Forwarder) HandlePacket(r *stack.Route, id stack.TransportEndpointID, pkt stack.PacketBuffer) bool { f.handler(&ForwarderRequest{ stack: f.stack, route: r, @@ -61,7 +61,7 @@ type ForwarderRequest struct { stack *stack.Stack route *stack.Route id stack.TransportEndpointID - pkt tcpip.PacketBuffer + pkt stack.PacketBuffer } // ID returns the 4-tuple (src address, src port, dst address, dst port) that diff --git a/pkg/tcpip/transport/udp/protocol.go b/pkg/tcpip/transport/udp/protocol.go index 259c3072a..6e31a9bac 100644 --- a/pkg/tcpip/transport/udp/protocol.go +++ b/pkg/tcpip/transport/udp/protocol.go @@ -66,7 +66,7 @@ func (*protocol) ParsePorts(v buffer.View) (src, dst uint16, err *tcpip.Error) { // HandleUnknownDestinationPacket handles packets targeted at this protocol but // that don't match any existing endpoint. -func (p *protocol) HandleUnknownDestinationPacket(r *stack.Route, id stack.TransportEndpointID, pkt tcpip.PacketBuffer) bool { +func (p *protocol) HandleUnknownDestinationPacket(r *stack.Route, id stack.TransportEndpointID, pkt stack.PacketBuffer) bool { // Get the header then trim it from the view. hdr := header.UDP(pkt.Data.First()) if int(hdr.Length()) > pkt.Data.Size() { @@ -135,7 +135,7 @@ func (p *protocol) HandleUnknownDestinationPacket(r *stack.Route, id stack.Trans pkt.SetType(header.ICMPv4DstUnreachable) pkt.SetCode(header.ICMPv4PortUnreachable) pkt.SetChecksum(header.ICMPv4Checksum(pkt, payload)) - r.WritePacket(nil /* gso */, stack.NetworkHeaderParams{Protocol: header.ICMPv4ProtocolNumber, TTL: r.DefaultTTL(), TOS: stack.DefaultTOS}, tcpip.PacketBuffer{ + r.WritePacket(nil /* gso */, stack.NetworkHeaderParams{Protocol: header.ICMPv4ProtocolNumber, TTL: r.DefaultTTL(), TOS: stack.DefaultTOS}, stack.PacketBuffer{ Header: hdr, Data: payload, }) @@ -172,7 +172,7 @@ func (p *protocol) HandleUnknownDestinationPacket(r *stack.Route, id stack.Trans pkt.SetType(header.ICMPv6DstUnreachable) pkt.SetCode(header.ICMPv6PortUnreachable) pkt.SetChecksum(header.ICMPv6Checksum(pkt, r.LocalAddress, r.RemoteAddress, payload)) - r.WritePacket(nil /* gso */, stack.NetworkHeaderParams{Protocol: header.ICMPv6ProtocolNumber, TTL: r.DefaultTTL(), TOS: stack.DefaultTOS}, tcpip.PacketBuffer{ + r.WritePacket(nil /* gso */, stack.NetworkHeaderParams{Protocol: header.ICMPv6ProtocolNumber, TTL: r.DefaultTTL(), TOS: stack.DefaultTOS}, stack.PacketBuffer{ Header: hdr, Data: payload, }) @@ -180,16 +180,22 @@ func (p *protocol) HandleUnknownDestinationPacket(r *stack.Route, id stack.Trans return true } -// SetOption implements TransportProtocol.SetOption. -func (p *protocol) SetOption(option interface{}) *tcpip.Error { +// SetOption implements stack.TransportProtocol.SetOption. +func (*protocol) SetOption(option interface{}) *tcpip.Error { return tcpip.ErrUnknownProtocolOption } -// Option implements TransportProtocol.Option. -func (p *protocol) Option(option interface{}) *tcpip.Error { +// Option implements stack.TransportProtocol.Option. +func (*protocol) Option(option interface{}) *tcpip.Error { return tcpip.ErrUnknownProtocolOption } +// Close implements stack.TransportProtocol.Close. +func (*protocol) Close() {} + +// Wait implements stack.TransportProtocol.Wait. +func (*protocol) Wait() {} + // NewProtocol returns a UDP transport protocol. func NewProtocol() stack.TransportProtocol { return &protocol{} diff --git a/pkg/tcpip/transport/udp/udp_test.go b/pkg/tcpip/transport/udp/udp_test.go index f0ff3fe71..8acaa607a 100644 --- a/pkg/tcpip/transport/udp/udp_test.go +++ b/pkg/tcpip/transport/udp/udp_test.go @@ -343,11 +343,11 @@ func (c *testContext) createEndpointForFlow(flow testFlow) { c.createEndpoint(flow.sockProto()) if flow.isV6Only() { if err := c.ep.SetSockOptBool(tcpip.V6OnlyOption, true); err != nil { - c.t.Fatalf("SetSockOpt failed: %v", err) + c.t.Fatalf("SetSockOptBool failed: %s", err) } } else if flow.isBroadcast() { - if err := c.ep.SetSockOpt(tcpip.BroadcastOption(1)); err != nil { - c.t.Fatal("SetSockOpt failed:", err) + if err := c.ep.SetSockOptBool(tcpip.BroadcastOption, true); err != nil { + c.t.Fatalf("SetSockOptBool failed: %s", err) } } } @@ -358,7 +358,8 @@ func (c *testContext) createEndpointForFlow(flow testFlow) { func (c *testContext) getPacketAndVerify(flow testFlow, checkers ...checker.NetworkChecker) []byte { c.t.Helper() - ctx, _ := context.WithTimeout(context.Background(), 2*time.Second) + ctx, cancel := context.WithTimeout(context.Background(), 2*time.Second) + defer cancel() p, ok := c.linkEP.ReadContext(ctx) if !ok { c.t.Fatalf("Packet wasn't written out") @@ -409,6 +410,7 @@ func (c *testContext) injectV6Packet(payload []byte, h *header4Tuple, valid bool // Initialize the IP header. ip := header.IPv6(buf) ip.Encode(&header.IPv6Fields{ + TrafficClass: testTOS, PayloadLength: uint16(header.UDPMinimumSize + len(payload)), NextHeader: uint8(udp.ProtocolNumber), HopLimit: 65, @@ -438,7 +440,7 @@ func (c *testContext) injectV6Packet(payload []byte, h *header4Tuple, valid bool u.SetChecksum(^u.CalculateChecksum(xsum)) // Inject packet. - c.linkEP.InjectInbound(ipv6.ProtocolNumber, tcpip.PacketBuffer{ + c.linkEP.InjectInbound(ipv6.ProtocolNumber, stack.PacketBuffer{ Data: buf.ToVectorisedView(), NetworkHeader: buffer.View(ip), TransportHeader: buffer.View(u), @@ -485,7 +487,7 @@ func (c *testContext) injectV4Packet(payload []byte, h *header4Tuple, valid bool // Inject packet. - c.linkEP.InjectInbound(ipv4.ProtocolNumber, tcpip.PacketBuffer{ + c.linkEP.InjectInbound(ipv4.ProtocolNumber, stack.PacketBuffer{ Data: buf.ToVectorisedView(), NetworkHeader: buffer.View(ip), TransportHeader: buffer.View(u), @@ -606,7 +608,7 @@ func testReadInternal(c *testContext, flow testFlow, packetShouldBeDropped, expe // Check the peer address. h := flow.header4Tuple(incoming) if addr.Addr != h.srcAddr.Addr { - c.t.Fatalf("unexpected remote address: got %s, want %s", addr.Addr, h.srcAddr) + c.t.Fatalf("unexpected remote address: got %s, want %v", addr.Addr, h.srcAddr) } // Check the payload. @@ -1270,8 +1272,8 @@ func TestTTL(t *testing.T) { c.createEndpointForFlow(flow) const multicastTTL = 42 - if err := c.ep.SetSockOpt(tcpip.MulticastTTLOption(multicastTTL)); err != nil { - c.t.Fatalf("SetSockOpt failed: %v", err) + if err := c.ep.SetSockOptInt(tcpip.MulticastTTLOption, multicastTTL); err != nil { + c.t.Fatalf("SetSockOptInt failed: %s", err) } var wantTTL uint8 @@ -1310,8 +1312,8 @@ func TestSetTTL(t *testing.T) { c.createEndpointForFlow(flow) - if err := c.ep.SetSockOpt(tcpip.TTLOption(wantTTL)); err != nil { - c.t.Fatalf("SetSockOpt failed: %v", err) + if err := c.ep.SetSockOptInt(tcpip.TTLOption, int(wantTTL)); err != nil { + c.t.Fatalf("SetSockOptInt(TTLOption, %d) failed: %s", wantTTL, err) } var p stack.NetworkProtocol @@ -1336,7 +1338,7 @@ func TestSetTTL(t *testing.T) { } } -func TestTOSV4(t *testing.T) { +func TestSetTOS(t *testing.T) { for _, flow := range []testFlow{unicastV4, multicastV4, broadcast} { t.Run(fmt.Sprintf("flow:%s", flow), func(t *testing.T) { c := newDualTestContext(t, defaultMTU) @@ -1345,25 +1347,26 @@ func TestTOSV4(t *testing.T) { c.createEndpointForFlow(flow) const tos = testTOS - var v tcpip.IPv4TOSOption - if err := c.ep.GetSockOpt(&v); err != nil { - c.t.Errorf("GetSockopt failed: %s", err) + v, err := c.ep.GetSockOptInt(tcpip.IPv4TOSOption) + if err != nil { + c.t.Errorf("GetSockOptInt(IPv4TOSOption) failed: %s", err) } // Test for expected default value. if v != 0 { - c.t.Errorf("got GetSockOpt(...) = %#v, want = %#v", v, 0) + c.t.Errorf("got GetSockOpt(IPv4TOSOption) = 0x%x, want = 0x%x", v, 0) } - if err := c.ep.SetSockOpt(tcpip.IPv4TOSOption(tos)); err != nil { - c.t.Errorf("SetSockOpt(%#v) failed: %s", tcpip.IPv4TOSOption(tos), err) + if err := c.ep.SetSockOptInt(tcpip.IPv4TOSOption, tos); err != nil { + c.t.Errorf("SetSockOptInt(IPv4TOSOption, 0x%x) failed: %s", tos, err) } - if err := c.ep.GetSockOpt(&v); err != nil { - c.t.Errorf("GetSockopt failed: %s", err) + v, err = c.ep.GetSockOptInt(tcpip.IPv4TOSOption) + if err != nil { + c.t.Errorf("GetSockOptInt(IPv4TOSOption) failed: %s", err) } - if want := tcpip.IPv4TOSOption(tos); v != want { - c.t.Errorf("got GetSockOpt(...) = %#v, want = %#v", v, want) + if v != tos { + c.t.Errorf("got GetSockOptInt(IPv4TOSOption) = 0x%x, want = 0x%x", v, tos) } testWrite(c, flow, checker.TOS(tos, 0)) @@ -1371,7 +1374,7 @@ func TestTOSV4(t *testing.T) { } } -func TestTOSV6(t *testing.T) { +func TestSetTClass(t *testing.T) { for _, flow := range []testFlow{unicastV4in6, unicastV6, unicastV6Only, multicastV4in6, multicastV6, broadcastIn6} { t.Run(fmt.Sprintf("flow:%s", flow), func(t *testing.T) { c := newDualTestContext(t, defaultMTU) @@ -1379,71 +1382,93 @@ func TestTOSV6(t *testing.T) { c.createEndpointForFlow(flow) - const tos = testTOS - var v tcpip.IPv6TrafficClassOption - if err := c.ep.GetSockOpt(&v); err != nil { - c.t.Errorf("GetSockopt failed: %s", err) + const tClass = testTOS + v, err := c.ep.GetSockOptInt(tcpip.IPv6TrafficClassOption) + if err != nil { + c.t.Errorf("GetSockOptInt(IPv6TrafficClassOption) failed: %s", err) } // Test for expected default value. if v != 0 { - c.t.Errorf("got GetSockOpt(...) = %#v, want = %#v", v, 0) + c.t.Errorf("got GetSockOptInt(IPv6TrafficClassOption) = 0x%x, want = 0x%x", v, 0) } - if err := c.ep.SetSockOpt(tcpip.IPv6TrafficClassOption(tos)); err != nil { - c.t.Errorf("SetSockOpt failed: %s", err) + if err := c.ep.SetSockOptInt(tcpip.IPv6TrafficClassOption, tClass); err != nil { + c.t.Errorf("SetSockOptInt(IPv6TrafficClassOption, 0x%x) failed: %s", tClass, err) } - if err := c.ep.GetSockOpt(&v); err != nil { - c.t.Errorf("GetSockopt failed: %s", err) + v, err = c.ep.GetSockOptInt(tcpip.IPv6TrafficClassOption) + if err != nil { + c.t.Errorf("GetSockOptInt(IPv6TrafficClassOption) failed: %s", err) } - if want := tcpip.IPv6TrafficClassOption(tos); v != want { - c.t.Errorf("got GetSockOpt(...) = %#v, want = %#v", v, want) + if v != tClass { + c.t.Errorf("got GetSockOptInt(IPv6TrafficClassOption) = 0x%x, want = 0x%x", v, tClass) } - testWrite(c, flow, checker.TOS(tos, 0)) + // The header getter for TClass is called TOS, so use that checker. + testWrite(c, flow, checker.TOS(tClass, 0)) }) } } -func TestReceiveTOSV4(t *testing.T) { - for _, flow := range []testFlow{unicastV4, broadcast} { - t.Run(fmt.Sprintf("flow:%s", flow), func(t *testing.T) { - c := newDualTestContext(t, defaultMTU) - defer c.cleanup() +func TestReceiveTosTClass(t *testing.T) { + testCases := []struct { + name string + getReceiveOption tcpip.SockOptBool + tests []testFlow + }{ + {"ReceiveTosOption", tcpip.ReceiveTOSOption, []testFlow{unicastV4, broadcast}}, + {"ReceiveTClassOption", tcpip.ReceiveTClassOption, []testFlow{unicastV4in6, unicastV6, unicastV6Only, broadcastIn6}}, + } + for _, testCase := range testCases { + for _, flow := range testCase.tests { + t.Run(fmt.Sprintf("%s:flow:%s", testCase.name, flow), func(t *testing.T) { + c := newDualTestContext(t, defaultMTU) + defer c.cleanup() - c.createEndpointForFlow(flow) + c.createEndpointForFlow(flow) + option := testCase.getReceiveOption + name := testCase.name - // Verify that setting and reading the option works. - v, err := c.ep.GetSockOptBool(tcpip.ReceiveTOSOption) - if err != nil { - c.t.Fatal("GetSockOptBool(tcpip.ReceiveTOSOption) failed:", err) - } - // Test for expected default value. - if v != false { - c.t.Errorf("got GetSockOptBool(tcpip.ReceiveTOSOption) = %t, want = %t", v, false) - } + // Verify that setting and reading the option works. + v, err := c.ep.GetSockOptBool(option) + if err != nil { + c.t.Errorf("GetSockOptBool(%s) failed: %s", name, err) + } + // Test for expected default value. + if v != false { + c.t.Errorf("got GetSockOptBool(%s) = %t, want = %t", name, v, false) + } - want := true - if err := c.ep.SetSockOptBool(tcpip.ReceiveTOSOption, want); err != nil { - c.t.Fatalf("SetSockOptBool(tcpip.ReceiveTOSOption, %t) failed: %s", want, err) - } + want := true + if err := c.ep.SetSockOptBool(option, want); err != nil { + c.t.Fatalf("SetSockOptBool(%s, %t) failed: %s", name, want, err) + } - got, err := c.ep.GetSockOptBool(tcpip.ReceiveTOSOption) - if err != nil { - c.t.Fatal("GetSockOptBool(tcpip.ReceiveTOSOption) failed:", err) - } - if got != want { - c.t.Fatalf("got GetSockOptBool(tcpip.ReceiveTOSOption) = %t, want = %t", got, want) - } + got, err := c.ep.GetSockOptBool(option) + if err != nil { + c.t.Errorf("GetSockOptBool(%s) failed: %s", name, err) + } - // Verify that the correct received TOS is handed through as - // ancillary data to the ControlMessages struct. - if err := c.ep.Bind(tcpip.FullAddress{Port: stackPort}); err != nil { - c.t.Fatal("Bind failed:", err) - } - testRead(c, flow, checker.ReceiveTOS(testTOS)) - }) + if got != want { + c.t.Errorf("got GetSockOptBool(%s) = %t, want = %t", name, got, want) + } + + // Verify that the correct received TOS or TClass is handed through as + // ancillary data to the ControlMessages struct. + if err := c.ep.Bind(tcpip.FullAddress{Port: stackPort}); err != nil { + c.t.Fatalf("Bind failed: %s", err) + } + switch option { + case tcpip.ReceiveTClassOption: + testRead(c, flow, checker.ReceiveTClass(testTOS)) + case tcpip.ReceiveTOSOption: + testRead(c, flow, checker.ReceiveTOS(testTOS)) + default: + t.Fatalf("unknown test variant: %s", name) + } + }) + } } } @@ -1541,7 +1566,8 @@ func TestV4UnknownDestination(t *testing.T) { } c.injectPacket(tc.flow, payload) if !tc.icmpRequired { - ctx, _ := context.WithTimeout(context.Background(), time.Second) + ctx, cancel := context.WithTimeout(context.Background(), time.Second) + defer cancel() if p, ok := c.linkEP.ReadContext(ctx); ok { t.Fatalf("unexpected packet received: %+v", p) } @@ -1549,7 +1575,8 @@ func TestV4UnknownDestination(t *testing.T) { } // ICMP required. - ctx, _ := context.WithTimeout(context.Background(), time.Second) + ctx, cancel := context.WithTimeout(context.Background(), time.Second) + defer cancel() p, ok := c.linkEP.ReadContext(ctx) if !ok { t.Fatalf("packet wasn't written out") @@ -1617,7 +1644,8 @@ func TestV6UnknownDestination(t *testing.T) { } c.injectPacket(tc.flow, payload) if !tc.icmpRequired { - ctx, _ := context.WithTimeout(context.Background(), time.Second) + ctx, cancel := context.WithTimeout(context.Background(), time.Second) + defer cancel() if p, ok := c.linkEP.ReadContext(ctx); ok { t.Fatalf("unexpected packet received: %+v", p) } @@ -1625,7 +1653,8 @@ func TestV6UnknownDestination(t *testing.T) { } // ICMP required. - ctx, _ := context.WithTimeout(context.Background(), time.Second) + ctx, cancel := context.WithTimeout(context.Background(), time.Second) + defer cancel() p, ok := c.linkEP.ReadContext(ctx) if !ok { t.Fatalf("packet wasn't written out") diff --git a/pkg/usermem/BUILD b/pkg/usermem/BUILD index ff8b9e91a..6c9ada9c7 100644 --- a/pkg/usermem/BUILD +++ b/pkg/usermem/BUILD @@ -25,7 +25,6 @@ go_library( "bytes_io_unsafe.go", "usermem.go", "usermem_arm64.go", - "usermem_unsafe.go", "usermem_x86.go", ], visibility = ["//:sandbox"], @@ -33,6 +32,7 @@ go_library( "//pkg/atomicbitops", "//pkg/binary", "//pkg/context", + "//pkg/gohacks", "//pkg/log", "//pkg/safemem", "//pkg/syserror", diff --git a/pkg/usermem/usermem.go b/pkg/usermem/usermem.go index 71fd4e155..cd6a0ea6b 100644 --- a/pkg/usermem/usermem.go +++ b/pkg/usermem/usermem.go @@ -23,14 +23,12 @@ import ( "gvisor.dev/gvisor/pkg/binary" "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/gohacks" "gvisor.dev/gvisor/pkg/safemem" "gvisor.dev/gvisor/pkg/syserror" ) // IO provides access to the contents of a virtual memory space. -// -// FIXME(b/38173783): Implementations of IO cannot expect ctx to contain any -// meaningful data. type IO interface { // CopyOut copies len(src) bytes from src to the memory mapped at addr. It // returns the number of bytes copied. If the number of bytes copied is < @@ -251,7 +249,7 @@ func CopyStringIn(ctx context.Context, uio IO, addr Addr, maxlen int, opts IOOpt } end, ok := addr.AddLength(uint64(readlen)) if !ok { - return stringFromImmutableBytes(buf[:done]), syserror.EFAULT + return gohacks.StringFromImmutableBytes(buf[:done]), syserror.EFAULT } // Shorten the read to avoid crossing page boundaries, since faulting // in a page unnecessarily is expensive. This also ensures that partial @@ -272,16 +270,16 @@ func CopyStringIn(ctx context.Context, uio IO, addr Addr, maxlen int, opts IOOpt // Look for the terminating zero byte, which may have occurred before // hitting err. if i := bytes.IndexByte(buf[done:done+n], byte(0)); i >= 0 { - return stringFromImmutableBytes(buf[:done+i]), nil + return gohacks.StringFromImmutableBytes(buf[:done+i]), nil } done += n if err != nil { - return stringFromImmutableBytes(buf[:done]), err + return gohacks.StringFromImmutableBytes(buf[:done]), err } addr = end } - return stringFromImmutableBytes(buf), syserror.ENAMETOOLONG + return gohacks.StringFromImmutableBytes(buf), syserror.ENAMETOOLONG } // CopyOutVec copies bytes from src to the memory mapped at ars in uio. The diff --git a/pkg/usermem/usermem_x86.go b/pkg/usermem/usermem_x86.go index 8059b72d2..d96f829fb 100644 --- a/pkg/usermem/usermem_x86.go +++ b/pkg/usermem/usermem_x86.go @@ -12,7 +12,7 @@ // See the License for the specific language governing permissions and // limitations under the License. -// +build amd64 i386 +// +build amd64 386 package usermem |