diff options
Diffstat (limited to 'pkg')
439 files changed, 24597 insertions, 7894 deletions
diff --git a/pkg/abi/linux/BUILD b/pkg/abi/linux/BUILD index 322d1ccc4..59b0e138a 100644 --- a/pkg/abi/linux/BUILD +++ b/pkg/abi/linux/BUILD @@ -10,6 +10,7 @@ go_library( name = "linux", srcs = [ "aio.go", + "arch_amd64.go", "audit.go", "bpf.go", "capability.go", diff --git a/pkg/abi/linux/arch_amd64.go b/pkg/abi/linux/arch_amd64.go new file mode 100644 index 000000000..0be31e755 --- /dev/null +++ b/pkg/abi/linux/arch_amd64.go @@ -0,0 +1,23 @@ +// 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 linux + +// Start and end addresses of the vsyscall page. +const ( + VSyscallStartAddr uint64 = 0xffffffffff600000 + VSyscallEndAddr uint64 = 0xffffffffff601000 +) 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_amd64.go b/pkg/abi/linux/epoll_amd64.go index 34ff18009..7e74b1143 100644 --- a/pkg/abi/linux/epoll_amd64.go +++ b/pkg/abi/linux/epoll_amd64.go @@ -12,11 +12,13 @@ // See the License for the specific language governing permissions and // limitations under the License. +// +build amd64 + package linux // EpollEvent is equivalent to struct epoll_event from epoll(2). // -// +marshal +// +marshal slice:EpollEventSlice type EpollEvent struct { Events uint32 // Linux makes struct epoll_event::data a __u64. We represent it as diff --git a/pkg/abi/linux/epoll_arm64.go b/pkg/abi/linux/epoll_arm64.go index f86c35329..a35939cc9 100644 --- a/pkg/abi/linux/epoll_arm64.go +++ b/pkg/abi/linux/epoll_arm64.go @@ -12,11 +12,13 @@ // See the License for the specific language governing permissions and // limitations under the License. +// +build arm64 + package linux // EpollEvent is equivalent to struct epoll_event from epoll(2). // -// +marshal +// +marshal slice:EpollEventSlice type EpollEvent struct { Events uint32 // Linux makes struct epoll_event a __u64, necessitating 4 bytes of padding diff --git a/pkg/abi/linux/file.go b/pkg/abi/linux/file.go index e229ac21c..055ac1d7c 100644 --- a/pkg/abi/linux/file.go +++ b/pkg/abi/linux/file.go @@ -266,6 +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 @@ -284,6 +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/mm.go b/pkg/abi/linux/mm.go index cd043dac3..07cc1895e 100644 --- a/pkg/abi/linux/mm.go +++ b/pkg/abi/linux/mm.go @@ -90,14 +90,19 @@ const ( MS_SYNC = 1 << 2 ) +// NumaPolicy is the NUMA memory policy for a memory range. See numa(7). +// +// +marshal +type NumaPolicy int32 + // Policies for get_mempolicy(2)/set_mempolicy(2). const ( - MPOL_DEFAULT = 0 - MPOL_PREFERRED = 1 - MPOL_BIND = 2 - MPOL_INTERLEAVE = 3 - MPOL_LOCAL = 4 - MPOL_MAX = 5 + MPOL_DEFAULT NumaPolicy = 0 + MPOL_PREFERRED NumaPolicy = 1 + MPOL_BIND NumaPolicy = 2 + MPOL_INTERLEAVE NumaPolicy = 3 + MPOL_LOCAL NumaPolicy = 4 + MPOL_MAX NumaPolicy = 5 ) // Flags for get_mempolicy(2). diff --git a/pkg/abi/linux/netfilter.go b/pkg/abi/linux/netfilter.go index bd2e13ba1..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 @@ -253,6 +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 { @@ -443,3 +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/seccomp.go b/pkg/abi/linux/seccomp.go index 4eeb5cd7a..d0607e256 100644 --- a/pkg/abi/linux/seccomp.go +++ b/pkg/abi/linux/seccomp.go @@ -63,3 +63,10 @@ func (a BPFAction) String() string { func (a BPFAction) Data() uint16 { return uint16(a & SECCOMP_RET_DATA) } + +// SockFprog is sock_fprog taken from <linux/filter.h>. +type SockFprog struct { + Len uint16 + pad [6]byte + Filter *BPFInstruction +} 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 index a77a3beea..dcd086298 100644 --- a/pkg/buffer/BUILD +++ b/pkg/buffer/BUILD @@ -10,8 +10,8 @@ go_template_instance( prefix = "buffer", template = "//pkg/ilist:generic_list", types = { - "Element": "*Buffer", - "Linker": "*Buffer", + "Element": "*buffer", + "Linker": "*buffer", }, ) @@ -34,6 +34,10 @@ go_library( go_test( name = "buffer_test", size = "small", - srcs = ["view_test.go"], + srcs = [ + "safemem_test.go", + "view_test.go", + ], library = ":buffer", + deps = ["//pkg/safemem"], ) diff --git a/pkg/buffer/buffer.go b/pkg/buffer/buffer.go index d5f64609b..c6d089fd9 100644 --- a/pkg/buffer/buffer.go +++ b/pkg/buffer/buffer.go @@ -13,6 +13,10 @@ // 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 ( @@ -21,7 +25,7 @@ import ( const bufferSize = 8144 // See below. -// Buffer encapsulates a queueable byte buffer. +// 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 @@ -30,38 +34,61 @@ const bufferSize = 8144 // See below. // large enough chunk to limit excessive segmentation. // // +stateify savable -type Buffer struct { +type buffer struct { data [bufferSize]byte read int write int bufferEntry } -// Reset resets internal data. +// reset resets internal data. // -// This must be called before use. -func (b *Buffer) Reset() { +// This must be called before returning the buffer to the pool. +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 { +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) + return new(buffer) }, } diff --git a/pkg/buffer/safemem.go b/pkg/buffer/safemem.go index 071aaa488..0e5b86344 100644 --- a/pkg/buffer/safemem.go +++ b/pkg/buffer/safemem.go @@ -15,19 +15,17 @@ package buffer import ( - "io" - "gvisor.dev/gvisor/pkg/safemem" ) // WriteBlock returns this buffer as a write Block. -func (b *Buffer) WriteBlock() safemem.Block { - return safemem.BlockFromSafeSlice(b.data[b.write:]) +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.data[b.read:b.write]) +func (b *buffer) ReadBlock() safemem.Block { + return safemem.BlockFromSafeSlice(b.ReadSlice()) } // WriteFromBlocks implements safemem.Writer.WriteFromBlocks. @@ -47,21 +45,21 @@ func (v *View) WriteFromBlocks(srcs safemem.BlockSeq) (uint64, error) { // Need at least one buffer. firstBuf := v.data.Back() if firstBuf == nil { - firstBuf = bufferPool.Get().(*Buffer) + firstBuf = bufferPool.Get().(*buffer) v.data.PushBack(firstBuf) } // Does the last block have sufficient capacity alone? - if l := len(firstBuf.data) - firstBuf.write; l >= need { + 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) + emptyBuf := bufferPool.Get().(*buffer) v.data.PushBack(emptyBuf) - need -= len(emptyBuf.data) // Full block. + need -= emptyBuf.WriteSize() blocks = append(blocks, emptyBuf.WriteBlock()) } dst = safemem.BlockSeqFromSlice(blocks) @@ -73,11 +71,11 @@ func (v *View) WriteFromBlocks(srcs safemem.BlockSeq) (uint64, error) { // Update all indices. for left := int(n); left > 0; firstBuf = firstBuf.Next() { - if l := len(firstBuf.data) - firstBuf.write; left >= l { - firstBuf.write += l // Whole block. + if l := firstBuf.WriteSize(); left >= l { + firstBuf.WriteMove(l) // Whole block. left -= l } else { - firstBuf.write += left // Partial block. + firstBuf.WriteMove(left) // Partial block. left = 0 } } @@ -103,18 +101,18 @@ func (v *View) ReadToBlocks(dsts safemem.BlockSeq) (uint64, error) { firstBuf := v.data.Front() if firstBuf == nil { - return 0, io.EOF + return 0, nil // No EOF. } // Is all the data in a single block? - if l := firstBuf.write - firstBuf.read; l >= need { + 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.write - buf.read + need -= buf.ReadSize() blocks = append(blocks, buf.ReadBlock()) } src = safemem.BlockSeqFromSlice(blocks) 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 index 00fc11e9c..e6901eadb 100644 --- a/pkg/buffer/view.go +++ b/pkg/buffer/view.go @@ -38,14 +38,6 @@ func (v *View) TrimFront(count int64) { } } -// Read implements io.Reader.Read. -// -// Note that reading does not advance the read index. This must be done -// manually using TrimFront or other methods. -func (v *View) Read(p []byte) (int, error) { - return v.ReadAt(p, 0) -} - // ReadAt implements io.ReaderAt.ReadAt. func (v *View) ReadAt(p []byte, offset int64) (int, error) { var ( @@ -54,54 +46,46 @@ func (v *View) ReadAt(p []byte, offset int64) (int, error) { ) for buf := v.data.Front(); buf != nil && done < int64(len(p)); buf = buf.Next() { needToSkip := int(offset - skipped) - if l := buf.write - buf.read; l <= needToSkip { - skipped += int64(l) + if sz := buf.ReadSize(); sz <= needToSkip { + skipped += int64(sz) continue } // Actually read data. - n := copy(p[done:], buf.data[buf.read+needToSkip:buf.write]) + n := copy(p[done:], buf.ReadSlice()[needToSkip:]) skipped += int64(needToSkip) done += int64(n) } - if int(done) < len(p) { + if int(done) < len(p) || offset+done == v.size { return int(done), io.EOF } return int(done), nil } -// Write implements io.Writer.Write. -func (v *View) Write(p []byte) (int, error) { - v.Append(p) // Does not fail. - return len(p), 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; { - l := int64(buf.write - buf.read) - if l > count { + sz := int64(buf.ReadSize()) + if sz > count { // There is still data for reading. - buf.read += int(count) + buf.ReadMove(int(count)) v.size -= count count = 0 break } - // Read from this buffer. - buf.read += int(l) - count -= l - v.size -= l - - // When all data has been read from a buffer, we push - // it into the empty buffer pool for reuse. + // 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)) @@ -109,37 +93,39 @@ func (v *View) advanceRead(count int64) { } // 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 || length >= v.size { + 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() { - l := int64(buf.write - buf.read) // Local bytes. - switch { - case v.size-l >= length: - // Drop the buffer completely; see above. - v.data.Remove(buf) - v.size -= l - buf.Reset() - bufferPool.Put(buf) - - case v.size > length && v.size-l < length: - // Just truncate the buffer locally. - delta := (length - (v.size - l)) - buf.write = buf.read + int(delta) + 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 - - default: - // Should never happen. - panic("invalid buffer during truncation") + break } + + // Drop the buffer completely; see above. + v.data.Remove(buf) + buf.Reset() + bufferPool.Put(buf) + v.size -= sz } - v.size = length // Save the new size. } -// Grow grows the given view to the number of bytes. If zero -// is true, all these bytes will be zero. If zero is false, -// then this is the caller's responsibility. +// 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) { @@ -149,29 +135,29 @@ func (v *View) Grow(length int64, zero bool) { for v.size < length { buf := v.data.Back() - // Is there at least one buffer? + // Is there some space in the last buffer? if buf == nil || buf.Full() { - buf = bufferPool.Get().(*Buffer) + buf = bufferPool.Get().(*buffer) v.data.PushBack(buf) } // Write up to length bytes. - l := len(buf.data) - buf.write - if int64(l) > length-v.size { - l = int(length - v.size) + 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+l; i++ { + for i := buf.write; i < buf.write+sz; i++ { buf.data[i] = 0 } } // Advance the index. - buf.write += l - v.size += int64(l) + buf.WriteMove(sz) + v.size += int64(sz) } } @@ -181,31 +167,40 @@ func (v *View) Prepend(data []byte) { if buf := v.data.Front(); buf != nil && buf.read > 0 { // Fill up before the first write. avail := buf.read - copy(buf.data[0:], data[len(data)-avail:]) - data = data[:len(data)-avail] - v.size += int64(avail) + 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) + buf := bufferPool.Get().(*buffer) v.data.PushFront(buf) // The buffer is empty; copy last chunk. - start := len(data) - len(buf.data) - if start < 0 { - start = 0 // Everything. + 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. - bStart := len(buf.data) - len(data[start:]) - n := copy(buf.data[bStart:], data[start:]) - buf.read = bStart - buf.write = len(buf.data) - data = data[:start] + 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) } } @@ -214,16 +209,16 @@ func (v *View) Append(data []byte) { for done := 0; done < len(data); { buf := v.data.Back() - // Find the first empty buffer. + // Ensure there's a buffer with space. if buf == nil || buf.Full() { - buf = bufferPool.Get().(*Buffer) + buf = bufferPool.Get().(*buffer) v.data.PushBack(buf) } // Copy in to the given buffer. - n := copy(buf.data[buf.write:], data[done:]) + n := copy(buf.WriteSlice(), data[done:]) done += n - buf.write += n + buf.WriteMove(n) v.size += int64(n) } } @@ -232,52 +227,52 @@ func (v *View) Append(data []byte) { // // 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. +// 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.Next() == nil { - return buf.data[buf.read:buf.write] // Only one buffer. + 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.data[buf.read:buf.write]...) + data = append(data, buf.ReadSlice()...) } return data } // Size indicates the total amount of data available in this view. -func (v *View) Size() (sz int64) { - sz = v.size // Pre-calculated. - return sz +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.data[buf.read:buf.write]) + other.Append(buf.ReadSlice()) } - return other + 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() { - if l := int64(buf.write - buf.read); l > 0 { - fn(buf.data[buf.read:buf.write]) - } + fn(buf.ReadSlice()) } } // Merge merges the provided View with this one. // -// The other view will be empty after this operation. +// 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.Empty(); buf = other.data.Front() { + for buf := other.data.Front(); buf != nil; buf = other.data.Front() { other.data.Remove(buf) v.data.PushBack(buf) } @@ -288,6 +283,9 @@ func (v *View) Merge(other *View) { } // 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 @@ -297,17 +295,17 @@ func (v *View) WriteFromReader(r io.Reader, count int64) (int64, error) { for done < count { buf := v.data.Back() - // Find the first empty buffer. + // Ensure we have an empty buffer. if buf == nil || buf.Full() { - buf = bufferPool.Get().(*Buffer) + buf = bufferPool.Get().(*buffer) v.data.PushBack(buf) } // Is this less than the minimum batch? - if len(buf.data[buf.write:]) < minBatch && (count-done) >= int64(minBatch) { + if buf.WriteSize() < minBatch && (count-done) >= int64(minBatch) { tmp := make([]byte, minBatch) n, err = r.Read(tmp) - v.Write(tmp[:n]) + v.Append(tmp[:n]) done += int64(n) if err != nil { break @@ -316,14 +314,14 @@ func (v *View) WriteFromReader(r io.Reader, count int64) (int64, error) { } // Limit the read, if necessary. - end := len(buf.data) - if int64(end-buf.write) > (count - done) { - end = buf.write + int(count-done) + 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.data[buf.write:end]) - buf.write += n + n, err = r.Read(buf.WriteSlice()[:sz]) + buf.WriteMove(n) done += int64(n) v.size += int64(n) if err == io.EOF { @@ -340,6 +338,9 @@ func (v *View) WriteFromReader(r io.Reader, count int64) (int64, error) { // // 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 @@ -348,15 +349,22 @@ func (v *View) ReadToWriter(w io.Writer, count int64) (int64, error) { ) offset := 0 // Spill-over for batching. for buf := v.data.Front(); buf != nil && done < count; buf = buf.Next() { - l := buf.write - buf.read - offset + // Has this been consumed? Skip it. + sz := buf.ReadSize() + if sz <= offset { + offset -= sz + continue + } + sz -= offset // Is this less than the minimum batch? - if l < minBatch && (count-done) >= int64(minBatch) && (v.size-done) >= int64(minBatch) { + 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 - l // Reset below. + offset = n - sz // Reset below. if err != nil { break } @@ -364,12 +372,12 @@ func (v *View) ReadToWriter(w io.Writer, count int64) (int64, error) { } // Limit the write if necessary. - if int64(l) >= (count - done) { - l = int(count - done) + if int64(sz) >= left { + sz = int(left) } // Perform the actual write. - n, err = w.Write(buf.data[buf.read+offset : buf.read+offset+l]) + n, err = w.Write(buf.ReadSlice()[offset : offset+sz]) done += int64(n) if err != nil { break diff --git a/pkg/buffer/view_test.go b/pkg/buffer/view_test.go index 37e652f16..3db1bc6ee 100644 --- a/pkg/buffer/view_test.go +++ b/pkg/buffer/view_test.go @@ -16,218 +16,452 @@ 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 - ops []func(*View) + op func(*testing.T, *View) }{ - // Prepend. + // 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: "prepend", - input: "world", - ops: []func(*View){ - func(v *View) { - v.Prepend([]byte("hello ")) - }, + 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 fill", - input: strings.Repeat("1", bufferSize-1), - ops: []func(*View){ - func(v *View) { - v.Prepend([]byte("0")) - }, + name: "prepend-backfill-full", + input: "hello world", + output: "jello world", + op: func(t *testing.T, v *View) { + v.TrimFront(1) + v.Prepend([]byte("j")) }, - output: "0" + strings.Repeat("1", bufferSize-1), }, { - name: "prepend overflow", - input: strings.Repeat("1", bufferSize), - ops: []func(*View){ - func(v *View) { - v.Prepend([]byte("0")) - }, + name: "prepend-backfill-under", + input: "hello world", + output: "hola world", + op: func(t *testing.T, v *View) { + v.TrimFront(5) + v.Prepend([]byte("hola")) }, - output: "0" + strings.Repeat("1", bufferSize), }, { - name: "prepend multiple buffers", - input: strings.Repeat("1", bufferSize-1), - ops: []func(*View){ - func(v *View) { - v.Prepend([]byte(strings.Repeat("0", bufferSize*3))) - }, + 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. + // Append and write. { - name: "append", - input: "hello", - ops: []func(*View){ - func(v *View) { - v.Append([]byte(" world")) - }, - }, + 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), - ops: []func(*View){ - func(v *View) { - v.Append([]byte("0")) - }, - }, + 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), - ops: []func(*View){ - func(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), - ops: []func(*View){ - func(v *View) { - v.Append([]byte(strings.Repeat("0", bufferSize*3))) - }, - }, + 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", - ops: []func(*View){ - func(v *View) { - v.Truncate(5) - }, - }, + name: "truncate", + input: "hello world", output: "hello", + op: func(t *testing.T, v *View) { + v.Truncate(5) + }, }, { - name: "truncate multiple buffers", - input: strings.Repeat("1", bufferSize*2), - ops: []func(*View){ - func(v *View) { - v.Truncate(bufferSize*2 - 1) - }, + name: "truncate-noop", + input: "hello world", + output: "hello world", + op: func(t *testing.T, v *View) { + v.Truncate(v.Size() + 1) }, - output: strings.Repeat("1", bufferSize*2-1), }, { - name: "truncate multiple buffers to one buffer", - input: strings.Repeat("1", bufferSize*2), - ops: []func(*View){ - func(v *View) { - v.Truncate(5) - }, + 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", - ops: []func(*View){ - func(v *View) { - v.TrimFront(6) - }, - }, + name: "trim", + input: "hello world", output: "world", + op: func(t *testing.T, v *View) { + v.TrimFront(6) + }, }, { - name: "trim multiple buffers", - input: strings.Repeat("1", bufferSize*2), - ops: []func(*View){ - func(v *View) { - v.TrimFront(1) - }, + name: "trim-too-large", + input: "hello world", + output: "", + op: func(t *testing.T, v *View) { + v.TrimFront(v.Size() + 1) }, - output: strings.Repeat("1", bufferSize*2-1), }, { - name: "trim multiple buffers to one buffer", - input: strings.Repeat("1", bufferSize*2), - ops: []func(*View){ - func(v *View) { - v.TrimFront(bufferSize*2 - 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", - ops: []func(*View){ - func(v *View) { - v.Grow(1, true) - }, - }, + name: "grow", + input: "hello world", output: "hello world", + op: func(t *testing.T, v *View) { + v.Grow(1, true) + }, }, { - name: "grow from zero", - ops: []func(*View){ - func(v *View) { - v.Grow(1024, 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), - ops: []func(*View){ - func(v *View) { - v.Grow(bufferSize*2, 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 { - t.Run(tc.name, func(t *testing.T) { - // Construct the new view. - var view View - view.Append([]byte(tc.input)) - - // Run all operations. - for _, op := range tc.ops { - op(&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()) - } - }) + 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/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 a0bc55ea1..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 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/flipcall/packet_window_allocator.go b/pkg/flipcall/packet_window_allocator.go index ccb918fab..af9cc3d21 100644 --- a/pkg/flipcall/packet_window_allocator.go +++ b/pkg/flipcall/packet_window_allocator.go @@ -134,7 +134,7 @@ func (pwa *PacketWindowAllocator) Allocate(size int) (PacketWindowDescriptor, er start := pwa.nextAlloc pwa.nextAlloc = end return PacketWindowDescriptor{ - FD: pwa.fd, + FD: pwa.FD(), Offset: start, Length: size, }, nil @@ -158,7 +158,7 @@ func (pwa *PacketWindowAllocator) ensureFileSize(min int64) error { } newSize = newNewSize } - if err := syscall.Ftruncate(pwa.fd, newSize); err != nil { + if err := syscall.Ftruncate(pwa.FD(), newSize); err != nil { return fmt.Errorf("ftruncate failed: %v", err) } pwa.fileSize = newSize 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/ilist/list.go b/pkg/ilist/list.go index 8f93e4d6d..0d07da3b1 100644 --- a/pkg/ilist/list.go +++ b/pkg/ilist/list.go @@ -86,12 +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) { linker := ElementMapper{}.linkerFor(e) linker.SetNext(l.head) linker.SetPrev(nil) - if l.head != nil { ElementMapper{}.linkerFor(l.head).SetPrev(e) } else { @@ -106,7 +115,6 @@ func (l *List) PushBack(e Element) { linker := ElementMapper{}.linkerFor(e) linker.SetNext(nil) linker.SetPrev(l.tail) - if l.tail != nil { ElementMapper{}.linkerFor(l.tail).SetNext(e) } else { @@ -127,7 +135,6 @@ func (l *List) PushBackList(m *List) { l.tail = m.tail } - m.head = nil m.tail = nil } diff --git a/pkg/log/glog.go b/pkg/log/glog.go index b4f7bb5a4..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(depth int, 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 { @@ -81,5 +81,5 @@ func (g *GoogleEmitter) Emit(depth int, level Level, timestamp time.Time, format 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 0943db1cc..bdf9d691e 100644 --- a/pkg/log/json.go +++ b/pkg/log/json.go @@ -58,7 +58,7 @@ func (lv *Level) UnmarshalJSON(b []byte) error { // JSONEmitter logs messages in json format. type JSONEmitter struct { - Writer + *Writer } // Emit implements Emitter.Emit. diff --git a/pkg/log/json_k8s.go b/pkg/log/json_k8s.go index 6c6fc8b6f..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(_ int, 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 a794da1aa..37e0605ad 100644 --- a/pkg/log/log.go +++ b/pkg/log/log.go @@ -374,5 +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/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 3863ad1f5..57b89ad7d 100644 --- a/pkg/p9/messages.go +++ b/pkg/p9/messages.go @@ -1926,19 +1926,17 @@ func (r *Rreaddir) decode(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) { + 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. diff --git a/pkg/p9/messages_test.go b/pkg/p9/messages_test.go index c20324404..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{ diff --git a/pkg/p9/transport_flipcall.go b/pkg/p9/transport_flipcall.go index a0d274f3b..38038abdf 100644 --- a/pkg/p9/transport_flipcall.go +++ b/pkg/p9/transport_flipcall.go @@ -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/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/seccomp/seccomp_unsafe.go b/pkg/seccomp/seccomp_unsafe.go index be328db12..f7e986589 100644 --- a/pkg/seccomp/seccomp_unsafe.go +++ b/pkg/seccomp/seccomp_unsafe.go @@ -21,13 +21,6 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" ) -// sockFprog is sock_fprog taken from <linux/filter.h>. -type sockFprog struct { - Len uint16 - pad [6]byte - Filter *linux.BPFInstruction -} - // SetFilter installs the given BPF program. // // This is safe to call from an afterFork context. @@ -39,7 +32,7 @@ func SetFilter(instrs []linux.BPFInstruction) syscall.Errno { return errno } - sockProg := sockFprog{ + sockProg := linux.SockFprog{ Len: uint16(len(instrs)), Filter: (*linux.BPFInstruction)(unsafe.Pointer(&instrs[0])), } 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 5053393c1..c29e1b841 100644 --- a/pkg/sentry/arch/arch_aarch64.go +++ b/pkg/sentry/arch/arch_aarch64.go @@ -95,8 +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. @@ -145,7 +151,9 @@ func (s *State) Fork() State { return State{ Regs: s.Regs, aarch64FPState: s.aarch64FPState.fork(), + TPValue: s.TPValue, FeatureSet: s.FeatureSet, + OrigR0: s.OrigR0, } } @@ -255,6 +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. diff --git a/pkg/sentry/arch/arch_arm64.go b/pkg/sentry/arch/arch_arm64.go index 885115ae2..db99c5acb 100644 --- a/pkg/sentry/arch/arch_arm64.go +++ b/pkg/sentry/arch/arch_arm64.go @@ -140,16 +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 0c1db4b13..1cb1adf8c 100644 --- a/pkg/sentry/arch/signal_arm64.go +++ b/pkg/sentry/arch/signal_arm64.go @@ -98,9 +98,12 @@ func (c *context64) SignalSetup(st *Stack, act *SignalAct, info *SignalInfo, alt if ucSize < 0 { panic("can't get size of UContext64") } - // st.Arch.Width() is for the restorer address. sizeof(siginfo) == 128. - frameSize := int(st.Arch.Width()) + ucSize + 128 - frameBottom := (sp-usermem.Addr(frameSize)) & ^usermem.Addr(15) - 8 + + // frameSize = ucSize + sizeof(siginfo). + // sizeof(siginfo) == 128. + // R30 stores the restorer address. + frameSize := ucSize + 128 + frameBottom := (sp - usermem.Addr(frameSize)) & ^usermem.Addr(15) sp = frameBottom + usermem.Addr(frameSize) st.Bottom = sp @@ -130,12 +133,27 @@ func (c *context64) SignalSetup(st *Stack, act *SignalAct, info *SignalInfo, alt c.Regs.Regs[0] = uint64(info.Signo) c.Regs.Regs[1] = uint64(infoAddr) c.Regs.Regs[2] = uint64(ucAddr) - + c.Regs.Regs[30] = uint64(act.Restorer) return nil } // SignalRestore implements Context.SignalRestore. -// Only used on intel. func (c *context64) SignalRestore(st *Stack, rt bool) (linux.SignalSet, SignalStack, error) { - return 0, SignalStack{}, nil + // Copy out the stack frame. + var uc UContext64 + if _, err := st.Pop(&uc); err != nil { + return 0, SignalStack{}, err + } + var info SignalInfo + if _, err := st.Pop(&info); err != nil { + return 0, SignalStack{}, err + } + + // Restore registers. + c.Regs.Regs = uc.MContext.Regs + c.Regs.Pc = uc.MContext.Pc + c.Regs.Sp = uc.MContext.Sp + c.Regs.Pstate = uc.MContext.Pstate + + return uc.Sigset, uc.Stack, nil } 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/proc.go b/pkg/sentry/control/proc.go index 5457ba5e7..b51fb3959 100644 --- a/pkg/sentry/control/proc.go +++ b/pkg/sentry/control/proc.go @@ -224,8 +224,6 @@ func (proc *Proc) execAsync(args *ExecArgs) (*kernel.ThreadGroup, kernel.ThreadI } } - 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 { @@ -235,7 +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 } @@ -254,7 +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 } 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/dirent.go b/pkg/sentry/fs/dirent.go index e0b32e1c1..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). @@ -1519,15 +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 { 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/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..39299b7e4 100644 --- a/pkg/sentry/fs/host/control.go +++ b/pkg/sentry/fs/host/control.go @@ -23,6 +23,8 @@ import ( "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" ) +// LINT.IfChange + type scmRights struct { fds []int } @@ -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 @@ -91,3 +93,5 @@ func fdsToFiles(ctx context.Context, fds []int) []*fs.File { } return files } + +// LINT.ThenChange(../../fsimpl/host/control.go) 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 7c60dc1db..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" @@ -24,49 +23,10 @@ import ( "gvisor.dev/gvisor/pkg/sentry/device" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" - "gvisor.dev/gvisor/pkg/sentry/kernel/time" + ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" "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: @@ -107,54 +67,22 @@ 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)} - } - - // 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 - } +func owner(s *syscall.Stat_t) fs.FileOwner { + return fs.FileOwner{ + UID: auth.KUID(s.Uid), + GID: auth.KGID(s.Gid), } - 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), - AccessTime: time.FromUnix(s.Atim.Sec, s.Atim.Nsec), - ModificationTime: time.FromUnix(s.Mtim.Sec, s.Mtim.Nsec), - StatusChangeTime: time.FromUnix(s.Ctim.Sec, s.Ctim.Nsec), + 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), Links: uint64(s.Nlink), } } 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/mounts.go b/pkg/sentry/fs/mounts.go index c7981f66e..b414ddaee 100644 --- a/pkg/sentry/fs/mounts.go +++ b/pkg/sentry/fs/mounts.go @@ -273,19 +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/mounts.go b/pkg/sentry/fs/proc/mounts.go index 94deb553b..1fc9c703c 100644 --- a/pkg/sentry/fs/proc/mounts.go +++ b/pkg/sentry/fs/proc/mounts.go @@ -170,7 +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" { diff --git a/pkg/sentry/fs/proc/net.go b/pkg/sentry/fs/proc/net.go index 95d5817ff..bd18177d4 100644 --- a/pkg/sentry/fs/proc/net.go +++ b/pkg/sentry/fs/proc/net.go @@ -40,47 +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 - // TODO(gvisor.dev/issue/1833): Support for using the network stack in the - // network namespace of the calling process. We should make this per-process, - // a.k.a. /proc/PID/net, and make /proc/net a symlink to /proc/self/net. - if s := p.k.RootNetworkNamespace().Stack(); 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. @@ -837,4 +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 d4c4b533d..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 } diff --git a/pkg/sentry/fs/proc/task.go b/pkg/sentry/fs/proc/task.go index 4e9b0fc00..4bbe90198 100644 --- a/pkg/sentry/fs/proc/task.go +++ b/pkg/sentry/fs/proc/task.go @@ -57,14 +57,23 @@ 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 type taskDir struct { ramfs.Dir - t *kernel.Task - pidns *kernel.PIDNamespace + t *kernel.Task } var _ fs.InodeOperations = (*taskDir)(nil) @@ -84,6 +93,7 @@ func (p *proc) newTaskDir(t *kernel.Task, msrc *fs.MountSource, isThreadGroup bo "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), @@ -253,11 +263,12 @@ func newExe(t *kernel.Task, msrc *fs.MountSource) *fs.Inode { } 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 } @@ -267,7 +278,7 @@ func (e *exe) executable() (file fsbridge.File, err error) { // (with locks held). file = mm.Executable() if file == nil { - err = syserror.ENOENT + err = syserror.ESRCH } }) return @@ -312,11 +323,22 @@ 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) @@ -852,15 +874,15 @@ func (o *oomScoreAdj) GetFile(ctx context.Context, dirent *fs.Dirent, flags fs.F // Read implements fs.FileOperations.Read. func (f *oomScoreAdjFile) Read(ctx context.Context, _ *fs.File, dst usermem.IOSequence, offset int64) (int64, error) { - if offset != 0 { - return 0, io.EOF + if f.t.ExitState() == kernel.TaskExitDead { + return 0, syserror.ESRCH } - adj, err := f.t.OOMScoreAdj() - if err != nil { - return 0, err + var buf bytes.Buffer + fmt.Fprintf(&buf, "%d\n", f.t.OOMScoreAdj()) + if offset >= int64(buf.Len()) { + return 0, io.EOF } - adjBytes := []byte(strconv.FormatInt(int64(adj), 10) + "\n") - n, err := dst.CopyOut(ctx, adjBytes) + n, err := dst.CopyOut(ctx, buf.Bytes()[offset:]) return int64(n), err } @@ -879,6 +901,9 @@ func (f *oomScoreAdjFile) Write(ctx context.Context, _ *fs.File, src usermem.IOS return 0, err } + if f.t.ExitState() == kernel.TaskExitDead { + return 0, syserror.ESRCH + } if err := f.t.SetOOMScoreAdj(v); err != nil { return 0, err } 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/line_discipline.go b/pkg/sentry/fs/tty/line_discipline.go index 12b1c6097..2e9dd2d55 100644 --- a/pkg/sentry/fs/tty/line_discipline.go +++ b/pkg/sentry/fs/tty/line_discipline.go @@ -27,6 +27,8 @@ import ( "gvisor.dev/gvisor/pkg/waiter" ) +// LINT.IfChange + const ( // canonMaxBytes is the number of bytes that fit into a single line of // terminal input in canonical mode. This corresponds to N_TTY_BUF_SIZE @@ -443,3 +445,5 @@ func (l *lineDiscipline) peek(b []byte) int { } return size } + +// LINT.ThenChange(../../fsimpl/devpts/line_discipline.go) diff --git a/pkg/sentry/fs/tty/master.go b/pkg/sentry/fs/tty/master.go index f62da49bd..fe07fa929 100644 --- a/pkg/sentry/fs/tty/master.go +++ b/pkg/sentry/fs/tty/master.go @@ -26,6 +26,8 @@ import ( "gvisor.dev/gvisor/pkg/waiter" ) +// LINT.IfChange + // masterInodeOperations are the fs.InodeOperations for the master end of the // Terminal (ptmx file). // @@ -232,3 +234,5 @@ func maybeEmitUnimplementedEvent(ctx context.Context, cmd uint32) { unimpl.EmitUnimplementedEvent(ctx) } } + +// LINT.ThenChange(../../fsimpl/devpts/master.go) diff --git a/pkg/sentry/fs/tty/queue.go b/pkg/sentry/fs/tty/queue.go index 1ca79c0b2..ceabb9b1e 100644 --- a/pkg/sentry/fs/tty/queue.go +++ b/pkg/sentry/fs/tty/queue.go @@ -25,6 +25,8 @@ import ( "gvisor.dev/gvisor/pkg/waiter" ) +// LINT.IfChange + // waitBufMaxBytes is the maximum size of a wait buffer. It is based on // TTYB_DEFAULT_MEM_LIMIT. const waitBufMaxBytes = 131072 @@ -234,3 +236,5 @@ func (q *queue) waitBufAppend(b []byte) { q.waitBuf = append(q.waitBuf, b) q.waitBufLen += uint64(len(b)) } + +// LINT.ThenChange(../../fsimpl/devpts/queue.go) diff --git a/pkg/sentry/fs/tty/slave.go b/pkg/sentry/fs/tty/slave.go index 6a2dbc576..9871f6fc6 100644 --- a/pkg/sentry/fs/tty/slave.go +++ b/pkg/sentry/fs/tty/slave.go @@ -25,6 +25,8 @@ import ( "gvisor.dev/gvisor/pkg/waiter" ) +// LINT.IfChange + // slaveInodeOperations are the fs.InodeOperations for the slave end of the // Terminal (pts file). // @@ -172,3 +174,5 @@ func (sf *slaveFileOperations) Ioctl(ctx context.Context, _ *fs.File, io usermem return 0, syserror.ENOTTY } } + +// LINT.ThenChange(../../fsimpl/devpts/slave.go) diff --git a/pkg/sentry/fs/tty/terminal.go b/pkg/sentry/fs/tty/terminal.go index 5883f26db..ddcccf4da 100644 --- a/pkg/sentry/fs/tty/terminal.go +++ b/pkg/sentry/fs/tty/terminal.go @@ -23,6 +23,8 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) +// LINT.IfChange + // Terminal is a pseudoterminal. // // +stateify savable @@ -126,3 +128,5 @@ func (tm *Terminal) tty(isMaster bool) *kernel.TTY { } return tm.slaveKTTY } + +// LINT.ThenChange(../../fsimpl/devpts/terminal.go) diff --git a/pkg/sentry/fs/user/BUILD b/pkg/sentry/fs/user/BUILD new file mode 100644 index 000000000..f37f979f1 --- /dev/null +++ b/pkg/sentry/fs/user/BUILD @@ -0,0 +1,34 @@ +load("//tools:defs.bzl", "go_library", "go_test") + +package(licenses = ["notice"]) + +go_library( + name = "user", + srcs = ["user.go"], + visibility = ["//pkg/sentry:internal"], + deps = [ + "//pkg/abi/linux", + "//pkg/context", + "//pkg/fspath", + "//pkg/sentry/fs", + "//pkg/sentry/kernel/auth", + "//pkg/sentry/vfs", + "//pkg/usermem", + ], +) + +go_test( + name = "user_test", + size = "small", + srcs = ["user_test.go"], + library = ":user", + deps = [ + "//pkg/abi/linux", + "//pkg/context", + "//pkg/sentry/fs", + "//pkg/sentry/fs/tmpfs", + "//pkg/sentry/kernel/auth", + "//pkg/sentry/kernel/contexttest", + "//pkg/usermem", + ], +) diff --git a/pkg/sentry/fs/user/user.go b/pkg/sentry/fs/user/user.go new file mode 100644 index 000000000..fe7f67c00 --- /dev/null +++ b/pkg/sentry/fs/user/user.go @@ -0,0 +1,237 @@ +// 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 user + +import ( + "bufio" + "fmt" + "io" + "strconv" + "strings" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/usermem" +) + +type fileReader struct { + // Ctx is the context for the file reader. + Ctx context.Context + + // File is the file to read from. + File *fs.File +} + +// Read implements io.Reader.Read. +func (r *fileReader) Read(buf []byte) (int, error) { + n, err := r.File.Readv(r.Ctx, usermem.BytesIOSequence(buf)) + return int(n), err +} + +// getExecUserHome returns the home directory of the executing user read from +// /etc/passwd as read from the container filesystem. +func getExecUserHome(ctx context.Context, rootMns *fs.MountNamespace, uid auth.KUID) (string, error) { + // The default user home directory to return if no user matching the user + // if found in the /etc/passwd found in the image. + const defaultHome = "/" + + // Open the /etc/passwd file from the dirent via the root mount namespace. + mnsRoot := rootMns.Root() + maxTraversals := uint(linux.MaxSymlinkTraversals) + dirent, err := rootMns.FindInode(ctx, mnsRoot, nil, "/etc/passwd", &maxTraversals) + if err != nil { + // NOTE: Ignore errors opening the passwd file. If the passwd file + // doesn't exist we will return the default home directory. + return defaultHome, nil + } + defer dirent.DecRef() + + // Check read permissions on the file. + if err := dirent.Inode.CheckPermission(ctx, fs.PermMask{Read: true}); err != nil { + // NOTE: Ignore permissions errors here and return default root dir. + return defaultHome, nil + } + + // Only open regular files. We don't open other files like named pipes as + // they may block and might present some attack surface to the container. + // Note that runc does not seem to do this kind of checking. + if !fs.IsRegular(dirent.Inode.StableAttr) { + return defaultHome, nil + } + + f, err := dirent.Inode.GetFile(ctx, dirent, fs.FileFlags{Read: true, Directory: false}) + if err != nil { + return "", err + } + defer f.DecRef() + + r := &fileReader{ + Ctx: ctx, + File: f, + } + + return findHomeInPasswd(uint32(uid), r, defaultHome) +} + +type fileReaderVFS2 struct { + ctx context.Context + fd *vfs.FileDescription +} + +func (r *fileReaderVFS2) Read(buf []byte) (int, error) { + n, err := r.fd.Read(r.ctx, usermem.BytesIOSequence(buf), vfs.ReadOptions{}) + return int(n), err +} + +func getExecUserHomeVFS2(ctx context.Context, mns *vfs.MountNamespace, uid auth.KUID) (string, error) { + const defaultHome = "/" + + root := mns.Root() + defer root.DecRef() + + creds := auth.CredentialsFromContext(ctx) + + target := &vfs.PathOperation{ + Root: root, + Start: root, + Path: fspath.Parse("/etc/passwd"), + } + + opts := &vfs.OpenOptions{ + Flags: linux.O_RDONLY, + } + + fd, err := root.Mount().Filesystem().VirtualFilesystem().OpenAt(ctx, creds, target, opts) + if err != nil { + return defaultHome, nil + } + defer fd.DecRef() + + r := &fileReaderVFS2{ + ctx: ctx, + fd: fd, + } + + homeDir, err := findHomeInPasswd(uint32(uid), r, defaultHome) + if err != nil { + return "", err + } + + return homeDir, nil +} + +// MaybeAddExecUserHome returns a new slice with the HOME enviroment variable +// set if the slice does not already contain it, otherwise it returns the +// original slice unmodified. +func MaybeAddExecUserHome(ctx context.Context, mns *fs.MountNamespace, uid auth.KUID, envv []string) ([]string, error) { + // Check if the envv already contains HOME. + for _, env := range envv { + if strings.HasPrefix(env, "HOME=") { + // We have it. Return the original slice unmodified. + return envv, nil + } + } + + // Read /etc/passwd for the user's HOME directory and set the HOME + // environment variable as required by POSIX if it is not overridden by + // the user. + homeDir, err := getExecUserHome(ctx, mns, uid) + if err != nil { + return nil, fmt.Errorf("error reading exec user: %v", err) + } + + return append(envv, "HOME="+homeDir), nil +} + +// MaybeAddExecUserHomeVFS2 returns a new slice with the HOME enviroment +// variable set if the slice does not already contain it, otherwise it returns +// the original slice unmodified. +func MaybeAddExecUserHomeVFS2(ctx context.Context, vmns *vfs.MountNamespace, uid auth.KUID, envv []string) ([]string, error) { + // Check if the envv already contains HOME. + for _, env := range envv { + if strings.HasPrefix(env, "HOME=") { + // We have it. Return the original slice unmodified. + return envv, nil + } + } + + // Read /etc/passwd for the user's HOME directory and set the HOME + // environment variable as required by POSIX if it is not overridden by + // the user. + homeDir, err := getExecUserHomeVFS2(ctx, vmns, uid) + if err != nil { + return nil, fmt.Errorf("error reading exec user: %v", err) + } + return append(envv, "HOME="+homeDir), nil +} + +// findHomeInPasswd parses a passwd file and returns the given user's home +// directory. This function does it's best to replicate the runc's behavior. +func findHomeInPasswd(uid uint32, passwd io.Reader, defaultHome string) (string, error) { + s := bufio.NewScanner(passwd) + + for s.Scan() { + if err := s.Err(); err != nil { + return "", err + } + + line := strings.TrimSpace(s.Text()) + if line == "" { + continue + } + + // Pull out part of passwd entry. Loosely parse the passwd entry as some + // passwd files could be poorly written and for compatibility with runc. + // + // Per 'man 5 passwd' + // /etc/passwd contains one line for each user account, with seven + // fields delimited by colons (“:”). These fields are: + // + // - login name + // - optional encrypted password + // - numerical user ID + // - numerical group ID + // - user name or comment field + // - user home directory + // - optional user command interpreter + parts := strings.Split(line, ":") + + found := false + homeDir := "" + for i, p := range parts { + switch i { + case 2: + parsedUID, err := strconv.ParseUint(p, 10, 32) + if err == nil && parsedUID == uint64(uid) { + found = true + } + case 5: + homeDir = p + } + } + if found { + // NOTE: If the uid is present but the home directory is not + // present in the /etc/passwd entry we return an empty string. This + // is, for better or worse, what runc does. + return homeDir, nil + } + } + + return defaultHome, nil +} diff --git a/pkg/sentry/fs/user/user_test.go b/pkg/sentry/fs/user/user_test.go new file mode 100644 index 000000000..7d8e9ac7c --- /dev/null +++ b/pkg/sentry/fs/user/user_test.go @@ -0,0 +1,198 @@ +// 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 user + +import ( + "fmt" + "strings" + "testing" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/fs/tmpfs" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/kernel/contexttest" + "gvisor.dev/gvisor/pkg/usermem" +) + +// createEtcPasswd creates /etc/passwd with the given contents and mode. If +// mode is empty, then no file will be created. If mode is not a regular file +// mode, then contents is ignored. +func createEtcPasswd(ctx context.Context, root *fs.Dirent, contents string, mode linux.FileMode) error { + if err := root.CreateDirectory(ctx, root, "etc", fs.FilePermsFromMode(0755)); err != nil { + return err + } + etc, err := root.Walk(ctx, root, "etc") + if err != nil { + return err + } + defer etc.DecRef() + switch mode.FileType() { + case 0: + // Don't create anything. + return nil + case linux.S_IFREG: + passwd, err := etc.Create(ctx, root, "passwd", fs.FileFlags{Write: true}, fs.FilePermsFromMode(mode)) + if err != nil { + return err + } + defer passwd.DecRef() + if _, err := passwd.Writev(ctx, usermem.BytesIOSequence([]byte(contents))); err != nil { + return err + } + return nil + case linux.S_IFDIR: + return etc.CreateDirectory(ctx, root, "passwd", fs.FilePermsFromMode(mode)) + case linux.S_IFIFO: + return etc.CreateFifo(ctx, root, "passwd", fs.FilePermsFromMode(mode)) + default: + return fmt.Errorf("unknown file type %x", mode.FileType()) + } +} + +// TestGetExecUserHome tests the getExecUserHome function. +func TestGetExecUserHome(t *testing.T) { + tests := map[string]struct { + uid auth.KUID + passwdContents string + passwdMode linux.FileMode + expected string + }{ + "success": { + uid: 1000, + passwdContents: "adin::1000:1111::/home/adin:/bin/sh", + passwdMode: linux.S_IFREG | 0666, + expected: "/home/adin", + }, + "no_perms": { + uid: 1000, + passwdContents: "adin::1000:1111::/home/adin:/bin/sh", + passwdMode: linux.S_IFREG, + expected: "/", + }, + "no_passwd": { + uid: 1000, + expected: "/", + }, + "directory": { + uid: 1000, + passwdMode: linux.S_IFDIR | 0666, + expected: "/", + }, + // Currently we don't allow named pipes. + "named_pipe": { + uid: 1000, + passwdMode: linux.S_IFIFO | 0666, + expected: "/", + }, + } + + for name, tc := range tests { + t.Run(name, func(t *testing.T) { + ctx := contexttest.Context(t) + msrc := fs.NewPseudoMountSource(ctx) + rootInode := tmpfs.NewDir(ctx, nil, fs.RootOwner, fs.FilePermsFromMode(0777), msrc) + + mns, err := fs.NewMountNamespace(ctx, rootInode) + if err != nil { + t.Fatalf("NewMountNamespace failed: %v", err) + } + defer mns.DecRef() + root := mns.Root() + defer root.DecRef() + ctx = fs.WithRoot(ctx, root) + + if err := createEtcPasswd(ctx, root, tc.passwdContents, tc.passwdMode); err != nil { + t.Fatalf("createEtcPasswd failed: %v", err) + } + + got, err := getExecUserHome(ctx, mns, tc.uid) + if err != nil { + t.Fatalf("failed to get user home: %v", err) + } + + if got != tc.expected { + t.Fatalf("expected %v, got: %v", tc.expected, got) + } + }) + } +} + +// TestFindHomeInPasswd tests the findHomeInPasswd function's passwd file parsing. +func TestFindHomeInPasswd(t *testing.T) { + tests := map[string]struct { + uid uint32 + passwd string + expected string + def string + }{ + "empty": { + uid: 1000, + passwd: "", + expected: "/", + def: "/", + }, + "whitespace": { + uid: 1000, + passwd: " ", + expected: "/", + def: "/", + }, + "full": { + uid: 1000, + passwd: "adin::1000:1111::/home/adin:/bin/sh", + expected: "/home/adin", + def: "/", + }, + // For better or worse, this is how runc works. + "partial": { + uid: 1000, + passwd: "adin::1000:1111:", + expected: "", + def: "/", + }, + "multiple": { + uid: 1001, + passwd: "adin::1000:1111::/home/adin:/bin/sh\nian::1001:1111::/home/ian:/bin/sh", + expected: "/home/ian", + def: "/", + }, + "duplicate": { + uid: 1000, + passwd: "adin::1000:1111::/home/adin:/bin/sh\nian::1000:1111::/home/ian:/bin/sh", + expected: "/home/adin", + def: "/", + }, + "empty_lines": { + uid: 1001, + passwd: "adin::1000:1111::/home/adin:/bin/sh\n\n\nian::1001:1111::/home/ian:/bin/sh", + expected: "/home/ian", + def: "/", + }, + } + + for name, tc := range tests { + t.Run(name, func(t *testing.T) { + got, err := findHomeInPasswd(tc.uid, strings.NewReader(tc.passwd), tc.def) + if err != nil { + t.Fatalf("error parsing passwd: %v", err) + } + if tc.expected != got { + t.Fatalf("expected %v, got: %v", tc.expected, got) + } + }) + } +} diff --git a/pkg/sentry/fsbridge/vfs.go b/pkg/sentry/fsbridge/vfs.go index 6aa17bfc1..89168220a 100644 --- a/pkg/sentry/fsbridge/vfs.go +++ b/pkg/sentry/fsbridge/vfs.go @@ -26,22 +26,22 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) -// fsFile implements File interface over vfs.FileDescription. +// VFSFile implements File interface over vfs.FileDescription. // // +stateify savable -type vfsFile struct { +type VFSFile struct { file *vfs.FileDescription } -var _ File = (*vfsFile)(nil) +var _ File = (*VFSFile)(nil) // NewVFSFile creates a new File over fs.File. func NewVFSFile(file *vfs.FileDescription) File { - return &vfsFile{file: file} + return &VFSFile{file: file} } // PathnameWithDeleted implements File. -func (f *vfsFile) PathnameWithDeleted(ctx context.Context) string { +func (f *VFSFile) PathnameWithDeleted(ctx context.Context) string { root := vfs.RootFromContext(ctx) defer root.DecRef() @@ -51,7 +51,7 @@ func (f *vfsFile) PathnameWithDeleted(ctx context.Context) string { } // ReadFull implements File. -func (f *vfsFile) ReadFull(ctx context.Context, dst usermem.IOSequence, offset int64) (int64, error) { +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{}) @@ -67,12 +67,12 @@ func (f *vfsFile) ReadFull(ctx context.Context, dst usermem.IOSequence, offset i } // ConfigureMMap implements File. -func (f *vfsFile) ConfigureMMap(ctx context.Context, opts *memmap.MMapOpts) error { +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) { +func (f *VFSFile) Type(ctx context.Context) (linux.FileMode, error) { stat, err := f.file.Stat(ctx, vfs.StatOptions{}) if err != nil { return 0, err @@ -81,15 +81,21 @@ func (f *vfsFile) Type(ctx context.Context) (linux.FileMode, error) { } // IncRef implements File. -func (f *vfsFile) IncRef() { +func (f *VFSFile) IncRef() { f.file.IncRef() } // DecRef implements File. -func (f *vfsFile) DecRef() { +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 @@ -115,8 +121,6 @@ func NewVFSLookup(mntns *vfs.MountNamespace, root, workingDir vfs.VirtualDentry) // // remainingTraversals is not configurable in VFS2, all callers are using the // default anyways. -// -// TODO(gvisor.dev/issue/1623): Check mount has read and exec permission. 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) @@ -134,5 +138,5 @@ func (l *vfsLookup) OpenPath(ctx context.Context, pathname string, opts vfs.Open if err != nil { return nil, err } - return &vfsFile{file: fd}, nil + return &VFSFile{file: fd}, nil } diff --git a/pkg/sentry/fsimpl/devpts/BUILD b/pkg/sentry/fsimpl/devpts/BUILD new file mode 100644 index 000000000..585764223 --- /dev/null +++ b/pkg/sentry/fsimpl/devpts/BUILD @@ -0,0 +1,43 @@ +load("//tools:defs.bzl", "go_library", "go_test") + +licenses(["notice"]) + +go_library( + name = "devpts", + srcs = [ + "devpts.go", + "line_discipline.go", + "master.go", + "queue.go", + "slave.go", + "terminal.go", + ], + visibility = ["//pkg/sentry:internal"], + deps = [ + "//pkg/abi/linux", + "//pkg/context", + "//pkg/safemem", + "//pkg/sentry/arch", + "//pkg/sentry/fsimpl/kernfs", + "//pkg/sentry/kernel", + "//pkg/sentry/kernel/auth", + "//pkg/sentry/unimpl", + "//pkg/sentry/vfs", + "//pkg/sync", + "//pkg/syserror", + "//pkg/usermem", + "//pkg/waiter", + ], +) + +go_test( + name = "devpts_test", + size = "small", + srcs = ["devpts_test.go"], + library = ":devpts", + deps = [ + "//pkg/abi/linux", + "//pkg/sentry/contexttest", + "//pkg/usermem", + ], +) diff --git a/pkg/sentry/fsimpl/devpts/devpts.go b/pkg/sentry/fsimpl/devpts/devpts.go new file mode 100644 index 000000000..181d765d3 --- /dev/null +++ b/pkg/sentry/fsimpl/devpts/devpts.go @@ -0,0 +1,209 @@ +// 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 devpts provides a filesystem implementation that behaves like +// devpts. +package devpts + +import ( + "fmt" + "math" + "sort" + "strconv" + "sync" + + "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" +) + +// Name is the filesystem name. +const Name = "devpts" + +// FilesystemType implements vfs.FilesystemType. +type FilesystemType struct{} + +// Name implements vfs.FilesystemType.Name. +func (FilesystemType) Name() string { + return Name +} + +var _ vfs.FilesystemType = (*FilesystemType)(nil) + +// 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) { + // No data allowed. + if opts.Data != "" { + return nil, nil, syserror.EINVAL + } + + fs, root := fstype.newFilesystem(vfsObj, creds) + return fs.VFSFilesystem(), root.VFSDentry(), nil +} + +// newFilesystem creates a new devpts filesystem with root directory and ptmx +// master inode. It returns the filesystem and root Dentry. +func (fstype FilesystemType) newFilesystem(vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials) (*kernfs.Filesystem, *kernfs.Dentry) { + fs := &kernfs.Filesystem{} + fs.Init(vfsObj, fstype) + + // Construct the root directory. This is always inode id 1. + root := &rootInode{ + slaves: make(map[uint32]*slaveInode), + } + root.InodeAttrs.Init(creds, 1, linux.ModeDirectory|0555) + root.OrderedChildren.Init(kernfs.OrderedChildrenOptions{}) + root.dentry.Init(root) + + // Construct the pts master inode and dentry. Linux always uses inode + // id 2 for ptmx. See fs/devpts/inode.c:mknod_ptmx. + master := &masterInode{ + root: root, + } + master.InodeAttrs.Init(creds, 2, linux.ModeCharacterDevice|0666) + master.dentry.Init(master) + + // Add the master as a child of the root. + links := root.OrderedChildren.Populate(&root.dentry, map[string]*kernfs.Dentry{ + "ptmx": &master.dentry, + }) + root.IncLinks(links) + + return fs, &root.dentry +} + +// rootInode is the root directory inode for the devpts mounts. +type rootInode struct { + kernfs.AlwaysValid + kernfs.InodeAttrs + kernfs.InodeDirectoryNoNewChildren + kernfs.InodeNotSymlink + kernfs.OrderedChildren + + // Keep a reference to this inode's dentry. + dentry kernfs.Dentry + + // master is the master pty inode. Immutable. + master *masterInode + + // root is the root directory inode for this filesystem. Immutable. + root *rootInode + + // mu protects the fields below. + mu sync.Mutex + + // slaves maps pty ids to slave inodes. + slaves map[uint32]*slaveInode + + // nextIdx is the next pty index to use. Must be accessed atomically. + // + // TODO(b/29356795): reuse indices when ptys are closed. + nextIdx uint32 +} + +var _ kernfs.Inode = (*rootInode)(nil) + +// allocateTerminal creates a new Terminal and installs a pts node for it. +func (i *rootInode) allocateTerminal(creds *auth.Credentials) (*Terminal, error) { + i.mu.Lock() + defer i.mu.Unlock() + if i.nextIdx == math.MaxUint32 { + return nil, syserror.ENOMEM + } + idx := i.nextIdx + i.nextIdx++ + + // Sanity check that slave with idx does not exist. + if _, ok := i.slaves[idx]; ok { + panic(fmt.Sprintf("pty index collision; index %d already exists", idx)) + } + + // Create the new terminal and slave. + t := newTerminal(idx) + slave := &slaveInode{ + root: i, + t: t, + } + // Linux always uses pty index + 3 as the inode id. See + // fs/devpts/inode.c:devpts_pty_new(). + slave.InodeAttrs.Init(creds, uint64(idx+3), linux.ModeCharacterDevice|0600) + slave.dentry.Init(slave) + i.slaves[idx] = slave + + return t, nil +} + +// masterClose is called when the master end of t is closed. +func (i *rootInode) masterClose(t *Terminal) { + i.mu.Lock() + defer i.mu.Unlock() + + // Sanity check that slave with idx exists. + if _, ok := i.slaves[t.n]; !ok { + panic(fmt.Sprintf("pty with index %d does not exist", t.n)) + } + delete(i.slaves, t.n) +} + +// Open implements kernfs.Inode.Open. +func (i *rootInode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), vfsd, &i.OrderedChildren, &opts) + if err != nil { + return nil, err + } + return fd.VFSFileDescription(), nil +} + +// Lookup implements kernfs.Inode.Lookup. +func (i *rootInode) Lookup(ctx context.Context, name string) (*vfs.Dentry, error) { + idx, err := strconv.ParseUint(name, 10, 32) + if err != nil { + return nil, syserror.ENOENT + } + i.mu.Lock() + defer i.mu.Unlock() + if si, ok := i.slaves[uint32(idx)]; ok { + si.dentry.IncRef() + return si.dentry.VFSDentry(), nil + + } + return nil, syserror.ENOENT +} + +// IterDirents implements kernfs.Inode.IterDirents. +func (i *rootInode) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback, offset, relOffset int64) (int64, error) { + i.mu.Lock() + defer i.mu.Unlock() + ids := make([]int, 0, len(i.slaves)) + for id := range i.slaves { + ids = append(ids, int(id)) + } + sort.Ints(ids) + for _, id := range ids[relOffset:] { + dirent := vfs.Dirent{ + Name: strconv.FormatUint(uint64(id), 10), + Type: linux.DT_CHR, + Ino: i.slaves[uint32(id)].InodeAttrs.Ino(), + NextOff: offset + 1, + } + if err := cb.Handle(dirent); err != nil { + return offset, err + } + offset++ + } + return offset, nil +} diff --git a/pkg/sentry/fsimpl/devpts/devpts_test.go b/pkg/sentry/fsimpl/devpts/devpts_test.go new file mode 100644 index 000000000..b7c149047 --- /dev/null +++ b/pkg/sentry/fsimpl/devpts/devpts_test.go @@ -0,0 +1,56 @@ +// 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 devpts + +import ( + "testing" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/contexttest" + "gvisor.dev/gvisor/pkg/usermem" +) + +func TestSimpleMasterToSlave(t *testing.T) { + ld := newLineDiscipline(linux.DefaultSlaveTermios) + ctx := contexttest.Context(t) + inBytes := []byte("hello, tty\n") + src := usermem.BytesIOSequence(inBytes) + outBytes := make([]byte, 32) + dst := usermem.BytesIOSequence(outBytes) + + // Write to the input queue. + nw, err := ld.inputQueueWrite(ctx, src) + if err != nil { + t.Fatalf("error writing to input queue: %v", err) + } + if nw != int64(len(inBytes)) { + t.Fatalf("wrote wrong length: got %d, want %d", nw, len(inBytes)) + } + + // Read from the input queue. + nr, err := ld.inputQueueRead(ctx, dst) + if err != nil { + t.Fatalf("error reading from input queue: %v", err) + } + if nr != int64(len(inBytes)) { + t.Fatalf("read wrong length: got %d, want %d", nr, len(inBytes)) + } + + outStr := string(outBytes[:nr]) + inStr := string(inBytes) + if outStr != inStr { + t.Fatalf("written and read strings do not match: got %q, want %q", outStr, inStr) + } +} diff --git a/pkg/sentry/fsimpl/devpts/line_discipline.go b/pkg/sentry/fsimpl/devpts/line_discipline.go new file mode 100644 index 000000000..e201801d6 --- /dev/null +++ b/pkg/sentry/fsimpl/devpts/line_discipline.go @@ -0,0 +1,449 @@ +// 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 devpts + +import ( + "bytes" + "unicode/utf8" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" + "gvisor.dev/gvisor/pkg/waiter" +) + +// LINT.IfChange + +const ( + // canonMaxBytes is the number of bytes that fit into a single line of + // terminal input in canonical mode. This corresponds to N_TTY_BUF_SIZE + // in include/linux/tty.h. + canonMaxBytes = 4096 + + // nonCanonMaxBytes is the maximum number of bytes that can be read at + // a time in noncanonical mode. + nonCanonMaxBytes = canonMaxBytes - 1 + + spacesPerTab = 8 +) + +// lineDiscipline dictates how input and output are handled between the +// pseudoterminal (pty) master and slave. It can be configured to alter I/O, +// modify control characters (e.g. Ctrl-C for SIGINT), etc. The following man +// pages are good resources for how to affect the line discipline: +// +// * termios(3) +// * tty_ioctl(4) +// +// This file corresponds most closely to drivers/tty/n_tty.c. +// +// lineDiscipline has a simple structure but supports a multitude of options +// (see the above man pages). It consists of two queues of bytes: one from the +// terminal master to slave (the input queue) and one from slave to master (the +// output queue). When bytes are written to one end of the pty, the line +// discipline reads the bytes, modifies them or takes special action if +// required, and enqueues them to be read by the other end of the pty: +// +// input from terminal +-------------+ input to process (e.g. bash) +// +------------------------>| input queue |---------------------------+ +// | (inputQueueWrite) +-------------+ (inputQueueRead) | +// | | +// | v +// masterFD slaveFD +// ^ | +// | | +// | output to terminal +--------------+ output from process | +// +------------------------| output queue |<--------------------------+ +// (outputQueueRead) +--------------+ (outputQueueWrite) +// +// Lock order: +// termiosMu +// inQueue.mu +// outQueue.mu +// +// +stateify savable +type lineDiscipline struct { + // sizeMu protects size. + sizeMu sync.Mutex `state:"nosave"` + + // size is the terminal size (width and height). + size linux.WindowSize + + // inQueue is the input queue of the terminal. + inQueue queue + + // outQueue is the output queue of the terminal. + outQueue queue + + // termiosMu protects termios. + termiosMu sync.RWMutex `state:"nosave"` + + // termios is the terminal configuration used by the lineDiscipline. + termios linux.KernelTermios + + // column is the location in a row of the cursor. This is important for + // handling certain special characters like backspace. + column int + + // masterWaiter is used to wait on the master end of the TTY. + masterWaiter waiter.Queue `state:"zerovalue"` + + // slaveWaiter is used to wait on the slave end of the TTY. + slaveWaiter waiter.Queue `state:"zerovalue"` +} + +func newLineDiscipline(termios linux.KernelTermios) *lineDiscipline { + ld := lineDiscipline{termios: termios} + ld.inQueue.transformer = &inputQueueTransformer{} + ld.outQueue.transformer = &outputQueueTransformer{} + return &ld +} + +// getTermios gets the linux.Termios for the tty. +func (l *lineDiscipline) getTermios(ctx context.Context, io usermem.IO, args arch.SyscallArguments) (uintptr, error) { + l.termiosMu.RLock() + defer l.termiosMu.RUnlock() + // We must copy a Termios struct, not KernelTermios. + t := l.termios.ToTermios() + _, err := usermem.CopyObjectOut(ctx, io, args[2].Pointer(), t, usermem.IOOpts{ + AddressSpaceActive: true, + }) + return 0, err +} + +// setTermios sets a linux.Termios for the tty. +func (l *lineDiscipline) setTermios(ctx context.Context, io usermem.IO, args arch.SyscallArguments) (uintptr, error) { + l.termiosMu.Lock() + defer l.termiosMu.Unlock() + oldCanonEnabled := l.termios.LEnabled(linux.ICANON) + // We must copy a Termios struct, not KernelTermios. + var t linux.Termios + _, err := usermem.CopyObjectIn(ctx, io, args[2].Pointer(), &t, usermem.IOOpts{ + AddressSpaceActive: true, + }) + l.termios.FromTermios(t) + + // If canonical mode is turned off, move bytes from inQueue's wait + // buffer to its read buffer. Anything already in the read buffer is + // now readable. + if oldCanonEnabled && !l.termios.LEnabled(linux.ICANON) { + l.inQueue.mu.Lock() + l.inQueue.pushWaitBufLocked(l) + l.inQueue.readable = true + l.inQueue.mu.Unlock() + l.slaveWaiter.Notify(waiter.EventIn) + } + + return 0, err +} + +func (l *lineDiscipline) windowSize(ctx context.Context, io usermem.IO, args arch.SyscallArguments) error { + l.sizeMu.Lock() + defer l.sizeMu.Unlock() + _, err := usermem.CopyObjectOut(ctx, io, args[2].Pointer(), l.size, usermem.IOOpts{ + AddressSpaceActive: true, + }) + return err +} + +func (l *lineDiscipline) setWindowSize(ctx context.Context, io usermem.IO, args arch.SyscallArguments) error { + l.sizeMu.Lock() + defer l.sizeMu.Unlock() + _, err := usermem.CopyObjectIn(ctx, io, args[2].Pointer(), &l.size, usermem.IOOpts{ + AddressSpaceActive: true, + }) + return err +} + +func (l *lineDiscipline) masterReadiness() waiter.EventMask { + // We don't have to lock a termios because the default master termios + // is immutable. + return l.inQueue.writeReadiness(&linux.MasterTermios) | l.outQueue.readReadiness(&linux.MasterTermios) +} + +func (l *lineDiscipline) slaveReadiness() waiter.EventMask { + l.termiosMu.RLock() + defer l.termiosMu.RUnlock() + return l.outQueue.writeReadiness(&l.termios) | l.inQueue.readReadiness(&l.termios) +} + +func (l *lineDiscipline) inputQueueReadSize(ctx context.Context, io usermem.IO, args arch.SyscallArguments) error { + return l.inQueue.readableSize(ctx, io, args) +} + +func (l *lineDiscipline) inputQueueRead(ctx context.Context, dst usermem.IOSequence) (int64, error) { + l.termiosMu.RLock() + defer l.termiosMu.RUnlock() + n, pushed, err := l.inQueue.read(ctx, dst, l) + if err != nil { + return 0, err + } + if n > 0 { + l.masterWaiter.Notify(waiter.EventOut) + if pushed { + l.slaveWaiter.Notify(waiter.EventIn) + } + return n, nil + } + return 0, syserror.ErrWouldBlock +} + +func (l *lineDiscipline) inputQueueWrite(ctx context.Context, src usermem.IOSequence) (int64, error) { + l.termiosMu.RLock() + defer l.termiosMu.RUnlock() + n, err := l.inQueue.write(ctx, src, l) + if err != nil { + return 0, err + } + if n > 0 { + l.slaveWaiter.Notify(waiter.EventIn) + return n, nil + } + return 0, syserror.ErrWouldBlock +} + +func (l *lineDiscipline) outputQueueReadSize(ctx context.Context, io usermem.IO, args arch.SyscallArguments) error { + return l.outQueue.readableSize(ctx, io, args) +} + +func (l *lineDiscipline) outputQueueRead(ctx context.Context, dst usermem.IOSequence) (int64, error) { + l.termiosMu.RLock() + defer l.termiosMu.RUnlock() + n, pushed, err := l.outQueue.read(ctx, dst, l) + if err != nil { + return 0, err + } + if n > 0 { + l.slaveWaiter.Notify(waiter.EventOut) + if pushed { + l.masterWaiter.Notify(waiter.EventIn) + } + return n, nil + } + return 0, syserror.ErrWouldBlock +} + +func (l *lineDiscipline) outputQueueWrite(ctx context.Context, src usermem.IOSequence) (int64, error) { + l.termiosMu.RLock() + defer l.termiosMu.RUnlock() + n, err := l.outQueue.write(ctx, src, l) + if err != nil { + return 0, err + } + if n > 0 { + l.masterWaiter.Notify(waiter.EventIn) + return n, nil + } + return 0, syserror.ErrWouldBlock +} + +// transformer is a helper interface to make it easier to stateify queue. +type transformer interface { + // transform functions require queue's mutex to be held. + transform(*lineDiscipline, *queue, []byte) int +} + +// outputQueueTransformer implements transformer. It performs line discipline +// transformations on the output queue. +// +// +stateify savable +type outputQueueTransformer struct{} + +// transform does output processing for one end of the pty. See +// drivers/tty/n_tty.c:do_output_char for an analogous kernel function. +// +// Preconditions: +// * l.termiosMu must be held for reading. +// * q.mu must be held. +func (*outputQueueTransformer) transform(l *lineDiscipline, q *queue, buf []byte) int { + // transformOutput is effectively always in noncanonical mode, as the + // master termios never has ICANON set. + + if !l.termios.OEnabled(linux.OPOST) { + q.readBuf = append(q.readBuf, buf...) + if len(q.readBuf) > 0 { + q.readable = true + } + return len(buf) + } + + var ret int + for len(buf) > 0 { + size := l.peek(buf) + cBytes := append([]byte{}, buf[:size]...) + ret += size + buf = buf[size:] + // We're guaranteed that cBytes has at least one element. + switch cBytes[0] { + case '\n': + if l.termios.OEnabled(linux.ONLRET) { + l.column = 0 + } + if l.termios.OEnabled(linux.ONLCR) { + q.readBuf = append(q.readBuf, '\r', '\n') + continue + } + case '\r': + if l.termios.OEnabled(linux.ONOCR) && l.column == 0 { + continue + } + if l.termios.OEnabled(linux.OCRNL) { + cBytes[0] = '\n' + if l.termios.OEnabled(linux.ONLRET) { + l.column = 0 + } + break + } + l.column = 0 + case '\t': + spaces := spacesPerTab - l.column%spacesPerTab + if l.termios.OutputFlags&linux.TABDLY == linux.XTABS { + l.column += spaces + q.readBuf = append(q.readBuf, bytes.Repeat([]byte{' '}, spacesPerTab)...) + continue + } + l.column += spaces + case '\b': + if l.column > 0 { + l.column-- + } + default: + l.column++ + } + q.readBuf = append(q.readBuf, cBytes...) + } + if len(q.readBuf) > 0 { + q.readable = true + } + return ret +} + +// inputQueueTransformer implements transformer. It performs line discipline +// transformations on the input queue. +// +// +stateify savable +type inputQueueTransformer struct{} + +// transform does input processing for one end of the pty. Characters read are +// transformed according to flags set in the termios struct. See +// drivers/tty/n_tty.c:n_tty_receive_char_special for an analogous kernel +// function. +// +// Preconditions: +// * l.termiosMu must be held for reading. +// * q.mu must be held. +func (*inputQueueTransformer) transform(l *lineDiscipline, q *queue, buf []byte) int { + // If there's a line waiting to be read in canonical mode, don't write + // anything else to the read buffer. + if l.termios.LEnabled(linux.ICANON) && q.readable { + return 0 + } + + maxBytes := nonCanonMaxBytes + if l.termios.LEnabled(linux.ICANON) { + maxBytes = canonMaxBytes + } + + var ret int + for len(buf) > 0 && len(q.readBuf) < canonMaxBytes { + size := l.peek(buf) + cBytes := append([]byte{}, buf[:size]...) + // We're guaranteed that cBytes has at least one element. + switch cBytes[0] { + case '\r': + if l.termios.IEnabled(linux.IGNCR) { + buf = buf[size:] + ret += size + continue + } + if l.termios.IEnabled(linux.ICRNL) { + cBytes[0] = '\n' + } + case '\n': + if l.termios.IEnabled(linux.INLCR) { + cBytes[0] = '\r' + } + } + + // In canonical mode, we discard non-terminating characters + // after the first 4095. + if l.shouldDiscard(q, cBytes) { + buf = buf[size:] + ret += size + continue + } + + // Stop if the buffer would be overfilled. + if len(q.readBuf)+size > maxBytes { + break + } + buf = buf[size:] + ret += size + + // If we get EOF, make the buffer available for reading. + if l.termios.LEnabled(linux.ICANON) && l.termios.IsEOF(cBytes[0]) { + q.readable = true + break + } + + q.readBuf = append(q.readBuf, cBytes...) + + // Anything written to the readBuf will have to be echoed. + if l.termios.LEnabled(linux.ECHO) { + l.outQueue.writeBytes(cBytes, l) + l.masterWaiter.Notify(waiter.EventIn) + } + + // If we finish a line, make it available for reading. + if l.termios.LEnabled(linux.ICANON) && l.termios.IsTerminating(cBytes) { + q.readable = true + break + } + } + + // In noncanonical mode, everything is readable. + if !l.termios.LEnabled(linux.ICANON) && len(q.readBuf) > 0 { + q.readable = true + } + + return ret +} + +// shouldDiscard returns whether c should be discarded. In canonical mode, if +// too many bytes are enqueued, we keep reading input and discarding it until +// we find a terminating character. Signal/echo processing still occurs. +// +// Precondition: +// * l.termiosMu must be held for reading. +// * q.mu must be held. +func (l *lineDiscipline) shouldDiscard(q *queue, cBytes []byte) bool { + return l.termios.LEnabled(linux.ICANON) && len(q.readBuf)+len(cBytes) >= canonMaxBytes && !l.termios.IsTerminating(cBytes) +} + +// peek returns the size in bytes of the next character to process. As long as +// b isn't empty, peek returns a value of at least 1. +func (l *lineDiscipline) peek(b []byte) int { + size := 1 + // If UTF-8 support is enabled, runes might be multiple bytes. + if l.termios.IEnabled(linux.IUTF8) { + _, size = utf8.DecodeRune(b) + } + return size +} + +// LINT.ThenChange(../../fs/tty/line_discipline.go) diff --git a/pkg/sentry/fsimpl/devpts/master.go b/pkg/sentry/fsimpl/devpts/master.go new file mode 100644 index 000000000..04a292927 --- /dev/null +++ b/pkg/sentry/fsimpl/devpts/master.go @@ -0,0 +1,226 @@ +// 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 devpts + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/unimpl" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" + "gvisor.dev/gvisor/pkg/waiter" +) + +// LINT.IfChange + +// masterInode is the inode for the master end of the Terminal. +type masterInode struct { + kernfs.InodeAttrs + kernfs.InodeNoopRefCount + kernfs.InodeNotDirectory + kernfs.InodeNotSymlink + + // Keep a reference to this inode's dentry. + dentry kernfs.Dentry + + // root is the devpts root inode. + root *rootInode +} + +var _ kernfs.Inode = (*masterInode)(nil) + +// Open implements kernfs.Inode.Open. +func (mi *masterInode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + t, err := mi.root.allocateTerminal(rp.Credentials()) + if err != nil { + return nil, err + } + + mi.IncRef() + fd := &masterFileDescription{ + inode: mi, + t: t, + } + if err := fd.vfsfd.Init(fd, opts.Flags, rp.Mount(), vfsd, &vfs.FileDescriptionOptions{}); err != nil { + mi.DecRef() + return nil, err + } + return &fd.vfsfd, nil +} + +// Stat implements kernfs.Inode.Stat. +func (mi *masterInode) Stat(vfsfs *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) { + statx, err := mi.InodeAttrs.Stat(vfsfs, opts) + if err != nil { + return linux.Statx{}, err + } + statx.Blksize = 1024 + statx.RdevMajor = linux.TTYAUX_MAJOR + statx.RdevMinor = linux.PTMX_MINOR + return statx, nil +} + +// SetStat implements kernfs.Inode.SetStat +func (mi *masterInode) SetStat(ctx context.Context, vfsfs *vfs.Filesystem, creds *auth.Credentials, opts vfs.SetStatOptions) error { + if opts.Stat.Mask&linux.STATX_SIZE != 0 { + return syserror.EINVAL + } + return mi.InodeAttrs.SetStat(ctx, vfsfs, creds, opts) +} + +type masterFileDescription struct { + vfsfd vfs.FileDescription + vfs.FileDescriptionDefaultImpl + + inode *masterInode + t *Terminal +} + +var _ vfs.FileDescriptionImpl = (*masterFileDescription)(nil) + +// Release implements vfs.FileDescriptionImpl.Release. +func (mfd *masterFileDescription) Release() { + mfd.inode.root.masterClose(mfd.t) + mfd.inode.DecRef() +} + +// EventRegister implements waiter.Waitable.EventRegister. +func (mfd *masterFileDescription) EventRegister(e *waiter.Entry, mask waiter.EventMask) { + mfd.t.ld.masterWaiter.EventRegister(e, mask) +} + +// EventUnregister implements waiter.Waitable.EventUnregister. +func (mfd *masterFileDescription) EventUnregister(e *waiter.Entry) { + mfd.t.ld.masterWaiter.EventUnregister(e) +} + +// Readiness implements waiter.Waitable.Readiness. +func (mfd *masterFileDescription) Readiness(mask waiter.EventMask) waiter.EventMask { + return mfd.t.ld.masterReadiness() +} + +// Read implements vfs.FileDescriptionImpl.Read. +func (mfd *masterFileDescription) Read(ctx context.Context, dst usermem.IOSequence, _ vfs.ReadOptions) (int64, error) { + return mfd.t.ld.outputQueueRead(ctx, dst) +} + +// Write implements vfs.FileDescriptionImpl.Write. +func (mfd *masterFileDescription) Write(ctx context.Context, src usermem.IOSequence, _ vfs.WriteOptions) (int64, error) { + return mfd.t.ld.inputQueueWrite(ctx, src) +} + +// Ioctl implements vfs.FileDescriptionImpl.Ioctl. +func (mfd *masterFileDescription) Ioctl(ctx context.Context, io usermem.IO, args arch.SyscallArguments) (uintptr, error) { + switch cmd := args[1].Uint(); cmd { + case linux.FIONREAD: // linux.FIONREAD == linux.TIOCINQ + // Get the number of bytes in the output queue read buffer. + return 0, mfd.t.ld.outputQueueReadSize(ctx, io, args) + case linux.TCGETS: + // N.B. TCGETS on the master actually returns the configuration + // of the slave end. + return mfd.t.ld.getTermios(ctx, io, args) + case linux.TCSETS: + // N.B. TCSETS on the master actually affects the configuration + // of the slave end. + return mfd.t.ld.setTermios(ctx, io, args) + case linux.TCSETSW: + // TODO(b/29356795): This should drain the output queue first. + return mfd.t.ld.setTermios(ctx, io, args) + case linux.TIOCGPTN: + _, err := usermem.CopyObjectOut(ctx, io, args[2].Pointer(), uint32(mfd.t.n), usermem.IOOpts{ + AddressSpaceActive: true, + }) + return 0, err + case linux.TIOCSPTLCK: + // TODO(b/29356795): Implement pty locking. For now just pretend we do. + return 0, nil + case linux.TIOCGWINSZ: + return 0, mfd.t.ld.windowSize(ctx, io, args) + case linux.TIOCSWINSZ: + return 0, mfd.t.ld.setWindowSize(ctx, io, args) + case linux.TIOCSCTTY: + // Make the given terminal the controlling terminal of the + // calling process. + return 0, mfd.t.setControllingTTY(ctx, io, args, true /* isMaster */) + case linux.TIOCNOTTY: + // Release this process's controlling terminal. + return 0, mfd.t.releaseControllingTTY(ctx, io, args, true /* isMaster */) + case linux.TIOCGPGRP: + // Get the foreground process group. + return mfd.t.foregroundProcessGroup(ctx, io, args, true /* isMaster */) + case linux.TIOCSPGRP: + // Set the foreground process group. + return mfd.t.setForegroundProcessGroup(ctx, io, args, true /* isMaster */) + default: + maybeEmitUnimplementedEvent(ctx, cmd) + return 0, syserror.ENOTTY + } +} + +// SetStat implements vfs.FileDescriptionImpl.SetStat. +func (mfd *masterFileDescription) SetStat(ctx context.Context, opts vfs.SetStatOptions) error { + creds := auth.CredentialsFromContext(ctx) + fs := mfd.vfsfd.VirtualDentry().Mount().Filesystem() + return mfd.inode.SetStat(ctx, fs, creds, opts) +} + +// Stat implements vfs.FileDescriptionImpl.Stat. +func (mfd *masterFileDescription) Stat(ctx context.Context, opts vfs.StatOptions) (linux.Statx, error) { + fs := mfd.vfsfd.VirtualDentry().Mount().Filesystem() + return mfd.inode.Stat(fs, opts) +} + +// maybeEmitUnimplementedEvent emits unimplemented event if cmd is valid. +func maybeEmitUnimplementedEvent(ctx context.Context, cmd uint32) { + switch cmd { + case linux.TCGETS, + linux.TCSETS, + linux.TCSETSW, + linux.TCSETSF, + linux.TIOCGWINSZ, + linux.TIOCSWINSZ, + linux.TIOCSETD, + linux.TIOCSBRK, + linux.TIOCCBRK, + linux.TCSBRK, + linux.TCSBRKP, + linux.TIOCSTI, + linux.TIOCCONS, + linux.FIONBIO, + linux.TIOCEXCL, + linux.TIOCNXCL, + linux.TIOCGEXCL, + 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) + } +} + +// LINT.ThenChange(../../fs/tty/master.go) diff --git a/pkg/sentry/fsimpl/devpts/queue.go b/pkg/sentry/fsimpl/devpts/queue.go new file mode 100644 index 000000000..29a6be858 --- /dev/null +++ b/pkg/sentry/fsimpl/devpts/queue.go @@ -0,0 +1,240 @@ +// 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 devpts + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/safemem" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" + "gvisor.dev/gvisor/pkg/waiter" +) + +// LINT.IfChange + +// waitBufMaxBytes is the maximum size of a wait buffer. It is based on +// TTYB_DEFAULT_MEM_LIMIT. +const waitBufMaxBytes = 131072 + +// queue represents one of the input or output queues between a pty master and +// slave. Bytes written to a queue are added to the read buffer until it is +// full, at which point they are written to the wait buffer. Bytes are +// processed (i.e. undergo termios transformations) as they are added to the +// read buffer. The read buffer is readable when its length is nonzero and +// readable is true. +// +// +stateify savable +type queue struct { + // mu protects everything in queue. + mu sync.Mutex `state:"nosave"` + + // readBuf is buffer of data ready to be read when readable is true. + // This data has been processed. + readBuf []byte + + // waitBuf contains data that can't fit into readBuf. It is put here + // until it can be loaded into the read buffer. waitBuf contains data + // that hasn't been processed. + waitBuf [][]byte + waitBufLen uint64 + + // readable indicates whether the read buffer can be read from. In + // canonical mode, there can be an unterminated line in the read buffer, + // so readable must be checked. + readable bool + + // transform is the the queue's function for transforming bytes + // entering the queue. For example, transform might convert all '\r's + // entering the queue to '\n's. + transformer +} + +// readReadiness returns whether q is ready to be read from. +func (q *queue) readReadiness(t *linux.KernelTermios) waiter.EventMask { + q.mu.Lock() + defer q.mu.Unlock() + if len(q.readBuf) > 0 && q.readable { + return waiter.EventIn + } + return waiter.EventMask(0) +} + +// writeReadiness returns whether q is ready to be written to. +func (q *queue) writeReadiness(t *linux.KernelTermios) waiter.EventMask { + q.mu.Lock() + defer q.mu.Unlock() + if q.waitBufLen < waitBufMaxBytes { + return waiter.EventOut + } + return waiter.EventMask(0) +} + +// readableSize writes the number of readable bytes to userspace. +func (q *queue) readableSize(ctx context.Context, io usermem.IO, args arch.SyscallArguments) error { + q.mu.Lock() + defer q.mu.Unlock() + var size int32 + if q.readable { + size = int32(len(q.readBuf)) + } + + _, err := usermem.CopyObjectOut(ctx, io, args[2].Pointer(), size, usermem.IOOpts{ + AddressSpaceActive: true, + }) + return err + +} + +// read reads from q to userspace. It returns the number of bytes read as well +// as whether the read caused more readable data to become available (whether +// data was pushed from the wait buffer to the read buffer). +// +// Preconditions: +// * l.termiosMu must be held for reading. +func (q *queue) read(ctx context.Context, dst usermem.IOSequence, l *lineDiscipline) (int64, bool, error) { + q.mu.Lock() + defer q.mu.Unlock() + + if !q.readable { + return 0, false, syserror.ErrWouldBlock + } + + if dst.NumBytes() > canonMaxBytes { + dst = dst.TakeFirst(canonMaxBytes) + } + + n, err := dst.CopyOutFrom(ctx, safemem.ReaderFunc(func(dst safemem.BlockSeq) (uint64, error) { + src := safemem.BlockSeqOf(safemem.BlockFromSafeSlice(q.readBuf)) + n, err := safemem.CopySeq(dst, src) + if err != nil { + return 0, err + } + q.readBuf = q.readBuf[n:] + + // If we read everything, this queue is no longer readable. + if len(q.readBuf) == 0 { + q.readable = false + } + + return n, nil + })) + if err != nil { + return 0, false, err + } + + // Move data from the queue's wait buffer to its read buffer. + nPushed := q.pushWaitBufLocked(l) + + return int64(n), nPushed > 0, nil +} + +// write writes to q from userspace. +// +// Preconditions: +// * l.termiosMu must be held for reading. +func (q *queue) write(ctx context.Context, src usermem.IOSequence, l *lineDiscipline) (int64, error) { + q.mu.Lock() + defer q.mu.Unlock() + + // Copy data into the wait buffer. + n, err := src.CopyInTo(ctx, safemem.WriterFunc(func(src safemem.BlockSeq) (uint64, error) { + copyLen := src.NumBytes() + room := waitBufMaxBytes - q.waitBufLen + // If out of room, return EAGAIN. + if room == 0 && copyLen > 0 { + return 0, syserror.ErrWouldBlock + } + // Cap the size of the wait buffer. + if copyLen > room { + copyLen = room + src = src.TakeFirst64(room) + } + buf := make([]byte, copyLen) + + // Copy the data into the wait buffer. + dst := safemem.BlockSeqOf(safemem.BlockFromSafeSlice(buf)) + n, err := safemem.CopySeq(dst, src) + if err != nil { + return 0, err + } + q.waitBufAppend(buf) + + return n, nil + })) + if err != nil { + return 0, err + } + + // Push data from the wait to the read buffer. + q.pushWaitBufLocked(l) + + return n, nil +} + +// writeBytes writes to q from b. +// +// Preconditions: +// * l.termiosMu must be held for reading. +func (q *queue) writeBytes(b []byte, l *lineDiscipline) { + q.mu.Lock() + defer q.mu.Unlock() + + // Write to the wait buffer. + q.waitBufAppend(b) + q.pushWaitBufLocked(l) +} + +// pushWaitBufLocked fills the queue's read buffer with data from the wait +// buffer. +// +// Preconditions: +// * l.termiosMu must be held for reading. +// * q.mu must be locked. +func (q *queue) pushWaitBufLocked(l *lineDiscipline) int { + if q.waitBufLen == 0 { + return 0 + } + + // Move data from the wait to the read buffer. + var total int + var i int + for i = 0; i < len(q.waitBuf); i++ { + n := q.transform(l, q, q.waitBuf[i]) + total += n + if n != len(q.waitBuf[i]) { + // The read buffer filled up without consuming the + // entire buffer. + q.waitBuf[i] = q.waitBuf[i][n:] + break + } + } + + // Update wait buffer based on consumed data. + q.waitBuf = q.waitBuf[i:] + q.waitBufLen -= uint64(total) + + return total +} + +// Precondition: q.mu must be locked. +func (q *queue) waitBufAppend(b []byte) { + q.waitBuf = append(q.waitBuf, b) + q.waitBufLen += uint64(len(b)) +} + +// LINT.ThenChange(../../fs/tty/queue.go) diff --git a/pkg/sentry/fsimpl/devpts/slave.go b/pkg/sentry/fsimpl/devpts/slave.go new file mode 100644 index 000000000..0a98dc896 --- /dev/null +++ b/pkg/sentry/fsimpl/devpts/slave.go @@ -0,0 +1,186 @@ +// 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 devpts + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/arch" + "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" + "gvisor.dev/gvisor/pkg/usermem" + "gvisor.dev/gvisor/pkg/waiter" +) + +// LINT.IfChange + +// slaveInode is the inode for the slave end of the Terminal. +type slaveInode struct { + kernfs.InodeAttrs + kernfs.InodeNoopRefCount + kernfs.InodeNotDirectory + kernfs.InodeNotSymlink + + // Keep a reference to this inode's dentry. + dentry kernfs.Dentry + + // root is the devpts root inode. + root *rootInode + + // t is the connected Terminal. + t *Terminal +} + +var _ kernfs.Inode = (*slaveInode)(nil) + +// Open implements kernfs.Inode.Open. +func (si *slaveInode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + si.IncRef() + fd := &slaveFileDescription{ + inode: si, + } + if err := fd.vfsfd.Init(fd, opts.Flags, rp.Mount(), vfsd, &vfs.FileDescriptionOptions{}); err != nil { + si.DecRef() + return nil, err + } + return &fd.vfsfd, nil + +} + +// Valid implements kernfs.Inode.Valid. +func (si *slaveInode) Valid(context.Context) bool { + // Return valid if the slave still exists. + si.root.mu.Lock() + defer si.root.mu.Unlock() + _, ok := si.root.slaves[si.t.n] + return ok +} + +// Stat implements kernfs.Inode.Stat. +func (si *slaveInode) Stat(vfsfs *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) { + statx, err := si.InodeAttrs.Stat(vfsfs, opts) + if err != nil { + return linux.Statx{}, err + } + statx.Blksize = 1024 + statx.RdevMajor = linux.UNIX98_PTY_SLAVE_MAJOR + statx.RdevMinor = si.t.n + return statx, nil +} + +// SetStat implements kernfs.Inode.SetStat +func (si *slaveInode) SetStat(ctx context.Context, vfsfs *vfs.Filesystem, creds *auth.Credentials, opts vfs.SetStatOptions) error { + if opts.Stat.Mask&linux.STATX_SIZE != 0 { + return syserror.EINVAL + } + return si.InodeAttrs.SetStat(ctx, vfsfs, creds, opts) +} + +type slaveFileDescription struct { + vfsfd vfs.FileDescription + vfs.FileDescriptionDefaultImpl + + inode *slaveInode +} + +var _ vfs.FileDescriptionImpl = (*slaveFileDescription)(nil) + +// Release implements fs.FileOperations.Release. +func (sfd *slaveFileDescription) Release() { + sfd.inode.DecRef() +} + +// EventRegister implements waiter.Waitable.EventRegister. +func (sfd *slaveFileDescription) EventRegister(e *waiter.Entry, mask waiter.EventMask) { + sfd.inode.t.ld.slaveWaiter.EventRegister(e, mask) +} + +// EventUnregister implements waiter.Waitable.EventUnregister. +func (sfd *slaveFileDescription) EventUnregister(e *waiter.Entry) { + sfd.inode.t.ld.slaveWaiter.EventUnregister(e) +} + +// Readiness implements waiter.Waitable.Readiness. +func (sfd *slaveFileDescription) Readiness(mask waiter.EventMask) waiter.EventMask { + return sfd.inode.t.ld.slaveReadiness() +} + +// Read implements vfs.FileDescriptionImpl.Read. +func (sfd *slaveFileDescription) Read(ctx context.Context, dst usermem.IOSequence, _ vfs.ReadOptions) (int64, error) { + return sfd.inode.t.ld.inputQueueRead(ctx, dst) +} + +// Write implements vfs.FileDescriptionImpl.Write. +func (sfd *slaveFileDescription) Write(ctx context.Context, src usermem.IOSequence, _ vfs.WriteOptions) (int64, error) { + return sfd.inode.t.ld.outputQueueWrite(ctx, src) +} + +// Ioctl implements vfs.FileDescripionImpl.Ioctl. +func (sfd *slaveFileDescription) Ioctl(ctx context.Context, io usermem.IO, args arch.SyscallArguments) (uintptr, error) { + switch cmd := args[1].Uint(); cmd { + case linux.FIONREAD: // linux.FIONREAD == linux.TIOCINQ + // Get the number of bytes in the input queue read buffer. + return 0, sfd.inode.t.ld.inputQueueReadSize(ctx, io, args) + case linux.TCGETS: + return sfd.inode.t.ld.getTermios(ctx, io, args) + case linux.TCSETS: + return sfd.inode.t.ld.setTermios(ctx, io, args) + case linux.TCSETSW: + // TODO(b/29356795): This should drain the output queue first. + return sfd.inode.t.ld.setTermios(ctx, io, args) + case linux.TIOCGPTN: + _, err := usermem.CopyObjectOut(ctx, io, args[2].Pointer(), uint32(sfd.inode.t.n), usermem.IOOpts{ + AddressSpaceActive: true, + }) + return 0, err + case linux.TIOCGWINSZ: + return 0, sfd.inode.t.ld.windowSize(ctx, io, args) + case linux.TIOCSWINSZ: + return 0, sfd.inode.t.ld.setWindowSize(ctx, io, args) + case linux.TIOCSCTTY: + // Make the given terminal the controlling terminal of the + // calling process. + return 0, sfd.inode.t.setControllingTTY(ctx, io, args, false /* isMaster */) + case linux.TIOCNOTTY: + // Release this process's controlling terminal. + return 0, sfd.inode.t.releaseControllingTTY(ctx, io, args, false /* isMaster */) + case linux.TIOCGPGRP: + // Get the foreground process group. + return sfd.inode.t.foregroundProcessGroup(ctx, io, args, false /* isMaster */) + case linux.TIOCSPGRP: + // Set the foreground process group. + return sfd.inode.t.setForegroundProcessGroup(ctx, io, args, false /* isMaster */) + default: + maybeEmitUnimplementedEvent(ctx, cmd) + return 0, syserror.ENOTTY + } +} + +// SetStat implements vfs.FileDescriptionImpl.SetStat. +func (sfd *slaveFileDescription) SetStat(ctx context.Context, opts vfs.SetStatOptions) error { + creds := auth.CredentialsFromContext(ctx) + fs := sfd.vfsfd.VirtualDentry().Mount().Filesystem() + return sfd.inode.SetStat(ctx, fs, creds, opts) +} + +// Stat implements vfs.FileDescriptionImpl.Stat. +func (sfd *slaveFileDescription) Stat(ctx context.Context, opts vfs.StatOptions) (linux.Statx, error) { + fs := sfd.vfsfd.VirtualDentry().Mount().Filesystem() + return sfd.inode.Stat(fs, opts) +} + +// LINT.ThenChange(../../fs/tty/slave.go) diff --git a/pkg/sentry/fsimpl/devpts/terminal.go b/pkg/sentry/fsimpl/devpts/terminal.go new file mode 100644 index 000000000..b44e673d8 --- /dev/null +++ b/pkg/sentry/fsimpl/devpts/terminal.go @@ -0,0 +1,124 @@ +// 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 devpts + +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/usermem" +) + +// LINT.IfChanges + +// Terminal is a pseudoterminal. +// +// +stateify savable +type Terminal struct { + // n is the terminal index. It is immutable. + n uint32 + + // ld is the line discipline of the terminal. It is immutable. + ld *lineDiscipline + + // masterKTTY contains the controlling process of the master end of + // this terminal. This field is immutable. + masterKTTY *kernel.TTY + + // slaveKTTY contains the controlling process of the slave end of this + // terminal. This field is immutable. + slaveKTTY *kernel.TTY +} + +func newTerminal(n uint32) *Terminal { + termios := linux.DefaultSlaveTermios + t := Terminal{ + n: n, + ld: newLineDiscipline(termios), + masterKTTY: &kernel.TTY{Index: n}, + slaveKTTY: &kernel.TTY{Index: n}, + } + return &t +} + +// setControllingTTY makes tm the controlling terminal of the calling thread +// group. +func (tm *Terminal) setControllingTTY(ctx context.Context, io usermem.IO, args arch.SyscallArguments, isMaster bool) error { + task := kernel.TaskFromContext(ctx) + if task == nil { + panic("setControllingTTY must be called from a task context") + } + + return task.ThreadGroup().SetControllingTTY(tm.tty(isMaster), args[2].Int()) +} + +// releaseControllingTTY removes tm as the controlling terminal of the calling +// thread group. +func (tm *Terminal) releaseControllingTTY(ctx context.Context, io usermem.IO, args arch.SyscallArguments, isMaster bool) error { + task := kernel.TaskFromContext(ctx) + if task == nil { + panic("releaseControllingTTY must be called from a task context") + } + + return task.ThreadGroup().ReleaseControllingTTY(tm.tty(isMaster)) +} + +// foregroundProcessGroup gets the process group ID of tm's foreground process. +func (tm *Terminal) foregroundProcessGroup(ctx context.Context, io usermem.IO, args arch.SyscallArguments, isMaster bool) (uintptr, error) { + task := kernel.TaskFromContext(ctx) + if task == nil { + panic("foregroundProcessGroup must be called from a task context") + } + + ret, err := task.ThreadGroup().ForegroundProcessGroup(tm.tty(isMaster)) + if err != nil { + return 0, err + } + + // Write it out to *arg. + _, err = usermem.CopyObjectOut(ctx, io, args[2].Pointer(), int32(ret), usermem.IOOpts{ + AddressSpaceActive: true, + }) + return 0, err +} + +// foregroundProcessGroup sets tm's foreground process. +func (tm *Terminal) setForegroundProcessGroup(ctx context.Context, io usermem.IO, args arch.SyscallArguments, isMaster bool) (uintptr, error) { + task := kernel.TaskFromContext(ctx) + if task == nil { + panic("setForegroundProcessGroup must be called from a task context") + } + + // Read in the process group ID. + var pgid int32 + if _, err := usermem.CopyObjectIn(ctx, io, args[2].Pointer(), &pgid, usermem.IOOpts{ + AddressSpaceActive: true, + }); err != nil { + return 0, err + } + + ret, err := task.ThreadGroup().SetForegroundProcessGroup(tm.tty(isMaster), kernel.ProcessGroupID(pgid)) + return uintptr(ret), err +} + +func (tm *Terminal) tty(isMaster bool) *kernel.TTY { + if isMaster { + return tm.masterKTTY + } + return tm.slaveKTTY +} + +// LINT.ThenChange(../../fs/tty/terminal.go) diff --git a/pkg/sentry/fsimpl/devtmpfs/devtmpfs.go b/pkg/sentry/fsimpl/devtmpfs/devtmpfs.go index abd4f24e7..142ee53b0 100644 --- a/pkg/sentry/fsimpl/devtmpfs/devtmpfs.go +++ b/pkg/sentry/fsimpl/devtmpfs/devtmpfs.go @@ -42,6 +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() { @@ -158,16 +163,25 @@ func (a *Accessor) CreateDeviceFile(ctx context.Context, pathname string, kind v func (a *Accessor) UserspaceInit(ctx context.Context) error { actx := a.wrapContext(ctx) - // systemd: src/shared/dev-setup.c:dev_setup() + // Initialize symlinks. for _, symlink := range []struct { source string target string }{ - // /proc/kcore is not implemented. + // systemd: src/shared/dev-setup.c:dev_setup() {source: "fd", target: "/proc/self/fd"}, {source: "stdin", target: "/proc/self/fd/0"}, {source: "stdout", target: "/proc/self/fd/1"}, {source: "stderr", target: "/proc/self/fd/2"}, + // /proc/kcore is not implemented. + + // Linux implements /dev/ptmx as a device node, but advises + // container implementations to create /dev/ptmx as a symlink + // to pts/ptmx (Documentation/filesystems/devpts.txt). Systemd + // follows this advice (src/nspawn/nspawn.c:setup_pts()), while + // LXC tries to create a bind mount and falls back to a symlink + // (src/lxc/conf.c:lxc_setup_devpts()). + {source: "ptmx", target: "pts/ptmx"}, } { if err := a.vfsObj.SymlinkAt(actx, a.creds, a.pathOperationAt(symlink.source), symlink.target); err != nil { return fmt.Errorf("failed to create symlink %q => %q: %v", symlink.source, symlink.target, err) diff --git a/pkg/sentry/fsimpl/ext/BUILD b/pkg/sentry/fsimpl/ext/BUILD index 6f78f478f..ff861d0fe 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", @@ -80,8 +93,8 @@ go_test( "//pkg/sentry/kernel/auth", "//pkg/sentry/vfs", "//pkg/syserror", + "//pkg/test/testutil", "//pkg/usermem", - "//runsc/testutil", "@com_github_google_go-cmp//cmp:go_default_library", "@com_github_google_go-cmp//cmp/cmpopts:go_default_library", ], 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 bd6ede995..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. @@ -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 29bb73765..64e9a579f 100644 --- a/pkg/sentry/fsimpl/ext/ext_test.go +++ b/pkg/sentry/fsimpl/ext/ext_test.go @@ -32,9 +32,8 @@ import ( "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/test/testutil" "gvisor.dev/gvisor/pkg/usermem" - - "gvisor.dev/gvisor/runsc/testutil" ) const ( diff --git a/pkg/sentry/fsimpl/ext/filesystem.go b/pkg/sentry/fsimpl/ext/filesystem.go index e05429d41..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) @@ -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 6962083f5..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 } @@ -186,7 +186,7 @@ func (in *inode) open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts *vfs.OpenOpt } 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..b9c4beee4 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"]) @@ -15,15 +15,26 @@ 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", "p9file.go", "pagemath.go", "regular_file.go", @@ -41,11 +52,13 @@ go_library( "//pkg/p9", "//pkg/safemem", "//pkg/sentry/fs/fsutil", + "//pkg/sentry/hostfd", "//pkg/sentry/kernel/auth", "//pkg/sentry/kernel/time", "//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 5dbfc6250..c67766ab2 100644 --- a/pkg/sentry/fsimpl/gofer/directory.go +++ b/pkg/sentry/fsimpl/gofer/directory.go @@ -21,21 +21,66 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/p9" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" ) func (d *dentry) isDir() bool { return d.fileType() == linux.S_IFDIR } -// 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{}) +// 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.negativeChildren[name] = struct{}{} + d.children[name] = child +} + +// Preconditions: d.dirMu must be locked. d.isDir(). +func (d *dentry) cacheNegativeLookupLocked(name string) { + // Don't cache negative lookups if InteropModeShared is in effect (since + // this makes remote lookup unavoidable), or if d.isSynthetic() (in which + // case the only files in the directory are those for which a dentry exists + // in d.children). Instead, just delete any previously-cached dentry. + if d.fs.opts.interop == InteropModeShared || d.isSynthetic() { + delete(d.children, name) + return + } + if d.children == nil { + d.children = make(map[string]*dentry) + } + d.children[name] = nil +} + +// createSyntheticDirectory creates a synthetic directory with the given name +// in d. +// +// Preconditions: d.dirMu must be locked. d.isDir(). d does not already contain +// a child with the given name. +func (d *dentry) createSyntheticDirectoryLocked(name string, mode linux.FileMode, kuid auth.KUID, kgid auth.KGID) { + d2 := &dentry{ + refs: 1, // held by d + fs: d.fs, + mode: uint32(mode) | linux.S_IFDIR, + uid: uint32(kuid), + gid: uint32(kgid), + blockSize: usermem.PageSize, // arbitrary + handle: handle{ + fd: -1, + }, + } + d2.pf.dentry = d2 + d2.vfsd.Init(d2) + + d.cacheNewChildLocked(d2, name) + d.syntheticChildren++ } type directoryFD struct { @@ -56,14 +101,19 @@ 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.cachedMetadataAuthoritative() { + d.touchAtime(fd.vfsfd.Mount()) + } + for fd.off < int64(len(fd.dirents)) { if err := cb.Handle(fd.dirents[fd.off]); err != nil { return err @@ -75,23 +125,21 @@ 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() @@ -102,7 +150,7 @@ func (d *dentry) getDirents(ctx context.Context) ([]vfs.Dirent, error) { // 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: ".", @@ -117,50 +165,81 @@ func (d *dentry) getDirents(ctx context.Context) ([]vfs.Dirent, error) { NextOff: 2, }, } - off := uint64(0) - const count = 64 * 1024 // for consistency with the vfs1 client - d.handleMu.RLock() - defer d.handleMu.RUnlock() - if !d.handleReadable { - // This should not be possible because a readable handle should have - // been opened when the calling directoryFD was opened. - panic("gofer.dentry.getDirents called without a readable handle") - } - for { - p9ds, err := d.handle.file.readdir(ctx, off, count) - if err != nil { - return nil, err + var realChildren map[string]struct{} + if !d.isSynthetic() { + if d.syntheticChildren != 0 && d.fs.opts.interop == InteropModeShared { + // Record the set of children d actually has so that we don't emit + // duplicate entries for synthetic children. + realChildren = make(map[string]struct{}) + } + off := uint64(0) + const count = 64 * 1024 // for consistency with the vfs1 client + d.handleMu.RLock() + if !d.handleReadable { + // This should not be possible because a readable handle should + // have been opened when the calling directoryFD was opened. + d.handleMu.RUnlock() + panic("gofer.dentry.getDirents called without a readable handle") } - if len(p9ds) == 0 { - // Cache dirents for future directoryFDs if permitted. - if d.fs.opts.interop != InteropModeShared { - d.dirents = dirents + for { + p9ds, err := d.handle.file.readdir(ctx, off, count) + if err != nil { + d.handleMu.RUnlock() + return nil, err } - return dirents, nil + if len(p9ds) == 0 { + d.handleMu.RUnlock() + break + } + for _, p9d := range p9ds { + if p9d.Name == "." || p9d.Name == ".." { + continue + } + dirent := vfs.Dirent{ + Name: p9d.Name, + Ino: p9d.QID.Path, + NextOff: int64(len(dirents) + 1), + } + // p9 does not expose 9P2000.U's DMDEVICE, DMNAMEDPIPE, or + // DMSOCKET. + switch p9d.Type { + case p9.TypeSymlink: + dirent.Type = linux.DT_LNK + case p9.TypeDir: + dirent.Type = linux.DT_DIR + default: + dirent.Type = linux.DT_REG + } + dirents = append(dirents, dirent) + if realChildren != nil { + realChildren[p9d.Name] = struct{}{} + } + } + off = p9ds[len(p9ds)-1].Offset } - for _, p9d := range p9ds { - if p9d.Name == "." || p9d.Name == ".." { + } + // Emit entries for synthetic children. + if d.syntheticChildren != 0 { + for _, child := range d.children { + if child == nil || !child.isSynthetic() { continue } - dirent := vfs.Dirent{ - Name: p9d.Name, - Ino: p9d.QID.Path, - NextOff: int64(len(dirents) + 1), - } - // p9 does not expose 9P2000.U's DMDEVICE, DMNAMEDPIPE, or - // DMSOCKET. - switch p9d.Type { - case p9.TypeSymlink: - dirent.Type = linux.DT_LNK - case p9.TypeDir: - dirent.Type = linux.DT_DIR - default: - dirent.Type = linux.DT_REG + if _, ok := realChildren[child.name]; ok { + continue } - dirents = append(dirents, dirent) + dirents = append(dirents, vfs.Dirent{ + Name: child.name, + Type: uint8(atomic.LoadUint32(&child.mode) >> 12), + Ino: child.ino, + NextOff: int64(len(dirents) + 1), + }) } - off = p9ds[len(p9ds)-1].Offset } + // Cache dirents for future directoryFDs if permitted. + if d.cachedMetadataAuthoritative() { + d.dirents = dirents + } + return dirents, nil } // Seek implements vfs.FileDescriptionImpl.Seek. diff --git a/pkg/sentry/fsimpl/gofer/filesystem.go b/pkg/sentry/fsimpl/gofer/filesystem.go index 5cfb0dc4c..98ccb42fd 100644 --- a/pkg/sentry/fsimpl/gofer/filesystem.go +++ b/pkg/sentry/fsimpl/gofer/filesystem.go @@ -21,20 +21,24 @@ 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" ) // Sync implements vfs.FilesystemImpl.Sync. func (fs *filesystem) Sync(ctx context.Context) error { - // Snapshot current dentries and special files. + // Snapshot current syncable dentries and special files. fs.syncMu.Lock() - ds := make([]*dentry, 0, len(fs.dentries)) - for d := range fs.dentries { + ds := make([]*dentry, 0, len(fs.syncableDentries)) + for d := range fs.syncableDentries { + d.IncRef() ds = append(ds, d) } sffds := make([]*specialFileFD, 0, len(fs.specialFileFDs)) for sffd := range fs.specialFileFDs { + sffd.vfsfd.IncRef() sffds = append(sffds, sffd) } fs.syncMu.Unlock() @@ -45,9 +49,6 @@ func (fs *filesystem) Sync(ctx context.Context) error { // Sync regular files. for _, d := range ds { - if !d.TryIncRef() { - continue - } err := d.syncSharedHandle(ctx) d.DecRef() if err != nil && retErr == nil { @@ -58,9 +59,6 @@ func (fs *filesystem) Sync(ctx context.Context) error { // Sync special files, which may be writable but do not use dentry shared // handles (so they won't be synced by the above). for _, sffd := range sffds { - if !sffd.vfsfd.TryIncRef() { - continue - } err := sffd.Sync(ctx) sffd.vfsfd.DecRef() if err != nil && retErr == nil { @@ -112,13 +110,15 @@ func putDentrySlice(ds *[]*dentry) { // to *ds. // // 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. +// !rp.Done(). If !d.cachedMetadataAuthoritative(), 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 +128,42 @@ 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 } - 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()) + // 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 + } + if d != d.parent && !d.cachedMetadataAuthoritative() { + _, 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 +178,37 @@ 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) || parent.isSynthetic() { + // 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. !parent.isSynthetic(). +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. @@ -214,29 +216,44 @@ func (fs *filesystem) revalidateChildLocked(ctx context.Context, vfsObj *vfs.Vir child.updateFromP9Attrs(attrMask, &attr) return child, nil } - // 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) + if file.isNil() && child.isSynthetic() { + // We have a synthetic file, and no remote file has arisen to + // replace it. + return child, nil + } + // The file at this path has changed or no longer exists. Mark the + // dentry invalidated, and re-evaluate its caching status (i.e. if it + // has 0 references, drop it). Wait to update parent.children until we + // know what to replace the existing dentry with (i.e. one of the + // returns below), to avoid a redundant map access. + vfsObj.InvalidateDentry(&child.vfsd) + if child.isSynthetic() { + // Normally we don't mark invalidated dentries as deleted since + // they may still exist (but at a different path), and also for + // consistency with Linux. However, synthetic files are guaranteed + // to become unreachable if their dentries are invalidated, so + // treat their invalidation as deletion. + child.setDeleted() + parent.syntheticChildren-- + child.decRefLocked() + parent.dirents = nil + } *ds = appendDentry(*ds, child) - childVFSD = nil } if file.isNil() { // No file exists at this path now. Cache the negative lookup if // allowed. - if fs.opts.interop != InteropModeShared { - parent.cacheNegativeChildLocked(name) - } + parent.cacheNegativeLookupLocked(name) 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) + delete(parent.children, name) 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) @@ -248,8 +265,9 @@ func (fs *filesystem) revalidateChildLocked(ctx context.Context, vfsObj *vfs.Vir // rp.Start().Impl().(*dentry)). It does not check that the returned directory // is searchable by the provider of rp. // -// Preconditions: fs.renameMu must be locked. !rp.Done(). If fs.opts.interop == -// InteropModeShared, then d's cached metadata must be up to date. +// Preconditions: fs.renameMu must be locked. !rp.Done(). If +// !d.cachedMetadataAuthoritative(), then d's cached metadata must be up to +// date. func (fs *filesystem) walkParentDirLocked(ctx context.Context, rp *vfs.ResolvingPath, d *dentry, ds **[]*dentry) (*dentry, error) { for !rp.Final() { d.dirMu.Lock() @@ -271,7 +289,7 @@ func (fs *filesystem) walkParentDirLocked(ctx context.Context, rp *vfs.Resolving // Preconditions: fs.renameMu must be locked. func (fs *filesystem) resolveLocked(ctx context.Context, rp *vfs.ResolvingPath, ds **[]*dentry) (*dentry, error) { d := rp.Start().Impl().(*dentry) - if fs.opts.interop == InteropModeShared { + if !d.cachedMetadataAuthoritative() { // Get updated metadata for rp.Start() as required by fs.stepLocked(). if err := d.updateFromGetattr(ctx); err != nil { return nil, err @@ -293,16 +311,17 @@ func (fs *filesystem) resolveLocked(ctx context.Context, rp *vfs.ResolvingPath, } // doCreateAt checks that creating a file at rp is permitted, then invokes -// create to do so. +// createInRemoteDir (if the parent directory is a real remote directory) or +// createInSyntheticDir (if the parent directory is synthetic) to do so. // // Preconditions: !rp.Done(). For the final path component in rp, // !rp.ShouldFollowSymlink(). -func (fs *filesystem) doCreateAt(ctx context.Context, rp *vfs.ResolvingPath, dir bool, create func(parent *dentry, name string) error) error { +func (fs *filesystem) doCreateAt(ctx context.Context, rp *vfs.ResolvingPath, dir bool, createInRemoteDir func(parent *dentry, name string) error, createInSyntheticDir func(parent *dentry, name string) error) error { var ds *[]*dentry fs.renameMu.RLock() defer fs.renameMuRUnlockAndCheckCaching(&ds) start := rp.Start().Impl().(*dentry) - if fs.opts.interop == InteropModeShared { + if !start.cachedMetadataAuthoritative() { // Get updated metadata for start as required by // fs.walkParentDirLocked(). if err := start.updateFromGetattr(ctx); err != nil { @@ -313,12 +332,9 @@ 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 +345,9 @@ 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 @@ -336,6 +355,20 @@ func (fs *filesystem) doCreateAt(ctx context.Context, rp *vfs.ResolvingPath, dir defer mnt.EndWrite() parent.dirMu.Lock() defer parent.dirMu.Unlock() + if parent.isSynthetic() { + if child := parent.children[name]; child != nil { + return syserror.EEXIST + } + if createInSyntheticDir == nil { + return syserror.EPERM + } + if err := createInSyntheticDir(parent, name); err != nil { + return err + } + parent.touchCMtime() + parent.dirents = nil + return nil + } if fs.opts.interop == InteropModeShared { // The existence of a dentry at name would be inconclusive because the // file it represents may have been deleted from the remote filesystem, @@ -344,18 +377,21 @@ func (fs *filesystem) doCreateAt(ctx context.Context, rp *vfs.ResolvingPath, dir // will fail with EEXIST like we would have. If the RPC succeeds, and a // stale dentry exists, the dentry will fail revalidation next time // it's used. - return create(parent, name) + return createInRemoteDir(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 // at name. As above, we attempt the file creation RPC anyway. - if err := create(parent, name); err != nil { + if err := createInRemoteDir(parent, name); err != nil { return err } - parent.touchCMtime(ctx) - delete(parent.negativeChildren, name) + if child, ok := parent.children[name]; ok && child == nil { + // Delete the now-stale negative dentry. + delete(parent.children, name) + } + parent.touchCMtime() parent.dirents = nil return nil } @@ -366,7 +402,7 @@ func (fs *filesystem) unlinkAt(ctx context.Context, rp *vfs.ResolvingPath, dir b fs.renameMu.RLock() defer fs.renameMuRUnlockAndCheckCaching(&ds) start := rp.Start().Impl().(*dentry) - if fs.opts.interop == InteropModeShared { + if !start.cachedMetadataAuthoritative() { // Get updated metadata for start as required by // fs.walkParentDirLocked(). if err := start.updateFromGetattr(ctx); err != nil { @@ -377,7 +413,7 @@ 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 { @@ -403,64 +439,107 @@ func (fs *filesystem) unlinkAt(ctx context.Context, rp *vfs.ResolvingPath, dir b 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 { + child.dirMu.Lock() + defer child.dirMu.Unlock() + if err := vfsObj.PrepareDeleteDentry(mntns, &child.vfsd); err != nil { + if parent.cachedMetadataAuthoritative() { + 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 a dentry exists, use it for best-effort checks on its deletability. if dir { - if child != nil && !child.isDir() { - return syserror.ENOTDIR + if child != nil { + // child must be an empty directory. + if child.syntheticChildren != 0 { + // This is definitely not an empty directory, irrespective of + // fs.opts.interop. + vfsObj.AbortDeleteDentry(&child.vfsd) + return syserror.ENOTEMPTY + } + // If InteropModeShared is in effect and the first call to + // PrepareDeleteDentry above succeeded, then child wasn't + // revalidated (so we can't expect its file type to be correct) and + // individually revalidating its children (to confirm that they + // still exist) would be a waste of time. + if child.cachedMetadataAuthoritative() { + if !child.isDir() { + vfsObj.AbortDeleteDentry(&child.vfsd) + return syserror.ENOTDIR + } + for _, grandchild := range child.children { + if grandchild != nil { + vfsObj.AbortDeleteDentry(&child.vfsd) + return syserror.ENOTEMPTY + } + } + } } flags = linux.AT_REMOVEDIR } else { + // child must be a non-directory file. 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 parent.isSynthetic() { + if child == nil { + return syserror.ENOENT + } + } else { + err = parent.file.unlinkAt(ctx, name, flags) + if err != nil { + if child != nil { + vfsObj.AbortDeleteDentry(&child.vfsd) + } + return err } - return err - } - if fs.opts.interop != InteropModeShared { - parent.touchCMtime(ctx) - parent.cacheNegativeChildLocked(name) - parent.dirents = nil } if child != nil { + vfsObj.CommitDeleteDentry(&child.vfsd) child.setDeleted() - vfsObj.CommitDeleteDentry(childVFSD) + if child.isSynthetic() { + parent.syntheticChildren-- + child.decRefLocked() + } ds = appendDentry(ds, child) } + parent.cacheNegativeLookupLocked(name) + if parent.cachedMetadataAuthoritative() { + parent.dirents = nil + parent.touchCMtime() + if dir { + parent.decLinks() + } + } return nil } @@ -499,6 +578,18 @@ 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 @@ -512,7 +603,7 @@ 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 } } @@ -526,7 +617,7 @@ func (fs *filesystem) GetParentDentryAt(ctx context.Context, rp *vfs.ResolvingPa fs.renameMu.RLock() defer fs.renameMuRUnlockAndCheckCaching(&ds) start := rp.Start().Impl().(*dentry) - if fs.opts.interop == InteropModeShared { + if !start.cachedMetadataAuthoritative() { // Get updated metadata for start as required by // fs.walkParentDirLocked(). if err := start.updateFromGetattr(ctx); err != nil { @@ -549,15 +640,32 @@ func (fs *filesystem) LinkAt(ctx context.Context, rp *vfs.ResolvingPath, vd vfs. } // 9P2000.L supports hard links, but we don't. return syserror.EPERM - }) + }, nil) } // MkdirAt implements vfs.FilesystemImpl.MkdirAt. func (fs *filesystem) MkdirAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.MkdirOptions) error { + creds := rp.Credentials() 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 { + if !opts.ForSyntheticMountpoint || err == syserror.EEXIST { + return err + } + ctx.Infof("Failed to create remote directory %q: %v; falling back to synthetic directory", name, err) + parent.createSyntheticDirectoryLocked(name, opts.Mode, creds.EffectiveKUID, creds.EffectiveKGID) + } + if fs.opts.interop != InteropModeShared { + parent.incLinks() + } + return nil + }, func(parent *dentry, name string) error { + if !opts.ForSyntheticMountpoint { + // Can't create non-synthetic files in synthetic directories. + return syserror.EPERM + } + parent.createSyntheticDirectoryLocked(name, opts.Mode, creds.EffectiveKUID, creds.EffectiveKGID) + parent.incLinks() + return nil }) } @@ -567,7 +675,7 @@ func (fs *filesystem) MknodAt(ctx context.Context, rp *vfs.ResolvingPath, opts v creds := rp.Credentials() _, err := parent.file.mknod(ctx, name, (p9.FileMode)(opts.Mode), opts.DevMajor, opts.DevMinor, (p9.UID)(creds.EffectiveKUID), (p9.GID)(creds.EffectiveKGID)) return err - }) + }, nil) } // OpenAt implements vfs.FilesystemImpl.OpenAt. @@ -587,7 +695,7 @@ func (fs *filesystem) OpenAt(ctx context.Context, rp *vfs.ResolvingPath, opts vf defer fs.renameMuRUnlockAndCheckCaching(&ds) start := rp.Start().Impl().(*dentry) - if fs.opts.interop == InteropModeShared { + if !start.cachedMetadataAuthoritative() { // Get updated metadata for start as required by fs.stepLocked(). if err := start.updateFromGetattr(ctx); err != nil { return nil, err @@ -603,13 +711,17 @@ 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. parent.dirMu.Lock() child, err := fs.stepLocked(ctx, rp, parent, &ds) if err == syserror.ENOENT && mayCreate { + if parent.isSynthetic() { + parent.dirMu.Unlock() + return nil, syserror.EPERM + } fd, err := parent.createAndOpenChildLocked(ctx, rp, &opts) parent.dirMu.Unlock() return fd, err @@ -640,7 +752,7 @@ afterTrailingSymlink: // Preconditions: fs.renameMu must be locked. 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, d.isDir()); err != nil { + if err := d.checkPermissions(rp.Credentials(), ats); err != nil { return nil, err } mnt := rp.Mount() @@ -669,8 +781,10 @@ func (d *dentry) openLocked(ctx context.Context, rp *vfs.ResolvingPath, opts *vf 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 + if !d.isSynthetic() { + 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, opts.Flags, mnt, &d.vfsd, &vfs.FileDescriptionOptions{}); err != nil { @@ -700,8 +814,9 @@ func (d *dentry) openLocked(ctx context.Context, rp *vfs.ResolvingPath, opts *vf } // Preconditions: d.fs.renameMu must be locked. d.dirMu must be locked. +// !d.isSynthetic(). 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() { @@ -777,17 +892,14 @@ 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) - if d.fs.opts.interop != InteropModeShared { - d.touchCMtime(ctx) - delete(d.negativeChildren, name) + d.cacheNewChildLocked(child, name) + if d.cachedMetadataAuthoritative() { + 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{ @@ -858,12 +970,12 @@ func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa defer mnt.EndWrite() oldParent := oldParentVD.Dentry().Impl().(*dentry) - if fs.opts.interop == InteropModeShared { + if !oldParent.cachedMetadataAuthoritative() { if err := oldParent.updateFromGetattr(ctx); err != nil { 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() @@ -871,7 +983,7 @@ 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 } @@ -879,11 +991,11 @@ 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 } } @@ -894,7 +1006,7 @@ 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() @@ -903,32 +1015,21 @@ 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 - // This is similar to unlinkAt, except: - // - // - We revalidate the replaced dentry unconditionally 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 err != nil { - return err - } - if replaced != nil { - if replaced.isDir() { - if !renamed.isDir() { - return syserror.EISDIR - } - } else { - if rp.MustBeDir() || renamed.isDir() { - return syserror.ENOTDIR - } + replaced, err := fs.getChildLocked(ctx, rp.VirtualFilesystem(), newParent, newName, &ds) + if err != nil { + return err + } + var replacedVFSD *vfs.Dentry + if replaced != nil { + replacedVFSD = &replaced.vfsd + if replaced.isDir() { + if !renamed.isDir() { + return syserror.EISDIR } - replacedVFSD = &replaced.vfsd } else { - replacedVFSD = nil + if rp.MustBeDir() || renamed.isDir() { + return syserror.ENOTDIR + } } } @@ -940,17 +1041,73 @@ func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa if err := vfsObj.PrepareRenameDentry(mntns, &renamed.vfsd, replacedVFSD); err != nil { return err } - if err := renamed.file.rename(ctx, newParent.file, newName); err != nil { - vfsObj.AbortRenameDentry(&renamed.vfsd, replacedVFSD) - return err + + // Update the remote filesystem. + if !renamed.isSynthetic() { + if err := renamed.file.rename(ctx, newParent.file, newName); err != nil { + vfsObj.AbortRenameDentry(&renamed.vfsd, replacedVFSD) + return err + } + } else if replaced != nil && !replaced.isSynthetic() { + // We are replacing an existing real file with a synthetic one, so we + // need to unlink the former. + flags := uint32(0) + if replaced.isDir() { + flags = linux.AT_REMOVEDIR + } + if err := newParent.file.unlinkAt(ctx, newName, flags); err != nil { + vfsObj.AbortRenameDentry(&renamed.vfsd, replacedVFSD) + return err + } + } + + // Update the dentry tree. + vfsObj.CommitRenameReplaceDentry(&renamed.vfsd, replacedVFSD) + if replaced != nil { + replaced.setDeleted() + if replaced.isSynthetic() { + newParent.syntheticChildren-- + replaced.decRefLocked() + } + ds = appendDentry(ds, replaced) + } + oldParent.cacheNegativeLookupLocked(oldName) + // We don't use newParent.cacheNewChildLocked() since we don't want to mess + // with reference counts and queue oldParent for checkCachingLocked if the + // parent isn't actually changing. + if oldParent != newParent { + ds = appendDentry(ds, oldParent) + newParent.IncRef() + if renamed.isSynthetic() { + oldParent.syntheticChildren-- + newParent.syntheticChildren++ + } } - if fs.opts.interop != InteropModeShared { - oldParent.cacheNegativeChildLocked(oldName) + renamed.parent = newParent + renamed.name = newName + if newParent.children == nil { + newParent.children = make(map[string]*dentry) + } + newParent.children[newName] = renamed + + // Update metadata. + if renamed.cachedMetadataAuthoritative() { + renamed.touchCtime() + } + if oldParent.cachedMetadataAuthoritative() { oldParent.dirents = nil - delete(newParent.negativeChildren, newName) + oldParent.touchCMtime() + if renamed.isDir() { + oldParent.decLinks() + } + } + if newParent.cachedMetadataAuthoritative() { newParent.dirents = nil + newParent.touchCMtime() + if renamed.isDir() { + newParent.incLinks() + } } - vfsObj.CommitRenameReplaceDentry(&renamed.vfsd, &newParent.vfsd, newName, replacedVFSD) return nil } @@ -997,6 +1154,10 @@ func (fs *filesystem) StatFSAt(ctx context.Context, rp *vfs.ResolvingPath) (linu if err != nil { return linux.Statfs{}, err } + // If d is synthetic, invoke statfs on the first ancestor of d that isn't. + for d.isSynthetic() { + d = d.parent + } fsstat, err := d.file.statFS(ctx) if err != nil { return linux.Statfs{}, err @@ -1026,7 +1187,7 @@ func (fs *filesystem) SymlinkAt(ctx context.Context, rp *vfs.ResolvingPath, targ creds := rp.Credentials() _, err := parent.file.symlink(ctx, target, name, (p9.UID)(creds.EffectiveKUID), (p9.GID)(creds.EffectiveKGID)) return err - }) + }, nil) } // UnlinkAt implements vfs.FilesystemImpl.UnlinkAt. @@ -1034,8 +1195,21 @@ func (fs *filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error return fs.unlinkAt(ctx, rp, false /* dir */) } +// BoundEndpointAt implements FilesystemImpl.BoundEndpointAt. +func (fs *filesystem) BoundEndpointAt(ctx context.Context, rp *vfs.ResolvingPath) (transport.BoundEndpoint, error) { + var ds *[]*dentry + fs.renameMu.RLock() + defer fs.renameMuRUnlockAndCheckCaching(&ds) + _, err := fs.resolveLocked(ctx, rp, &ds) + if err != nil { + return nil, err + } + // TODO(gvisor.dev/issue/1476): Implement BoundEndpointAt. + 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) @@ -1043,11 +1217,11 @@ 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) @@ -1055,7 +1229,7 @@ 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. @@ -1067,7 +1241,7 @@ 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. @@ -1079,12 +1253,12 @@ 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 c4a8f0b38..8b4e91d17 100644 --- a/pkg/sentry/fsimpl/gofer/gofer.go +++ b/pkg/sentry/fsimpl/gofer/gofer.go @@ -27,13 +27,15 @@ // dentry.handleMu // dentry.dataMu // -// Locking dentry.dirMu in multiple dentries requires holding -// filesystem.renameMu for writing. +// Locking dentry.dirMu in multiple dentries requires that either ancestor +// dentries are locked before descendant dentries, or that filesystem.renameMu +// is locked for writing. package gofer import ( "fmt" "strconv" + "strings" "sync" "sync/atomic" "syscall" @@ -44,6 +46,7 @@ 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" @@ -72,6 +75,9 @@ 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. @@ -97,11 +103,12 @@ type filesystem struct { cachedDentries dentryList cachedDentriesLen uint64 - // dentries contains all dentries in this filesystem. specialFileFDs - // contains all open specialFileFDs. These fields are protected by syncMu. - syncMu sync.Mutex - dentries map[*dentry]struct{} - specialFileFDs map[*specialFileFD]struct{} + // syncableDentries contains all dentries in this filesystem for which + // !dentry.file.isNil(). specialFileFDs contains all open specialFileFDs. + // These fields are protected by syncMu. + syncMu sync.Mutex + syncableDentries map[*dentry]struct{} + specialFileFDs map[*specialFileFD]struct{} } type filesystemOptions struct { @@ -182,7 +189,8 @@ const ( // InteropModeShared is appropriate when there are users of the remote // filesystem that may mutate its state other than the client. // - // - The client must verify cached filesystem state before using it. + // - The client must verify ("revalidate") cached filesystem state before + // using it. // // - Client changes to filesystem state must be sent to the remote // filesystem synchronously. @@ -199,6 +207,11 @@ 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) @@ -366,15 +379,16 @@ func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virt // Construct the filesystem object. fs := &filesystem{ - mfp: mfp, - opts: fsopts, - uid: creds.EffectiveKUID, - gid: creds.EffectiveKGID, - client: client, - dentries: make(map[*dentry]struct{}), - specialFileFDs: make(map[*specialFileFD]struct{}), + mfp: mfp, + opts: fsopts, + uid: creds.EffectiveKUID, + gid: creds.EffectiveKGID, + client: client, + clock: ktime.RealtimeClockFromContext(ctx), + syncableDentries: 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) @@ -398,7 +412,7 @@ func (fs *filesystem) Release() { mf := fs.mfp.MemoryFile() fs.syncMu.Lock() - for d := range fs.dentries { + for d := range fs.syncableDentries { d.handleMu.Lock() d.dataMu.Lock() if d.handleWritable { @@ -433,16 +447,35 @@ type dentry struct { vfsd vfs.Dentry // refs is the reference count. Each dentry holds a reference on its - // parent, even if disowned. refs is accessed using atomic memory - // operations. + // parent, even if disowned. An additional reference is held on all + // synthetic dentries until they are unlinked or invalidated. When refs + // reaches 0, the dentry may be added to the cache or destroyed. If refs == + // -1, the dentry has already been destroyed. refs is accessed using atomic + // memory operations. 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. + // + // If file.isNil(), this dentry represents a synthetic file, i.e. a file + // that does not exist on the remote filesystem. As of this writing, this + // is only possible for a directory created with + // MkdirOptions.ForSyntheticMountpoint == true. + // TODO(gvisor.dev/issue/1476): Support synthetic sockets (and pipes). file p9file // If deleted is non-zero, the file represented by this dentry has been @@ -457,15 +490,26 @@ 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, InteropModeShared is not in - // effect, and dirents is not nil, it is a cache of all entries in the - // directory, in the order they were returned by the server. dirents is - // protected by dirMu. + // 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 and the directory + // is not synthetic). + // + // children is protected by dirMu. + children map[string]*dentry + + // If this dentry represents a directory, syntheticChildren is the number + // of child dentries for which dentry.isSynthetic() == true. + // syntheticChildren is protected by dirMu. + syntheticChildren int + + // If this dentry represents a directory, + // dentry.cachedMetadataAuthoritative() == true, and dirents is not nil, it + // is a cache of all entries in the directory, in the order they were + // returned by the server. dirents is protected by dirMu. dirents []vfs.Dirent // Cached metadata; protected by metadataMu and accessed using atomic @@ -485,6 +529,11 @@ 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 @@ -557,6 +606,8 @@ func dentryAttrMask() p9.AttrMask { // initially has no references, but is not cached; it is the caller's // responsibility to set the dentry's reference count and/or call // dentry.checkCachingLocked() as appropriate. +// +// Preconditions: !file.isNil(). func (fs *filesystem) newDentry(ctx context.Context, file p9file, qid p9.QID, mask p9.AttrMask, attr *p9.Attr) (*dentry, error) { if !mask.Mode { ctx.Warningf("can't create gofer.dentry without file type") @@ -580,10 +631,10 @@ func (fs *filesystem) newDentry(ctx context.Context, file p9file, qid p9.QID, ma }, } d.pf.dentry = d - if mask.UID { + if mask.UID && attr.UID != auth.NoID { d.uid = uint32(attr.UID) } - if mask.GID { + if mask.GID && attr.GID != auth.NoID { d.gid = uint32(attr.GID) } if mask.Size { @@ -604,14 +655,25 @@ 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() - fs.dentries[d] = struct{}{} + fs.syncableDentries[d] = struct{}{} fs.syncMu.Unlock() return d, nil } +func (d *dentry) isSynthetic() bool { + return d.file.isNil() +} + +func (d *dentry) cachedMetadataAuthoritative() bool { + return d.fs.opts.interop != InteropModeShared || d.isSynthetic() +} + // updateFromP9Attrs is called to update d's metadata after an update from the // remote filesystem. func (d *dentry) updateFromP9Attrs(mask p9.AttrMask, attr *p9.Attr) { @@ -645,6 +707,9 @@ 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) @@ -653,6 +718,7 @@ func (d *dentry) updateFromP9Attrs(mask p9.AttrMask, attr *p9.Attr) { d.metadataMu.Unlock() } +// Preconditions: !d.isSynthetic() func (d *dentry) updateFromGetattr(ctx context.Context) error { // Use d.handle.file, which represents a 9P fid that has been opened, in // preference to d.file, which represents a 9P fid that has not. This may @@ -687,10 +753,7 @@ 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)) @@ -703,7 +766,7 @@ 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 { @@ -713,7 +776,8 @@ 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 { @@ -722,7 +786,7 @@ func (d *dentry) setStat(ctx context.Context, creds *auth.Credentials, stat *lin defer mnt.EndWrite() setLocalAtime := false setLocalMtime := false - if d.fs.opts.interop != InteropModeShared { + if d.cachedMetadataAuthoritative() { // Timestamp updates will be handled locally. setLocalAtime = stat.Mask&linux.STATX_ATIME != 0 setLocalMtime = stat.Mask&linux.STATX_MTIME != 0 @@ -735,40 +799,39 @@ func (d *dentry) setStat(ctx context.Context, creds *auth.Credentials, stat *lin } d.metadataMu.Lock() defer d.metadataMu.Unlock() - if stat.Mask != 0 { - if err := d.file.setAttr(ctx, p9.SetAttrMask{ - Permissions: stat.Mask&linux.STATX_MODE != 0, - UID: stat.Mask&linux.STATX_UID != 0, - GID: stat.Mask&linux.STATX_GID != 0, - Size: stat.Mask&linux.STATX_SIZE != 0, - ATime: stat.Mask&linux.STATX_ATIME != 0, - MTime: stat.Mask&linux.STATX_MTIME != 0, - ATimeNotSystemTime: stat.Atime.Nsec != linux.UTIME_NOW, - MTimeNotSystemTime: stat.Mtime.Nsec != linux.UTIME_NOW, - }, p9.SetAttr{ - Permissions: p9.FileMode(stat.Mode), - UID: p9.UID(stat.UID), - GID: p9.GID(stat.GID), - Size: stat.Size, - ATimeSeconds: uint64(stat.Atime.Sec), - ATimeNanoSeconds: uint64(stat.Atime.Nsec), - MTimeSeconds: uint64(stat.Mtime.Sec), - MTimeNanoSeconds: uint64(stat.Mtime.Nsec), - }); err != nil { - return err + if !d.isSynthetic() { + if stat.Mask != 0 { + if err := d.file.setAttr(ctx, p9.SetAttrMask{ + Permissions: stat.Mask&linux.STATX_MODE != 0, + UID: stat.Mask&linux.STATX_UID != 0, + GID: stat.Mask&linux.STATX_GID != 0, + Size: stat.Mask&linux.STATX_SIZE != 0, + ATime: stat.Mask&linux.STATX_ATIME != 0, + MTime: stat.Mask&linux.STATX_MTIME != 0, + ATimeNotSystemTime: stat.Atime.Nsec != linux.UTIME_NOW, + MTimeNotSystemTime: stat.Mtime.Nsec != linux.UTIME_NOW, + }, p9.SetAttr{ + Permissions: p9.FileMode(stat.Mode), + UID: p9.UID(stat.UID), + GID: p9.GID(stat.GID), + Size: stat.Size, + ATimeSeconds: uint64(stat.Atime.Sec), + ATimeNanoSeconds: uint64(stat.Atime.Nsec), + MTimeSeconds: uint64(stat.Mtime.Sec), + MTimeNanoSeconds: uint64(stat.Mtime.Nsec), + }); err != nil { + return err + } + } + if d.fs.opts.interop == InteropModeShared { + // There's no point to updating d's metadata in this case since + // it'll be overwritten by revalidation before the next time it's + // used anyway. (InteropModeShared inhibits client caching of + // regular file data, so there's no cache to truncate either.) + return nil } } - if d.fs.opts.interop == InteropModeShared { - // There's no point to updating d's metadata in this case since it'll - // be overwritten by revalidation before the next time it's used - // anyway. (InteropModeShared inhibits client caching of regular file - // 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)) } @@ -780,25 +843,19 @@ 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 @@ -835,8 +892,8 @@ 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. @@ -850,7 +907,7 @@ 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) { @@ -870,16 +927,32 @@ func (d *dentry) DecRef() { } } +// decRefLocked decrements d's reference count without calling +// d.checkCachingLocked, even if d's reference count reaches 0; callers are +// responsible for ensuring that d.checkCachingLocked will be called later. +func (d *dentry) decRefLocked() { + if refs := atomic.AddInt64(&d.refs, -1); refs < 0 { + panic("gofer.dentry.decRefLocked() called without holding a reference") + } +} + // 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-- @@ -887,9 +960,13 @@ 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-- @@ -914,34 +991,45 @@ 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() { @@ -961,17 +1049,19 @@ func (d *dentry) destroyLocked() { d.handle.close(ctx) } d.handleMu.Unlock() - d.file.close(ctx) - // 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() + if !d.file.isNil() { + d.file.close(ctx) + d.file = p9file{} + // Remove d from the set of syncable dentries. + d.fs.syncMu.Lock() + delete(d.fs.syncableDentries, d) + d.fs.syncMu.Unlock() + } + // 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") } @@ -986,24 +1076,65 @@ 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) { + if d.file.isNil() { + return nil, nil + } + 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 d.file.isNil() { + return "", syserror.ENODATA + } + 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 d.file.isNil() { + return syserror.EPERM + } + 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 d.file.isNil() { + return syserror.EPERM + } + 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(). +// Preconditions: !d.file.isNil(). d.isRegularFile() || d.isDirectory(). func (d *dentry) ensureSharedHandle(ctx context.Context, read, write, trunc bool) error { // O_TRUNC unconditionally requires us to obtain a new handle (opened with // O_TRUNC). @@ -1051,7 +1182,7 @@ func (d *dentry) ensureSharedHandle(ctx context.Context, read, write, trunc bool // 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.Dup3(int(h.fd), int(d.handle.fd), 0); 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) @@ -1094,6 +1225,26 @@ 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 { @@ -1112,7 +1263,8 @@ 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.cachedMetadataAuthoritative() && 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 { @@ -1130,21 +1282,21 @@ 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..adff39490 --- /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{ + syncableDentries: 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/handle.go b/pkg/sentry/fsimpl/gofer/handle.go index cfe66f797..724a3f1f7 100644 --- a/pkg/sentry/fsimpl/gofer/handle.go +++ b/pkg/sentry/fsimpl/gofer/handle.go @@ -20,6 +20,7 @@ import ( "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/p9" "gvisor.dev/gvisor/pkg/safemem" + "gvisor.dev/gvisor/pkg/sentry/hostfd" ) // handle represents a remote "open file descriptor", consisting of an opened @@ -77,7 +78,7 @@ func (h *handle) readToBlocksAt(ctx context.Context, dsts safemem.BlockSeq, offs } if h.fd >= 0 { ctx.UninterruptibleSleepStart(false) - n, err := hostPreadv(h.fd, dsts, int64(offset)) + n, err := hostfd.Preadv2(h.fd, dsts, int64(offset), 0 /* flags */) ctx.UninterruptibleSleepFinish(false) return n, err } @@ -103,7 +104,7 @@ func (h *handle) writeFromBlocksAt(ctx context.Context, srcs safemem.BlockSeq, o } if h.fd >= 0 { ctx.UninterruptibleSleepStart(false) - n, err := hostPwritev(h.fd, srcs, int64(offset)) + n, err := hostfd.Pwritev2(h.fd, srcs, int64(offset), 0 /* flags */) ctx.UninterruptibleSleepFinish(false) return n, err } diff --git a/pkg/sentry/fsimpl/gofer/handle_unsafe.go b/pkg/sentry/fsimpl/gofer/handle_unsafe.go deleted file mode 100644 index 19560ab26..000000000 --- a/pkg/sentry/fsimpl/gofer/handle_unsafe.go +++ /dev/null @@ -1,66 +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 gofer - -import ( - "syscall" - "unsafe" - - "gvisor.dev/gvisor/pkg/safemem" -) - -// Preconditions: !dsts.IsEmpty(). -func hostPreadv(fd int32, dsts safemem.BlockSeq, off int64) (uint64, error) { - // No buffering is necessary regardless of safecopy; host syscalls will - // return EFAULT if appropriate, instead of raising SIGBUS. - if dsts.NumBlocks() == 1 { - // Use pread() instead of preadv() to avoid iovec allocation and - // copying. - dst := dsts.Head() - n, _, e := syscall.Syscall6(syscall.SYS_PREAD64, uintptr(fd), dst.Addr(), uintptr(dst.Len()), uintptr(off), 0, 0) - if e != 0 { - return 0, e - } - return uint64(n), nil - } - iovs := safemem.IovecsFromBlockSeq(dsts) - n, _, e := syscall.Syscall6(syscall.SYS_PREADV, uintptr(fd), uintptr((unsafe.Pointer)(&iovs[0])), uintptr(len(iovs)), uintptr(off), 0, 0) - if e != 0 { - return 0, e - } - return uint64(n), nil -} - -// Preconditions: !srcs.IsEmpty(). -func hostPwritev(fd int32, srcs safemem.BlockSeq, off int64) (uint64, error) { - // No buffering is necessary regardless of safecopy; host syscalls will - // return EFAULT if appropriate, instead of raising SIGBUS. - if srcs.NumBlocks() == 1 { - // Use pwrite() instead of pwritev() to avoid iovec allocation and - // copying. - src := srcs.Head() - n, _, e := syscall.Syscall6(syscall.SYS_PWRITE64, uintptr(fd), src.Addr(), uintptr(src.Len()), uintptr(off), 0, 0) - if e != 0 { - return 0, e - } - return uint64(n), nil - } - iovs := safemem.IovecsFromBlockSeq(srcs) - n, _, e := syscall.Syscall6(syscall.SYS_PWRITEV, uintptr(fd), uintptr((unsafe.Pointer)(&iovs[0])), uintptr(len(iovs)), uintptr(off), 0, 0) - if e != 0 { - return 0, e - } - return uint64(n), nil -} 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 e95209661..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 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 index 731f192b3..2dcb03a73 100644 --- a/pkg/sentry/fsimpl/host/BUILD +++ b/pkg/sentry/fsimpl/host/BUILD @@ -5,19 +5,28 @@ licenses(["notice"]) go_library( name = "host", srcs = [ - "default_file.go", + "control.go", "host.go", + "ioctl_unsafe.go", + "tty.go", "util.go", + "util_unsafe.go", ], + visibility = ["//pkg/sentry:internal"], deps = [ "//pkg/abi/linux", "//pkg/context", "//pkg/log", "//pkg/refs", - "//pkg/safemem", + "//pkg/sentry/arch", "//pkg/sentry/fsimpl/kernfs", + "//pkg/sentry/hostfd", + "//pkg/sentry/kernel", "//pkg/sentry/kernel/auth", "//pkg/sentry/memmap", + "//pkg/sentry/socket/control", + "//pkg/sentry/socket/unix/transport", + "//pkg/sentry/unimpl", "//pkg/sentry/vfs", "//pkg/sync", "//pkg/syserror", diff --git a/pkg/sentry/fsimpl/host/control.go b/pkg/sentry/fsimpl/host/control.go new file mode 100644 index 000000000..b9082a20f --- /dev/null +++ b/pkg/sentry/fsimpl/host/control.go @@ -0,0 +1,96 @@ +// 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" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/socket/control" + "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" + "gvisor.dev/gvisor/pkg/sentry/vfs" +) + +type scmRights struct { + fds []int +} + +func newSCMRights(fds []int) control.SCMRightsVFS2 { + return &scmRights{fds} +} + +// Files implements control.SCMRights.Files. +func (c *scmRights) Files(ctx context.Context, max int) (control.RightsFilesVFS2, bool) { + n := max + var trunc bool + if l := len(c.fds); n > l { + n = l + } else if n < l { + trunc = true + } + + rf := control.RightsFilesVFS2(fdsToFiles(ctx, c.fds[:n])) + + // Only consume converted FDs (fdsToFiles may convert fewer than n FDs). + c.fds = c.fds[len(rf):] + return rf, trunc +} + +// Clone implements transport.RightsControlMessage.Clone. +func (c *scmRights) Clone() transport.RightsControlMessage { + // Host rights never need to be cloned. + return nil +} + +// Release implements transport.RightsControlMessage.Release. +func (c *scmRights) Release() { + for _, fd := range c.fds { + syscall.Close(fd) + } + c.fds = nil +} + +// If an error is encountered, only files created before the error will be +// returned. This is what Linux does. +func fdsToFiles(ctx context.Context, fds []int) []*vfs.FileDescription { + files := make([]*vfs.FileDescription, 0, len(fds)) + for _, fd := range fds { + // Get flags. We do it here because they may be modified + // by subsequent functions. + fileFlags, _, errno := syscall.Syscall(syscall.SYS_FCNTL, uintptr(fd), syscall.F_GETFL, 0) + if errno != 0 { + ctx.Warningf("Error retrieving host FD flags: %v", error(errno)) + break + } + + // Create the file backed by hostFD. + file, err := ImportFD(ctx, kernel.KernelFromContext(ctx).HostMount(), fd, false /* isTTY */) + if err != nil { + ctx.Warningf("Error creating file from host FD: %v", err) + break + } + + if err := file.SetStatusFlags(ctx, auth.CredentialsFromContext(ctx), uint32(fileFlags&linux.O_NONBLOCK)); err != nil { + ctx.Warningf("Error setting flags on host FD file: %v", err) + break + } + + files = append(files, file) + } + return files +} diff --git a/pkg/sentry/fsimpl/host/default_file.go b/pkg/sentry/fsimpl/host/default_file.go deleted file mode 100644 index 172cdb161..000000000 --- a/pkg/sentry/fsimpl/host/default_file.go +++ /dev/null @@ -1,233 +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 host - -import ( - "math" - "syscall" - - "golang.org/x/sys/unix" - "gvisor.dev/gvisor/pkg/abi/linux" - "gvisor.dev/gvisor/pkg/context" - "gvisor.dev/gvisor/pkg/safemem" - "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" -) - -// defaultFileFD implements FileDescriptionImpl for non-socket, non-TTY files. -type defaultFileFD struct { - fileDescription - - // canMap specifies whether we allow the file to be memory mapped. - canMap bool - - // mu protects the fields below. - mu sync.Mutex - - // offset specifies the current file offset. - offset int64 -} - -// TODO(gvisor.dev/issue/1672): Implement Waitable interface. - -// PRead implements FileDescriptionImpl. -func (f *defaultFileFD) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { - // TODO(b/34716638): Some char devices do support offsets, e.g. /dev/null. - if f.inode.isStream { - return 0, syserror.ESPIPE - } - - return readFromHostFD(ctx, f.inode.hostFD, dst, offset, int(opts.Flags)) -} - -// Read implements FileDescriptionImpl. -func (f *defaultFileFD) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { - // TODO(b/34716638): Some char devices do support offsets, e.g. /dev/null. - if f.inode.isStream { - // These files can't be memory mapped, assert this. - if f.canMap { - panic("files that can return EWOULDBLOCK (sockets, pipes, etc.) cannot be memory mapped") - } - - f.mu.Lock() - n, err := readFromHostFD(ctx, f.inode.hostFD, dst, -1, int(opts.Flags)) - f.mu.Unlock() - 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. - f.mu.Lock() - n, err := readFromHostFD(ctx, f.inode.hostFD, dst, f.offset, int(opts.Flags)) - f.offset += n - f.mu.Unlock() - return n, err -} - -func readFromHostFD(ctx context.Context, fd int, dst usermem.IOSequence, offset int64, flags int) (int64, error) { - if flags&^(linux.RWF_VALID) != 0 { - return 0, syserror.EOPNOTSUPP - } - - reader := safemem.FromVecReaderFunc{ - func(srcs [][]byte) (int64, error) { - n, err := unix.Preadv2(fd, srcs, offset, flags) - return int64(n), err - }, - } - n, err := dst.CopyOutFrom(ctx, reader) - return int64(n), err -} - -// PWrite implements FileDescriptionImpl. -func (f *defaultFileFD) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) { - // TODO(b/34716638): Some char devices do support offsets, e.g. /dev/null. - if f.inode.isStream { - return 0, syserror.ESPIPE - } - - return writeToHostFD(ctx, f.inode.hostFD, src, offset, int(opts.Flags)) -} - -// Write implements FileDescriptionImpl. -func (f *defaultFileFD) Write(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) { - // TODO(b/34716638): Some char devices do support offsets, e.g. /dev/null. - if f.inode.isStream { - // These files can't be memory mapped, assert this. - if f.canMap { - panic("files that can return EWOULDBLOCK (sockets, pipes, etc.) cannot be memory mapped") - } - - f.mu.Lock() - n, err := writeToHostFD(ctx, f.inode.hostFD, src, -1, int(opts.Flags)) - f.mu.Unlock() - 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. - f.mu.Lock() - n, err := writeToHostFD(ctx, f.inode.hostFD, src, f.offset, int(opts.Flags)) - f.offset += n - f.mu.Unlock() - return n, err -} - -func writeToHostFD(ctx context.Context, fd int, src usermem.IOSequence, offset int64, flags int) (int64, error) { - if flags&^(linux.RWF_VALID) != 0 { - return 0, syserror.EOPNOTSUPP - } - - writer := safemem.FromVecWriterFunc{ - func(srcs [][]byte) (int64, error) { - n, err := unix.Pwritev2(fd, srcs, offset, flags) - 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 *defaultFileFD) Seek(_ context.Context, offset int64, whence int32) (int64, error) { - // TODO(b/34716638): Some char devices do support seeking, e.g. /dev/null. - if f.inode.isStream { - return 0, syserror.ESPIPE - } - - f.mu.Lock() - defer f.mu.Unlock() - - switch whence { - case linux.SEEK_SET: - if offset < 0 { - return f.offset, syserror.EINVAL - } - f.offset = offset - - case linux.SEEK_CUR: - // Check for overflow. Note that underflow cannot occur, since f.offset >= 0. - if offset > math.MaxInt64-f.offset { - return f.offset, syserror.EOVERFLOW - } - if f.offset+offset < 0 { - return f.offset, syserror.EINVAL - } - f.offset += offset - - case linux.SEEK_END: - var s syscall.Stat_t - if err := syscall.Fstat(f.inode.hostFD, &s); err != nil { - return f.offset, err - } - size := s.Size - - // Check for overflow. Note that underflow cannot occur, since size >= 0. - if offset > math.MaxInt64-size { - return f.offset, syserror.EOVERFLOW - } - if size+offset < 0 { - return f.offset, syserror.EINVAL - } - f.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(f.inode.hostFD, offset, int(whence)) - if err != nil { - return f.offset, err - } - f.offset = n - - default: - // Invalid whence. - return f.offset, syserror.EINVAL - } - - return f.offset, nil -} - -// Sync implements FileDescriptionImpl. -func (f *defaultFileFD) 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 *defaultFileFD) ConfigureMMap(_ context.Context, opts *memmap.MMapOpts) error { - if !f.canMap { - return syserror.ENODEV - } - // TODO(gvisor.dev/issue/1672): Implement ConfigureMMap and Mappable interface. - return syserror.ENODEV -} diff --git a/pkg/sentry/fsimpl/host/host.go b/pkg/sentry/fsimpl/host/host.go index c205e6a0b..1e53b5c1b 100644 --- a/pkg/sentry/fsimpl/host/host.go +++ b/pkg/sentry/fsimpl/host/host.go @@ -19,6 +19,7 @@ package host import ( "errors" "fmt" + "math" "syscall" "golang.org/x/sys/unix" @@ -27,45 +28,79 @@ import ( "gvisor.dev/gvisor/pkg/log" "gvisor.dev/gvisor/pkg/refs" "gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs" + "gvisor.dev/gvisor/pkg/sentry/hostfd" "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("host.filesystemType.GetFilesystem should never be called") +} + +// Name implements FilesystemType.Name. +func (filesystemType) Name() string { + return "none" +} + // filesystem implements vfs.FilesystemImpl. type filesystem struct { kernfs.Filesystem } +// NewFilesystem sets up and returns a new hostfs filesystem. +// +// Note that there should only ever be one instance of host.filesystem, +// a global mount for host fds. +func NewFilesystem(vfsObj *vfs.VirtualFilesystem) *vfs.Filesystem { + fs := &filesystem{} + fs.Init(vfsObj, filesystemType{}) + return fs.VFSFilesystem() +} + // ImportFD sets up and returns a vfs.FileDescription from a donated fd. -func ImportFD(mnt *vfs.Mount, hostFD int, ownerUID auth.KUID, ownerGID auth.KGID, isTTY bool) (*vfs.FileDescription, error) { - // Must be importing to a mount of host.filesystem. - fs, ok := mnt.Filesystem().Impl().(*filesystem) +func ImportFD(ctx context.Context, 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 syscall.Stat_t - if err := syscall.Fstat(hostFD, &s); err != nil { + var s unix.Stat_t + if err := unix.Fstat(hostFD, &s); err != nil { return nil, err } fileMode := linux.FileMode(s.Mode) fileType := fileMode.FileType() - // Pipes, character devices, and sockets can return EWOULDBLOCK for - // operations that would block. - isStream := fileType == syscall.S_IFIFO || fileType == syscall.S_IFCHR || fileType == syscall.S_IFSOCK + + // 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, - isStream: isStream, + seekable: seekable, isTTY: isTTY, + canMap: canMap(uint32(fileType)), ino: fs.NextIno(), - mode: fileMode, - uid: ownerUID, - gid: ownerGID, + // 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{} @@ -73,7 +108,7 @@ func ImportFD(mnt *vfs.Mount, hostFD int, ownerUID auth.KUID, ownerGID auth.KGID // i.open will take a reference on d. defer d.DecRef() - return i.open(d.VFSDentry(), mnt) + return i.open(ctx, d.VFSDentry(), mnt) } // inode implements kernfs.Inode. @@ -90,32 +125,33 @@ type inode struct { // This field is initialized at creation time and is immutable. hostFD int - // isStream is true if the host fd points to a file representing a stream, + // 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. - isStream bool + 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 - // mu protects the inode metadata below. - mu sync.Mutex + // offsetMu protects offset. + offsetMu sync.Mutex - // mode is the file mode of this inode. Note that this value may become out - // of date if the mode is changed on the host, e.g. with chmod. - mode linux.FileMode - - // uid and gid of the file owner. Note that these refer to the owner of the - // file created on import, not the fd on the host. - uid auth.KUID - gid auth.KGID + // offset specifies the current file offset. + offset int64 } // Note that these flags may become out of date, since they can be modified @@ -132,55 +168,167 @@ func fileFlagsFromHostFD(fd int) (int, error) { } // CheckPermissions implements kernfs.Inode. -func (i *inode) CheckPermissions(ctx context.Context, creds *auth.Credentials, atx vfs.AccessTypes) error { - return vfs.GenericCheckPermissions(creds, atx, false /* isDir */, uint16(i.mode), i.uid, i.gid) +func (i *inode) CheckPermissions(ctx context.Context, creds *auth.Credentials, ats vfs.AccessTypes) error { + var s syscall.Stat_t + if err := syscall.Fstat(i.hostFD, &s); err != nil { + return err + } + return vfs.GenericCheckPermissions(creds, ats, linux.FileMode(s.Mode), auth.KUID(s.Uid), auth.KGID(s.Gid)) } // Mode implements kernfs.Inode. func (i *inode) Mode() linux.FileMode { - return i.mode + var s syscall.Stat_t + if err := syscall.Fstat(i.hostFD, &s); err != nil { + // Retrieving the mode from the host fd using fstat(2) should not fail. + // If the syscall does not succeed, something is fundamentally wrong. + panic(fmt.Sprintf("failed to retrieve mode from host fd %d: %v", i.hostFD, err)) + } + return linux.FileMode(s.Mode) } // 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 - if err := unix.Statx(i.hostFD, "", int(unix.AT_EMPTY_PATH|opts.Sync), int(opts.Mask), &s); err != nil { + 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 := unixToLinuxStatx(s) - // Use our own internal inode number and file owner. + 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 - ls.UID = uint32(i.uid) - ls.GID = uint32(i.gid) - - // Update file mode from the host. - i.mode = linux.FileMode(ls.Mode) return ls, nil } // SetStat implements kernfs.Inode. -func (i *inode) SetStat(_ *vfs.Filesystem, opts vfs.SetStatOptions) error { +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_UID|linux.STATX_GID) != 0 { + if m&^(linux.STATX_MODE|linux.STATX_SIZE|linux.STATX_ATIME|linux.STATX_MTIME) != 0 { return syserror.EPERM } + var hostStat syscall.Stat_t + if err := syscall.Fstat(i.hostFD, &hostStat); err != nil { + return err + } + if err := vfs.CheckSetStat(ctx, creds, &s, linux.FileMode(hostStat.Mode&linux.PermissionsMask), auth.KUID(hostStat.Uid), auth.KGID(hostStat.Gid)); err != nil { + return err + } + if m&linux.STATX_MODE != 0 { if err := syscall.Fchmod(i.hostFD, uint32(s.Mode)); err != nil { return err } - i.mode = linux.FileMode(s.Mode) } if m&linux.STATX_SIZE != 0 { if err := syscall.Ftruncate(i.hostFD, int64(s.Size)); err != nil { @@ -188,11 +336,11 @@ func (i *inode) SetStat(_ *vfs.Filesystem, opts vfs.SetStatOptions) error { } } if m&(linux.STATX_ATIME|linux.STATX_MTIME) != 0 { - timestamps := []unix.Timespec{ + ts := [2]syscall.Timespec{ toTimespec(s.Atime, m&linux.STATX_ATIME == 0), toTimespec(s.Mtime, m&linux.STATX_MTIME == 0), } - if err := unix.UtimesNanoAt(i.hostFD, "", timestamps, unix.AT_EMPTY_PATH); err != nil { + if err := setTimestamps(i.hostFD, &ts); err != nil { return err } } @@ -212,13 +360,16 @@ func (i *inode) Destroy() { } // 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(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + return i.open(ctx, vfsd, rp.Mount()) } -func (i *inode) open(d *vfs.Dentry, mnt *vfs.Mount) (*vfs.FileDescription, error) { - - fileType := i.mode.FileType() +func (i *inode) open(ctx context.Context, d *vfs.Dentry, mnt *vfs.Mount) (*vfs.FileDescription, error) { + var s syscall.Stat_t + if err := syscall.Fstat(i.hostFD, &s); err != nil { + return nil, err + } + fileType := s.Mode & linux.FileTypeMask if fileType == syscall.S_IFSOCK { if i.isTTY { return nil, errors.New("cannot use host socket as TTY") @@ -227,36 +378,42 @@ func (i *inode) open(d *vfs.Dentry, mnt *vfs.Mount) (*vfs.FileDescription, error 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 { - // TODO(gvisor.dev/issue/1672): support importing host fd as TTY. - return nil, errors.New("importing host fd as TTY not supported") - } - - // 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 := &defaultFileFD{ - fileDescription: fileDescription{ - inode: i, - }, - canMap: canMap(uint32(fileType)), - mu: sync.Mutex{}, - offset: 0, + 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 } - vfsfd := &fd.vfsfd flags, err := fileFlagsFromHostFD(i.hostFD) if err != nil { return nil, err } - if err := vfsfd.Init(fd, uint32(flags), mnt, d, &vfs.FileDescriptionOptions{}); err != nil { + 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 @@ -271,8 +428,9 @@ type fileDescription struct { } // SetStat implements vfs.FileDescriptionImpl. -func (f *fileDescription) SetStat(_ context.Context, opts vfs.SetStatOptions) error { - return f.inode.SetStat(nil, opts) +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. @@ -284,3 +442,168 @@ func (f *fileDescription) Stat(_ context.Context, opts vfs.StatOptions) (linux.S 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 + } + reader := hostfd.GetReadWriterAt(int32(hostFD), offset, flags) + n, err := dst.CopyOutFrom(ctx, reader) + hostfd.PutReadWriterAt(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 + } + writer := hostfd.GetReadWriterAt(int32(hostFD), offset, flags) + n, err := src.CopyInTo(ctx, writer) + hostfd.PutReadWriterAt(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 index e1ccacb4d..2bc757b1a 100644 --- a/pkg/sentry/fsimpl/host/util.go +++ b/pkg/sentry/fsimpl/host/util.go @@ -22,47 +22,27 @@ import ( "gvisor.dev/gvisor/pkg/syserror" ) -func toTimespec(ts linux.StatxTimestamp, omit bool) unix.Timespec { +func toTimespec(ts linux.StatxTimestamp, omit bool) syscall.Timespec { if omit { - return unix.Timespec{ + return syscall.Timespec{ Sec: 0, Nsec: unix.UTIME_OMIT, } } - return unix.Timespec{ - Sec: int64(ts.Sec), + return syscall.Timespec{ + Sec: ts.Sec, Nsec: int64(ts.Nsec), } } -func unixToLinuxStatx(s unix.Statx_t) linux.Statx { - return linux.Statx{ - Mask: s.Mask, - Blksize: s.Blksize, - Attributes: s.Attributes, - Nlink: s.Nlink, - UID: s.Uid, - GID: s.Gid, - Mode: s.Mode, - Ino: s.Ino, - Size: s.Size, - Blocks: s.Blocks, - AttributesMask: s.Attributes_mask, - Atime: unixToLinuxStatxTimestamp(s.Atime), - Btime: unixToLinuxStatxTimestamp(s.Btime), - Ctime: unixToLinuxStatxTimestamp(s.Ctime), - Mtime: unixToLinuxStatxTimestamp(s.Mtime), - RdevMajor: s.Rdev_major, - RdevMinor: s.Rdev_minor, - DevMajor: s.Dev_major, - DevMinor: s.Dev_minor, - } -} - 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 { diff --git a/pkg/sentry/fsimpl/host/util_unsafe.go b/pkg/sentry/fsimpl/host/util_unsafe.go new file mode 100644 index 000000000..5136ac844 --- /dev/null +++ b/pkg/sentry/fsimpl/host/util_unsafe.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 host + +import ( + "syscall" + "unsafe" +) + +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 1c026f4d8..c7779fc11 100644 --- a/pkg/sentry/fsimpl/kernfs/dynamic_bytes_file.go +++ b/pkg/sentry/fsimpl/kernfs/dynamic_bytes_file.go @@ -53,7 +53,7 @@ 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, opts vfs.OpenOptions) (*vfs.FileDescription, error) { +func (f *DynamicBytesFile) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { fd := &DynamicBytesFD{} if err := fd.Init(rp.Mount(), vfsd, f.data, opts.Flags); err != nil { return nil, err @@ -61,9 +61,10 @@ func (f *DynamicBytesFile) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vf 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 } diff --git a/pkg/sentry/fsimpl/kernfs/fd_impl_util.go b/pkg/sentry/fsimpl/kernfs/fd_impl_util.go index da821d524..dd5806301 100644 --- a/pkg/sentry/fsimpl/kernfs/fd_impl_util.go +++ b/pkg/sentry/fsimpl/kernfs/fd_impl_util.go @@ -15,8 +15,11 @@ package kernfs 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/memmap" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserror" @@ -42,15 +45,27 @@ type GenericDirectoryFD struct { off int64 } -// Init initializes a GenericDirectoryFD. -func (fd *GenericDirectoryFD) Init(m *vfs.Mount, d *vfs.Dentry, children *OrderedChildren, opts *vfs.OpenOptions) error { +// NewGenericDirectoryFD creates a new GenericDirectoryFD and returns its +// dentry. +func NewGenericDirectoryFD(m *vfs.Mount, d *vfs.Dentry, children *OrderedChildren, opts *vfs.OpenOptions) (*GenericDirectoryFD, error) { + fd := &GenericDirectoryFD{} + if err := fd.Init(children, opts); err != nil { + return nil, err + } + if err := fd.vfsfd.Init(fd, opts.Flags, m, d, &vfs.FileDescriptionOptions{}); err != nil { + return nil, err + } + return fd, nil +} + +// Init initializes a GenericDirectoryFD. Use it when overriding +// GenericDirectoryFD. Caller must call fd.VFSFileDescription.Init() with the +// correct implementation. +func (fd *GenericDirectoryFD) Init(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, opts.Flags, m, d, &vfs.FileDescriptionOptions{}); err != nil { - return err - } fd.children = children return nil } @@ -128,7 +143,7 @@ func (fd *GenericDirectoryFD) IterDirents(ctx context.Context, cb vfs.IterDirent // Handle "..". if fd.off == 1 { - parentInode := vfsd.ParentOrSelf().Impl().(*Dentry).inode + parentInode := genericParentOrSelf(vfsd.Impl().(*Dentry)).inode stat, err := parentInode.Stat(vfsFS, opts) if err != nil { return err @@ -186,6 +201,10 @@ func (fd *GenericDirectoryFD) Seek(ctx context.Context, offset int64, whence int // Use offset as given. case linux.SEEK_CUR: offset += fd.off + case linux.SEEK_END: + // TODO(gvisor.dev/issue/1193): This can prevent new files from showing up + // if they are added after SEEK_END. + offset = math.MaxInt64 default: return 0, syserror.EINVAL } @@ -205,7 +224,7 @@ func (fd *GenericDirectoryFD) Stat(ctx context.Context, opts vfs.StatOptions) (l // 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 1d7e04ad4..9e8d80414 100644 --- a/pkg/sentry/fsimpl/kernfs/filesystem.go +++ b/pkg/sentry/fsimpl/kernfs/filesystem.go @@ -22,6 +22,8 @@ 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" "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, rp.Mount()) 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.processDeferredDecRefs() + defer fs.mu.RUnlock() + + _, 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 } @@ -364,7 +391,7 @@ func (fs *Filesystem) OpenAt(ctx context.Context, rp *vfs.ResolvingPath, opts vf // O_NOFOLLOW have no effect here (they're handled by VFS by setting // 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 + opts.Flags &= linux.O_ACCMODE | linux.O_CREAT | linux.O_EXCL | linux.O_TRUNC | linux.O_DIRECTORY | linux.O_NOFOLLOW | linux.O_NONBLOCK ats := vfs.AccessTypesForOpenFlags(&opts) // Do not create new file. @@ -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) + return inode.Open(ctx, 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) + return inode.Open(ctx, 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) + return child.inode.Open(ctx, 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, rp.Mount()) + 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) + return child.inode.Open(ctx, 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: @@ -535,38 +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 } @@ -586,14 +635,15 @@ 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() @@ -622,7 +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. @@ -646,7 +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 } @@ -670,11 +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 } @@ -694,11 +744,12 @@ 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() mntns := vfs.MountNamespaceFromContext(ctx) @@ -714,8 +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() @@ -728,7 +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() @@ -770,5 +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 d50018b18..615592d5f 100644 --- a/pkg/sentry/fsimpl/kernfs/inode_impl_util.go +++ b/pkg/sentry/fsimpl/kernfs/inode_impl_util.go @@ -181,6 +181,11 @@ func (InodeNotSymlink) Readlink(context.Context) (string, error) { return "", syserror.EINVAL } +// Getlink implements Inode.Getlink. +func (InodeNotSymlink) Getlink(context.Context, *vfs.Mount) (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. @@ -211,6 +216,11 @@ func (a *InodeAttrs) Init(creds *auth.Credentials, ino uint64, mode linux.FileMo atomic.StoreUint32(&a.nlink, nlink) } +// Ino returns the inode id. +func (a *InodeAttrs) Ino() uint64 { + return atomic.LoadUint64(&a.ino) +} + // Mode implements Inode.Mode. func (a *InodeAttrs) Mode() linux.FileMode { return linux.FileMode(atomic.LoadUint32(&a.mode)) @@ -228,13 +238,23 @@ func (a *InodeAttrs) Stat(*vfs.Filesystem, vfs.StatOptions) (linux.Statx, error) 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, 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 +276,17 @@ 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)), ) @@ -346,8 +364,8 @@ func (o *OrderedChildren) Destroy() { // cache. Populate returns the number of directories inserted, which the caller // may use to update the link count for the parent directory. // -// Precondition: d.Impl() must be a kernfs Dentry. d must represent a directory -// inode. children must not contain any conflicting entries already in o. +// Precondition: d must represent a directory inode. children must not contain +// any conflicting entries already in o. func (o *OrderedChildren) Populate(d *Dentry, children map[string]*Dentry) uint32 { var links uint32 for name, child := range children { @@ -357,7 +375,7 @@ 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 +525,7 @@ type InodeSymlink struct { } // Open implements Inode.Open. -func (InodeSymlink) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { +func (InodeSymlink) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { return nil, syserror.ELOOP } @@ -549,8 +567,23 @@ 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, opts vfs.OpenOptions) (*vfs.FileDescription, error) { - fd := &GenericDirectoryFD{} - fd.Init(rp.Mount(), vfsd, &s.OrderedChildren, &opts) +func (s *StaticDirectory) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd, err := NewGenericDirectoryFD(rp.Mount(), vfsd, &s.OrderedChildren, &opts) + if err != nil { + return nil, err + } 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 a8ab2a2ba..732837933 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. @@ -171,17 +168,22 @@ const ( // // Must be initialized by Init prior to first use. type Dentry struct { - refs.AtomicRefCount + vfsd vfs.Dentry - vfsd vfs.Dentry - inode Inode + 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. @@ -225,8 +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. } } @@ -236,7 +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() @@ -246,13 +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 @@ -302,7 +308,7 @@ type Inode interface { // // 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) + Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) } type inodeRefs interface { @@ -319,7 +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. @@ -330,8 +336,10 @@ type inodeMetadata interface { 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 @@ -415,7 +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, mnt *vfs.Mount) (vfs.VirtualDentry, string, error) } diff --git a/pkg/sentry/fsimpl/kernfs/kernfs_test.go b/pkg/sentry/fsimpl/kernfs/kernfs_test.go index 0459fb305..a9f671bc8 100644 --- a/pkg/sentry/fsimpl/kernfs/kernfs_test.go +++ b/pkg/sentry/fsimpl/kernfs/kernfs_test.go @@ -91,7 +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 } @@ -116,9 +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, opts vfs.OpenOptions) (*vfs.FileDescription, error) { - fd := &kernfs.GenericDirectoryFD{} - if err := fd.Init(rp.Mount(), vfsd, &d.OrderedChildren, &opts); err != nil { +func (d *readonlyDir) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), vfsd, &d.OrderedChildren, &opts) + if err != nil { return nil, err } return fd.VFSFileDescription(), nil @@ -146,9 +146,11 @@ func (fs *filesystem) newDir(creds *auth.Credentials, mode linux.FileMode, conte return &dir.dentry } -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, &opts) +func (d *dir) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), vfsd, &d.OrderedChildren, &opts) + if err != nil { + return nil, err + } return fd.VFSFileDescription(), nil } @@ -187,9 +189,13 @@ 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..0aa6dc979 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.Mount) (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..d6bd67467 --- /dev/null +++ b/pkg/sentry/fsimpl/pipefs/pipefs.go @@ -0,0 +1,147 @@ +// 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(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + return i.pipe.Open(ctx, 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 a83245866..17c1342b5 100644 --- a/pkg/sentry/fsimpl/proc/BUILD +++ b/pkg/sentry/fsimpl/proc/BUILD @@ -8,10 +8,11 @@ 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"], @@ -19,8 +20,9 @@ go_library( "//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", @@ -53,6 +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 5c19d5522..104fc9030 100644 --- a/pkg/sentry/fsimpl/proc/filesystem.go +++ b/pkg/sentry/fsimpl/proc/filesystem.go @@ -36,8 +36,13 @@ type FilesystemType struct{} var _ vfs.FilesystemType = (*FilesystemType)(nil) -// GetFilesystem implements vfs.FilesystemType. -func (ft *FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { +// Name implements vfs.FilesystemType.Name. +func (FilesystemType) Name() string { + return Name +} + +// 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") @@ -48,7 +53,7 @@ func (ft *FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virtual } procfs := &kernfs.Filesystem{} - procfs.VFSFilesystem().Init(vfsObj, procfs) + procfs.VFSFilesystem().Init(vfsObj, &ft, procfs) var cgroups map[string]string if opts.InternalData != nil { diff --git a/pkg/sentry/fsimpl/proc/subtasks.go b/pkg/sentry/fsimpl/proc/subtasks.go index 611645f3f..a5cfa8333 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) @@ -91,6 +88,9 @@ func (i *subtasksInode) IterDirents(ctx context.Context, cb vfs.IterDirentsCallb if len(tasks) == 0 { return offset, syserror.ENOENT } + if relOffset >= int64(len(tasks)) { + return offset, nil + } tids := make([]int, 0, len(tasks)) for _, tid := range tasks { @@ -113,10 +113,52 @@ func (i *subtasksInode) IterDirents(ctx context.Context, cb vfs.IterDirentsCallb return offset, nil } +type subtasksFD struct { + kernfs.GenericDirectoryFD + + task *kernel.Task +} + +func (fd *subtasksFD) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback) error { + if fd.task.ExitState() >= kernel.TaskExitZombie { + return syserror.ENOENT + } + return fd.GenericDirectoryFD.IterDirents(ctx, cb) +} + +// Seek implements vfs.FileDecriptionImpl.Seek. +func (fd *subtasksFD) Seek(ctx context.Context, offset int64, whence int32) (int64, error) { + if fd.task.ExitState() >= kernel.TaskExitZombie { + return 0, syserror.ENOENT + } + return fd.GenericDirectoryFD.Seek(ctx, offset, whence) +} + +// Stat implements vfs.FileDescriptionImpl.Stat. +func (fd *subtasksFD) Stat(ctx context.Context, opts vfs.StatOptions) (linux.Statx, error) { + if fd.task.ExitState() >= kernel.TaskExitZombie { + return linux.Statx{}, syserror.ENOENT + } + return fd.GenericDirectoryFD.Stat(ctx, opts) +} + +// SetStat implements vfs.FileDescriptionImpl.SetStat. +func (fd *subtasksFD) SetStat(ctx context.Context, opts vfs.SetStatOptions) error { + if fd.task.ExitState() >= kernel.TaskExitZombie { + return syserror.ENOENT + } + return fd.GenericDirectoryFD.SetStat(ctx, opts) +} + // Open implements kernfs.Inode. -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, &opts) +func (i *subtasksInode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd := &subtasksFD{task: i.task} + if err := fd.Init(&i.OrderedChildren, &opts); err != nil { + return nil, err + } + if err := fd.VFSFileDescription().Init(fd, opts.Flags, rp.Mount(), vfsd, &vfs.FileDescriptionOptions{}); err != nil { + return nil, err + } return fd.VFSFileDescription(), nil } @@ -131,3 +173,8 @@ func (i *subtasksInode) Stat(vsfs *vfs.Filesystem, opts vfs.StatOptions) (linux. } 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 c0d643f51..66419d91b 100644 --- a/pkg/sentry/fsimpl/proc/task.go +++ b/pkg/sentry/fsimpl/proc/task.go @@ -44,19 +44,21 @@ type taskInode struct { var _ kernfs.Inode = (*taskInode)(nil) func newTaskInode(inoGen InoGenerator, task *kernel.Task, pidns *kernel.PIDNamespace, isThreadGroup bool, cgroupControllers map[string]string) *kernfs.Dentry { + // TODO(gvisor.dev/issue/164): Fail with ESRCH if task exited. 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"), @@ -100,19 +102,17 @@ func (i *taskInode) Valid(ctx context.Context) bool { } // Open implements kernfs.Inode. -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, &opts) +func (i *taskInode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), vfsd, &i.OrderedChildren, &opts) + if err != nil { + return nil, err + } 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 @@ -175,14 +175,7 @@ func (i *taskOwnedInode) Stat(fs *vfs.Filesystem, opts vfs.StatOptions) (linux.S 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) { @@ -224,22 +217,6 @@ 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: diff --git a/pkg/sentry/fsimpl/proc/task_fds.go b/pkg/sentry/fsimpl/proc/task_fds.go new file mode 100644 index 000000000..8ad976073 --- /dev/null +++ b/pkg/sentry/fsimpl/proc/task_fds.go @@ -0,0 +1,306 @@ +// 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(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), vfsd, &i.OrderedChildren, &opts) + if err != nil { + return nil, err + } + 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, mnt *vfs.Mount) (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(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), vfsd, &i.OrderedChildren, &opts) + if err != nil { + return nil, err + } + 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 5a231ac86..515f25327 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 } @@ -100,17 +111,18 @@ func (d *auxvData) Generate(ctx context.Context, buf *bytes.Buffer) error { } defer m.DecUsers(ctx) - // Space for buffer with AT_NULL (0) terminator at the end. auxv := m.Auxv() + // Space for buffer with AT_NULL (0) terminator at the end. buf.Grow((len(auxv) + 1) * 16) for _, e := range auxv { - var tmp [8]byte - usermem.ByteOrder.PutUint64(tmp[:], e.Key) - buf.Write(tmp[:]) - - usermem.ByteOrder.PutUint64(tmp[:], uint64(e.Value)) + var tmp [16]byte + usermem.ByteOrder.PutUint64(tmp[:8], e.Key) + usermem.ByteOrder.PutUint64(tmp[8:], uint64(e.Value)) buf.Write(tmp[:]) } + var atNull [16]byte + buf.Write(atNull[:]) + return nil } @@ -496,7 +508,7 @@ 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 @@ -539,11 +551,10 @@ var _ vfs.WritableDynamicBytesSource = (*oomScoreAdj)(nil) // Generate implements vfs.DynamicBytesSource.Generate. func (o *oomScoreAdj) Generate(ctx context.Context, buf *bytes.Buffer) error { - adj, err := o.task.OOMScoreAdj() - if err != nil { - return err + if o.task.ExitState() == kernel.TaskExitDead { + return syserror.ESRCH } - fmt.Fprintf(buf, "%d\n", adj) + fmt.Fprintf(buf, "%d\n", o.task.OOMScoreAdj()) return nil } @@ -562,9 +573,260 @@ func (o *oomScoreAdj) Write(ctx context.Context, src usermem.IOSequence, offset 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.Mount) (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, mnt *vfs.Mount) (vfs.VirtualDentry, string, error) { + if err := checkTaskState(s.task); err != nil { + return vfs.VirtualDentry{}, "", err + } + + // Create a synthetic inode to represent the namespace. + dentry := &kernfs.Dentry{} + dentry.Init(&namespaceInode{}) + vd := vfs.MakeVirtualDentry(mnt, dentry.VFSDentry()) + vd.IncRef() + dentry.DecRef() + return vd, "", nil +} + +// namespaceInode is a synthetic inode created to represent a namespace in +// /proc/[pid]/ns/*. +type namespaceInode struct { + kernfs.InodeAttrs + kernfs.InodeNoopRefCount + kernfs.InodeNotDirectory + kernfs.InodeNotSymlink +} + +var _ kernfs.Inode = (*namespaceInode)(nil) + +// Init initializes a namespace inode. +func (i *namespaceInode) Init(creds *auth.Credentials, ino uint64, perm linux.FileMode) { + if perm&^linux.PermissionsMask != 0 { + panic(fmt.Sprintf("Only permission mask must be set: %x", perm&linux.PermissionsMask)) + } + i.InodeAttrs.Init(creds, ino, linux.ModeRegular|perm) +} + +// Open implements Inode.Open. +func (i *namespaceInode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd := &namespaceFD{inode: i} + i.IncRef() + if err := fd.vfsfd.Init(fd, opts.Flags, rp.Mount(), vfsd, &vfs.FileDescriptionOptions{}); err != nil { + return nil, err + } + return &fd.vfsfd, nil +} + +// namespace FD is a synthetic file that represents a namespace in +// /proc/[pid]/ns/*. +type namespaceFD struct { + vfs.FileDescriptionDefaultImpl + + vfsfd vfs.FileDescription + inode *namespaceInode +} + +var _ vfs.FileDescriptionImpl = (*namespaceFD)(nil) + +// Stat implements FileDescriptionImpl. +func (fd *namespaceFD) Stat(ctx context.Context, opts vfs.StatOptions) (linux.Statx, error) { + vfs := fd.vfsfd.VirtualDentry().Mount().Filesystem() + return fd.inode.Stat(vfs, opts) +} + +// SetStat implements FileDescriptionImpl. +func (fd *namespaceFD) SetStat(ctx context.Context, opts vfs.SetStatOptions) error { + vfs := fd.vfsfd.VirtualDentry().Mount().Filesystem() + creds := auth.CredentialsFromContext(ctx) + return fd.inode.SetStat(ctx, vfs, creds, opts) +} + +// Release implements FileDescriptionImpl. +func (fd *namespaceFD) Release() { + fd.inode.DecRef() +} + +// OnClose implements FileDescriptionImpl. +func (*namespaceFD) OnClose(context.Context) error { + return nil +} diff --git a/pkg/sentry/fsimpl/proc/tasks_net.go b/pkg/sentry/fsimpl/proc/task_net.go index d4e1812d8..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" @@ -32,17 +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 - // TODO(gvisor.dev/issue/1833): Support for using the network stack in the - // network namespace of the calling process. We should make this per-process, - // a.k.a. /proc/PID/net, and make /proc/net a symlink to /proc/self/net. - if stack := k.RootNetworkNamespace().Stack(); 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" @@ -53,6 +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}), @@ -84,7 +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. @@ -203,22 +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>" } @@ -231,6 +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 @@ -249,14 +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 @@ -338,15 +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() @@ -395,14 +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. @@ -410,11 +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. @@ -496,15 +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() @@ -548,25 +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. @@ -667,9 +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 b1e39c82f..5aeda8c9b 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{k: k}), - "uptime": newDentry(root, inoGen.NextIno(), 0444, &uptimeData{}), - "version": newDentry(root, inoGen.NextIno(), 0444, &versionData{k: k}), + "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 @@ -205,9 +201,11 @@ func (i *tasksInode) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback } // Open implements kernfs.Inode. -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, &opts) +func (i *tasksInode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), vfsd, &i.OrderedChildren, &opts) + if err != nil { + return nil, err + } return fd.VFSFileDescription(), nil } @@ -229,6 +227,20 @@ func (i *tasksInode) Stat(vsfs *vfs.Filesystem, opts vfs.StatOptions) (linux.Sta 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 { features := k.FeatureSet() if features == nil { diff --git a/pkg/sentry/fsimpl/proc/tasks_files.go b/pkg/sentry/fsimpl/proc/tasks_files.go index 434998910..92007df81 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.Mount) (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.Mount) (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 3d5dc463c..f08668ca2 100644 --- a/pkg/sentry/fsimpl/proc/tasks_sys.go +++ b/pkg/sentry/fsimpl/proc/tasks_sys.go @@ -39,7 +39,7 @@ func newSysDir(root *auth.Credentials, inoGen InoGenerator, k *kernel.Kernel) *k "shmmni": newDentry(root, inoGen.NextIno(), 0444, shmData(linux.SHMMNI)), }), "vm": kernfs.NewStaticDir(root, inoGen.NextIno(), 0555, map[string]*kernfs.Dentry{ - "mmap_min_addr": newDentry(root, inoGen.NextIno(), 0444, &mmapMinAddrData{}), + "mmap_min_addr": newDentry(root, inoGen.NextIno(), 0444, &mmapMinAddrData{k: k}), "overcommit_memory": newDentry(root, inoGen.NextIno(), 0444, newStaticFile("0\n")), }), "net": newSysNetDir(root, inoGen, k), diff --git a/pkg/sentry/fsimpl/proc/tasks_test.go b/pkg/sentry/fsimpl/proc/tasks_test.go index 0eb401619..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, @@ -68,9 +70,15 @@ var ( "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, @@ -95,17 +103,37 @@ func setup(t *testing.T) *testutil.System { 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 := k.VFS().NewMountNamespace(ctx, creds, "", Name, &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, k.VFS(), mntns) } @@ -114,7 +142,7 @@ 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) } @@ -140,7 +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) @@ -180,7 +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, @@ -190,6 +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 { @@ -200,10 +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) } } @@ -301,12 +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) } @@ -343,7 +373,7 @@ func TestTask(t *testing.T) { t.Fatalf("CreateTask(): %v", err) } - collector := s.ListDirents(s.PathOpAtRoot("/1")) + collector := s.ListDirents(s.PathOpAtRoot("/proc/1")) s.AssertAllDirentTypes(collector, taskStaticFiles) } @@ -361,14 +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 { @@ -385,35 +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) @@ -431,6 +462,22 @@ 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()) @@ -438,6 +485,8 @@ func TestTree(t *testing.T) { 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) } @@ -445,11 +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..5ce50625b --- /dev/null +++ b/pkg/sentry/fsimpl/sockfs/sockfs.go @@ -0,0 +1,98 @@ +// 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" +) + +// 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) { + panic("sockfs.filesystemType.GetFilesystem should never be called") +} + +// 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 +} + +// NewFilesystem sets up and returns 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 := &filesystem{} + fs.Init(vfsObj, filesystemType{}) + return fs.VFSFilesystem() +} + +// 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(ctx context.Context, 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/sys.go b/pkg/sentry/fsimpl/sys/sys.go index c36c4fa11..f8d25d35e 100644 --- a/pkg/sentry/fsimpl/sys/sys.go +++ b/pkg/sentry/fsimpl/sys/sys.go @@ -39,10 +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) @@ -94,15 +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, opts vfs.OpenOptions) (*vfs.FileDescription, error) { - fd := &kernfs.GenericDirectoryFD{} - fd.Init(rp.Mount(), vfsd, &d.OrderedChildren, &opts) +func (d *dir) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), vfsd, &d.OrderedChildren, &opts) + if err != nil { + return nil, err + } return fd.VFSFileDescription(), nil } diff --git a/pkg/sentry/fsimpl/testutil/BUILD b/pkg/sentry/fsimpl/testutil/BUILD index e4f36f4ae..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/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 488478e29..c16a36cdb 100644 --- a/pkg/sentry/fsimpl/testutil/kernel.go +++ b/pkg/sentry/fsimpl/testutil/kernel.go @@ -23,13 +23,16 @@ 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/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" @@ -123,10 +126,17 @@ func Boot() (*kernel.Kernel, error) { // CreateTask creates a new bare bones task for tests. 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()), @@ -135,10 +145,25 @@ func CreateTask(ctx context.Context, name string, tc *kernel.ThreadGroup, mntns 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 e16808c63..0556af877 100644 --- a/pkg/sentry/fsimpl/testutil/testutil.go +++ b/pkg/sentry/fsimpl/testutil/testutil.go @@ -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 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 383133e44..2fb5c4d84 100644 --- a/pkg/sentry/fsimpl/tmpfs/benchmark_test.go +++ b/pkg/sentry/fsimpl/tmpfs/benchmark_test.go @@ -168,7 +168,7 @@ 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) @@ -362,7 +362,7 @@ 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) @@ -438,13 +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 b4380af38..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 } } @@ -68,24 +104,30 @@ 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 err := cb.Handle(vfs.Dirent{ Name: ".", Type: linux.DT_DIR, - Ino: vfsd.Impl().(*dentry).inode.ino, + Ino: dir.inode.ino, NextOff: 1, }); err != nil { return err } fd.off++ } + if fd.off == 1 { - parentInode := vfsd.ParentOrSelf().Impl().(*dentry).inode + parentInode := genericParentOrSelf(&dir.dentry).inode if err := cb.Handle(vfs.Dirent{ Name: "..", Type: parentInode.direntType(), @@ -97,7 +139,6 @@ func (fd *directoryFD) IterDirents(ctx context.Context, cb vfs.IterDirentsCallba fd.off++ } - dir := vfsd.Impl().(*dentry).inode.impl.(*directory) var child *dentry if fd.iter == nil { // Start iteration at the beginning of dir. @@ -112,7 +153,7 @@ func (fd *directoryFD) IterDirents(ctx context.Context, cb vfs.IterDirentsCallba // Skip other directoryFD iterators. if child.inode != nil { if err := cb.Handle(vfs.Dirent{ - Name: child.vfsd.Name(), + Name: child.name, Type: child.inode.direntType(), Ino: child.inode.ino, NextOff: fd.off + 1, @@ -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 e1b551422..5b62f9ebb 100644 --- a/pkg/sentry/fsimpl/tmpfs/filesystem.go +++ b/pkg/sentry/fsimpl/tmpfs/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" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserror" ) @@ -37,25 +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 @@ -63,7 +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 @@ -75,7 +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 { @@ -83,10 +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. @@ -117,33 +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() @@ -151,7 +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. @@ -166,7 +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 } } @@ -178,17 +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 } @@ -203,30 +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: @@ -238,14 +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 }) } @@ -284,12 +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. @@ -300,11 +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 { @@ -313,16 +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. @@ -335,7 +379,7 @@ afterTrailingSymlink: func (d *dentry) open(ctx context.Context, rp *vfs.ResolvingPath, opts *vfs.OpenOptions, afterCreate bool) (*vfs.FileDescription, error) { 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 } } @@ -365,9 +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)) } @@ -385,6 +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 } @@ -398,7 +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 } @@ -415,25 +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 } } @@ -443,18 +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 { @@ -466,11 +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 } @@ -478,33 +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) mntns := vfs.MountNamespaceFromContext(ctx) defer mntns.DecRef() - if err := vfsObj.PrepareRenameDentry(mntns, renamedVFSD, replacedVFSD); err != nil { + 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() @@ -545,13 +597,14 @@ func (fs *filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error vfsObj := rp.VirtualFilesystem() mntns := vfs.MountNamespaceFromContext(ctx) defer mntns.DecRef() - if err := vfsObj.PrepareDeleteDentry(mntns, childVFSD); err != nil { + 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 } @@ -563,7 +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. @@ -587,16 +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 }) } @@ -605,22 +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 } @@ -635,66 +686,79 @@ func (fs *filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error vfsObj := rp.VirtualFilesystem() mntns := vfs.MountNamespaceFromContext(ctx) defer mntns.DecRef() - if err := vfsObj.PrepareDeleteDentry(mntns, childVFSD); err != nil { + 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/regular_file.go b/pkg/sentry/fsimpl/tmpfs/regular_file.go index 711442424..57e5e28ec 100644 --- a/pkg/sentry/fsimpl/tmpfs/regular_file.go +++ b/pkg/sentry/fsimpl/tmpfs/regular_file.go @@ -89,7 +89,7 @@ func (fs *filesystem) newRegularFile(creds *auth.Credentials, mode linux.FileMod 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 } @@ -286,7 +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. @@ -308,14 +309,22 @@ 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 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 521206305..efc931468 100644 --- a/pkg/sentry/fsimpl/tmpfs/tmpfs.go +++ b/pkg/sentry/fsimpl/tmpfs/tmpfs.go @@ -12,21 +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 // 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" @@ -37,8 +41,10 @@ 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. @@ -63,6 +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) @@ -74,8 +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 } @@ -87,20 +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 { @@ -141,10 +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 ... @@ -155,10 +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 @@ -168,6 +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) @@ -175,7 +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 @@ -192,7 +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) } @@ -228,22 +292,19 @@ func (i *inode) tryIncRef() bool { func (i *inode) decRef() { if refs := atomic.AddInt64(&i.refs, -1); refs == 0 { if regFile, ok := i.impl.(*regularFile); ok { - // Hold inode.mu and regFile.dataMu while mutating - // size. - i.mu.Lock() - regFile.dataMu.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.dataMu.Unlock() - i.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 @@ -254,44 +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: @@ -299,18 +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 { @@ -338,29 +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 } @@ -419,6 +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: @@ -433,6 +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 { @@ -457,5 +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/hostfd/BUILD b/pkg/sentry/hostfd/BUILD new file mode 100644 index 000000000..364a78306 --- /dev/null +++ b/pkg/sentry/hostfd/BUILD @@ -0,0 +1,17 @@ +load("//tools:defs.bzl", "go_library") + +licenses(["notice"]) + +go_library( + name = "hostfd", + srcs = [ + "hostfd.go", + "hostfd_unsafe.go", + ], + visibility = ["//pkg/sentry:internal"], + deps = [ + "//pkg/safemem", + "//pkg/sync", + "@org_golang_x_sys//unix:go_default_library", + ], +) diff --git a/pkg/sentry/hostfd/hostfd.go b/pkg/sentry/hostfd/hostfd.go new file mode 100644 index 000000000..70dd9cafb --- /dev/null +++ b/pkg/sentry/hostfd/hostfd.go @@ -0,0 +1,84 @@ +// 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 hostfd provides efficient I/O with host file descriptors. +package hostfd + +import ( + "gvisor.dev/gvisor/pkg/safemem" + "gvisor.dev/gvisor/pkg/sync" +) + +// ReadWriterAt implements safemem.Reader and safemem.Writer by reading from +// and writing to a host file descriptor respectively. ReadWriterAts should be +// obtained by calling GetReadWriterAt. +// +// Clients should usually prefer to use Preadv2 and Pwritev2 directly. +type ReadWriterAt struct { + fd int32 + offset int64 + flags uint32 +} + +var rwpool = sync.Pool{ + New: func() interface{} { + return &ReadWriterAt{} + }, +} + +// GetReadWriterAt returns a ReadWriterAt that reads from / writes to the given +// host file descriptor, starting at the given offset and using the given +// preadv2(2)/pwritev2(2) flags. If offset is -1, the host file descriptor's +// offset is used instead. Users are responsible for ensuring that fd remains +// valid for the lifetime of the returned ReadWriterAt, and must call +// PutReadWriterAt when it is no longer needed. +func GetReadWriterAt(fd int32, offset int64, flags uint32) *ReadWriterAt { + rw := rwpool.Get().(*ReadWriterAt) + *rw = ReadWriterAt{ + fd: fd, + offset: offset, + flags: flags, + } + return rw +} + +// PutReadWriterAt releases a ReadWriterAt returned by a previous call to +// GetReadWriterAt that is no longer in use. +func PutReadWriterAt(rw *ReadWriterAt) { + rwpool.Put(rw) +} + +// ReadToBlocks implements safemem.Reader.ReadToBlocks. +func (rw *ReadWriterAt) ReadToBlocks(dsts safemem.BlockSeq) (uint64, error) { + if dsts.IsEmpty() { + return 0, nil + } + n, err := Preadv2(rw.fd, dsts, rw.offset, rw.flags) + if rw.offset >= 0 { + rw.offset += int64(n) + } + return n, err +} + +// WriteFromBlocks implements safemem.Writer.WriteFromBlocks. +func (rw *ReadWriterAt) WriteFromBlocks(srcs safemem.BlockSeq) (uint64, error) { + if srcs.IsEmpty() { + return 0, nil + } + n, err := Pwritev2(rw.fd, srcs, rw.offset, rw.flags) + if rw.offset >= 0 { + rw.offset += int64(n) + } + return n, err +} diff --git a/pkg/sentry/hostfd/hostfd_unsafe.go b/pkg/sentry/hostfd/hostfd_unsafe.go new file mode 100644 index 000000000..5e9e60fc4 --- /dev/null +++ b/pkg/sentry/hostfd/hostfd_unsafe.go @@ -0,0 +1,107 @@ +// 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 hostfd + +import ( + "io" + "syscall" + "unsafe" + + "golang.org/x/sys/unix" + "gvisor.dev/gvisor/pkg/safemem" +) + +// Preadv2 reads up to dsts.NumBytes() bytes from host file descriptor fd into +// dsts. offset and flags are interpreted as for preadv2(2). +// +// Preconditions: !dsts.IsEmpty(). +func Preadv2(fd int32, dsts safemem.BlockSeq, offset int64, flags uint32) (uint64, error) { + // No buffering is necessary regardless of safecopy; host syscalls will + // return EFAULT if appropriate, instead of raising SIGBUS. + var ( + n uintptr + e syscall.Errno + ) + // Avoid preadv2(2) if possible, since it's relatively new and thus least + // likely to be supported by the host kernel. + if flags == 0 { + if dsts.NumBlocks() == 1 { + // Use read() or pread() to avoid iovec allocation and copying. + dst := dsts.Head() + if offset == -1 { + n, _, e = syscall.Syscall(unix.SYS_READ, uintptr(fd), dst.Addr(), uintptr(dst.Len())) + } else { + n, _, e = syscall.Syscall6(unix.SYS_PREAD64, uintptr(fd), dst.Addr(), uintptr(dst.Len()), uintptr(offset), 0 /* pos_h */, 0 /* unused */) + } + } else { + iovs := safemem.IovecsFromBlockSeq(dsts) + if offset == -1 { + n, _, e = syscall.Syscall(unix.SYS_READV, uintptr(fd), uintptr((unsafe.Pointer)(&iovs[0])), uintptr(len(iovs))) + } else { + n, _, e = syscall.Syscall6(unix.SYS_PREADV, uintptr(fd), uintptr((unsafe.Pointer)(&iovs[0])), uintptr(len(iovs)), uintptr(offset), 0 /* pos_h */, 0 /* unused */) + } + } + } else { + iovs := safemem.IovecsFromBlockSeq(dsts) + n, _, e = syscall.Syscall6(unix.SYS_PREADV2, uintptr(fd), uintptr((unsafe.Pointer)(&iovs[0])), uintptr(len(iovs)), uintptr(offset), 0 /* pos_h */, uintptr(flags)) + } + if e != 0 { + return 0, e + } + if n == 0 { + return 0, io.EOF + } + return uint64(n), nil +} + +// Pwritev2 writes up to srcs.NumBytes() from srcs into host file descriptor +// fd. offset and flags are interpreted as for pwritev2(2). +// +// Preconditions: !srcs.IsEmpty(). +func Pwritev2(fd int32, srcs safemem.BlockSeq, offset int64, flags uint32) (uint64, error) { + // No buffering is necessary regardless of safecopy; host syscalls will + // return EFAULT if appropriate, instead of raising SIGBUS. + var ( + n uintptr + e syscall.Errno + ) + // Avoid pwritev2(2) if possible, since it's relatively new and thus least + // likely to be supported by the host kernel. + if flags == 0 { + if srcs.NumBlocks() == 1 { + // Use write() or pwrite() to avoid iovec allocation and copying. + src := srcs.Head() + if offset == -1 { + n, _, e = syscall.Syscall(unix.SYS_WRITE, uintptr(fd), src.Addr(), uintptr(src.Len())) + } else { + n, _, e = syscall.Syscall6(unix.SYS_PWRITE64, uintptr(fd), src.Addr(), uintptr(src.Len()), uintptr(offset), 0 /* pos_h */, 0 /* unused */) + } + } else { + iovs := safemem.IovecsFromBlockSeq(srcs) + if offset == -1 { + n, _, e = syscall.Syscall(unix.SYS_WRITEV, uintptr(fd), uintptr((unsafe.Pointer)(&iovs[0])), uintptr(len(iovs))) + } else { + n, _, e = syscall.Syscall6(unix.SYS_PWRITEV, uintptr(fd), uintptr((unsafe.Pointer)(&iovs[0])), uintptr(len(iovs)), uintptr(offset), 0 /* pos_h */, 0 /* unused */) + } + } + } else { + iovs := safemem.IovecsFromBlockSeq(srcs) + n, _, e = syscall.Syscall6(unix.SYS_PWRITEV2, uintptr(fd), uintptr((unsafe.Pointer)(&iovs[0])), uintptr(len(iovs)), uintptr(offset), 0 /* pos_h */, uintptr(flags)) + } + if e != 0 { + return 0, e + } + return uint64(n), nil +} diff --git a/pkg/sentry/kernel/BUILD b/pkg/sentry/kernel/BUILD index beba29a09..e47af66d6 100644 --- a/pkg/sentry/kernel/BUILD +++ b/pkg/sentry/kernel/BUILD @@ -169,6 +169,9 @@ go_library( "//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", diff --git a/pkg/sentry/kernel/epoll/BUILD b/pkg/sentry/kernel/epoll/BUILD index dedf0fa15..75eedd5a2 100644 --- a/pkg/sentry/kernel/epoll/BUILD +++ b/pkg/sentry/kernel/epoll/BUILD @@ -24,6 +24,7 @@ go_library( ], visibility = ["//pkg/sentry:internal"], deps = [ + "//pkg/abi/linux", "//pkg/context", "//pkg/refs", "//pkg/sentry/fs", diff --git a/pkg/sentry/kernel/epoll/epoll.go b/pkg/sentry/kernel/epoll/epoll.go index 8bffb78fc..3d78cd48f 100644 --- a/pkg/sentry/kernel/epoll/epoll.go +++ b/pkg/sentry/kernel/epoll/epoll.go @@ -20,6 +20,7 @@ import ( "fmt" "syscall" + "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/refs" "gvisor.dev/gvisor/pkg/sentry/fs" @@ -30,19 +31,6 @@ import ( "gvisor.dev/gvisor/pkg/waiter" ) -// Event describes the event mask that was observed and the user data to be -// returned when one of the events occurs. It has this format to match the linux -// format to avoid extra copying/allocation when writing events to userspace. -type Event struct { - // Events is the event mask containing the set of events that have been - // observed on an entry. - Events uint32 - - // Data is an opaque 64-bit value provided by the caller when adding the - // entry, and returned to the caller when the entry reports an event. - Data [2]int32 -} - // EntryFlags is a bitmask that holds an entry's flags. type EntryFlags int @@ -227,9 +215,9 @@ func (e *EventPoll) Readiness(mask waiter.EventMask) waiter.EventMask { } // ReadEvents returns up to max available events. -func (e *EventPoll) ReadEvents(max int) []Event { +func (e *EventPoll) ReadEvents(max int) []linux.EpollEvent { var local pollEntryList - var ret []Event + var ret []linux.EpollEvent e.listsMu.Lock() @@ -251,7 +239,7 @@ func (e *EventPoll) ReadEvents(max int) []Event { } // Add event to the array that will be returned to caller. - ret = append(ret, Event{ + ret = append(ret, linux.EpollEvent{ Events: uint32(ready), Data: entry.userData, }) @@ -296,8 +284,10 @@ 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/fd_table.go b/pkg/sentry/kernel/fd_table.go index 58001d56c..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 +307,61 @@ 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) { @@ -327,7 +393,7 @@ func (f *FDTable) NewFDVFS2(ctx context.Context, minfd int32, file *vfs.FileDesc fd = f.next } for fd < end { - if d, _, _ := f.get(fd); d == nil { + if d, _, _ := f.getVFS2(fd); d == nil { f.setVFS2(fd, file, flags) if fd == f.next { // Update next search start position. @@ -447,7 +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) { @@ -522,7 +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/kernel.go b/pkg/sentry/kernel/kernel.go index 1d627564f..c91b9dce2 100644 --- a/pkg/sentry/kernel/kernel.go +++ b/pkg/sentry/kernel/kernel.go @@ -50,6 +50,8 @@ import ( "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" @@ -248,6 +250,22 @@ type Kernel struct { // VFS keeps the filesystem state used across the kernel. vfs vfs.VirtualFilesystem + // hostMount is the Mount used for file descriptors that were imported + // from the host. + hostMount *vfs.Mount + + // 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 + + // socketMount is the Mount used for sockets created by the socket() and + // socketpair() syscalls. There are several cases where a socket dentry will + // not be contained in socketMount: + // 1. Socket files created by mknod() + // 2. Socket fds imported from the host (Kernel.hostMount is used for these) + // 3. Socket files created by binding Unix sockets to a file path + socketMount *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 @@ -348,6 +366,29 @@ 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 create sockfs mount: %v", err) + } + k.socketMount = socketMount + } + return nil } @@ -467,6 +508,11 @@ 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 { @@ -484,7 +530,7 @@ 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 @@ -533,17 +579,32 @@ 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{}{} } } @@ -1005,9 +1066,15 @@ 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() + } } }) } @@ -1035,9 +1102,15 @@ 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() + } } }) } @@ -1400,9 +1473,10 @@ 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. @@ -1425,7 +1499,30 @@ 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 @@ -1533,3 +1630,26 @@ func (k *Kernel) EmitUnimplementedEvent(ctx context.Context) { func (k *Kernel) VFS() *vfs.VirtualFilesystem { return &k.vfs } + +// SetHostMount sets the hostfs mount. +func (k *Kernel) SetHostMount(mnt *vfs.Mount) { + if k.hostMount != nil { + panic("Kernel.hostMount cannot be set more than once") + } + k.hostMount = mnt +} + +// HostMount returns the hostfs mount. +func (k *Kernel) HostMount() *vfs.Mount { + return k.hostMount +} + +// PipeMount returns the pipefs mount. +func (k *Kernel) PipeMount() *vfs.Mount { + return k.pipeMount +} + +// SocketMount returns the sockfs mount. +func (k *Kernel) SocketMount() *vfs.Mount { + return k.socketMount +} diff --git a/pkg/sentry/kernel/pipe/pipe.go b/pkg/sentry/kernel/pipe/pipe.go index 725e9db7d..62c8691f1 100644 --- a/pkg/sentry/kernel/pipe/pipe.go +++ b/pkg/sentry/kernel/pipe/pipe.go @@ -255,7 +255,8 @@ 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.view.Size(); wanted > avail { + avail := p.max - p.view.Size() + if wanted > avail { if wanted <= p.atomicIOBytes { return 0, syserror.ErrWouldBlock } @@ -268,8 +269,14 @@ func (p *Pipe) write(ctx context.Context, ops writeOps) (int64, error) { return done, err } - if wanted > done { - // Partial write due to full pipe. + if done < avail { + // Non-failure, but short write. + return done, nil + } + 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 } 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 ded95f532..18416643b 100644 --- a/pkg/sentry/kernel/rseq.go +++ b/pkg/sentry/kernel/rseq.go @@ -304,7 +304,7 @@ func (t *Task) rseqAddrInterrupt() { } var cs linux.RSeqCriticalSection - if err := cs.CopyIn(t, critAddr); err != nil { + 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)) 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 c0dbbe890..f48247c94 100644 --- a/pkg/sentry/kernel/task.go +++ b/pkg/sentry/kernel/task.go @@ -484,7 +484,7 @@ type Task struct { // bit. // // numaPolicy and numaNodeMask are protected by mu. - numaPolicy int32 + numaPolicy linux.NumaPolicy numaNodeMask uint64 // netns is the task's network namespace. netns is never nil. @@ -555,13 +555,6 @@ type Task struct { // // startTime is protected by mu. startTime ktime.Time - - // oomScoreAdj is the task's OOM score adjustment. This is currently not - // used but is maintained for consistency. - // TODO(gvisor.dev/issue/1967) - // - // oomScoreAdj is protected by mu, and is owned by the task goroutine. - oomScoreAdj int32 } func (t *Task) savePtraceTracer() *Task { @@ -784,6 +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. @@ -856,27 +858,29 @@ func (t *Task) ContainerID() string { return t.containerID } -// OOMScoreAdj gets the task's OOM score adjustment. -func (t *Task) OOMScoreAdj() (int32, error) { - t.mu.Lock() - defer t.mu.Unlock() - if t.ExitState() == TaskExitDead { - return 0, syserror.ESRCH - } - return t.oomScoreAdj, nil +// 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 OOM score adjustment. The value should be -// between -1000 and 1000 inclusive. +// 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 { - t.mu.Lock() - defer t.mu.Unlock() - if t.ExitState() == TaskExitDead { - return syserror.ESRCH - } if adj > 1000 || adj < -1000 { return syserror.EINVAL } - t.oomScoreAdj = adj + 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 dda502bb8..e1ecca99e 100644 --- a/pkg/sentry/kernel/task_clone.go +++ b/pkg/sentry/kernel/task_clone.go @@ -15,6 +15,8 @@ package kernel import ( + "sync/atomic" + "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/bpf" "gvisor.dev/gvisor/pkg/sentry/inet" @@ -260,15 +262,11 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) { 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 } - adj, err := t.OOMScoreAdj() - if err != nil { - return 0, nil, err - } - cfg := &TaskConfig{ Kernel: t.k, ThreadGroup: tg, @@ -287,7 +285,6 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) { RSeqAddr: rseqAddr, RSeqSignature: rseqSignature, ContainerID: t.ContainerID(), - OOMScoreAdj: adj, } if opts.NewThreadGroup { cfg.Parent = t diff --git a/pkg/sentry/kernel/task_context.go b/pkg/sentry/kernel/task_context.go index 0158b1788..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 } 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_run.go b/pkg/sentry/kernel/task_run.go index 799cbcd93..d654dd997 100644 --- a/pkg/sentry/kernel/task_run.go +++ b/pkg/sentry/kernel/task_run.go @@ -96,6 +96,7 @@ func (t *Task) run(threadID uintptr) { t.tg.liveGoroutines.Done() t.tg.pidns.owner.liveGoroutines.Done() t.tg.pidns.owner.runningGoroutines.Done() + t.p.Release() // Keep argument alive because stack trace for dead variables may not be correct. runtime.KeepAlive(threadID) @@ -353,7 +354,7 @@ func (app *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_sched.go b/pkg/sentry/kernel/task_sched.go index 8b148db35..09366b60c 100644 --- a/pkg/sentry/kernel/task_sched.go +++ b/pkg/sentry/kernel/task_sched.go @@ -653,14 +653,14 @@ func (t *Task) SetNiceness(n int) { } // NumaPolicy returns t's current numa policy. -func (t *Task) NumaPolicy() (policy int32, nodeMask uint64) { +func (t *Task) NumaPolicy() (policy linux.NumaPolicy, nodeMask uint64) { t.mu.Lock() defer t.mu.Unlock() return t.numaPolicy, t.numaNodeMask } // SetNumaPolicy sets t's numa policy. -func (t *Task) SetNumaPolicy(policy int32, nodeMask uint64) { +func (t *Task) SetNumaPolicy(policy linux.NumaPolicy, nodeMask uint64) { t.mu.Lock() defer t.mu.Unlock() t.numaPolicy = policy diff --git a/pkg/sentry/kernel/task_signals.go b/pkg/sentry/kernel/task_signals.go index 8802db142..7d25e98f7 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() @@ -263,6 +263,19 @@ func (t *Task) deliverSignalToHandler(info *arch.SignalInfo, act arch.SignalAct) if t.haveSavedSignalMask { mask = t.savedSignalMask } + + // Set up the restorer. + // x86-64 should always uses SA_RESTORER, but this flag is optional on other platforms. + // Please see the linux code as reference: + // linux/arch/x86/kernel/signal.c:__setup_rt_frame() + // If SA_RESTORER is not configured, we can use the sigreturn trampolines + // the vdso provides instead. + // Please see the linux code as reference: + // linux/arch/arm64/kernel/signal.c:setup_return() + if act.Flags&linux.SA_RESTORER == 0 { + act.Restorer = t.MemoryManager().VDSOSigReturn() + } + if err := t.Arch().SignalSetup(st, &act, info, &alt, mask); err != nil { return err } @@ -513,8 +526,6 @@ 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 2bbf48bb8..a5035bb7f 100644 --- a/pkg/sentry/kernel/task_start.go +++ b/pkg/sentry/kernel/task_start.go @@ -93,9 +93,6 @@ type TaskConfig struct { // ContainerID is the container the new task belongs to. ContainerID string - - // oomScoreAdj is the task's OOM score adjustment. - OOMScoreAdj int32 } // NewTask creates a new task defined by cfg. @@ -146,7 +143,6 @@ func (ts *TaskSet) newTask(cfg *TaskConfig) (*Task, error) { rseqSignature: cfg.RSeqSignature, futexWaiter: futex.NewWaiter(), containerID: cfg.ContainerID, - oomScoreAdj: cfg.OOMScoreAdj, } 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/thread_group.go b/pkg/sentry/kernel/thread_group.go index 268f62e9d..52849f5b3 100644 --- a/pkg/sentry/kernel/thread_group.go +++ b/pkg/sentry/kernel/thread_group.go @@ -254,6 +254,13 @@ 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 pidns. The 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/loader.go b/pkg/sentry/loader/loader.go index d6675b8f0..88449fe95 100644 --- a/pkg/sentry/loader/loader.go +++ b/pkg/sentry/loader/loader.go @@ -311,6 +311,15 @@ func Load(ctx context.Context, args LoadArgs, extraAuxv []arch.AuxEntry, vdso *V m.SetAuxv(auxv) m.SetExecutable(file) + symbolValue, err := getSymbolValueFromVDSO("rt_sigreturn") + if err != nil { + return 0, nil, "", syserr.NewDynamic(fmt.Sprintf("Failed to find rt_sigreturn in vdso: %v", err), syserr.FromError(err).ToLinux()) + } + + // Found rt_sigretrun. + addr := uint64(vdsoAddr) + symbolValue - vdsoPrelink + m.SetVDSOSigReturn(addr) + ac.SetIP(uintptr(loaded.entry)) ac.SetStack(uintptr(stack.Bottom)) diff --git a/pkg/sentry/loader/vdso.go b/pkg/sentry/loader/vdso.go index 161b28c2c..00977fc08 100644 --- a/pkg/sentry/loader/vdso.go +++ b/pkg/sentry/loader/vdso.go @@ -15,9 +15,11 @@ package loader import ( + "bytes" "debug/elf" "fmt" "io" + "strings" "gvisor.dev/gvisor/pkg/abi" "gvisor.dev/gvisor/pkg/context" @@ -38,6 +40,8 @@ import ( "gvisor.dev/gvisor/pkg/waiter" ) +const vdsoPrelink = 0xffffffffff700000 + type fileContext struct { context.Context } @@ -221,6 +225,27 @@ type VDSO struct { phdrs []elf.ProgHeader `state:".([]elfProgHeader)"` } +// getSymbolValueFromVDSO returns the specific symbol value in vdso.so. +func getSymbolValueFromVDSO(symbol string) (uint64, error) { + f, err := elf.NewFile(bytes.NewReader(vdsoBin)) + if err != nil { + return 0, err + } + syms, err := f.Symbols() + if err != nil { + return 0, err + } + + for _, sym := range syms { + if elf.ST_BIND(sym.Info) != elf.STB_LOCAL && sym.Section != elf.SHN_UNDEF { + if strings.Contains(sym.Name, symbol) { + return sym.Value, nil + } + } + } + return 0, fmt.Errorf("no %v in vdso.so", symbol) +} + // 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) { diff --git a/pkg/sentry/mm/address_space.go b/pkg/sentry/mm/address_space.go index 0332fc71c..5c667117c 100644 --- a/pkg/sentry/mm/address_space.go +++ b/pkg/sentry/mm/address_space.go @@ -201,8 +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 d8a5b9d29..aac56679b 100644 --- a/pkg/sentry/mm/lifecycle.go +++ b/pkg/sentry/mm/lifecycle.go @@ -84,6 +84,7 @@ func (mm *MemoryManager) Fork(ctx context.Context) (*MemoryManager, error) { dumpability: mm.dumpability, aioManager: aioManager{contexts: make(map[uint64]*AIOContext)}, sleepForActivation: mm.sleepForActivation, + vdsoSigReturnAddr: mm.vdsoSigReturnAddr, } // Copy vmas. diff --git a/pkg/sentry/mm/metadata.go b/pkg/sentry/mm/metadata.go index 6a49334f4..28e5057f7 100644 --- a/pkg/sentry/mm/metadata.go +++ b/pkg/sentry/mm/metadata.go @@ -167,3 +167,17 @@ func (mm *MemoryManager) SetExecutable(file fsbridge.File) { 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 c2195ae11..6db7c3d40 100644 --- a/pkg/sentry/mm/mm.go +++ b/pkg/sentry/mm/mm.go @@ -35,6 +35,7 @@ package mm import ( + "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/safemem" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/fsbridge" @@ -231,6 +232,9 @@ type MemoryManager struct { // 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. @@ -283,7 +287,7 @@ type vma struct { mlockMode memmap.MLockMode // numaPolicy is the NUMA policy for this vma set by mbind(). - numaPolicy int32 + numaPolicy linux.NumaPolicy // numaNodemask is the NUMA nodemask for this vma set by mbind(). numaNodemask uint64 diff --git a/pkg/sentry/mm/procfs.go b/pkg/sentry/mm/procfs.go index 1ab92f046..6efe5102b 100644 --- a/pkg/sentry/mm/procfs.go +++ b/pkg/sentry/mm/procfs.go @@ -148,7 +148,7 @@ func (mm *MemoryManager) appendVMAMapsEntryLocked(ctx context.Context, vseg vmaI // Do not include the guard page: fs/proc/task_mmu.c:show_map_vma() => // stack_guard_page_start(). - fmt.Fprintf(b, "%08x-%08x %s%s %08x %02x:%02x %d ", + lineLen, _ := fmt.Fprintf(b, "%08x-%08x %s%s %08x %02x:%02x %d ", vseg.Start(), vseg.End(), vma.realPerms, private, vma.off, devMajor, devMinor, ino) // Figure out our filename or hint. @@ -165,7 +165,7 @@ func (mm *MemoryManager) appendVMAMapsEntryLocked(ctx context.Context, vseg vmaI } if s != "" { // Per linux, we pad until the 74th character. - if pad := 73 - b.Len(); pad > 0 { + if pad := 73 - lineLen; pad > 0 { b.WriteString(strings.Repeat(" ", pad)) } b.WriteString(s) diff --git a/pkg/sentry/mm/syscalls.go b/pkg/sentry/mm/syscalls.go index c5dfa5972..3f496aa9f 100644 --- a/pkg/sentry/mm/syscalls.go +++ b/pkg/sentry/mm/syscalls.go @@ -974,7 +974,7 @@ func (mm *MemoryManager) MLockAll(ctx context.Context, opts MLockAllOpts) error } // NumaPolicy implements the semantics of Linux's get_mempolicy(MPOL_F_ADDR). -func (mm *MemoryManager) NumaPolicy(addr usermem.Addr) (int32, uint64, error) { +func (mm *MemoryManager) NumaPolicy(addr usermem.Addr) (linux.NumaPolicy, uint64, error) { mm.mappingMu.RLock() defer mm.mappingMu.RUnlock() vseg := mm.vmas.FindSegment(addr) @@ -986,7 +986,7 @@ func (mm *MemoryManager) NumaPolicy(addr usermem.Addr) (int32, uint64, error) { } // SetNumaPolicy implements the semantics of Linux's mbind(). -func (mm *MemoryManager) SetNumaPolicy(addr usermem.Addr, length uint64, policy int32, nodemask uint64) error { +func (mm *MemoryManager) SetNumaPolicy(addr usermem.Addr, length uint64, policy linux.NumaPolicy, nodemask uint64) error { if !addr.IsPageAligned() { return syserror.EINVAL } diff --git a/pkg/sentry/platform/kvm/context.go b/pkg/sentry/platform/kvm/context.go index c769ac7b4..6507121ea 100644 --- a/pkg/sentry/platform/kvm/context.go +++ b/pkg/sentry/platform/kvm/context.go @@ -85,3 +85,6 @@ func (c *context) Switch(as platform.AddressSpace, ac arch.Context, _ int32) (*a func (c *context) Interrupt() { c.interrupt.NotifyInterrupt() } + +// Release implements platform.Context.Release(). +func (c *context) Release() {} diff --git a/pkg/sentry/platform/kvm/kvm.go b/pkg/sentry/platform/kvm/kvm.go index a9b4af43e..ae813e24e 100644 --- a/pkg/sentry/platform/kvm/kvm.go +++ b/pkg/sentry/platform/kvm/kvm.go @@ -191,6 +191,11 @@ func (*constructor) OpenDevice() (*os.File, error) { return OpenDevice() } +// Flags implements platform.Constructor.Flags(). +func (*constructor) Requirements() platform.Requirements { + return platform.Requirements{} +} + func init() { platform.Register("kvm", &constructor{}) } diff --git a/pkg/sentry/platform/kvm/kvm_arm64.go b/pkg/sentry/platform/kvm/kvm_arm64.go index 79045651e..716198712 100644 --- a/pkg/sentry/platform/kvm/kvm_arm64.go +++ b/pkg/sentry/platform/kvm/kvm_arm64.go @@ -18,6 +18,8 @@ package kvm import ( "syscall" + + "gvisor.dev/gvisor/pkg/sentry/platform/ring0" ) type kvmOneReg struct { @@ -46,6 +48,6 @@ type userRegs struct { func updateGlobalOnce(fd int) error { physicalInit() err := updateSystemValues(int(fd)) - updateVectorTable() + ring0.Init() return err } diff --git a/pkg/sentry/platform/kvm/machine_arm64_unsafe.go b/pkg/sentry/platform/kvm/machine_arm64_unsafe.go index 00801dee6..3c02cef7c 100644 --- a/pkg/sentry/platform/kvm/machine_arm64_unsafe.go +++ b/pkg/sentry/platform/kvm/machine_arm64_unsafe.go @@ -48,69 +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/platform.go b/pkg/sentry/platform/platform.go index 2ca696382..171513f3f 100644 --- a/pkg/sentry/platform/platform.go +++ b/pkg/sentry/platform/platform.go @@ -148,6 +148,9 @@ type Context interface { // Interrupt interrupts a concurrent call to Switch(), causing it to return // ErrContextInterrupt. Interrupt() + + // Release() releases any resources associated with this context. + Release() } var ( @@ -353,10 +356,28 @@ func (fr FileRange) String() string { return fmt.Sprintf("[%#x, %#x)", fr.Start, fr.End) } +// Requirements is used to specify platform specific requirements. +type Requirements struct { + // RequiresCurrentPIDNS indicates that the sandbox has to be started in the + // current pid namespace. + RequiresCurrentPIDNS bool + // RequiresCapSysPtrace indicates that the sandbox has to be started with + // the CAP_SYS_PTRACE capability. + RequiresCapSysPtrace bool +} + // Constructor represents a platform type. type Constructor interface { + // New returns a new platform instance. + // + // Arguments: + // + // * deviceFile - the device file (e.g. /dev/kvm for the KVM platform). New(deviceFile *os.File) (Platform, error) OpenDevice() (*os.File, error) + + // Requirements returns platform specific requirements. + Requirements() Requirements } // platforms contains all available platform types. 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.go b/pkg/sentry/platform/ptrace/ptrace.go index 03adb624b..08d055e05 100644 --- a/pkg/sentry/platform/ptrace/ptrace.go +++ b/pkg/sentry/platform/ptrace/ptrace.go @@ -177,6 +177,9 @@ func (c *context) Interrupt() { c.interrupt.NotifyInterrupt() } +// Release implements platform.Context.Release(). +func (c *context) Release() {} + // PTrace represents a collection of ptrace subprocesses. type PTrace struct { platform.MMapMinAddr @@ -248,6 +251,16 @@ func (*constructor) OpenDevice() (*os.File, error) { return nil, nil } +// Flags implements platform.Constructor.Flags(). +func (*constructor) Requirements() platform.Requirements { + // TODO(b/75837838): Also set a new PID namespace so that we limit + // access to other host processes. + return platform.Requirements{ + RequiresCapSysPtrace: true, + RequiresCurrentPIDNS: true, + } +} + func init() { platform.Register("ptrace", &constructor{}) } 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..773ddb1ed 100644 --- a/pkg/sentry/platform/ptrace/subprocess.go +++ b/pkg/sentry/platform/ptrace/subprocess.go @@ -332,7 +332,7 @@ func (t *thread) unexpectedStubExit() { msg, err := t.getEventMessage() status := syscall.WaitStatus(msg) if status.Signaled() && status.Signal() == syscall.SIGKILL { - // SIGKILL can be only sent by an user or OOM-killer. In both + // SIGKILL can be only sent by a user or OOM-killer. In both // these cases, we don't need to panic. There is no reasons to // think that something wrong in gVisor. log.Warningf("The ptrace stub process %v has been killed by SIGKILL.", t.tgid) @@ -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/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/entry_arm64.s b/pkg/sentry/platform/ring0/entry_arm64.s index d42eda37b..db6465663 100644 --- a/pkg/sentry/platform/ring0/entry_arm64.s +++ b/pkg/sentry/platform/ring0/entry_arm64.s @@ -394,6 +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) @@ -464,7 +466,23 @@ 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. diff --git a/pkg/sentry/platform/ring0/lib_arm64.go b/pkg/sentry/platform/ring0/lib_arm64.go index 80922f43d..444a83913 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 581841555..16d5f478b 100644 --- a/pkg/sentry/platform/ring0/pagetables/BUILD +++ b/pkg/sentry/platform/ring0/pagetables/BUILD @@ -81,6 +81,9 @@ go_library( "pagetables_arm64.go", "pagetables_x86.go", "pcids.go", + "pcids_aarch64.go", + "pcids_aarch64.s", + "pcids_x86.go", "walker_amd64.go", "walker_arm64.go", "walker_empty.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 index 9206030bf..964496aac 100644 --- a/pkg/sentry/platform/ring0/pagetables/pcids.go +++ b/pkg/sentry/platform/ring0/pagetables/pcids.go @@ -18,9 +18,6 @@ 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 @@ -44,7 +41,7 @@ type PCIDs struct { // // Nil is returned iff the start and size are out of range. func NewPCIDs(start, size uint16) *PCIDs { - if start+uint16(size) >= limitPCID { + if start+uint16(size) > limitPCID { return nil // See comment. } p := &PCIDs{ 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 new file mode 100644 index 000000000..91fc5e8dd --- /dev/null +++ b/pkg/sentry/platform/ring0/pagetables/pcids_x86.go @@ -0,0 +1,20 @@ +// 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 i386 amd64 + +package pagetables + +// 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 4d42d29cb..ca16d0381 100644 --- a/pkg/sentry/socket/control/BUILD +++ b/pkg/sentry/socket/control/BUILD @@ -4,7 +4,10 @@ package(licenses = ["notice"]) go_library( name = "control", - srcs = ["control.go"], + srcs = [ + "control.go", + "control_vfs2.go", + ], imports = [ "gvisor.dev/gvisor/pkg/sentry/fs", ], @@ -18,6 +21,7 @@ go_library( "//pkg/sentry/kernel/auth", "//pkg/sentry/socket", "//pkg/sentry/socket/unix/transport", + "//pkg/sentry/vfs", "//pkg/syserror", "//pkg/tcpip", "//pkg/usermem", diff --git a/pkg/sentry/socket/control/control.go b/pkg/sentry/socket/control/control.go index 8834a1e1a..8b439a078 100644 --- a/pkg/sentry/socket/control/control.go +++ b/pkg/sentry/socket/control/control.go @@ -41,6 +41,8 @@ type SCMCredentials interface { Credentials(t *kernel.Task) (kernel.ThreadID, auth.UID, auth.GID) } +// LINT.IfChange + // SCMRights represents a SCM_RIGHTS socket control message. type SCMRights interface { transport.RightsControlMessage @@ -142,6 +144,8 @@ func PackRights(t *kernel.Task, rights SCMRights, cloexec bool, buf []byte, flag return putCmsg(buf, flags, linux.SCM_RIGHTS, align, fds) } +// LINT.ThenChange(./control_vfs2.go) + // scmCredentials represents an SCM_CREDENTIALS socket control message. // // +stateify savable @@ -537,11 +541,19 @@ func Parse(t *kernel.Task, socketOrEndpoint interface{}, buf []byte) (socket.Con } if len(fds) > 0 { - rights, err := NewSCMRights(t, fds) - if err != nil { - return socket.ControlMessages{}, err + if kernel.VFS2Enabled { + rights, err := NewSCMRightsVFS2(t, fds) + if err != nil { + return socket.ControlMessages{}, err + } + cmsgs.Unix.Rights = rights + } else { + rights, err := NewSCMRights(t, fds) + if err != nil { + return socket.ControlMessages{}, err + } + cmsgs.Unix.Rights = rights } - cmsgs.Unix.Rights = rights } return cmsgs, nil @@ -566,6 +578,8 @@ func MakeCreds(t *kernel.Task) SCMCredentials { return &scmCredentials{t, tcred.EffectiveKUID, tcred.EffectiveKGID} } +// LINT.IfChange + // New creates default control messages if needed. func New(t *kernel.Task, socketOrEndpoint interface{}, rights SCMRights) transport.ControlMessages { return transport.ControlMessages{ @@ -573,3 +587,5 @@ func New(t *kernel.Task, socketOrEndpoint interface{}, rights SCMRights) transpo Rights: rights, } } + +// LINT.ThenChange(./control_vfs2.go) diff --git a/pkg/sentry/socket/control/control_vfs2.go b/pkg/sentry/socket/control/control_vfs2.go new file mode 100644 index 000000000..fd08179be --- /dev/null +++ b/pkg/sentry/socket/control/control_vfs2.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 control + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" +) + +// SCMRightsVFS2 represents a SCM_RIGHTS socket control message. +type SCMRightsVFS2 interface { + transport.RightsControlMessage + + // Files returns up to max RightsFiles. + // + // Returned files are consumed and ownership is transferred to the caller. + // Subsequent calls to Files will return the next files. + Files(ctx context.Context, max int) (rf RightsFilesVFS2, truncated bool) +} + +// RightsFiles represents a SCM_RIGHTS socket control message. A reference is +// maintained for each vfs.FileDescription and is release either when an FD is created or +// when the Release method is called. +type RightsFilesVFS2 []*vfs.FileDescription + +// NewSCMRightsVFS2 creates a new SCM_RIGHTS socket control message +// representation using local sentry FDs. +func NewSCMRightsVFS2(t *kernel.Task, fds []int32) (SCMRightsVFS2, error) { + files := make(RightsFilesVFS2, 0, len(fds)) + for _, fd := range fds { + file := t.GetFileVFS2(fd) + if file == nil { + files.Release() + return nil, syserror.EBADF + } + files = append(files, file) + } + return &files, nil +} + +// Files implements SCMRights.Files. +func (fs *RightsFilesVFS2) Files(ctx context.Context, max int) (RightsFilesVFS2, bool) { + n := max + var trunc bool + if l := len(*fs); n > l { + n = l + } else if n < l { + trunc = true + } + rf := (*fs)[:n] + *fs = (*fs)[n:] + return rf, trunc +} + +// Clone implements transport.RightsControlMessage.Clone. +func (fs *RightsFilesVFS2) Clone() transport.RightsControlMessage { + nfs := append(RightsFilesVFS2(nil), *fs...) + for _, nf := range nfs { + nf.IncRef() + } + return &nfs +} + +// Release implements transport.RightsControlMessage.Release. +func (fs *RightsFilesVFS2) Release() { + for _, f := range *fs { + f.DecRef() + } + *fs = nil +} + +// rightsFDsVFS2 gets up to the specified maximum number of FDs. +func rightsFDsVFS2(t *kernel.Task, rights SCMRightsVFS2, cloexec bool, max int) ([]int32, bool) { + files, trunc := rights.Files(t, max) + fds := make([]int32, 0, len(files)) + for i := 0; i < max && len(files) > 0; i++ { + fd, err := t.NewFDFromVFS2(0, files[0], kernel.FDFlags{ + CloseOnExec: cloexec, + }) + files[0].DecRef() + files = files[1:] + if err != nil { + t.Warningf("Error inserting FD: %v", err) + // This is what Linux does. + break + } + + fds = append(fds, int32(fd)) + } + return fds, trunc +} + +// PackRightsVFS2 packs as many FDs as will fit into the unused capacity of buf. +func PackRightsVFS2(t *kernel.Task, rights SCMRightsVFS2, cloexec bool, buf []byte, flags int) ([]byte, int) { + maxFDs := (cap(buf) - len(buf) - linux.SizeOfControlMessageHeader) / 4 + // Linux does not return any FDs if none fit. + if maxFDs <= 0 { + flags |= linux.MSG_CTRUNC + return buf, flags + } + fds, trunc := rightsFDsVFS2(t, rights, cloexec, maxFDs) + if trunc { + flags |= linux.MSG_CTRUNC + } + align := t.Arch().Width() + return putCmsg(buf, flags, linux.SCM_RIGHTS, align, fds) +} + +// NewVFS2 creates default control messages if needed. +func NewVFS2(t *kernel.Task, socketOrEndpoint interface{}, rights SCMRightsVFS2) transport.ControlMessages { + return transport.ControlMessages{ + Credentials: makeCreds(t, socketOrEndpoint), + Rights: rights, + } +} diff --git a/pkg/sentry/socket/netfilter/BUILD b/pkg/sentry/socket/netfilter/BUILD index 7cd2ce55b..721094bbf 100644 --- a/pkg/sentry/socket/netfilter/BUILD +++ b/pkg/sentry/socket/netfilter/BUILD @@ -7,6 +7,7 @@ go_library( srcs = [ "extensions.go", "netfilter.go", + "owner_matcher.go", "targets.go", "tcp_matcher.go", "udp_matcher.go", @@ -22,7 +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 b4b244abf..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)) @@ -86,7 +86,7 @@ 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()) diff --git a/pkg/sentry/socket/netfilter/netfilter.go b/pkg/sentry/socket/netfilter/netfilter.go index 2ec11f6ac..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 @@ -123,19 +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. @@ -148,14 +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 @@ -228,18 +233,20 @@ func convertNetstackToBinary(tablename string, table iptables.Table) (linux.Kern return entries, meta, nil } -func marshalTarget(target iptables.Target) []byte { +func marshalTarget(target stack.Target) []byte { switch tg := target.(type) { - case iptables.AcceptTarget: - return marshalStandardTarget(iptables.RuleAccept) - case iptables.DropTarget: - return marshalStandardTarget(iptables.RuleDrop) - case iptables.ErrorTarget: + case stack.AcceptTarget: + return marshalStandardTarget(stack.RuleAccept) + case stack.DropTarget: + return marshalStandardTarget(stack.RuleDrop) + case stack.ErrorTarget: return marshalErrorTarget(errorTargetName) - case iptables.UserChainTarget: + case stack.UserChainTarget: return marshalErrorTarget(tg.Name) - case iptables.ReturnTarget: - return marshalStandardTarget(iptables.RuleReturn) + case stack.ReturnTarget: + return marshalStandardTarget(stack.RuleReturn) + case stack.RedirectTarget: + return marshalRedirectTarget() case JumpTarget: return marshalJumpTarget(tg) default: @@ -247,7 +254,7 @@ func marshalTarget(target iptables.Target) []byte { } } -func marshalStandardTarget(verdict iptables.RuleVerdict) []byte { +func marshalStandardTarget(verdict stack.RuleVerdict) []byte { nflog("convert to binary: marshalling standard target") // The target's name will be the empty string. @@ -276,6 +283,19 @@ func marshalErrorTarget(errorName string) []byte { 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") @@ -295,13 +315,13 @@ func marshalJumpTarget(jt JumpTarget) []byte { // translateFromStandardVerdict translates verdicts the same way as the iptables // tool. -func translateFromStandardVerdict(verdict iptables.RuleVerdict) int32 { +func translateFromStandardVerdict(verdict stack.RuleVerdict) int32 { switch verdict { - case iptables.RuleAccept: + case stack.RuleAccept: return -linux.NF_ACCEPT - 1 - case iptables.RuleDrop: + case stack.RuleDrop: return -linux.NF_DROP - 1 - case iptables.RuleReturn: + case stack.RuleReturn: return linux.NF_RETURN default: // TODO(gvisor.dev/issue/170): Support Jump. @@ -310,18 +330,18 @@ func translateFromStandardVerdict(verdict iptables.RuleVerdict) int32 { } // translateToStandardTarget translates from the value in a -// linux.XTStandardTarget to an iptables.Verdict. -func translateToStandardTarget(val int32) (iptables.Target, error) { +// 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.AcceptTarget{}, nil + return stack.AcceptTarget{}, nil case -linux.NF_DROP - 1: - return iptables.DropTarget{}, nil + return stack.DropTarget{}, nil case -linux.NF_QUEUE - 1: return nil, errors.New("unsupported iptables verdict QUEUE") case linux.NF_RETURN: - return iptables.ReturnTarget{}, nil + return stack.ReturnTarget{}, nil default: return nil, fmt.Errorf("unknown iptables verdict %d", val) } @@ -329,7 +349,7 @@ func translateToStandardTarget(val int32) (iptables.Target, error) { // 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)) @@ -341,10 +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 @@ -404,14 +426,14 @@ 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, @@ -442,11 +464,11 @@ func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { 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 } @@ -455,7 +477,7 @@ func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { // Add the user chains. for ruleIdx, rule := range table.Rules { - target, ok := rule.Target.(iptables.UserChainTarget) + target, ok := rule.Target.(stack.UserChainTarget) if !ok { continue } @@ -495,11 +517,11 @@ func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { } // 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.AcceptTarget); !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 } @@ -511,7 +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, @@ -519,16 +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)) @@ -570,7 +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)) @@ -614,67 +636,125 @@ func parseTarget(optVal []byte) (iptables.Target, error) { switch name := errorTarget.Name.String(); name { case errorTargetName: nflog("set entries: error target") - return iptables.ErrorTarget{}, nil + return stack.ErrorTarget{}, nil default: // User defined chain. nflog("set entries: user-defined target %q", name) - return iptables.UserChainTarget{Name: name}, nil + 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 iptables.Rule) bool { +func validUnderflow(rule stack.Rule) bool { if len(rule.Matchers) != 0 { return false } switch rule.Target.(type) { - case iptables.AcceptTarget, iptables.DropTarget: + 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 index c421b87cf..c948de876 100644 --- a/pkg/sentry/socket/netfilter/targets.go +++ b/pkg/sentry/socket/netfilter/targets.go @@ -15,11 +15,10 @@ package netfilter import ( - "gvisor.dev/gvisor/pkg/tcpip" - "gvisor.dev/gvisor/pkg/tcpip/iptables" + "gvisor.dev/gvisor/pkg/tcpip/stack" ) -// JumpTarget implements iptables.Target. +// 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. @@ -29,7 +28,7 @@ type JumpTarget struct { RuleNum int } -// Action implements iptables.Target.Action. -func (jt JumpTarget) Action(tcpip.PacketBuffer) (iptables.RuleVerdict, int) { - return iptables.RuleJump, jt.RuleNum +// 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..55c0f04f3 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 { @@ -115,19 +114,20 @@ 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 { tcpHeader = header.TCP(pkt.TransportHeader) } else { // The TCP header hasn't been parsed yet. We have to do it here. - if len(pkt.Data.First()) < header.TCPMinimumSize { + hdr, ok := pkt.Data.PullUp(header.TCPMinimumSize) + if !ok { // There's no valid TCP header here, so we hotdrop the // packet. return false, true } - tcpHeader = header.TCP(pkt.Data.First()) + tcpHeader = header.TCP(hdr) } // Check whether the source and destination ports are within the diff --git a/pkg/sentry/socket/netfilter/udp_matcher.go b/pkg/sentry/socket/netfilter/udp_matcher.go index 86aa11696..04d03d494 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 } @@ -114,19 +113,20 @@ 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 { udpHeader = header.UDP(pkt.TransportHeader) } else { // The UDP header hasn't been parsed yet. We have to do it here. - if len(pkt.Data.First()) < header.UDPMinimumSize { + hdr, ok := pkt.Data.PullUp(header.UDPMinimumSize) + if !ok { // There's no valid UDP header here, so we hotdrop the // packet. return false, true } - udpHeader = header.UDP(pkt.Data.First()) + udpHeader = header.UDP(hdr) } // Check whether the source and destination ports are within the 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 13a9a60b4..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" @@ -932,7 +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) @@ -957,8 +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: @@ -990,12 +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 { @@ -1034,24 +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 @@ -1081,24 +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 { @@ -1148,47 +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) } @@ -1320,11 +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) @@ -1334,8 +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) } @@ -1357,12 +1363,7 @@ 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 default: emitUnimplementedEventIPv6(t, name) @@ -1378,8 +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) } @@ -1395,8 +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) } @@ -1421,23 +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 { @@ -1454,11 +1451,7 @@ 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 o, nil + return boolToInt32(v), nil case linux.IP_PKTINFO: if outLen < sizeOfInt32 { @@ -1469,11 +1462,7 @@ 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 o, nil + return boolToInt32(v), nil default: emitUnimplementedEventIP(t, name) @@ -1533,7 +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) @@ -1560,7 +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 { @@ -1584,7 +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 { @@ -1592,7 +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) @@ -1620,7 +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 { @@ -1628,7 +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 { @@ -1636,7 +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 { @@ -1708,11 +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 { @@ -1720,7 +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 { @@ -1728,7 +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 { @@ -1736,7 +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 { @@ -1847,7 +1832,7 @@ 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) @@ -1932,7 +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 */) @@ -1979,9 +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. @@ -2000,7 +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 { @@ -2010,7 +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) @@ -2334,6 +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) @@ -2380,9 +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 } @@ -2456,6 +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 diff --git a/pkg/sentry/socket/netstack/provider.go b/pkg/sentry/socket/netstack/provider.go index 5f181f017..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) { @@ -126,6 +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) @@ -135,10 +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/strace.go b/pkg/sentry/strace/strace.go index 77655558e..68ca537c8 100644 --- a/pkg/sentry/strace/strace.go +++ b/pkg/sentry/strace/strace.go @@ -719,7 +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) @@ -778,9 +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/syscalls/epoll.go b/pkg/sentry/syscalls/epoll.go index 87dcad18b..d9fb808c0 100644 --- a/pkg/sentry/syscalls/epoll.go +++ b/pkg/sentry/syscalls/epoll.go @@ -17,6 +17,7 @@ package syscalls import ( "time" + "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/kernel/epoll" ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" @@ -118,7 +119,7 @@ func RemoveEpoll(t *kernel.Task, epfd int32, fd int32) error { } // WaitEpoll implements the epoll_wait(2) linux syscall. -func WaitEpoll(t *kernel.Task, fd int32, max int, timeout int) ([]epoll.Event, error) { +func WaitEpoll(t *kernel.Task, fd int32, max int, timeout int) ([]linux.EpollEvent, error) { // Get epoll from the file descriptor. epollfile := t.GetFile(fd) if epollfile == nil { 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 3ab93fbde..51bf205cf 100644 --- a/pkg/sentry/syscalls/linux/sys_epoll.go +++ b/pkg/sentry/syscalls/linux/sys_epoll.go @@ -21,7 +21,6 @@ import ( "gvisor.dev/gvisor/pkg/sentry/kernel/epoll" "gvisor.dev/gvisor/pkg/sentry/syscalls" "gvisor.dev/gvisor/pkg/syserror" - "gvisor.dev/gvisor/pkg/usermem" "gvisor.dev/gvisor/pkg/waiter" ) @@ -72,7 +71,7 @@ func EpollCtl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysc var data [2]int32 if op != linux.EPOLL_CTL_DEL { var e linux.EpollEvent - if _, err := t.CopyIn(eventAddr, &e); err != nil { + if _, err := e.CopyIn(t, eventAddr); err != nil { return 0, nil, err } @@ -105,28 +104,6 @@ func EpollCtl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysc } } -// copyOutEvents copies epoll events from the kernel to user memory. -func copyOutEvents(t *kernel.Task, addr usermem.Addr, e []epoll.Event) error { - const itemLen = 12 - buffLen := len(e) * itemLen - if _, ok := addr.AddLength(uint64(buffLen)); !ok { - return syserror.EFAULT - } - - b := t.CopyScratchBuffer(buffLen) - for i := range e { - usermem.ByteOrder.PutUint32(b[i*itemLen:], e[i].Events) - usermem.ByteOrder.PutUint32(b[i*itemLen+4:], uint32(e[i].Data[0])) - usermem.ByteOrder.PutUint32(b[i*itemLen+8:], uint32(e[i].Data[1])) - } - - if _, err := t.CopyOutBytes(addr, b); err != nil { - return err - } - - return nil -} - // EpollWait implements the epoll_wait(2) linux syscall. func EpollWait(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { epfd := args[0].Int() @@ -140,7 +117,7 @@ func EpollWait(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sys } if len(r) != 0 { - if err := copyOutEvents(t, eventsAddr, r); err != nil { + if _, err := linux.CopyEpollEventSliceOut(t, eventsAddr, r); err != nil { return 0, nil, err } } diff --git a/pkg/sentry/syscalls/linux/sys_file.go b/pkg/sentry/syscalls/linux/sys_file.go index d10a9bed8..35a98212a 100644 --- a/pkg/sentry/syscalls/linux/sys_file.go +++ b/pkg/sentry/syscalls/linux/sys_file.go @@ -514,7 +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 @@ -529,7 +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. // @@ -564,17 +564,23 @@ 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) diff --git a/pkg/sentry/syscalls/linux/sys_mempolicy.go b/pkg/sentry/syscalls/linux/sys_mempolicy.go index ac934dc6f..9b4a5c3f1 100644 --- a/pkg/sentry/syscalls/linux/sys_mempolicy.go +++ b/pkg/sentry/syscalls/linux/sys_mempolicy.go @@ -162,10 +162,10 @@ func GetMempolicy(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel. if err != nil { return 0, nil, err } - policy = 0 // maxNodes == 1 + policy = linux.MPOL_DEFAULT // maxNodes == 1 } if mode != 0 { - if _, err := t.CopyOut(mode, policy); err != nil { + if _, err := policy.CopyOut(t, mode); err != nil { return 0, nil, err } } @@ -199,10 +199,10 @@ func GetMempolicy(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel. if policy&^linux.MPOL_MODE_FLAGS != linux.MPOL_INTERLEAVE { return 0, nil, syserror.EINVAL } - policy = 0 // maxNodes == 1 + policy = linux.MPOL_DEFAULT // maxNodes == 1 } if mode != 0 { - if _, err := t.CopyOut(mode, policy); err != nil { + if _, err := policy.CopyOut(t, mode); err != nil { return 0, nil, err } } @@ -216,7 +216,7 @@ func GetMempolicy(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel. // SetMempolicy implements the syscall set_mempolicy(2). func SetMempolicy(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { - modeWithFlags := args[0].Int() + modeWithFlags := linux.NumaPolicy(args[0].Int()) nodemask := args[1].Pointer() maxnode := args[2].Uint() @@ -233,7 +233,7 @@ func SetMempolicy(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel. func Mbind(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { addr := args[0].Pointer() length := args[1].Uint64() - mode := args[2].Int() + mode := linux.NumaPolicy(args[2].Int()) nodemask := args[3].Pointer() maxnode := args[4].Uint() flags := args[5].Uint() @@ -258,9 +258,9 @@ func Mbind(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall return 0, nil, err } -func copyInMempolicyNodemask(t *kernel.Task, modeWithFlags int32, nodemask usermem.Addr, maxnode uint32) (int32, uint64, error) { - flags := modeWithFlags & linux.MPOL_MODE_FLAGS - mode := modeWithFlags &^ linux.MPOL_MODE_FLAGS +func copyInMempolicyNodemask(t *kernel.Task, modeWithFlags linux.NumaPolicy, nodemask usermem.Addr, maxnode uint32) (linux.NumaPolicy, uint64, error) { + flags := linux.NumaPolicy(modeWithFlags & linux.MPOL_MODE_FLAGS) + mode := linux.NumaPolicy(modeWithFlags &^ linux.MPOL_MODE_FLAGS) if flags == linux.MPOL_MODE_FLAGS { // Can't specify both mode flags simultaneously. return 0, 0, syserror.EINVAL 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 9c6728530..f92bf8096 100644 --- a/pkg/sentry/syscalls/linux/sys_prctl.go +++ b/pkg/sentry/syscalls/linux/sys_prctl.go @@ -161,8 +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 78a2cb750..071b4bacc 100644 --- a/pkg/sentry/syscalls/linux/sys_read.go +++ b/pkg/sentry/syscalls/linux/sys_read.go @@ -96,8 +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 } @@ -120,8 +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 } 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 9bd2df104..46ebf27a2 100644 --- a/pkg/sentry/syscalls/linux/sys_stat.go +++ b/pkg/sentry/syscalls/linux/sys_stat.go @@ -115,7 +115,8 @@ func stat(t *kernel.Task, d *fs.Dirent, dirPath bool, statAddr usermem.Addr) err return err } s := statFromAttrs(t, d.Inode.StableAttr, uattr) - return s.CopyOut(t, statAddr) + _, err = s.CopyOut(t, statAddr) + return err } // fstat implements fstat for the given *fs.File. @@ -125,7 +126,8 @@ func fstat(t *kernel.Task, f *fs.File, statAddr usermem.Addr) error { return err } s := statFromAttrs(t, f.Dirent.Inode.StableAttr, uattr) - return s.CopyOut(t, statAddr) + _, err = s.CopyOut(t, statAddr) + return err } // Statx implements linux syscall statx(2). @@ -136,7 +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 { diff --git a/pkg/sentry/syscalls/linux/sys_write.go b/pkg/sentry/syscalls/linux/sys_write.go index 506ee54ce..6ec0de96e 100644 --- a/pkg/sentry/syscalls/linux/sys_write.go +++ b/pkg/sentry/syscalls/linux/sys_write.go @@ -87,8 +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 } diff --git a/pkg/sentry/syscalls/linux/vfs2/BUILD b/pkg/sentry/syscalls/linux/vfs2/BUILD index e7695e995..f6fb0f219 100644 --- a/pkg/sentry/syscalls/linux/vfs2/BUILD +++ b/pkg/sentry/syscalls/linux/vfs2/BUILD @@ -6,7 +6,6 @@ go_library( name = "vfs2", srcs = [ "epoll.go", - "epoll_unsafe.go", "execve.go", "fd.go", "filesystem.go", @@ -18,33 +17,44 @@ go_library( "linux64_override_arm64.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 index d6cb0e79a..34c90ae3e 100644 --- a/pkg/sentry/syscalls/linux/vfs2/epoll.go +++ b/pkg/sentry/syscalls/linux/vfs2/epoll.go @@ -28,6 +28,8 @@ import ( "gvisor.dev/gvisor/pkg/waiter" ) +var sizeofEpollEvent = (*linux.EpollEvent)(nil).SizeBytes() + // EpollCreate1 implements Linux syscall epoll_create1(2). func EpollCreate1(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { flags := args[0].Int() @@ -101,14 +103,14 @@ func EpollCtl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysc var event linux.EpollEvent switch op { case linux.EPOLL_CTL_ADD: - if err := event.CopyIn(t, eventAddr); err != nil { + 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 { + if _, err := event.CopyIn(t, eventAddr); err != nil { return 0, nil, err } return 0, nil, ep.ModifyInterest(file, fd, event) @@ -124,7 +126,7 @@ func EpollWait(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sys maxEvents := int(args[2].Int()) timeout := int(args[3].Int()) - const _EP_MAX_EVENTS = math.MaxInt32 / sizeofEpollEvent // Linux: fs/eventpoll.c:EP_MAX_EVENTS + var _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 } @@ -157,7 +159,8 @@ func EpollWait(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sys maxEvents -= n if n != 0 { // Copy what we read out. - copiedEvents, err := copyOutEvents(t, eventsAddr, events[:n]) + copiedBytes, err := linux.CopyEpollEventSliceOut(t, eventsAddr, events[:n]) + copiedEvents := copiedBytes / sizeofEpollEvent // rounded down eventsAddr += usermem.Addr(copiedEvents * sizeofEpollEvent) total += copiedEvents if err != nil { diff --git a/pkg/sentry/syscalls/linux/vfs2/epoll_unsafe.go b/pkg/sentry/syscalls/linux/vfs2/epoll_unsafe.go deleted file mode 100644 index 825f325bf..000000000 --- a/pkg/sentry/syscalls/linux/vfs2/epoll_unsafe.go +++ /dev/null @@ -1,44 +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 ( - "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/fd.go b/pkg/sentry/syscalls/linux/vfs2/fd.go index 3afcea665..8181d80f4 100644 --- a/pkg/sentry/syscalls/linux/vfs2/fd.go +++ b/pkg/sentry/syscalls/linux/vfs2/fd.go @@ -18,6 +18,7 @@ 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" ) @@ -140,6 +141,22 @@ func Fcntl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall 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 index fc5ceea4c..46d3e189c 100644 --- a/pkg/sentry/syscalls/linux/vfs2/filesystem.go +++ b/pkg/sentry/syscalls/linux/vfs2/filesystem.go @@ -172,7 +172,7 @@ func openat(t *kernel.Task, dirfd int32, pathAddr usermem.Addr, flags uint32, mo defer tpop.Release() file, err := t.Kernel().VFS().OpenAt(t, t.Credentials(), &tpop.pop, &vfs.OpenOptions{ - Flags: flags, + 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 { @@ -250,7 +250,7 @@ func rmdirat(t *kernel.Task, dirfd int32, pathAddr usermem.Addr) error { if err != nil { return err } - tpop, err := getTaskPathOperation(t, dirfd, path, disallowEmptyPath, followFinalSymlink) + tpop, err := getTaskPathOperation(t, dirfd, path, disallowEmptyPath, nofollowFinalSymlink) if err != nil { return err } diff --git a/pkg/sentry/syscalls/linux/vfs2/getdents.go b/pkg/sentry/syscalls/linux/vfs2/getdents.go index ddc140b65..c7c7bf7ce 100644 --- a/pkg/sentry/syscalls/linux/vfs2/getdents.go +++ b/pkg/sentry/syscalls/linux/vfs2/getdents.go @@ -97,7 +97,8 @@ func (cb *getdentsCallback) Handle(dirent vfs.Dirent) error { // char d_name[]; /* Filename (null-terminated) */ // }; size := 8 + 8 + 2 + 1 + 1 + len(dirent.Name) - if size < cb.remaining { + size = (size + 7) &^ 7 // round up to multiple of 8 + if size > cb.remaining { return syserror.EINVAL } buf = cb.t.CopyScratchBuffer(size) @@ -106,7 +107,12 @@ func (cb *getdentsCallback) Handle(dirent vfs.Dirent) error { usermem.ByteOrder.PutUint16(buf[16:18], uint16(size)) buf[18] = dirent.Type copy(buf[19:], dirent.Name) - buf[size-1] = 0 // NUL terminator + // 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 */ @@ -124,8 +130,9 @@ func (cb *getdentsCallback) Handle(dirent vfs.Dirent) error { 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) - if size < cb.remaining { + size := 8 + 8 + 2 + 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) @@ -133,8 +140,13 @@ func (cb *getdentsCallback) Handle(dirent vfs.Dirent) error { usermem.ByteOrder.PutUint64(buf[8:16], uint64(dirent.NextOff)) usermem.ByteOrder.PutUint16(buf[16:18], uint16(size)) copy(buf[18:], dirent.Name) - buf[size-3] = 0 // NUL terminator - buf[size-2] = 0 // zero padding byte + // 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) : size-1] + for i := range bufTail { + bufTail[i] = 0 + } buf[size-1] = dirent.Type } n, err := cb.t.CopyOutBytes(cb.addr, buf) diff --git a/pkg/sentry/syscalls/linux/vfs2/linux64_override_amd64.go b/pkg/sentry/syscalls/linux/vfs2/linux64_override_amd64.go index 7d220bc20..74920f785 100644 --- a/pkg/sentry/syscalls/linux/vfs2/linux64_override_amd64.go +++ b/pkg/sentry/syscalls/linux/vfs2/linux64_override_amd64.go @@ -39,26 +39,27 @@ func Override(table map[uintptr]kernel.Syscall) { table[19] = syscalls.Supported("readv", Readv) table[20] = syscalls.Supported("writev", Writev) table[21] = syscalls.Supported("access", Access) - delete(table, 22) // pipe + 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 - delete(table, 41) // socket - delete(table, 42) // connect - delete(table, 43) // accept - delete(table, 44) // sendto - delete(table, 45) // recvfrom - delete(table, 46) // sendmsg - delete(table, 47) // recvmsg - delete(table, 48) // shutdown - delete(table, 49) // bind - delete(table, 50) // listen - delete(table, 51) // getsockname - delete(table, 52) // getpeername - delete(table, 53) // socketpair - delete(table, 54) // setsockopt - delete(table, 55) // getsockopt + // 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("setsockopt", SetSockOpt, "In process of porting socket syscalls to VFS2.", nil) + table[55] = syscalls.PartiallySupported("getsockopt", GetSockOpt, "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 @@ -139,23 +140,26 @@ func Override(table map[uintptr]kernel.Syscall) { table[280] = syscalls.Supported("utimensat", Utimensat) table[281] = syscalls.Supported("epoll_pwait", EpollPwait) delete(table, 282) // signalfd - delete(table, 283) // timerfd_create + table[283] = syscalls.Supported("timerfd_create", TimerfdCreate) delete(table, 284) // eventfd delete(table, 285) // fallocate - delete(table, 286) // timerfd_settime - delete(table, 287) // timerfd_gettime - delete(table, 288) // accept4 + 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) - delete(table, 293) // pipe2 + table[293] = syscalls.Supported("pipe2", Pipe2) delete(table, 294) // inotify_init1 table[295] = syscalls.Supported("preadv", Preadv) table[296] = syscalls.Supported("pwritev", Pwritev) - delete(table, 299) // recvmmsg + // 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) - delete(table, 307) // sendmmsg + // 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) 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 index dbf4882da..ff1b25d7b 100644 --- a/pkg/sentry/syscalls/linux/vfs2/poll.go +++ b/pkg/sentry/syscalls/linux/vfs2/poll.go @@ -374,7 +374,8 @@ func copyOutTimespecRemaining(t *kernel.Task, startNs ktime.Time, timeout time.D } remaining := timeoutRemaining(t, startNs, timeout) tsRemaining := linux.NsecToTimespec(remaining.Nanoseconds()) - return tsRemaining.CopyOut(t, timespecAddr) + _, err := tsRemaining.CopyOut(t, timespecAddr) + return err } // copyOutTimevalRemaining copies the time remaining in timeout to timevalAddr. @@ -386,7 +387,8 @@ func copyOutTimevalRemaining(t *kernel.Task, startNs ktime.Time, timeout time.Du } remaining := timeoutRemaining(t, startNs, timeout) tvRemaining := linux.NsecToTimeval(remaining.Nanoseconds()) - return tvRemaining.CopyOut(t, timevalAddr) + _, err := tvRemaining.CopyOut(t, timevalAddr) + return err } // pollRestartBlock encapsulates the state required to restart poll(2) via @@ -477,7 +479,7 @@ func Select(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal timeout := time.Duration(-1) if timevalAddr != 0 { var timeval linux.Timeval - if err := timeval.CopyIn(t, timevalAddr); err != nil { + if _, err := timeval.CopyIn(t, timevalAddr); err != nil { return 0, nil, err } if timeval.Sec < 0 || timeval.Usec < 0 { @@ -519,7 +521,7 @@ func Pselect(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysca panic(fmt.Sprintf("unsupported sizeof(void*): %d", t.Arch().Width())) } var maskStruct sigSetWithSize - if err := maskStruct.CopyIn(t, maskWithSizeAddr); err != nil { + 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 { @@ -554,7 +556,7 @@ func copyTimespecInToDuration(t *kernel.Task, timespecAddr usermem.Addr) (time.D timeout := time.Duration(-1) if timespecAddr != 0 { var timespec linux.Timespec - if err := timespec.CopyIn(t, timespecAddr); err != nil { + if _, err := timespec.CopyIn(t, timespecAddr); err != nil { return 0, err } if !timespec.Valid() { @@ -573,7 +575,7 @@ func setTempSignalSet(t *kernel.Task, maskAddr usermem.Addr, maskSize uint) erro return syserror.EINVAL } var mask linux.SignalSet - if err := mask.CopyIn(t, maskAddr); err != nil { + if _, err := mask.CopyIn(t, maskAddr); err != nil { return err } mask &^= kernel.UnblockableSignals diff --git a/pkg/sentry/syscalls/linux/vfs2/read_write.go b/pkg/sentry/syscalls/linux/vfs2/read_write.go index 35f6308d6..6c6998f45 100644 --- a/pkg/sentry/syscalls/linux/vfs2/read_write.go +++ b/pkg/sentry/syscalls/linux/vfs2/read_write.go @@ -103,7 +103,7 @@ func read(t *kernel.Task, file *vfs.FileDescription, dst usermem.IOSequence, opt // Issue the request and break out if it completes with anything other than // "would block". - n, err := file.Read(t, dst, opts) + n, err = file.Read(t, dst, opts) total += n if err != syserror.ErrWouldBlock { break @@ -130,8 +130,8 @@ 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 } @@ -248,7 +248,7 @@ func pread(t *kernel.Task, file *vfs.FileDescription, dst usermem.IOSequence, of // Issue the request and break out if it completes with anything other than // "would block". - n, err := file.PRead(t, dst, offset+total, opts) + n, err = file.PRead(t, dst, offset+total, opts) total += n if err != syserror.ErrWouldBlock { break @@ -335,7 +335,7 @@ func write(t *kernel.Task, file *vfs.FileDescription, src usermem.IOSequence, op // Issue the request and break out if it completes with anything other than // "would block". - n, err := file.Write(t, src, opts) + n, err = file.Write(t, src, opts) total += n if err != syserror.ErrWouldBlock { break @@ -362,8 +362,8 @@ 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 } @@ -480,7 +480,7 @@ func pwrite(t *kernel.Task, file *vfs.FileDescription, src usermem.IOSequence, o // Issue the request and break out if it completes with anything other than // "would block". - n, err := file.PWrite(t, src, offset+total, opts) + n, err = file.PWrite(t, src, offset+total, opts) total += n if err != syserror.ErrWouldBlock { break diff --git a/pkg/sentry/syscalls/linux/vfs2/setstat.go b/pkg/sentry/syscalls/linux/vfs2/setstat.go index 9250659ff..4e61f1452 100644 --- a/pkg/sentry/syscalls/linux/vfs2/setstat.go +++ b/pkg/sentry/syscalls/linux/vfs2/setstat.go @@ -173,12 +173,13 @@ func Truncate(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysc return 0, nil, err } - return 0, nil, setstatat(t, linux.AT_FDCWD, path, disallowEmptyPath, followFinalSymlink, &vfs.SetStatOptions{ + 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). @@ -196,12 +197,13 @@ func Ftruncate(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sys } defer file.DecRef() - return 0, nil, file.SetStat(t, vfs.SetStatOptions{ + 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). @@ -224,7 +226,7 @@ func Utime(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall opts.Stat.Mtime.Nsec = linux.UTIME_NOW } else { var times linux.Utime - if err := times.CopyIn(t, timesAddr); err != nil { + if _, err := times.CopyIn(t, timesAddr); err != nil { return 0, nil, err } opts.Stat.Atime.Sec = times.Actime @@ -378,3 +380,12 @@ func setstatat(t *kernel.Task, dirfd int32, path fspath.Path, shouldAllowEmptyPa 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..10b668477 --- /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.PackRightsVFS2(t, cms.Unix.Rights.(control.SCMRightsVFS2), 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 index 12c532310..bb1d5cac4 100644 --- a/pkg/sentry/syscalls/linux/vfs2/stat.go +++ b/pkg/sentry/syscalls/linux/vfs2/stat.go @@ -16,6 +16,7 @@ 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" @@ -90,7 +91,8 @@ func fstatat(t *kernel.Task, dirfd int32, pathAddr, statAddr usermem.Addr, flags } var stat linux.Stat convertStatxToUserStat(t, &statx, &stat) - return stat.CopyOut(t, statAddr) + _, err = stat.CopyOut(t, statAddr) + return err } start = dirfile.VirtualDentry() start.IncRef() @@ -110,7 +112,8 @@ func fstatat(t *kernel.Task, dirfd int32, pathAddr, statAddr usermem.Addr, flags } var stat linux.Stat convertStatxToUserStat(t, &statx, &stat) - return stat.CopyOut(t, statAddr) + _, err = stat.CopyOut(t, statAddr) + return err } func timespecFromStatxTimestamp(sxts linux.StatxTimestamp) linux.Timespec { @@ -139,7 +142,8 @@ func Fstat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall } var stat linux.Stat convertStatxToUserStat(t, &statx, &stat) - return 0, nil, stat.CopyOut(t, statAddr) + _, err = stat.CopyOut(t, statAddr) + return 0, nil, err } // Statx implements Linux syscall statx(2). @@ -150,7 +154,15 @@ func Statx(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall mask := args[3].Uint() statxAddr := args[4].Pointer() - if flags&^(linux.AT_EMPTY_PATH|linux.AT_SYMLINK_NOFOLLOW) != 0 { + 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 } @@ -190,7 +202,8 @@ func Statx(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall return 0, nil, err } userifyStatx(t, &statx) - return 0, nil, statx.CopyOut(t, statxAddr) + _, err = statx.CopyOut(t, statxAddr) + return 0, nil, err } start = dirfile.VirtualDentry() start.IncRef() @@ -209,7 +222,8 @@ func Statx(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall return 0, nil, err } userifyStatx(t, &statx) - return 0, nil, statx.CopyOut(t, statxAddr) + _, err = statx.CopyOut(t, statxAddr) + return 0, nil, err } func userifyStatx(t *kernel.Task, statx *linux.Statx) { @@ -228,14 +242,65 @@ func Readlink(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysc // Access implements Linux syscall access(2). func Access(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { - // FIXME(jamieliu): actually implement - return 0, nil, nil + addr := args[0].Pointer() + mode := args[1].ModeT() + + return 0, nil, accessAt(t, linux.AT_FDCWD, addr, mode) } -// Faccessat implements Linux syscall access(2). +// 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) { - // FIXME(jamieliu): actually implement - return 0, nil, nil + 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). @@ -299,8 +364,8 @@ func Statfs(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal if err != nil { return 0, nil, err } - - return 0, nil, statfs.CopyOut(t, bufAddr) + _, err = statfs.CopyOut(t, bufAddr) + return 0, nil, err } // Fstatfs implements Linux syscall fstatfs(2). @@ -318,6 +383,6 @@ func Fstatfs(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysca if err != nil { return 0, nil, err } - - return 0, nil, statfs.CopyOut(t, bufAddr) + _, err = statfs.CopyOut(t, bufAddr) + return 0, nil, 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 index 89e9ff4d7..af455d5c1 100644 --- a/pkg/sentry/syscalls/linux/vfs2/xattr.go +++ b/pkg/sentry/syscalls/linux/vfs2/xattr.go @@ -51,7 +51,7 @@ func listxattr(t *kernel.Task, args arch.SyscallArguments, shouldFollowFinalSyml } defer tpop.Release() - names, err := t.Kernel().VFS().ListxattrAt(t, t.Credentials(), &tpop.pop) + names, err := t.Kernel().VFS().ListxattrAt(t, t.Credentials(), &tpop.pop, uint64(size)) if err != nil { return 0, nil, err } @@ -74,7 +74,7 @@ func Flistxattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sy } defer file.DecRef() - names, err := file.Listxattr(t) + names, err := file.Listxattr(t, uint64(size)) if err != nil { return 0, nil, err } @@ -116,7 +116,10 @@ func getxattr(t *kernel.Task, args arch.SyscallArguments, shouldFollowFinalSymli return 0, nil, err } - value, err := t.Kernel().VFS().GetxattrAt(t, t.Credentials(), &tpop.pop, name) + 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 } @@ -145,7 +148,7 @@ func Fgetxattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sys return 0, nil, err } - value, err := file.Getxattr(t, name) + value, err := file.Getxattr(t, &vfs.GetxattrOptions{Name: name, Size: uint64(size)}) if err != nil { return 0, nil, err } @@ -230,7 +233,7 @@ func Fsetxattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sys return 0, nil, err } - return 0, nil, file.Setxattr(t, vfs.SetxattrOptions{ + return 0, nil, file.Setxattr(t, &vfs.SetxattrOptions{ Name: name, Value: value, Flags: uint32(flags), diff --git a/pkg/sentry/vfs/BUILD b/pkg/sentry/vfs/BUILD index cb4deb068..9aeb83fb0 100644 --- a/pkg/sentry/vfs/BUILD +++ b/pkg/sentry/vfs/BUILD @@ -36,6 +36,7 @@ go_library( "pathname.go", "permissions.go", "resolving_path.go", + "timerfd.go", "vfs.go", ], visibility = ["//pkg/sentry:internal"], @@ -51,7 +52,10 @@ go_library( "//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", 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/dentry.go b/pkg/sentry/vfs/dentry.go index 35b208721..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 @@ -67,6 +47,10 @@ 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 @@ -74,36 +58,23 @@ import ( // // +stateify savable type Dentry struct { - // parent is this Dentry's parent in this Filesystem. If this Dentry is - // independent, parent is nil. - parent *Dentry - - // name is this Dentry's name in parent. - name string + // mu synchronizes deletion/invalidation and mounting over this Dentry. + mu sync.Mutex `state:"nosave"` - 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 `state:"nosave"` - // 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 @@ -134,20 +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() @@ -164,104 +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() @@ -280,42 +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 @@ -324,25 +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() @@ -376,24 +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) } } } @@ -401,25 +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/epoll.go b/pkg/sentry/vfs/epoll.go index 3da45d744..8e0b40841 100644 --- a/pkg/sentry/vfs/epoll.go +++ b/pkg/sentry/vfs/epoll.go @@ -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 diff --git a/pkg/sentry/vfs/file_description.go b/pkg/sentry/vfs/file_description.go index 9a1ad630c..418d69b96 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 @@ -100,10 +111,10 @@ type FileDescriptionOptions struct { } // Init must be called before first use of fd. If it succeeds, it takes -// references on mnt and d. statusFlags is the initial file description status -// flags, which is usually the full set of flags passed to open(2). -func (fd *FileDescription) Init(impl FileDescriptionImpl, statusFlags uint32, mnt *Mount, d *Dentry, opts *FileDescriptionOptions) error { - writable := MayWriteFileWithOpenFlags(statusFlags) +// references on mnt and d. flags is the initial file description flags, which +// is usually the full set of flags passed to open(2). +func (fd *FileDescription) Init(impl FileDescriptionImpl, flags uint32, mnt *Mount, d *Dentry, opts *FileDescriptionOptions) error { + writable := MayWriteFileWithOpenFlags(flags) if writable { if err := mnt.CheckBeginWrite(); err != nil { return err @@ -111,14 +122,18 @@ func (fd *FileDescription) Init(impl FileDescriptionImpl, statusFlags uint32, mn } fd.refs = 1 - fd.statusFlags = statusFlags | linux.O_LARGEFILE + + // Remove "file creation flags" to mirror the behavior from file.f_flags in + // fs/open.c:do_dentry_open + fd.statusFlags = flags &^ (linux.O_CREAT | linux.O_EXCL | linux.O_NOCTTY | linux.O_TRUNC) fd.vd = VirtualDentry{ mount: mnt, dentry: d, } - fd.vd.IncRef() + mnt.IncRef() + d.IncRef() fd.opts = *opts - fd.readable = MayReadFileWithOpenFlags(statusFlags) + fd.readable = MayReadFileWithOpenFlags(flags) fd.writable = writable fd.impl = impl return nil @@ -175,6 +190,12 @@ 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 +307,8 @@ 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 +327,7 @@ 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 +359,7 @@ 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 +405,11 @@ 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. @@ -514,6 +538,9 @@ 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 +559,9 @@ 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 +570,9 @@ 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 +609,23 @@ 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 +638,39 @@ 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 45191d1c3..f4c111926 100644 --- a/pkg/sentry/vfs/file_description_impl_util.go +++ b/pkg/sentry/vfs/file_description_impl_util.go @@ -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/filesystem.go b/pkg/sentry/vfs/filesystem.go index 556976d0b..20e5bb072 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 @@ -40,21 +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 @@ -144,6 +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. // @@ -332,7 +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(). @@ -362,7 +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: // @@ -424,9 +443,14 @@ type FilesystemImpl interface { // Errors: // // - 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) + // ListxattrAt returns ENOTSUP. + // + // - 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. @@ -435,7 +459,15 @@ 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. @@ -444,6 +476,10 @@ 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. @@ -452,8 +488,15 @@ 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 @@ -476,7 +519,7 @@ 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 bb9cada81..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. 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 31a4e5480..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 @@ -41,13 +48,16 @@ import ( // // +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. @@ -74,6 +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. @@ -81,6 +95,22 @@ 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 @@ -129,13 +159,7 @@ 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 } @@ -148,12 +172,7 @@ func (vfs *VirtualFilesystem) NewDisconnectedMount(fs *Filesystem, root *Dentry, if root != nil { root.IncRef() } - return &Mount{ - vfs: vfs, - fs: fs, - root: root, - refs: 1, - }, nil + return newMount(vfs, fs, root, nil /* mntns */, opts), nil } // MountAt creates and mounts a Filesystem configured by the given arguments. @@ -169,6 +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{}) @@ -180,7 +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() @@ -214,17 +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() @@ -261,9 +275,9 @@ func (vfs *VirtualFilesystem) UmountAt(ctx context.Context, creds *auth.Credenti } } - // 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 { @@ -630,12 +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 { @@ -646,3 +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/options.go b/pkg/sentry/vfs/options.go index 6af7fdac1..022bac127 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 @@ -32,6 +33,25 @@ type GetDentryOptions struct { type MkdirOptions struct { // Mode is the file mode bits for the created directory. Mode linux.FileMode + + // If ForSyntheticMountpoint is true, FilesystemImpl.MkdirAt() may create + // the given directory in memory only (as opposed to persistent storage). + // The created directory should be able to support the creation of + // subdirectories with ForSyntheticMountpoint == true. It does not need to + // support the creation of subdirectories with ForSyntheticMountpoint == + // false, or files of other types. + // + // FilesystemImpls are permitted to ignore the ForSyntheticMountpoint + // option. + // + // The ForSyntheticMountpoint option exists because, unlike mount(2), the + // OCI Runtime Specification permits the specification of mount points that + // do not exist, under the expectation that container runtimes will create + // them. (More accurately, the OCI Runtime Specification completely fails + // to document this feature, but it's implemented by runc.) + // ForSyntheticMountpoint allows such mount points to be created even when + // the underlying persistent filesystem is immutable. + ForSyntheticMountpoint bool } // MknodOptions contains options to VirtualFilesystem.MknodAt() and @@ -44,10 +64,27 @@ 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 @@ -75,7 +112,8 @@ type OpenOptions struct { // FileExec is set when the file is being opened to be executed. // VirtualFilesystem.OpenAt() checks that the caller has execute permissions - // on the file, and that the file is a regular file. + // on the file, that the file is a regular file, and that the mount doesn't + // have MS_NOEXEC set. FileExec bool } @@ -113,6 +151,20 @@ 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 8e250998a..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 } @@ -147,7 +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 @@ -177,11 +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 } } @@ -205,3 +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 eb4ebb511..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 { @@ -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 bde81e1ef..9015f2cc1 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" ) @@ -133,7 +134,7 @@ func (vfs *VirtualFilesystem) Init() error { anonfs := anonFilesystem{ devMinor: anonfsDevMinor, } - anonfs.vfsfs.Init(vfs, &anonfs) + anonfs.vfsfs.Init(vfs, &anonFilesystemType{}, &anonfs) defer anonfs.vfsfs.DecRef() anonMount, err := vfs.NewDisconnectedMount(&anonfs.vfsfs, nil, &MountOptions{}) if err != nil { @@ -174,6 +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) { @@ -213,7 +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) { @@ -254,7 +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) { @@ -290,7 +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) { @@ -317,13 +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) { @@ -333,19 +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 @@ -385,9 +427,12 @@ func (vfs *VirtualFilesystem) OpenAt(ctx context.Context, creds *auth.Credential if err == nil { vfs.putResolvingPath(rp) - // TODO(gvisor.dev/issue/1193): Move inside fsimpl to avoid another call - // to FileDescription.Stat(). 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 { @@ -474,7 +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) { @@ -507,7 +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) { @@ -588,7 +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) { @@ -620,7 +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) { @@ -632,10 +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 @@ -657,10 +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 @@ -752,6 +797,14 @@ type VirtualDentry struct { dentry *Dentry } +// MakeVirtualDentry creates a VirtualDentry. +func MakeVirtualDentry(mount *Mount, dentry *Dentry) VirtualDentry { + return VirtualDentry{ + mount: mount, + dentry: dentry, + } +} + // Ok returns true if vd is not empty. It does not require that a reference is // held. func (vd VirtualDentry) Ok() bool { diff --git a/pkg/sentry/watchdog/watchdog.go b/pkg/sentry/watchdog/watchdog.go index f7d6009a0..fcc46420f 100644 --- a/pkg/sentry/watchdog/watchdog.go +++ b/pkg/sentry/watchdog/watchdog.go @@ -319,8 +319,8 @@ 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() { @@ -329,16 +329,15 @@ func (w *Watchdog) reportStuckWatchdog() { 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() 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/syserror/syserror.go b/pkg/syserror/syserror.go index 4b5a0fca6..f86db0999 100644 --- a/pkg/syserror/syserror.go +++ b/pkg/syserror/syserror.go @@ -27,6 +27,7 @@ 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) 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/buffer/view.go b/pkg/tcpip/buffer/view.go index 17e94c562..f01217c91 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}) @@ -65,7 +77,8 @@ func NewVectorisedView(size int, views []View) VectorisedView { return VectorisedView{views: views, size: size} } -// TrimFront removes the first "count" bytes of the vectorised view. +// TrimFront removes the first "count" bytes of the vectorised view. It panics +// if count > vv.Size(). func (vv *VectorisedView) TrimFront(count int) { for count > 0 && len(vv.views) > 0 { if count < len(vv.views[0]) { @@ -74,8 +87,49 @@ func (vv *VectorisedView) TrimFront(count int) { return } count -= len(vv.views[0]) - vv.RemoveFirst() + vv.removeFirst() + } +} + +// 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. @@ -105,29 +159,45 @@ 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 { +// PullUp returns the first "count" bytes of the vectorised view. If those +// bytes aren't already contiguous inside the vectorised view, PullUp will +// reallocate as needed to make them contiguous. PullUp fails and returns false +// when count > vv.Size(). +func (vv *VectorisedView) PullUp(count int) (View, bool) { if len(vv.views) == 0 { - return nil + return nil, count == 0 + } + if count <= len(vv.views[0]) { + return vv.views[0][:count], true + } + if count > vv.size { + return nil, false } - return vv.views[0] -} -// RemoveFirst removes the first view of the vectorised view. -func (vv *VectorisedView) RemoveFirst() { - if len(vv.views) == 0 { - return + newFirst := NewView(count) + i := 0 + for offset := 0; offset < count; i++ { + copy(newFirst[offset:], vv.views[i]) + if count-offset < len(vv.views[i]) { + vv.views[i].TrimFront(count - offset) + break + } + offset += len(vv.views[i]) + vv.views[i] = nil } - vv.size -= len(vv.views[0]) - vv.views = vv.views[1:] + // We're guaranteed that i > 0, since count is too large for the first + // view. + vv.views[i-1] = newFirst + vv.views = vv.views[i-1:] + return newFirst, true } // 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 +205,7 @@ 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 +217,7 @@ 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 } @@ -162,3 +232,19 @@ 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 +} + +// removeFirst panics when len(vv.views) < 1. +func (vv *VectorisedView) removeFirst() { + vv.size -= len(vv.views[0]) + vv.views[0] = nil + vv.views = vv.views[1:] +} diff --git a/pkg/tcpip/buffer/view_test.go b/pkg/tcpip/buffer/view_test.go index ebc3a17b7..c56795c7b 100644 --- a/pkg/tcpip/buffer/view_test.go +++ b/pkg/tcpip/buffer/view_test.go @@ -16,6 +16,7 @@ package buffer import ( + "bytes" "reflect" "testing" ) @@ -233,3 +234,252 @@ 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) + } + }) + } +} + +var pullUpTestCases = []struct { + comment string + in VectorisedView + count int + want []byte + result VectorisedView + ok bool +}{ + { + comment: "simple case", + in: vv(2, "12"), + count: 1, + want: []byte("1"), + result: vv(2, "12"), + ok: true, + }, + { + comment: "entire View", + in: vv(2, "1", "2"), + count: 1, + want: []byte("1"), + result: vv(2, "1", "2"), + ok: true, + }, + { + comment: "spanning across two Views", + in: vv(3, "1", "23"), + count: 2, + want: []byte("12"), + result: vv(3, "12", "3"), + ok: true, + }, + { + comment: "spanning across all Views", + in: vv(5, "1", "23", "45"), + count: 5, + want: []byte("12345"), + result: vv(5, "12345"), + ok: true, + }, + { + comment: "count = 0", + in: vv(1, "1"), + count: 0, + want: []byte{}, + result: vv(1, "1"), + ok: true, + }, + { + comment: "count = size", + in: vv(1, "1"), + count: 1, + want: []byte("1"), + result: vv(1, "1"), + ok: true, + }, + { + comment: "count too large", + in: vv(3, "1", "23"), + count: 4, + want: nil, + result: vv(3, "1", "23"), + ok: false, + }, + { + comment: "empty vv", + in: vv(0, ""), + count: 1, + want: nil, + result: vv(0, ""), + ok: false, + }, + { + comment: "empty vv, count = 0", + in: vv(0, ""), + count: 0, + want: nil, + result: vv(0, ""), + ok: true, + }, + { + comment: "empty views", + in: vv(3, "", "1", "", "23"), + count: 2, + want: []byte("12"), + result: vv(3, "12", "3"), + ok: true, + }, +} + +func TestPullUp(t *testing.T) { + for _, c := range pullUpTestCases { + got, ok := c.in.PullUp(c.count) + + // Is the return value right? + if ok != c.ok { + t.Errorf("Test %q failed when calling PullUp(%d) on %v. Got an ok of %t. Want %t", + c.comment, c.count, c.in, ok, c.ok) + } + if bytes.Compare(got, View(c.want)) != 0 { + t.Errorf("Test %q failed when calling PullUp(%d) on %v. Got %v. Want %v", + c.comment, c.count, c.in, got, c.want) + } + + // Is the underlying structure right? + if !reflect.DeepEqual(c.in, c.result) { + t.Errorf("Test %q failed when calling PullUp(%d). Got vv with structure %v. Wanted %v", + c.comment, c.count, c.in, c.result) + } + } +} diff --git a/pkg/tcpip/checker/checker.go b/pkg/tcpip/checker/checker.go index 8dc0f7c0e..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: @@ -310,6 +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) } @@ -336,6 +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 @@ -350,6 +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 @@ -381,6 +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 @@ -398,6 +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 @@ -494,6 +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 @@ -612,6 +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) } @@ -644,6 +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) @@ -658,6 +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) @@ -700,6 +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) @@ -714,6 +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) @@ -728,7 +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 { @@ -760,9 +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...) } @@ -780,7 +799,54 @@ 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) + } + } +} + +// 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 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) + na := header.NDPNeighborAdvert(icmp.NDPPayload()) + + if got := na.TargetAddress(); got != want { + t.Errorf("got %T.TargetAddress() = %s, want = %s", na, got, want) + } + } +} + +// 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() + + icmp := h.(header.ICMPv6) + na := header.NDPNeighborAdvert(icmp.NDPPayload()) + + if got := na.SolicitedFlag(); got != want { + t.Errorf("got %T.SolicitedFlag = %t, want = %t", na, got, want) } } } @@ -819,6 +885,13 @@ func ndpOptions(t *testing.T, optsBuf header.NDPOptions, opts []header.NDPOption } 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) } @@ -831,6 +904,21 @@ func ndpOptions(t *testing.T, optsBuf header.NDPOptions, opts []header.NDPOption } } +// 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. // @@ -849,7 +937,7 @@ func NDPNSOptions(opts []header.NDPOption) TransportChecker { // NDPRS creates a checker that checks that the packet contains a valid NDP // Router Solicitation message (as per the raw wire format). // -// 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 // 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 { 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 76e88e9b3..ba80b64a8 100644 --- a/pkg/tcpip/header/ipv6.go +++ b/pkg/tcpip/header/ipv6.go @@ -27,7 +27,9 @@ const ( // IPv6PayloadLenOffset is the offset of the PayloadLength field in // IPv6 header. IPv6PayloadLenOffset = 4 - nextHdr = 6 + // IPv6NextHeaderOffset is the offset of the NextHeader field in + // IPv6 header. + IPv6NextHeaderOffset = 6 hopLimit = 7 v6SrcAddr = 8 v6DstAddr = v6SrcAddr + IPv6AddressSize @@ -163,7 +165,7 @@ func (b IPv6) HopLimit() uint8 { // NextHeader returns the value of the "next header" field of the ipv6 header. func (b IPv6) NextHeader() uint8 { - return b[nextHdr] + return b[IPv6NextHeaderOffset] } // TransportProtocol implements Network.TransportProtocol. @@ -223,7 +225,7 @@ func (b IPv6) SetDestinationAddress(addr tcpip.Address) { // SetNextHeader sets the value of the "next header" field of the ipv6 header. func (b IPv6) SetNextHeader(v uint8) { - b[nextHdr] = v + b[IPv6NextHeaderOffset] = v } // SetChecksum implements Network.SetChecksum. Given that IPv6 doesn't have a @@ -235,7 +237,7 @@ func (IPv6) SetChecksum(uint16) { func (b IPv6) Encode(i *IPv6Fields) { b.SetTOS(i.TrafficClass, i.FlowLabel) b.SetPayloadLength(i.PayloadLength) - b[nextHdr] = i.NextHeader + b[IPv6NextHeaderOffset] = i.NextHeader b[hopLimit] = i.HopLimit b.SetSourceAddress(i.SrcAddr) b.SetDestinationAddress(i.DstAddr) 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/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/targets.go b/pkg/tcpip/iptables/targets.go deleted file mode 100644 index 81a2e39a2..000000000 --- a/pkg/tcpip/iptables/targets.go +++ /dev/null @@ -1,65 +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 - -import ( - "gvisor.dev/gvisor/pkg/log" - "gvisor.dev/gvisor/pkg/tcpip" -) - -// AcceptTarget accepts packets. -type AcceptTarget struct{} - -// Action implements Target.Action. -func (AcceptTarget) Action(packet tcpip.PacketBuffer) (RuleVerdict, int) { - return RuleAccept, 0 -} - -// DropTarget drops packets. -type DropTarget struct{} - -// Action implements Target.Action. -func (DropTarget) Action(packet tcpip.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 tcpip.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(tcpip.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(tcpip.PacketBuffer) (RuleVerdict, int) { - return RuleReturn, 0 -} diff --git a/pkg/tcpip/link/channel/channel.go b/pkg/tcpip/link/channel/channel.go index 5944ba190..9bf67686d 100644 --- a/pkg/tcpip/link/channel/channel.go +++ b/pkg/tcpip/link/channel/channel.go @@ -28,7 +28,7 @@ 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 @@ -50,13 +50,11 @@ type NotificationHandle struct { } type queue struct { + // c is the outbound packet channel. + c chan PacketInfo // mu protects fields below. - mu sync.RWMutex - // c is the outbound packet channel. Sending to c should hold mu. - c chan PacketInfo - numWrite int - numRead int - notify []*NotificationHandle + mu sync.RWMutex + notify []*NotificationHandle } func (q *queue) Close() { @@ -64,11 +62,8 @@ func (q *queue) Close() { } func (q *queue) Read() (PacketInfo, bool) { - q.mu.Lock() - defer q.mu.Unlock() select { case p := <-q.c: - q.numRead++ return p, true default: return PacketInfo{}, false @@ -76,15 +71,8 @@ func (q *queue) Read() (PacketInfo, bool) { } func (q *queue) ReadContext(ctx context.Context) (PacketInfo, bool) { - // We have to receive from channel without holding the lock, since it can - // block indefinitely. This will cause a window that numWrite - numRead - // produces a larger number, but won't go to negative. numWrite >= numRead - // still holds. select { case pkt := <-q.c: - q.mu.Lock() - defer q.mu.Unlock() - q.numRead++ return pkt, true case <-ctx.Done(): return PacketInfo{}, false @@ -93,16 +81,12 @@ func (q *queue) ReadContext(ctx context.Context) (PacketInfo, bool) { func (q *queue) Write(p PacketInfo) bool { wrote := false - - // It's important to make sure nobody can see numWrite until we increment it, - // so numWrite >= numRead holds. - q.mu.Lock() select { case q.c <- p: wrote = true - q.numWrite++ default: } + q.mu.Lock() notify := q.notify q.mu.Unlock() @@ -116,13 +100,7 @@ func (q *queue) Write(p PacketInfo) bool { } func (q *queue) Num() int { - q.mu.RLock() - defer q.mu.RUnlock() - n := q.numWrite - q.numRead - if n < 0 { - panic("numWrite < numRead") - } - return n + return len(q.c) } func (q *queue) AddNotify(notify Notification) *NotificationHandle { @@ -203,12 +181,12 @@ func (e *Endpoint) NumQueued() int { } // 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) } @@ -251,13 +229,13 @@ 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, @@ -269,21 +247,15 @@ func (e *Endpoint) WritePacket(r *stack.Route, gso *stack.GSO, protocol tcpip.Ne } // 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 - 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, @@ -301,7 +273,7 @@ func (e *Endpoint) WritePackets(r *stack.Route, gso *stack.GSO, pkts []tcpip.Pac // 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, } 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..073c84ef9 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), }) @@ -92,21 +92,21 @@ 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") } // WriteRawPacket implements stack.LinkEndpoint.WriteRawPacket. func (e *endpoint) WriteRawPacket(vv buffer.VectorisedView) *tcpip.Error { - // Reject the packet if it's shorter than an ethernet header. - if vv.Size() < header.EthernetMinimumSize { + // There should be an ethernet header at the beginning of vv. + hdr, ok := vv.PullUp(header.EthernetMinimumSize) + if !ok { + // Reject the packet if it's shorter than an ethernet header. return tcpip.ErrBadAddress } - - // There should be an ethernet header at the beginning of vv. - linkHeader := header.Ethernet(vv.First()[:header.EthernetMinimumSize]) + linkHeader := header.Ethernet(hdr) 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/rawfile/BUILD b/pkg/tcpip/link/rawfile/BUILD index 14b527bc2..9cc08d0e2 100644 --- a/pkg/tcpip/link/rawfile/BUILD +++ b/pkg/tcpip/link/rawfile/BUILD @@ -1,4 +1,4 @@ -load("//tools:defs.bzl", "go_library") +load("//tools:defs.bzl", "go_library", "go_test") package(licenses = ["notice"]) @@ -18,3 +18,10 @@ go_library( "@org_golang_x_sys//unix:go_default_library", ], ) + +go_test( + name = "rawfile_test", + size = "small", + srcs = ["rawfile_test.go"], + library = ":rawfile", +) diff --git a/pkg/tcpip/link/rawfile/rawfile_test.go b/pkg/tcpip/link/rawfile/rawfile_test.go new file mode 100644 index 000000000..8f14ba761 --- /dev/null +++ b/pkg/tcpip/link/rawfile/rawfile_test.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 linux + +package rawfile + +import ( + "syscall" + "testing" +) + +func TestNonBlockingWrite3ZeroLength(t *testing.T) { + fd, err := syscall.Open("/dev/null", syscall.O_WRONLY, 0) + if err != nil { + t.Fatalf("failed to open /dev/null: %v", err) + } + defer syscall.Close(fd) + + if err := NonBlockingWrite3(fd, []byte{}, []byte{0}, nil); err != nil { + t.Fatalf("failed to write: %v", err) + } +} + +func TestNonBlockingWrite3Nil(t *testing.T) { + fd, err := syscall.Open("/dev/null", syscall.O_WRONLY, 0) + if err != nil { + t.Fatalf("failed to open /dev/null: %v", err) + } + defer syscall.Close(fd) + + if err := NonBlockingWrite3(fd, nil, []byte{0}, nil); err != nil { + t.Fatalf("failed to write: %v", err) + } +} diff --git a/pkg/tcpip/link/rawfile/rawfile_unsafe.go b/pkg/tcpip/link/rawfile/rawfile_unsafe.go index 44e25d475..92efd0bf8 100644 --- a/pkg/tcpip/link/rawfile/rawfile_unsafe.go +++ b/pkg/tcpip/link/rawfile/rawfile_unsafe.go @@ -76,9 +76,13 @@ func NonBlockingWrite3(fd int, b1, b2, b3 []byte) *tcpip.Error { // We have two buffers. Build the iovec that represents them and issue // a writev syscall. + var base *byte + if len(b1) > 0 { + base = &b1[0] + } iovec := [3]syscall.Iovec{ { - Base: &b1[0], + Base: base, Len: uint64(len(b1)), }, { 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..33f640b85 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 { @@ -674,7 +674,7 @@ func TestSimpleReceive(t *testing.T) { // Wait for packet to be received, then check it. c.waitForPackets(1, time.After(5*time.Second), "Timeout waiting for packet") c.mu.Lock() - rcvd := []byte(c.packets[0].vv.First()) + rcvd := []byte(c.packets[0].vv.ToView()) c.packets = c.packets[:0] c.mu.Unlock() diff --git a/pkg/tcpip/link/sniffer/sniffer.go b/pkg/tcpip/link/sniffer/sniffer.go index 3392b7edd..0799c8f4d 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 +168,39 @@ 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 { + logPacket(prefix, protocol, pkt, 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) } } } @@ -232,71 +208,33 @@ 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) { +func logPacket(prefix string, protocol tcpip.NetworkProtocolNumber, pkt *stack.PacketBuffer, gso *stack.GSO) { // Figure out the network layer info. var transProto uint8 src := tcpip.Address("unknown") @@ -305,28 +243,49 @@ func logPacket(prefix string, protocol tcpip.NetworkProtocolNumber, b buffer.Vie size := uint16(0) var fragmentOffset uint16 var moreFragments bool + + // Create a clone of pkt, including any headers if present. Avoid allocating + // backing memory for the clone. + views := [8]buffer.View{} + vv := buffer.NewVectorisedView(0, views[:0]) + vv.AppendView(pkt.Header.View()) + vv.Append(pkt.Data) + switch protocol { case header.IPv4ProtocolNumber: - ipv4 := header.IPv4(b) + hdr, ok := vv.PullUp(header.IPv4MinimumSize) + if !ok { + return + } + ipv4 := header.IPv4(hdr) fragmentOffset = ipv4.FragmentOffset() moreFragments = ipv4.Flags()&header.IPv4FlagMoreFragments == header.IPv4FlagMoreFragments src = ipv4.SourceAddress() dst = ipv4.DestinationAddress() transProto = ipv4.Protocol() size = ipv4.TotalLength() - uint16(ipv4.HeaderLength()) - b = b[ipv4.HeaderLength():] + vv.TrimFront(int(ipv4.HeaderLength())) id = int(ipv4.ID()) case header.IPv6ProtocolNumber: - ipv6 := header.IPv6(b) + hdr, ok := vv.PullUp(header.IPv6MinimumSize) + if !ok { + return + } + ipv6 := header.IPv6(hdr) src = ipv6.SourceAddress() dst = ipv6.DestinationAddress() transProto = ipv6.NextHeader() size = ipv6.PayloadLength() - b = b[header.IPv6MinimumSize:] + vv.TrimFront(header.IPv6MinimumSize) case header.ARPProtocolNumber: - arp := header.ARP(b) + hdr, ok := vv.PullUp(header.ARPSize) + if !ok { + return + } + vv.TrimFront(header.ARPSize) + arp := header.ARP(hdr) log.Infof( "%s arp %v (%v) -> %v (%v) valid:%v", prefix, @@ -342,7 +301,7 @@ func logPacket(prefix string, protocol tcpip.NetworkProtocolNumber, b buffer.Vie // We aren't guaranteed to have a transport header - it's possible for // writes via raw endpoints to contain only network headers. - if minSize, ok := transportProtocolMinSizes[tcpip.TransportProtocolNumber(transProto)]; ok && len(b) < minSize { + if minSize, ok := transportProtocolMinSizes[tcpip.TransportProtocolNumber(transProto)]; ok && vv.Size() < minSize { log.Infof("%s %v -> %v transport protocol: %d, but no transport header found (possible raw packet)", prefix, src, dst, transProto) return } @@ -355,7 +314,11 @@ func logPacket(prefix string, protocol tcpip.NetworkProtocolNumber, b buffer.Vie switch tcpip.TransportProtocolNumber(transProto) { case header.ICMPv4ProtocolNumber: transName = "icmp" - icmp := header.ICMPv4(b) + hdr, ok := vv.PullUp(header.ICMPv4MinimumSize) + if !ok { + break + } + icmp := header.ICMPv4(hdr) icmpType := "unknown" if fragmentOffset == 0 { switch icmp.Type() { @@ -388,7 +351,11 @@ func logPacket(prefix string, protocol tcpip.NetworkProtocolNumber, b buffer.Vie case header.ICMPv6ProtocolNumber: transName = "icmp" - icmp := header.ICMPv6(b) + hdr, ok := vv.PullUp(header.ICMPv6MinimumSize) + if !ok { + break + } + icmp := header.ICMPv6(hdr) icmpType := "unknown" switch icmp.Type() { case header.ICMPv6DstUnreachable: @@ -419,7 +386,11 @@ func logPacket(prefix string, protocol tcpip.NetworkProtocolNumber, b buffer.Vie case header.UDPProtocolNumber: transName = "udp" - udp := header.UDP(b) + hdr, ok := vv.PullUp(header.UDPMinimumSize) + if !ok { + break + } + udp := header.UDP(hdr) if fragmentOffset == 0 && len(udp) >= header.UDPMinimumSize { srcPort = udp.SourcePort() dstPort = udp.DestinationPort() @@ -429,7 +400,11 @@ func logPacket(prefix string, protocol tcpip.NetworkProtocolNumber, b buffer.Vie case header.TCPProtocolNumber: transName = "tcp" - tcp := header.TCP(b) + hdr, ok := vv.PullUp(header.TCPMinimumSize) + if !ok { + break + } + tcp := header.TCP(hdr) if fragmentOffset == 0 && len(tcp) >= header.TCPMinimumSize { offset := int(tcp.DataOffset()) if offset < header.TCPMinimumSize { diff --git a/pkg/tcpip/link/tun/device.go b/pkg/tcpip/link/tun/device.go index 6ff47a742..617446ea2 100644 --- a/pkg/tcpip/link/tun/device.go +++ b/pkg/tcpip/link/tun/device.go @@ -98,7 +98,12 @@ func (d *Device) SetIff(s *stack.Stack, name string, flags uint16) error { prefix = "tap" } - endpoint, err := attachOrCreateNIC(s, name, prefix) + linkCaps := stack.CapabilityNone + if isTap { + linkCaps |= stack.CapabilityResolutionRequired + } + + endpoint, err := attachOrCreateNIC(s, name, prefix, linkCaps) if err != nil { return syserror.EINVAL } @@ -109,7 +114,7 @@ func (d *Device) SetIff(s *stack.Stack, name string, flags uint16) error { return nil } -func attachOrCreateNIC(s *stack.Stack, name, prefix string) (*tunEndpoint, error) { +func attachOrCreateNIC(s *stack.Stack, name, prefix string, linkCaps stack.LinkEndpointCapabilities) (*tunEndpoint, error) { for { // 1. Try to attach to an existing NIC. if name != "" { @@ -135,6 +140,7 @@ func attachOrCreateNIC(s *stack.Stack, name, prefix string) (*tunEndpoint, error nicID: id, name: name, } + endpoint.Endpoint.LinkEPCapabilities = linkCaps if endpoint.name == "" { endpoint.name = fmt.Sprintf("%s%d", prefix, id) } @@ -207,7 +213,7 @@ func (d *Device) Write(data []byte) (int64, error) { remote = tcpip.LinkAddress(zeroMAC[:]) } - pkt := tcpip.PacketBuffer{ + pkt := stack.PacketBuffer{ Data: buffer.View(data).ToVectorisedView(), } if ethHdr != nil { 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 e9fcc89a8..cf73a939e 100644 --- a/pkg/tcpip/network/arp/arp.go +++ b/pkg/tcpip/network/arp/arp.go @@ -79,21 +79,24 @@ 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) { - v := pkt.Data.First() +func (e *endpoint) HandlePacket(r *stack.Route, pkt stack.PacketBuffer) { + v, ok := pkt.Data.PullUp(header.ARPSize) + if !ok { + return + } h := header.ARP(v) if !h.IsValid() { return @@ -113,7 +116,7 @@ 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 @@ -167,7 +170,7 @@ 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, }) } 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/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..4cbefe5ab 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" @@ -25,8 +24,12 @@ 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) { - h := header.IPv4(pkt.Data.First()) +func (e *endpoint) handleControl(typ stack.ControlType, extra uint32, pkt stack.PacketBuffer) { + h, ok := pkt.Data.PullUp(header.IPv4MinimumSize) + if !ok { + return + } + hdr := header.IPv4(h) // We don't use IsValid() here because ICMP only requires that the IP // header plus 8 bytes of the transport header be included. So it's @@ -35,12 +38,12 @@ func (e *endpoint) handleControl(typ stack.ControlType, extra uint32, pkt tcpip. // // Drop packet if it doesn't have the basic IPv4 header or if the // original source address doesn't match the endpoint's address. - if len(h) < header.IPv4MinimumSize || h.SourceAddress() != e.id.LocalAddress { + if hdr.SourceAddress() != e.id.LocalAddress { return } - hlen := int(h.HeaderLength()) - if pkt.Data.Size() < hlen || h.FragmentOffset() != 0 { + hlen := int(hdr.HeaderLength()) + if pkt.Data.Size() < hlen || hdr.FragmentOffset() != 0 { // We won't be able to handle this if it doesn't contain the // full IPv4 header, or if it's a fragment not at offset 0 // (because it won't have the transport header). @@ -49,15 +52,15 @@ func (e *endpoint) handleControl(typ stack.ControlType, extra uint32, pkt tcpip. // Skip the ip header, then deliver control message. pkt.Data.TrimFront(hlen) - p := h.TransportProtocol() - e.dispatcher.DeliverTransportControlPacket(e.id.LocalAddress, h.DestinationAddress(), ProtocolNumber, p, typ, extra, pkt) + p := hdr.TransportProtocol() + e.dispatcher.DeliverTransportControlPacket(e.id.LocalAddress, hdr.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() - if len(v) < header.ICMPv4MinimumSize { + v, ok := pkt.Data.PullUp(header.ICMPv4MinimumSize) + if !ok { received.Invalid.Increment() return } @@ -85,7 +88,7 @@ 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 +102,7 @@ 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 4f1742938..17202cc7a 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), }) @@ -273,29 +280,59 @@ 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() + } + + // 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 } - for i := range pkts { - ip := e.addIPHeader(r, &pkts[i].Header, pkts[i].DataSize, params) - pkts[i].NetworkHeader = buffer.View(ip) + // 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()) + h, ok := pkt.Data.PullUp(header.IPv4MinimumSize) + if !ok { + return tcpip.ErrInvalidOptionValue + } + ip := header.IPv4(h) if !ip.IsValid(pkt.Data.Size()) { return tcpip.ErrInvalidOptionValue } @@ -344,8 +381,12 @@ 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) { - headerView := pkt.Data.First() +func (e *endpoint) HandlePacket(r *stack.Route, pkt stack.PacketBuffer) { + headerView, ok := pkt.Data.PullUp(header.IPv4MinimumSize) + if !ok { + r.Stats().IP.MalformedPacketsReceived.Increment() + return + } h := header.IPv4(headerView) if !h.IsValid(pkt.Data.Size()) { r.Stats().IP.MalformedPacketsReceived.Increment() @@ -361,7 +402,7 @@ 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 } 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..bdf3a0d25 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" @@ -27,8 +27,12 @@ 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) { - h := header.IPv6(pkt.Data.First()) +func (e *endpoint) handleControl(typ stack.ControlType, extra uint32, pkt stack.PacketBuffer) { + h, ok := pkt.Data.PullUp(header.IPv6MinimumSize) + if !ok { + return + } + hdr := header.IPv6(h) // We don't use IsValid() here because ICMP only requires that up to // 1280 bytes of the original packet be included. So it's likely that it @@ -36,17 +40,21 @@ func (e *endpoint) handleControl(typ stack.ControlType, extra uint32, pkt tcpip. // // Drop packet if it doesn't have the basic IPv6 header or if the // original source address doesn't match the endpoint's address. - if len(h) < header.IPv6MinimumSize || h.SourceAddress() != e.id.LocalAddress { + if hdr.SourceAddress() != e.id.LocalAddress { return } // Skip the IP header, then handle the fragmentation header if there // is one. pkt.Data.TrimFront(header.IPv6MinimumSize) - p := h.TransportProtocol() + p := hdr.TransportProtocol() if p == header.IPv6FragmentHeader { - f := header.IPv6Fragment(pkt.Data.First()) - if !f.IsValid() || f.FragmentOffset() != 0 { + f, ok := pkt.Data.PullUp(header.IPv6FragmentHeaderSize) + if !ok { + return + } + fragHdr := header.IPv6Fragment(f) + if !fragHdr.IsValid() || fragHdr.FragmentOffset() != 0 { // We can't handle fragments that aren't at offset 0 // because they don't have the transport headers. return @@ -55,19 +63,19 @@ func (e *endpoint) handleControl(typ stack.ControlType, extra uint32, pkt tcpip. // Skip fragmentation header and find out the actual protocol // number. pkt.Data.TrimFront(header.IPv6FragmentHeaderSize) - p = f.TransportProtocol() + p = fragHdr.TransportProtocol() } // Deliver the control packet to the transport endpoint. - e.dispatcher.DeliverTransportControlPacket(e.id.LocalAddress, h.DestinationAddress(), ProtocolNumber, p, typ, extra, pkt) + e.dispatcher.DeliverTransportControlPacket(e.id.LocalAddress, hdr.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 - v := pkt.Data.First() - if len(v) < header.ICMPv6MinimumSize { + v, ok := pkt.Data.PullUp(header.ICMPv6HeaderSize) + if !ok { received.Invalid.Increment() return } @@ -76,69 +84,63 @@ func (e *endpoint) handleICMP(r *stack.Route, netHeader buffer.View, pkt tcpip.P // Validate ICMPv6 checksum before processing the packet. // - // 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.RemoveFirst() + payload := pkt.Data.Clone(nil) + payload.TrimFront(len(h)) 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. switch h.Type() { case header.ICMPv6PacketTooBig: received.PacketTooBig.Increment() - if len(v) < header.ICMPv6PacketTooBigMinimumSize { + hdr, ok := pkt.Data.PullUp(header.ICMPv6PacketTooBigMinimumSize) + if !ok { received.Invalid.Increment() return } pkt.Data.TrimFront(header.ICMPv6PacketTooBigMinimumSize) - mtu := h.MTU() + mtu := header.ICMPv6(hdr).MTU() e.handleControl(stack.ControlPacketTooBig, calculateMTU(mtu), pkt) case header.ICMPv6DstUnreachable: received.DstUnreachable.Increment() - if len(v) < header.ICMPv6DstUnreachableMinimumSize { + hdr, ok := pkt.Data.PullUp(header.ICMPv6DstUnreachableMinimumSize) + if !ok { received.Invalid.Increment() return } pkt.Data.TrimFront(header.ICMPv6DstUnreachableMinimumSize) - switch h.Code() { + switch header.ICMPv6(hdr).Code() { case header.ICMPv6PortUnreachable: e.handleControl(stack.ControlPortUnreachable, 0, pkt) } case header.ICMPv6NeighborSolicit: received.NeighborSolicit.Increment() - if len(v) < header.ICMPv6NeighborSolicitMinimumSize { + if pkt.Data.Size() < header.ICMPv6NeighborSolicitMinimumSize || !isNDPValid() { received.Invalid.Increment() return } - ns := header.NDPNeighborSolicit(h.NDPPayload()) + // The remainder of payload must be only the neighbor solicitation, so + // payload.ToView() always returns the solicitation. Per RFC 6980 section 5, + // NDP messages cannot be fragmented. Also note that in the common case NDP + // datagrams are very small and ToView() will not incur allocations. + ns := header.NDPNeighborSolicit(payload.ToView()) it, err := ns.Options().Iter(true) if err != nil { // If we have a malformed NDP NS option, drop the packet. @@ -148,58 +150,53 @@ 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 +204,36 @@ 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 +246,40 @@ 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 +288,7 @@ 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() @@ -253,12 +298,16 @@ 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 pkt.Data.Size() < header.ICMPv6NeighborAdvertSize || !isNDPValid() { received.Invalid.Increment() return } - na := header.NDPNeighborAdvert(h.NDPPayload()) + // The remainder of payload must be only the neighbor advertisement, so + // payload.ToView() always returns the advertisement. Per RFC 6980 section + // 5, NDP messages cannot be fragmented. Also note that in the common case + // NDP datagrams are very small and ToView() will not incur allocations. + na := header.NDPNeighborAdvert(payload.ToView()) it, err := na.Options().Iter(true) if err != nil { // If we have a malformed NDP NA option, drop the packet. @@ -268,45 +317,43 @@ 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 @@ -314,23 +361,36 @@ 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 { + icmpHdr, ok := pkt.Data.PullUp(header.ICMPv6EchoMinimumSize) + if !ok { received.Invalid.Increment() return } pkt.Data.TrimFront(header.ICMPv6EchoMinimumSize) hdr := buffer.NewPrependable(int(r.MaxHeaderLength()) + header.ICMPv6EchoMinimumSize) packet := header.ICMPv6(hdr.Prepend(header.ICMPv6EchoMinimumSize)) - copy(packet, h) + copy(packet, icmpHdr) 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 { @@ -341,7 +401,7 @@ func (e *endpoint) handleICMP(r *stack.Route, netHeader buffer.View, pkt tcpip.P case header.ICMPv6EchoReply: received.EchoReply.Increment() - if len(v) < header.ICMPv6EchoMinimumSize { + if pkt.Data.Size() < header.ICMPv6EchoMinimumSize { received.Invalid.Increment() return } @@ -355,8 +415,21 @@ 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() + + // Is the NDP payload of sufficient size to hold a Router + // Advertisement? + if pkt.Data.Size()-header.ICMPv6HeaderSize < header.NDPRAMinimumSize || !isNDPValid() { + received.Invalid.Increment() + return + } + routerAddr := iph.SourceAddress() // @@ -370,17 +443,11 @@ 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) + // The remainder of payload must be only the router advertisement, so + // payload.ToView() always returns the advertisement. Per RFC 6980 section + // 5, NDP messages cannot be fragmented. Also note that in the common case + // NDP datagrams are very small and ToView() will not incur allocations. + ra := header.NDPRouterAdvert(payload.ToView()) opts := ra.Options() // Are options valid as per the wire format? @@ -395,8 +462,6 @@ 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 +469,10 @@ 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 +532,7 @@ 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..d412ff688 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 { @@ -165,7 +166,8 @@ func TestICMPCounts(t *testing.T) { }, { typ: header.ICMPv6NeighborSolicit, - size: header.ICMPv6NeighborSolicitMinimumSize}, + size: header.ICMPv6NeighborSolicitMinimumSize, + }, { typ: header.ICMPv6NeighborAdvert, size: header.ICMPv6NeighborAdvertMinimumSize, @@ -187,7 +189,7 @@ 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 +328,7 @@ 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 +563,7 @@ 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 +740,7 @@ 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 +918,7 @@ 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 9aef5234b..486725131 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 +163,22 @@ 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) { - headerView := pkt.Data.First() +func (e *endpoint) HandlePacket(r *stack.Route, pkt stack.PacketBuffer) { + headerView, ok := pkt.Data.PullUp(header.IPv6MinimumSize) + if !ok { + r.Stats().IP.MalformedPacketsReceived.Increment() + return + } h := header.IPv6(headerView) if !h.IsValid(pkt.Data.Size()) { + r.Stats().IP.MalformedPacketsReceived.Increment() return } @@ -179,14 +186,235 @@ 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 +457,7 @@ 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 } 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 8febd54c8..5e963a4af 100644 --- a/pkg/tcpip/stack/BUILD +++ b/pkg/tcpip/stack/BUILD @@ -15,15 +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", @@ -33,6 +52,7 @@ go_library( visibility = ["//visibility:public"], deps = [ "//pkg/ilist", + "//pkg/log", "//pkg/rand", "//pkg/sleep", "//pkg/sync", @@ -40,7 +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", @@ -64,7 +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", @@ -80,6 +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/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..c7c663498 --- /dev/null +++ b/pkg/tcpip/stack/forwarder_test.go @@ -0,0 +1,638 @@ +// 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, ok := pkt.Data.PullUp(fwdTestNetHeaderLen) + if !ok { + return + } + 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.Data.ToView() + 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.Data.ToView() + 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.Data.ToView() + 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.Data.ToView() + 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/iptables/iptables.go b/pkg/tcpip/stack/iptables.go index dbaccbb36..6b91159d4 100644 --- a/pkg/tcpip/iptables/iptables.go +++ b/pkg/tcpip/stack/iptables.go @@ -12,14 +12,11 @@ // 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 +package stack import ( "fmt" - "gvisor.dev/gvisor/pkg/tcpip" "gvisor.dev/gvisor/pkg/tcpip/header" ) @@ -135,6 +132,27 @@ func EmptyFilterTable() Table { } } +// 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 @@ -155,7 +173,7 @@ const ( // dropped. // // Precondition: pkt.NetworkHeader is set. -func (it *IPTables) Check(hook Hook, pkt tcpip.PacketBuffer) bool { +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] @@ -191,8 +209,32 @@ func (it *IPTables) Check(hook Hook, pkt tcpip.PacketBuffer) bool { return true } -// Precondition: pkt.NetworkHeader is set. -func (it *IPTables) checkChain(hook Hook, pkt tcpip.PacketBuffer, table Table, ruleIdx int) chainVerdict { +// CheckPackets runs pkts through the rules for hook and returns a map of packets that +// should not go forward. +// +// Precondition: pkt is a IPv4 packet of at least length header.IPv4MinimumSize. +// +// TODO(gvisor.dev/issue/170): pk.NetworkHeader will always be set as a +// precondition. +// +// 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 is a IPv4 packet of at least length header.IPv4MinimumSize. +// TODO(gvisor.dev/issue/170): pk.NetworkHeader will always be set as a +// precondition. +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) { @@ -236,13 +278,25 @@ func (it *IPTables) checkChain(hook Hook, pkt tcpip.PacketBuffer, table Table, r return chainDrop } -// Precondition: pk.NetworkHeader is set. -func (it *IPTables) checkRule(hook Hook, pkt tcpip.PacketBuffer, table Table, ruleIdx int) (RuleVerdict, int) { +// Precondition: pkt is a IPv4 packet of at least length header.IPv4MinimumSize. +// TODO(gvisor.dev/issue/170): pk.NetworkHeader will always be set as a +// precondition. +func (it *IPTables) checkRule(hook Hook, pkt PacketBuffer, table Table, ruleIdx int) (RuleVerdict, int) { 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() { + // If pkt.NetworkHeader hasn't been set yet, it will be contained in + // pkt.Data. + if pkt.NetworkHeader == nil { + var ok bool + pkt.NetworkHeader, ok = pkt.Data.PullUp(header.IPv4MinimumSize) + if !ok { + // Precondition has been violated. + panic(fmt.Sprintf("iptables checks require IPv4 headers of at least %d bytes", header.IPv4MinimumSize)) + } + } + + // 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 } @@ -263,3 +317,26 @@ func (it *IPTables) checkRule(hook Hook, pkt tcpip.PacketBuffer, table Table, ru // 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..8be61f4b1 --- /dev/null +++ b/pkg/tcpip/stack/iptables_targets.go @@ -0,0 +1,155 @@ +// 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, ok := newPkt.Data.PullUp(header.IPv4MinimumSize) + if !ok { + return RuleDrop, 0 + } + netHeader := header.IPv4(headerView) + newPkt.NetworkHeader = headerView + + 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 pkt.Data.Size() < header.UDPMinimumSize { + return RuleDrop, 0 + } + hdr, ok := newPkt.Data.PullUp(header.UDPMinimumSize) + if !ok { + return RuleDrop, 0 + } + udpHeader = header.UDP(hdr) + } + udpHeader.SetDestinationPort(rt.MinPort) + case header.TCPProtocolNumber: + var tcpHeader header.TCP + if newPkt.TransportHeader != nil { + tcpHeader = header.TCP(newPkt.TransportHeader) + } else { + if pkt.Data.Size() < header.TCPMinimumSize { + return RuleDrop, 0 + } + hdr, ok := newPkt.Data.PullUp(header.TCPMinimumSize) + if !ok { + return RuleDrop, 0 + } + tcpHeader = header.TCP(hdr) + } + // 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 7d032fd23..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" @@ -144,6 +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. @@ -156,7 +168,7 @@ 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. @@ -164,5 +176,5 @@ 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 index of the rule to jump to. - Action(packet tcpip.PacketBuffer) (RuleVerdict, int) + Action(packet PacketBuffer) (RuleVerdict, int) } diff --git a/pkg/tcpip/stack/ndp.go b/pkg/tcpip/stack/ndp.go index a9f4d5dad..c11d62f97 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 @@ -392,28 +412,54 @@ type dadState struct { // defaultRouterState holds data associated with a default router discovered by // a Router Advertisement (RA). type defaultRouterState struct { - invalidationTimer tcpip.CancellableTimer + // Timer to invalidate the default router. + // + // May not be nil. + invalidationTimer *tcpip.CancellableTimer } // onLinkPrefixState holds data associated with an on-link prefix discovered by // a Router Advertisement's Prefix Information option (PI) when the NDP // configurations was configured to do so. type onLinkPrefixState struct { - invalidationTimer tcpip.CancellableTimer + // Timer to invalidate the on-link prefix. + // + // May not be nil. + 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 { + // Timer to deprecate the prefix. + // + // May not be nil. + deprecationTimer *tcpip.CancellableTimer - deprecationTimer tcpip.CancellableTimer - invalidationTimer tcpip.CancellableTimer + // Timer to invalidate the prefix. + // + // May not be nil. + 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 +476,7 @@ 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 +488,7 @@ 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 +524,7 @@ 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 +594,9 @@ 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 +612,7 @@ 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 +734,16 @@ 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() @@ -733,7 +788,6 @@ func (ndp *ndpState) invalidateDefaultRouter(ip tcpip.Address) { } rtr.invalidationTimer.StopLocked() - delete(ndp.defaultRouters, ip) // Let the integrator know a discovered default router is invalidated. @@ -762,7 +816,7 @@ func (ndp *ndpState) rememberDefaultRouter(ip tcpip.Address, rl time.Duration) { } state := defaultRouterState{ - invalidationTimer: tcpip.MakeCancellableTimer(&ndp.nic.mu, func() { + invalidationTimer: tcpip.NewCancellableTimer(&ndp.nic.mu, func() { ndp.invalidateDefaultRouter(ip) }), } @@ -792,7 +846,7 @@ func (ndp *ndpState) rememberOnLinkPrefix(prefix tcpip.Subnet, l time.Duration) } state := onLinkPrefixState{ - invalidationTimer: tcpip.MakeCancellableTimer(&ndp.nic.mu, func() { + invalidationTimer: tcpip.NewCancellableTimer(&ndp.nic.mu, func() { ndp.invalidateOnLinkPrefix(prefix) }), } @@ -817,7 +871,6 @@ func (ndp *ndpState) invalidateOnLinkPrefix(prefix tcpip.Subnet) { } s.invalidationTimer.StopLocked() - delete(ndp.onLinkPrefixes, prefix) // Let the integrator know a discovered on-link prefix is invalidated. @@ -899,23 +952,15 @@ 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 +972,8 @@ 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 +989,83 @@ func (ndp *ndpState) doSLAAC(prefix tcpip.Subnet, pl, vl time.Duration) { return } + state := slaacPrefixState{ + deprecationTimer: tcpip.NewCancellableTimer(&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.NewCancellableTimer(&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 +1073,140 @@ 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 { @@ -1094,58 +1217,78 @@ 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) } // cleanupState cleans up ndp's state. @@ -1163,21 +1306,21 @@ func (ndp *ndpState) cleanupAutoGenAddrResourcesAndNotify(addr tcpip.Address) bo // The NIC that ndp belongs to MUST be locked. 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 we are cleaning up // host-only state. - if hostOnly && linkLocalSubnet.Contains(addr) { - linkLocalAddrs++ + 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 { @@ -1185,7 +1328,7 @@ func (ndp *ndpState) cleanupState(hostOnly bool) { } 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 { @@ -1193,7 +1336,7 @@ func (ndp *ndpState) cleanupState(hostOnly bool) { } 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)) } } @@ -1235,9 +1378,9 @@ func (ndp *ndpState) startSolicitingRouters() { // 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())) } // As per RFC 4861 section 4.1, an NDP RS SHOULD include the source @@ -1267,7 +1410,7 @@ 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 98b1c807c..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 { @@ -406,8 +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) @@ -416,10 +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) @@ -430,13 +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 != "" { @@ -476,7 +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), @@ -602,7 +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(), }) @@ -631,6 +642,12 @@ 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) + } }) } } @@ -639,8 +656,9 @@ func TestDADStop(t *testing.T) { const nicID = 1 tests := []struct { - name string - stopFn func(t *testing.T, s *stack.Stack) + name string + stopFn func(t *testing.T, s *stack.Stack) + skipFinalAddrCheck bool }{ // Tests to make sure that DAD stops when an address is removed. { @@ -661,6 +679,19 @@ func TestDADStop(t *testing.T) { } }, }, + + // 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, + }, } for _, test := range tests { @@ -710,12 +741,15 @@ func TestDADStop(t *testing.T) { 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.Errorf("got stack.GetMainNICAddress(%d, %d) = (%s, nil), want = (%s, nil)", nicID, header.IPv6ProtocolNumber, addr, want) + + 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. @@ -901,7 +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)) @@ -936,14 +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) } @@ -952,7 +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{}) } @@ -960,7 +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{}) } @@ -969,7 +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. @@ -1017,8 +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), } @@ -1057,8 +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), } @@ -1099,8 +1129,6 @@ func TestRouterDiscoveryDispatcherNoRemember(t *testing.T) { } func TestRouterDiscovery(t *testing.T) { - t.Parallel() - ndpDisp := ndpDispatcher{ routerC: make(chan ndpRouterEvent, 1), rememberRouter: true, @@ -1202,8 +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, @@ -1270,8 +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), } @@ -1311,8 +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{ @@ -1356,8 +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, "") @@ -1546,8 +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, @@ -1642,8 +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), } @@ -1968,7 +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 @@ -1981,7 +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. @@ -2093,7 +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 @@ -2106,7 +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. @@ -2130,7 +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. @@ -2393,8 +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 @@ -2447,24 +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") } }) @@ -2476,8 +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{ @@ -2534,8 +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) @@ -2582,7 +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. @@ -2647,8 +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{ @@ -2704,8 +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 @@ -2805,12 +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 @@ -2902,11 +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. @@ -2953,11 +3405,115 @@ func TestNDPRecursiveDNSServerDispatch(t *testing.T) { } } +// 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) { - t.Parallel() - const ( lifetimeSeconds = 5 maxRouterAndPrefixEvents = 4 @@ -2983,11 +3539,12 @@ func TestCleanupNDPState(t *testing.T) { 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: "Forwarding Enable", + name: "Enable forwarding", cleanupFn: func(t *testing.T, s *stack.Stack) { t.Helper() s.SetForwarding(true) @@ -2998,7 +3555,7 @@ func TestCleanupNDPState(t *testing.T) { // A NIC should cleanup all NDP state when it is disabled. { - name: "NIC Disable", + name: "Disable NIC", cleanupFn: func(t *testing.T, s *stack.Stack) { t.Helper() @@ -3012,6 +3569,26 @@ func TestCleanupNDPState(t *testing.T) { keepAutoGenLinkLocal: false, maxAutoGenAddrEvents: 6, }, + + // A NIC should cleanup all NDP state when it is removed. + { + name: "Remove NIC", + cleanupFn: func(t *testing.T, s *stack.Stack) { + t.Helper() + + 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, + }, } for _, test := range tests { @@ -3230,35 +3807,37 @@ func TestCleanupNDPState(t *testing.T) { t.Errorf("auto-generated address events mismatch (-want +got):\n%s", diff) } - // 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 !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(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) } - } - 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 @@ -3377,8 +3956,6 @@ 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 { @@ -3395,13 +3972,22 @@ 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", @@ -3409,8 +3995,8 @@ func TestRouterSolicitation(t *testing.T) { 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, }, @@ -3424,8 +4010,8 @@ func TestRouterSolicitation(t *testing.T) { header.NDPSourceLinkLayerAddressOption(linkAddr1), }, maxRtrSolicit: 1, - rtrSolicitInt: time.Second, - effectiveRtrSolicitInt: time.Second, + rtrSolicitInt: 2 * time.Second, + effectiveRtrSolicitInt: 2 * time.Second, maxRtrSolicitDelay: 0, effectiveMaxRtrSolicitDelay: 0, }, @@ -3475,6 +4061,7 @@ func TestRouterSolicitation(t *testing.T) { t.Run(test.name, func(t *testing.T) { t.Parallel() + e := channelLinkWithHeaderLength{ Endpoint: channel.New(int(test.maxRtrSolicit), 1280, test.linkAddr), headerLength: test.linkHeaderLen, @@ -3482,7 +4069,8 @@ func TestRouterSolicitation(t *testing.T) { 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") @@ -3512,7 +4100,8 @@ func TestRouterSolicitation(t *testing.T) { } 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") } @@ -3543,15 +4132,19 @@ func TestRouterSolicitation(t *testing.T) { } 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 @@ -3565,27 +4158,27 @@ func TestRouterSolicitation(t *testing.T) { } 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 tests := []struct { name string startFn func(t *testing.T, s *stack.Stack) - stopFn 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: "Forwarding enabled and disabled", + 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) { + stopFn: func(t *testing.T, s *stack.Stack, _ bool) { t.Helper() s.SetForwarding(true) }, @@ -3594,7 +4187,7 @@ func TestStopStartSolicitingRouters(t *testing.T) { // Tests that when a NIC is enabled or disabled, router solicitations // are started or stopped, respectively. { - name: "NIC disabled and enabled", + name: "Enable and disable NIC", startFn: func(t *testing.T, s *stack.Stack) { t.Helper() @@ -3602,7 +4195,7 @@ func TestStopStartSolicitingRouters(t *testing.T) { t.Fatalf("s.EnableNIC(%d): %s", nicID, err) } }, - stopFn: func(t *testing.T, s *stack.Stack) { + stopFn: func(t *testing.T, s *stack.Stack, _ bool) { t.Helper() if err := s.DisableNIC(nicID); err != nil { @@ -3610,6 +4203,25 @@ func TestStopStartSolicitingRouters(t *testing.T) { } }, }, + + // 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) + } + }, + }, } for _, test := range tests { @@ -3623,7 +4235,6 @@ func TestStopStartSolicitingRouters(t *testing.T) { p, ok := e.ReadContext(ctx) if !ok { t.Fatal("timed out waiting for packet") - return } if p.Proto != header.IPv6ProtocolNumber { @@ -3648,12 +4259,12 @@ func TestStopStartSolicitingRouters(t *testing.T) { } // Stop soliciting routers. - test.stopFn(t, s) - ctx, cancel := context.WithTimeout(context.Background(), delay+defaultTimeout) + 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 forwarding was enabled. - ctx, cancel := context.WithTimeout(context.Background(), interval+defaultTimeout) + // 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") @@ -3662,19 +4273,24 @@ func TestStopStartSolicitingRouters(t *testing.T) { // Stopping router solicitations after it has already been stopped should // do nothing. - test.stopFn(t, s) - ctx, cancel = context.WithTimeout(context.Background(), delay+defaultTimeout) + 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+defaultTimeout) + 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") @@ -3683,7 +4299,7 @@ func TestStopStartSolicitingRouters(t *testing.T) { // Starting router solicitations after it has already completed should do // nothing. test.startFn(t, s) - ctx, cancel = context.WithTimeout(context.Background(), delay+defaultTimeout) + 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 3e6196aee..0c2b1f36a 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" @@ -54,7 +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 @@ -121,15 +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. @@ -165,8 +165,17 @@ func (n *NIC) disable() *tcpip.Error { } n.mu.Lock() - defer n.mu.Unlock() + 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 } @@ -189,7 +198,7 @@ func (n *NIC) disable() *tcpip.Error { } // The NIC may have already left the multicast group. - if err := n.leaveGroupLocked(header.IPv6AllNodesMulticastAddress); err != nil && err != tcpip.ErrBadLocalAddress { + if err := n.leaveGroupLocked(header.IPv6AllNodesMulticastAddress, false /* force */); err != nil && err != tcpip.ErrBadLocalAddress { return err } } @@ -305,24 +314,33 @@ func (n *NIC) remove() *tcpip.Error { n.mu.Lock() defer n.mu.Unlock() - // Detach from link endpoint, so no packet comes in. - n.linkEP.Attach(nil) + 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. - var errs []*tcpip.Error for nid, ref := range n.mu.endpoints { switch ref.getKind() { case permanentTentative, permanent: - if err := n.removePermanentAddressLocked(nid.LocalAddress); err != nil { - errs = append(errs, err) + if tempErr := n.removePermanentAddressLocked(nid.LocalAddress); tempErr != nil && err == nil { + err = tempErr } } } - if len(errs) > 0 { - return errs[0] - } - return nil + // Detach from link endpoint, so no packet comes in. + n.linkEP.Attach(nil) + + return err } // becomeIPv6Router transitions n into an IPv6 router. @@ -461,7 +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{ @@ -473,7 +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. @@ -969,6 +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 } } @@ -993,35 +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 } } @@ -1081,26 +1107,31 @@ 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 } @@ -1113,9 +1144,10 @@ func (n *NIC) isInGroup(addr tcpip.Address) bool { return joins != 0 } -func handlePacket(protocol tcpip.NetworkProtocolNumber, dst, src tcpip.Address, localLinkAddr, remotelinkAddr tcpip.LinkAddress, ref *referencedNetworkEndpoint, pkt tcpip.PacketBuffer) { +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() } @@ -1126,7 +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. @@ -1171,12 +1203,12 @@ func (n *NIC) DeliverNetworkPacket(linkEP LinkEndpoint, remote, local tcpip.Link n.stack.stats.IP.PacketsReceived.Increment() } - if len(pkt.Data.First()) < netProto.MinimumPacketSize() { + netHeader, ok := pkt.Data.PullUp(netProto.MinimumPacketSize()) + if !ok { n.stack.stats.MalformedRcvdPackets.Increment() return } - - src, dst := netProto.ParseAddresses(pkt.Data.First()) + src, dst := netProto.ParseAddresses(netHeader) if n.stack.handleLocal && !n.isLoopback() && n.getRef(protocol, src) != nil { // The source address is one of our own, so we never should have gotten a @@ -1186,6 +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 @@ -1201,10 +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 @@ -1213,24 +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 } @@ -1240,9 +1287,24 @@ 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. + if linkHeaderLen := int(n.linkEP.MaxHeaderLength()); linkHeaderLen != 0 { + pkt.Header = buffer.NewPrependable(linkHeaderLen) + } + + 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() @@ -1256,12 +1318,13 @@ func (n *NIC) DeliverTransportPacket(r *Route, protocol tcpip.TransportProtocolN // validly formed. n.stack.demux.deliverRawPacket(r, protocol, pkt) - if len(pkt.Data.First()) < transProto.MinimumPacketSize() { + transHeader, ok := pkt.Data.PullUp(transProto.MinimumPacketSize()) + if !ok { n.stack.stats.MalformedRcvdPackets.Increment() return } - srcPort, dstPort, err := transProto.ParsePorts(pkt.Data.First()) + srcPort, dstPort, err := transProto.ParsePorts(transHeader) if err != nil { n.stack.stats.MalformedRcvdPackets.Increment() return @@ -1288,7 +1351,7 @@ 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 @@ -1299,11 +1362,12 @@ func (n *NIC) DeliverTransportControlPacket(local, remote tcpip.Address, net tcp // ICMPv4 only guarantees that 8 bytes of the transport protocol will // be present in the payload. We know that the ports are within the // first 8 bytes for all known transport protocols. - if len(pkt.Data.First()) < 8 { + transHeader, ok := pkt.Data.PullUp(8) + if !ok { return } - srcPort, dstPort, err := transProto.ParsePorts(pkt.Data.First()) + srcPort, dstPort, err := transProto.ParsePorts(transHeader) if err != nil { return } @@ -1351,10 +1415,12 @@ 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() @@ -1368,7 +1434,17 @@ 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. @@ -1496,6 +1572,13 @@ 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))) } @@ -1523,6 +1606,13 @@ func (r *referencedNetworkEndpoint) isValidForOutgoingRLocked() bool { 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 // zero. func (r *referencedNetworkEndpoint) decRef() { @@ -1532,14 +1622,11 @@ 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..7d36f8e84 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. @@ -31,16 +36,14 @@ 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. + // down the stack, each layer adds to Header. Note that forwarded + // packets don't populate Headers on their way out -- their headers and + // payload are never parsed out and remain in Data. + // + // TODO(gvisor.dev/issue/170): Forwarded packets don't currently + // populate Header, but should. This will be doable once early parsing + // (https://github.com/google/gvisor/pull/1995) is supported. Header buffer.Prependable // These fields are used by both inbound and outbound packets. They @@ -55,6 +58,14 @@ 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 f9fd8f18f..23ca9ee03 100644 --- a/pkg/tcpip/stack/registration.go +++ b/pkg/tcpip/stack/registration.go @@ -67,12 +67,12 @@ 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) // Abort initiates an expedited endpoint teardown. It puts the endpoint // in a closed state and frees all resources associated with it. This @@ -100,7 +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 @@ -118,7 +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 @@ -150,7 +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 @@ -180,7 +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. @@ -189,7 +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. @@ -242,15 +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 @@ -265,7 +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() @@ -322,7 +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 @@ -354,7 +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 @@ -385,7 +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. @@ -393,7 +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. @@ -401,6 +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 @@ -423,7 +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 13354d884..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 @@ -620,6 +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() @@ -641,6 +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. @@ -733,7 +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 @@ -1696,7 +1721,7 @@ func (s *Stack) IsInGroup(nicID tcpip.NICID, multicastAddr tcpip.Address) (bool, } // 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() @@ -1704,7 +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() @@ -1814,6 +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 { @@ -1821,3 +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 e15db40fb..d45d2cc1f 100644 --- a/pkg/tcpip/stack/stack_test.go +++ b/pkg/tcpip/stack/stack_test.go @@ -90,21 +90,23 @@ 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)]++ // Consume the network header. - b := pkt.Data.First() + b, ok := pkt.Data.PullUp(fakeNetHeaderLen) + if !ok { + return + } pkt.Data.TrimFront(fakeNetHeaderLen) // Handle control packets. if b[2] == uint8(fakeControlProtocol) { - nb := pkt.Data.First() - if len(nb) < fakeNetHeaderLen { + nb, ok := pkt.Data.PullUp(fakeNetHeaderLen) + if !ok { return } - pkt.Data.TrimFront(fakeNetHeaderLen) f.dispatcher.DeliverTransportControlPacket(tcpip.Address(nb[1:2]), tcpip.Address(nb[0:1]), fakeNetNumber, tcpip.TransportProtocolNumber(nb[2]), stack.ControlPortUnreachable, 0, pkt) return @@ -126,7 +128,7 @@ 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)]++ @@ -141,7 +143,7 @@ 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), }) } @@ -153,11 +155,11 @@ 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 } @@ -255,7 +257,7 @@ type linkEPWithMockedAttach struct { // Attach implements stack.LinkEndpoint.Attach. func (l *linkEPWithMockedAttach) Attach(d stack.NetworkDispatcher) { l.LinkEndpoint.Attach(d) - l.attached = true + l.attached = d != nil } func (l *linkEPWithMockedAttach) isAttached() bool { @@ -287,7 +289,7 @@ 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 { @@ -299,7 +301,7 @@ 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 { @@ -311,7 +313,7 @@ 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 { @@ -322,7 +324,7 @@ 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 { @@ -334,7 +336,7 @@ 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 { @@ -356,7 +358,7 @@ 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(), }) @@ -414,7 +416,7 @@ 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 { @@ -566,7 +568,7 @@ func TestAttachToLinkEndpointImmediately(t *testing.T) { t.Fatalf("CreateNICWithOptions(%d, _, %+v) = %s", nicID, test.nicOpts, err) } if !e.isAttached() { - t.Fatalf("link endpoint not attached to a network disatcher") + t.Fatal("link endpoint not attached to a network dispatcher") } }) } @@ -631,196 +633,240 @@ func TestDisabledNICsNICInfoAndCheckNIC(t *testing.T) { checkNIC(false) } -func TestRoutesWithDisabledNIC(t *testing.T) { - const unspecifiedNIC = 0 - const nicID1 = 1 - const nicID2 = 2 - +func TestRemoveUnknownNIC(t *testing.T) { s := stack.New(stack.Options{ NetworkProtocols: []stack.NetworkProtocol{fakeNetFactory()}, }) - ep1 := channel.New(0, defaultMTU, "") - if err := s.CreateNIC(nicID1, ep1); err != nil { - t.Fatalf("CreateNIC(%d, _): %s", nicID1, err) + if err := s.RemoveNIC(1); err != tcpip.ErrUnknownNICID { + t.Fatalf("got s.RemoveNIC(1) = %v, want = %s", err, tcpip.ErrUnknownNICID) } +} - addr1 := tcpip.Address("\x01") - if err := s.AddAddress(nicID1, fakeNetNumber, addr1); err != nil { - t.Fatalf("AddAddress(%d, %d, %s): %s", nicID1, fakeNetNumber, addr1, err) - } +func TestRemoveNIC(t *testing.T) { + const nicID = 1 - ep2 := channel.New(0, defaultMTU, "") - if err := s.CreateNIC(nicID2, ep2); err != nil { - t.Fatalf("CreateNIC(%d, _): %s", nicID2, err) + 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) } - addr2 := tcpip.Address("\x02") - if err := s.AddAddress(nicID2, fakeNetNumber, addr2); err != nil { - t.Fatalf("AddAddress(%d, %d, %s): %s", nicID2, fakeNetNumber, addr2, 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") } - // 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}, - }) + // 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") + } +} - // 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) +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, + }, + } - // 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) - - // Disabling NIC1 should result in no routes to odd addresses. Routes to even - // addresses should continue to be available as NIC2 is still enabled. - if err := s.DisableNIC(nicID1); err != nil { - t.Fatalf("s.DisableNIC(%d): %s", nicID1, err) - } - nic1Dst := tcpip.Address("\x05") - testNoRoute(t, s, unspecifiedNIC, "", nic1Dst) - testNoRoute(t, s, unspecifiedNIC, addr1, nic1Dst) - testNoRoute(t, s, nicID1, addr1, nic1Dst) - nic2Dst := tcpip.Address("\x06") - testRoute(t, s, unspecifiedNIC, "", nic2Dst, addr2) - testRoute(t, s, unspecifiedNIC, addr2, nic2Dst, addr2) - testRoute(t, s, nicID2, addr2, nic2Dst, addr2) - - // Disabling NIC2 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 disabling NIC1 above. - if err := s.DisableNIC(nicID2); err != nil { - t.Fatalf("s.DisableNIC(%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) - - // Enabling NIC1 should make routes to odd addresses available again. Routes - // to even addresses should continue to be unavailable as NIC2 is still - // disabled. - if err := s.EnableNIC(nicID1); err != nil { - t.Fatalf("s.EnableNIC(%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) -} - -func TestRouteWritePacketWithDisabledNIC(t *testing.T) { 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()}, + }) - 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) + } - 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) + } - addr1 := tcpip.Address("\x01") - 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) + } - 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}, + }) + } - addr2 := tcpip.Address("\x02") - if err := s.AddAddress(nicID2, fakeNetNumber, addr2); err != nil { - t.Fatalf("AddAddress(%d, %d, %s): %s", nicID2, fakeNetNumber, addr2, err) + return s, ep1, ep2 } - // 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) + // 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) + } + }) } - s.SetRouteTable([]tcpip.Route{ - {Destination: subnet1, Gateway: "\x00", NIC: nicID1}, - {Destination: subnet0, Gateway: "\x00", NIC: nicID2}, - }) - } + }) - nic1Dst := tcpip.Address("\x05") - 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() + // 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) - nic2Dst := tcpip.Address("\x06") - 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() + 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() - // If we failed to get routes r1 or r2, we cannot proceed with the test. - if t.Failed() { - t.FailNow() - } + 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() - buf := buffer.View([]byte{1}) - testSend(t, r1, ep1, buf) - testSend(t, r2, ep2, buf) + // If we failed to get routes r1 or r2, we cannot proceed with the test. + if t.Failed() { + t.FailNow() + } - // Writes with Routes that use the disabled NIC1 should fail. - if err := s.DisableNIC(nicID1); err != nil { - t.Fatalf("s.DisableNIC(%d): %s", nicID1, err) - } - testFailingSend(t, r1, ep1, buf, tcpip.ErrInvalidEndpointState) - testSend(t, r2, ep2, buf) + buf := buffer.View([]byte{1}) + testSend(t, r1, ep1, buf) + testSend(t, r2, ep2, buf) - // Writes with Routes that use the disabled NIC2 should fail. - if err := s.DisableNIC(nicID2); err != nil { - t.Fatalf("s.DisableNIC(%d): %s", nicID2, err) - } - testFailingSend(t, r1, ep1, buf, tcpip.ErrInvalidEndpointState) - testFailingSend(t, r2, ep2, buf, tcpip.ErrInvalidEndpointState) + // 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 the re-enabled NIC1 should succeed. - // TODO(b/147015577): Should we instead completely invalidate all Routes that - // were bound to a disabled NIC at some point? - if err := s.EnableNIC(nicID1); err != nil { - t.Fatalf("s.EnableNIC(%d): %s", nicID1, err) - } - testSend(t, r1, ep1, buf) - testFailingSend(t, r2, ep2, buf, tcpip.ErrInvalidEndpointState) + // 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) { @@ -1401,19 +1447,19 @@ 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) } } @@ -1439,12 +1485,12 @@ 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. @@ -1459,10 +1505,10 @@ 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. @@ -2213,7 +2259,7 @@ 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 { @@ -2240,56 +2286,84 @@ 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) + } + }) } } @@ -2327,7 +2401,7 @@ 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) } }) } @@ -2696,7 +2770,7 @@ 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}}) } @@ -2710,11 +2784,11 @@ 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) } // @@ -2726,7 +2800,7 @@ 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) } @@ -2734,7 +2808,7 @@ 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 { @@ -2767,11 +2841,11 @@ 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 { @@ -3010,6 +3084,50 @@ func TestAddRemoveIPv4BroadcastAddressOnNICEnableDisable(t *testing.T) { } } +// 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 @@ -3060,8 +3178,6 @@ func TestJoinLeaveAllNodesMulticastOnNICEnableDisable(t *testing.T) { // 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 778c0a4d6..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{ @@ -279,29 +375,26 @@ 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()) + } + 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. } @@ -312,26 +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 } @@ -340,21 +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 { @@ -371,19 +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 @@ -396,84 +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 @@ -497,99 +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 @@ -601,8 +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 } @@ -616,13 +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 5d1da2f8b..a611e44ab 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" @@ -57,6 +56,8 @@ 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} } @@ -87,7 +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 { @@ -214,7 +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 { @@ -231,7 +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++ } @@ -242,8 +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) {} @@ -288,7 +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 } @@ -368,7 +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 { @@ -379,7 +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 { @@ -390,7 +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 { @@ -445,7 +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 { @@ -456,7 +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 { @@ -467,7 +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 { @@ -622,7 +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(), }) @@ -641,10 +642,11 @@ func TestTransportForwarding(t *testing.T) { t.Fatal("Response packet not forwarded") } - if dst := p.Pkt.Header.View()[0]; dst != 3 { + hdrs := p.Pkt.Data.ToView() + if dst := hdrs[0]; dst != 3 { t.Errorf("Response packet has incorrect destination addresss: got = %d, want = 3", dst) } - if src := p.Pkt.Header.View()[1]; src != 1 { + if src := hdrs[1]; src != 1 { t.Errorf("Response packet has incorrect source addresss: got = %d, want = 3", src) } } diff --git a/pkg/tcpip/tcpip.go b/pkg/tcpip/tcpip.go index 3dc5d87d6..1ca4088c9 100644 --- a/pkg/tcpip/tcpip.go +++ b/pkg/tcpip/tcpip.go @@ -336,6 +336,15 @@ type ControlMessages struct { 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) // that exposes functionality like read, write, connect, etc. to users of the // networking stack. @@ -470,6 +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. @@ -508,34 +520,90 @@ 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 SockOptBool = iota + ReceiveTClassOption // ReceiveTOSOption is used by SetSockOpt/GetSockOpt to specify if the TOS // ancillary message is passed with incoming packets. ReceiveTOSOption - // V6OnlyOption is used by {G,S}etSockOptBool to specify whether an IPv6 - // socket is to be restricted to sending and receiving IPv6 packets only. - V6OnlyOption - // ReceiveIPPacketInfoOption is used by {G,S}etSockOptBool to specify // if more inforamtion is provided with incoming packets such // as interface index and address. ReceiveIPPacketInfoOption - // TODO(b/146901447): convert existing bool socket options to be handled via - // Get/SetSockOptBool + // 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. + V6OnlyOption ) // SockOptInt represents socket options which values have the int type. 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. @@ -549,44 +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. @@ -595,10 +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. @@ -608,11 +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. @@ -626,20 +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. @@ -656,9 +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. @@ -667,10 +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 { @@ -693,22 +719,10 @@ 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 - -// IPv6TrafficClassOption is used by SetSockOpt/GetSockOpt to specify TOS -// for all subsequent outgoing IPv6 packets from the endpoint. -type IPv6TrafficClassOption uint8 - // IPPacketInfo is the message struture for IP_PKTINFO. // // +stateify savable 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/timer.go b/pkg/tcpip/timer.go index f5f01f32f..59f3b391f 100644 --- a/pkg/tcpip/timer.go +++ b/pkg/tcpip/timer.go @@ -88,6 +88,9 @@ func (t *cancellableTimerInstance) stop() { // // The term "related work" is defined as some work that needs to be done while // holding some lock that the timer must also hold while doing some work. +// +// Note, it is not safe to copy a CancellableTimer as its timer instance creates +// a closure over the address of the CancellableTimer. type CancellableTimer struct { // The active instance of a cancellable timer. instance cancellableTimerInstance @@ -131,10 +134,14 @@ 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 @@ -150,12 +157,28 @@ func (t *CancellableTimer) Reset(d time.Duration) { } } -// MakeCancellableTimer returns an unscheduled CancellableTimer with the given +// Lock is a no-op used by the copylocks checker from go vet. +// +// See CancellableTimer for details about why it shouldn't be copied. +// +// See https://github.com/golang/go/issues/8005#issuecomment-190753527 for more +// details about the copylocks checker. +func (*CancellableTimer) Lock() {} + +// Unlock is a no-op used by the copylocks checker from go vet. +// +// See CancellableTimer for details about why it shouldn't be copied. +// +// See https://github.com/golang/go/issues/8005#issuecomment-190753527 for more +// details about the copylocks checker. +func (*CancellableTimer) Unlock() {} + +// NewCancellableTimer returns an unscheduled CancellableTimer with the given // locker and fn. // // fn MUST NOT attempt to lock locker. // // Callers must call Reset to schedule the timer to fire. -func MakeCancellableTimer(locker sync.Locker, fn func()) CancellableTimer { - return CancellableTimer{locker: locker, fn: fn} +func NewCancellableTimer(locker sync.Locker, fn func()) *CancellableTimer { + return &CancellableTimer{locker: locker, fn: fn} } diff --git a/pkg/tcpip/timer_test.go b/pkg/tcpip/timer_test.go index 2d20f7ef3..b4940e397 100644 --- a/pkg/tcpip/timer_test.go +++ b/pkg/tcpip/timer_test.go @@ -28,13 +28,38 @@ 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.NewCancellableTimer(&lock, func() { + wg.Done() + }) + timer.Reset(shortDuration) + lock.Unlock() + }() + } + wg.Wait() +} + func TestCancellableTimerFire(t *testing.T) { t.Parallel() ch := make(chan struct{}) var lock sync.Mutex - timer := tcpip.MakeCancellableTimer(&lock, func() { + timer := tcpip.NewCancellableTimer(&lock, func() { ch <- struct{}{} }) timer.Reset(shortDuration) @@ -60,7 +85,7 @@ func TestCancellableTimerResetFromLongDuration(t *testing.T) { ch := make(chan struct{}) var lock sync.Mutex - timer := tcpip.MakeCancellableTimer(&lock, func() { ch <- struct{}{} }) + timer := tcpip.NewCancellableTimer(&lock, func() { ch <- struct{}{} }) timer.Reset(middleDuration) lock.Lock() @@ -91,7 +116,7 @@ func TestCancellableTimerResetFromShortDuration(t *testing.T) { var lock sync.Mutex lock.Lock() - timer := tcpip.MakeCancellableTimer(&lock, func() { ch <- struct{}{} }) + timer := tcpip.NewCancellableTimer(&lock, func() { ch <- struct{}{} }) timer.Reset(shortDuration) timer.StopLocked() lock.Unlock() @@ -128,7 +153,7 @@ func TestCancellableTimerImmediatelyStop(t *testing.T) { for i := 0; i < 1000; i++ { lock.Lock() - timer := tcpip.MakeCancellableTimer(&lock, func() { ch <- struct{}{} }) + timer := tcpip.NewCancellableTimer(&lock, func() { ch <- struct{}{} }) timer.Reset(shortDuration) timer.StopLocked() lock.Unlock() @@ -149,7 +174,7 @@ func TestCancellableTimerStoppedResetWithoutLock(t *testing.T) { var lock sync.Mutex lock.Lock() - timer := tcpip.MakeCancellableTimer(&lock, func() { ch <- struct{}{} }) + timer := tcpip.NewCancellableTimer(&lock, func() { ch <- struct{}{} }) timer.Reset(shortDuration) timer.StopLocked() lock.Unlock() @@ -180,7 +205,7 @@ func TestManyCancellableTimerResetAfterBlockedOnLock(t *testing.T) { var lock sync.Mutex lock.Lock() - timer := tcpip.MakeCancellableTimer(&lock, func() { ch <- struct{}{} }) + timer := tcpip.NewCancellableTimer(&lock, func() { ch <- struct{}{} }) timer.Reset(shortDuration) for i := 0; i < 10; i++ { // Sleep until the timer fires and gets blocked trying to take the lock. @@ -212,7 +237,7 @@ func TestManyCancellableTimerResetUnderLock(t *testing.T) { var lock sync.Mutex lock.Lock() - timer := tcpip.MakeCancellableTimer(&lock, func() { ch <- struct{}{} }) + timer := tcpip.NewCancellableTimer(&lock, func() { ch <- struct{}{} }) timer.Reset(shortDuration) for i := 0; i < 10; i++ { timer.StopLocked() 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 2a396e9bc..b1d820372 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) { @@ -134,8 +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 } @@ -322,7 +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) @@ -342,13 +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 } @@ -359,12 +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. @@ -391,32 +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 } @@ -441,10 +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, }) } @@ -471,7 +481,7 @@ 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), @@ -733,19 +743,19 @@ 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: - h := header.ICMPv4(pkt.Data.First()) - if h.Type() != header.ICMPv4EchoReply { + h, ok := pkt.Data.PullUp(header.ICMPv4MinimumSize) + if !ok || header.ICMPv4(h).Type() != header.ICMPv4EchoReply { e.stack.Stats().DroppedPackets.Increment() e.stats.ReceiveErrors.MalformedPacketsReceived.Increment() return } case header.IPv6ProtocolNumber: - h := header.ICMPv6(pkt.Data.First()) - if h.Type() != header.ICMPv6EchoReply { + h, ok := pkt.Data.PullUp(header.ICMPv6MinimumSize) + if !ok || header.ICMPv6(h).Type() != header.ICMPv6EchoReply { e.stack.Stats().DroppedPackets.Increment() e.stats.ReceiveErrors.MalformedPacketsReceived.Increment() return @@ -795,7 +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 113d92901..3c47692b2 100644 --- a/pkg/tcpip/transport/icmp/protocol.go +++ b/pkg/tcpip/transport/icmp/protocol.go @@ -104,7 +104,7 @@ 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 (*protocol) HandleUnknownDestinationPacket(*stack.Route, stack.TransportEndpointID, tcpip.PacketBuffer) bool { +func (*protocol) HandleUnknownDestinationPacket(*stack.Route, stack.TransportEndpointID, stack.PacketBuffer) bool { return true } 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 09a1cd436..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" ) @@ -100,8 +99,8 @@ func NewEndpoint(s *stack.Stack, cooked bool, netProto tcpip.NetworkProtocolNumb } // Abort implements stack.TransportEndpoint.Abort. -func (e *endpoint) Abort() { - e.Close() +func (ep *endpoint) Abort() { + ep.Close() } // Close implements tcpip.Endpoint.Close. @@ -134,7 +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 } @@ -299,7 +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. @@ -393,3 +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 2ef5fac76..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. @@ -160,8 +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 } @@ -342,17 +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 } @@ -525,14 +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 } @@ -540,7 +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. @@ -568,13 +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 a32f9eacf..f2aa69069 100644 --- a/pkg/tcpip/transport/tcp/BUILD +++ b/pkg/tcpip/transport/tcp/BUILD @@ -66,12 +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", ], @@ -89,7 +87,9 @@ go_test( "tcp_timestamp_test.go", ], # FIXME(b/68809571) - tags = ["flaky"], + tags = [ + "flaky", + ], deps = [ ":tcp", "//pkg/sync", @@ -105,6 +105,17 @@ go_test( "//pkg/tcpip/seqnum", "//pkg/tcpip/stack", "//pkg/tcpip/transport/tcp/testing/context", + "//pkg/test/testutil", "//pkg/waiter", ], ) + +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 85049e54e..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 @@ -229,7 +207,7 @@ 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() @@ -243,21 +221,6 @@ func (l *listenContext) createConnectingEndpoint(s *segment, iss seqnum.Value, i 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, @@ -269,12 +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()) @@ -282,6 +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) @@ -289,16 +269,33 @@ 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 @@ -310,11 +307,12 @@ func (l *listenContext) createEndpointAndPerformHandshake(s *segment, opts *head 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 @@ -348,30 +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 @@ -381,11 +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() @@ -399,40 +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 } @@ -442,7 +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 @@ -452,7 +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() @@ -480,7 +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() } @@ -497,7 +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 @@ -513,7 +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 } @@ -559,6 +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 @@ -592,15 +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. @@ -614,6 +633,8 @@ 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 | waiter.EventHUp | waiter.EventErr) }() @@ -622,7 +643,10 @@ func (e *endpoint) protocolListenLoop(rcvWnd seqnum.Size) *tcpip.Error { 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 { @@ -635,7 +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 c0f73ef16..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 @@ -293,11 +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, @@ -313,7 +295,15 @@ func (h *handshake) synSentState(s *segment) *tcpip.Error { 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 } @@ -357,10 +347,6 @@ func (h *handshake) synRcvdState(s *segment) *tcpip.Error { return tcpip.ErrInvalidEndpointState } - h.ep.mu.RLock() - amss := h.ep.amss - h.ep.mu.RUnlock() - h.resetState() synOpts := header.TCPSynOptions{ WS: h.rcvWndScale, @@ -368,9 +354,17 @@ func (h *handshake) synRcvdState(s *segment) *tcpip.Error { TSVal: h.ep.timestamp(), TSEcr: h.ep.recentTimestamp(), SACKPermitted: h.ep.sackPermitted, - MSS: amss, + 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 } @@ -399,15 +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 } @@ -493,7 +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() } } @@ -535,7 +530,6 @@ func (h *handshake) execute() *tcpip.Error { // Send the initial SYN segment and loop until the handshake is // completed. - h.ep.mu.Lock() h.ep.amss = calculateAdvertisedMSS(h.ep.userMSS, h.ep.route) synOpts := header.TCPSynOptions{ @@ -546,7 +540,6 @@ func (h *handshake) execute() *tcpip.Error { SACKPermitted: bool(sackEnabled), MSS: h.ep.amss, } - h.ep.mu.Unlock() // Execute is also called in a listen context so we want to make sure we // only send the TS/SACK option when we received the TS/SACK in the @@ -560,10 +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 { @@ -579,7 +585,15 @@ 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: @@ -600,7 +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: @@ -690,18 +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 } @@ -709,24 +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) @@ -738,48 +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)) } @@ -789,33 +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 @@ -867,7 +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 } @@ -882,7 +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 } @@ -1001,22 +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 @@ -1040,11 +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 @@ -1157,9 +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. @@ -1182,9 +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() { @@ -1248,6 +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 @@ -1269,7 +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) } @@ -1280,16 +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 @@ -1302,9 +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 @@ -1330,10 +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 }, }, @@ -1388,7 +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. @@ -1397,7 +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 { @@ -1417,7 +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() } } @@ -1460,7 +1507,6 @@ func (e *endpoint) protocolMainLoop(handshake bool, wakerInitDone chan<- struct{ } e.rcvListMu.Unlock() - e.mu.Lock() if e.workerCleanup { e.notifyProtocolGoroutine(notifyClose) } @@ -1468,7 +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) @@ -1480,16 +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 @@ -1502,21 +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() @@ -1526,14 +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() } @@ -1541,19 +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 { @@ -1639,6 +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) @@ -1648,9 +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() @@ -1673,7 +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/dispatcher.go b/pkg/tcpip/transport/tcp/dispatcher.go index d792b07d6..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" @@ -128,7 +127,7 @@ func (p *processor) handleSegments() { continue } - if !ep.workMu.TryLock() { + if !ep.mu.TryLock() { ep.newSegmentWaker.Assert() continue } @@ -138,12 +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 } @@ -151,7 +148,7 @@ func (p *processor) handleSegments() { p.epQ.enqueue(ep) } - ep.workMu.Unlock() + ep.mu.Unlock() } } } @@ -189,7 +186,7 @@ func (d *dispatcher) wait() { } } -func (d *dispatcher) queuePacket(r *stack.Route, stackEP stack.TransportEndpoint, id stack.TransportEndpointID, pkt tcpip.PacketBuffer) { +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 dc9c18b6f..07d3e64c8 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" ) @@ -283,6 +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 @@ -299,12 +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"` @@ -330,15 +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)"` @@ -513,6 +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. @@ -561,6 +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. @@ -583,14 +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 @@ -672,6 +796,7 @@ func newEndpoint(s *stack.Stack, netProto tcpip.NetworkProtocolNumber, waiterQue count: 9, }, uniqueID: s.UniqueID(), + txHash: s.Rand().Uint32(), } var ss SendBufferSizeOption @@ -696,7 +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 @@ -709,9 +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 } @@ -721,9 +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. @@ -735,9 +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() { @@ -823,25 +946,22 @@ func (e *endpoint) Abort() { // 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 @@ -860,66 +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) - switch e.EndpointState() { - // Sockets in StateSynRecv state(passive connections) are closed when - // the handshake fails or if the listening socket is closed while - // handshake was in progress. In such cases the handshake goroutine - // is already gone by the time Close is called and we need to cleanup - // here. - case StateInitial, StateBound, StateSynRecv: - e.cleanupLocked() - e.setEndpointState(StateClose) - case StateError, StateClose: - // do nothing. - default: + 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 @@ -953,24 +1062,37 @@ 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.mu.RLock() + e.LockUser() + defer e.UnlockUser() + e.rcvListMu.Lock() if e.rcvAutoParams.disabled { e.rcvListMu.Unlock() - e.mu.RUnlock() return } now := time.Now() if rtt := e.rcvAutoParams.rtt; rtt == 0 || now.Sub(e.rcvAutoParams.measureTime) < rtt { e.rcvAutoParams.copied += copied e.rcvListMu.Unlock() - e.mu.RUnlock() return } prevRTTCopied := e.rcvAutoParams.copied + copied @@ -1026,17 +1148,20 @@ func (e *endpoint) ModerateRecvBuf(copied int) { e.rcvAutoParams.measureTime = now e.rcvAutoParams.copied = 0 e.rcvListMu.Unlock() - e.mu.RUnlock() +} + +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 @@ -1046,7 +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 } @@ -1056,7 +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() @@ -1129,13 +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 } @@ -1147,113 +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. @@ -1345,21 +1425,71 @@ func (e *endpoint) windowCrossedACKThresholdLocked(deltaBefore int) (crossed boo // 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 @@ -1367,24 +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.mu.RLock() + e.LockUser() e.rcvListMu.Lock() // Make sure the receive buffer size allows us to send a @@ -1393,17 +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 @@ -1414,148 +1576,68 @@ func (e *endpoint) SetSockOptInt(opt tcpip.SockOptInt, v int) *tcpip.Error { if crossed, above := e.windowCrossedACKThresholdLocked(availAfter - availBefore); crossed && above { mask |= notifyNonZeroReceiveWindow } + e.rcvListMu.Unlock() - e.mu.RUnlock() + 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 @@ -1568,22 +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 } } @@ -1592,24 +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 @@ -1627,27 +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 { @@ -1663,25 +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() @@ -1697,12 +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 @@ -1719,159 +1841,61 @@ 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 } // checkV4MappedLocked determines the effective network protocol and converts @@ -1906,8 +1930,8 @@ 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 @@ -2076,13 +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 @@ -2091,72 +2119,56 @@ 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() if e.EndpointState() == StateTimeWait { - e.mu.Unlock() return tcpip.ErrNotConnected } - break + 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. + 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 @@ -2171,28 +2183,40 @@ func (e *endpoint) Listen(backlog int) *tcpip.Error { } func (e *endpoint) listen(backlog int) *tcpip.Error { - e.mu.Lock() - defer e.mu.Unlock() - - // 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 - // that point onward, acceptedChan is the responsibility of the cleanup() - // method (and should not be touched anywhere else, including here). - if e.EndpointState() == StateListen && !e.workerCleanup { - // Adjust the size of the channel iff we can fix existing - // pending connections into the new one. - if len(e.acceptedChan) > backlog { - return tcpip.ErrInvalidEndpointState - } - if cap(e.acceptedChan) == backlog { - return nil - } - origChan := e.acceptedChan - e.acceptedChan = make(chan *endpoint, backlog) - close(origChan) - for ep := range origChan { - e.acceptedChan <- ep + e.LockUser() + defer e.UnlockUser() + + if e.EndpointState() == StateListen && !e.closed { + e.acceptMu.Lock() + defer e.acceptMu.Unlock() + if e.acceptedChan == nil { + // listen is called after shutdown. + e.acceptedChan = make(chan *endpoint, backlog) + e.shutdownFlags = 0 + e.rcvListMu.Lock() + e.rcvClosed = false + e.rcvListMu.Unlock() + } else { + // Adjust the size of the channel iff we can fix + // existing pending connections into the new one. + if len(e.acceptedChan) > backlog { + return tcpip.ErrInvalidEndpointState + } + if cap(e.acceptedChan) == backlog { + return nil + } + origChan := e.acceptedChan + e.acceptedChan = make(chan *endpoint, backlog) + close(origChan) + for ep := range origChan { + e.acceptedChan <- ep + } } + + // Notify any blocked goroutines that they can attempt to + // deliver endpoints again. + e.acceptCond.Broadcast() + return nil } @@ -2222,9 +2246,12 @@ 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())) @@ -2234,7 +2261,6 @@ 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{}) @@ -2245,18 +2271,24 @@ 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 } @@ -2265,8 +2297,8 @@ 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) } @@ -2344,8 +2376,8 @@ 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, @@ -2356,8 +2388,8 @@ 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 @@ -2370,7 +2402,7 @@ 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 @@ -2389,7 +2421,7 @@ 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() @@ -2424,7 +2456,6 @@ func (e *endpoint) updateSndBufferUsage(v int) { // to be read, or when the connection is closed for receiving (in which case // s will be nil). func (e *endpoint) readyToRead(s *segment) { - e.mu.RLock() e.rcvListMu.Lock() if s != nil { s.incRef() @@ -2439,7 +2470,6 @@ func (e *endpoint) readyToRead(s *segment) { e.rcvClosed = true } e.rcvListMu.Unlock() - e.mu.RUnlock() e.waiterQueue.Notify(waiter.EventIn) } @@ -2583,9 +2613,7 @@ 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() @@ -2715,10 +2743,10 @@ 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 } @@ -2733,9 +2761,9 @@ 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..8b7562396 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: @@ -245,6 +247,11 @@ func (e *endpoint) Resume(s *stack.Stack) { if err := e.Listen(backlog); err != nil { panic("endpoint listening failed: " + err.String()) } + e.LockUser() + if e.shutdownFlags != 0 { + e.shutdownLocked(e.shutdownFlags) + } + e.UnlockUser() listenLoading.Done() tcpip.AsyncLoading.Done() }() 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 73098d904..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) @@ -191,7 +250,15 @@ 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 stack.TransportProtocol.SetOption. @@ -264,6 +331,21 @@ 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 } @@ -273,57 +355,69 @@ func (p *protocol) SetOption(option interface{}) *tcpip.Error { 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: @@ -341,6 +435,12 @@ 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{ @@ -350,6 +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 d80aff1b6..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 @@ -208,7 +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) @@ -222,7 +230,6 @@ func (r *receiver) consumeSegment(s *segment, segSeq seqnum.Value, segLen seqnum case StateLastAck: r.ep.transitionToStateCloseLocked() } - r.ep.mu.Unlock() } return true @@ -336,10 +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..7712ce652 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 } @@ -143,7 +144,11 @@ func (s *segment) logicalLen() seqnum.Size { // TCP checksum and stores the checksum and result of checksum verification in // the csum and csumValid fields of the segment. func (s *segment) parse() bool { - h := header.TCP(s.data.First()) + h, ok := s.data.PullUp(header.TCPMinimumSize) + if !ok { + return false + } + hdr := header.TCP(h) // h is the header followed by the payload. We check that the offset to // the data respects the following constraints: @@ -155,12 +160,16 @@ func (s *segment) parse() bool { // N.B. The segment has already been validated as having at least the // minimum TCP size before reaching here, so it's safe to read the // fields. - offset := int(h.DataOffset()) - if offset < header.TCPMinimumSize || offset > len(h) { + offset := int(hdr.DataOffset()) + if offset < header.TCPMinimumSize { + return false + } + hdrWithOpts, ok := s.data.PullUp(offset) + if !ok { return false } - s.options = []byte(h[header.TCPMinimumSize:offset]) + s.options = []byte(hdrWithOpts[header.TCPMinimumSize:]) s.parsedOptions = header.ParseTCPOptions(s.options) // Query the link capabilities to decide if checksum validation is @@ -172,18 +181,19 @@ func (s *segment) parse() bool { s.data.TrimFront(offset) } if verifyChecksum { - s.csum = h.Checksum() + hdr = header.TCP(hdrWithOpts) + s.csum = hdr.Checksum() xsum := s.route.PseudoHeaderChecksum(ProtocolNumber, uint16(s.data.Size())) - xsum = h.CalculateChecksum(xsum) + xsum = hdr.CalculateChecksum(xsum) s.data.TrimFront(offset) xsum = header.ChecksumVV(s.data, xsum) s.csumValid = xsum == 0xffff } - s.sequenceNumber = seqnum.Value(h.SequenceNumber()) - s.ackNumber = seqnum.Value(h.AckNumber()) - s.flags = h.Flags() - s.window = seqnum.Size(h.WindowSize()) + s.sequenceNumber = seqnum.Value(hdr.SequenceNumber()) + s.ackNumber = seqnum.Value(hdr.AckNumber()) + s.flags = hdr.Flags() + s.window = seqnum.Size(hdr.WindowSize()) return true } diff --git a/pkg/tcpip/transport/tcp/segment_heap.go b/pkg/tcpip/transport/tcp/segment_heap.go index e28f213ba..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. 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..5fe23113b 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/pkg/test/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 5b2b16afa..286c66cf5 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() @@ -590,6 +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) } @@ -728,7 +732,7 @@ func TestUserSuppliedMSSOnConnectV4(t *testing.T) { const maxMSS = mtu - header.IPv4MinimumSize - header.TCPMinimumSize tests := []struct { name string - setMSS uint16 + setMSS int expMSS uint16 }{ { @@ -756,15 +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)) @@ -818,15 +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)) @@ -1032,8 +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() @@ -1044,7 +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) } @@ -1052,9 +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, @@ -1063,7 +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) { @@ -1077,17 +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)) @@ -1125,17 +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. @@ -1711,7 +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) @@ -1984,7 +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 { @@ -2057,7 +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 { @@ -2221,10 +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) }, }, } @@ -2236,9 +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...) @@ -2248,8 +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() @@ -2257,7 +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)), ), @@ -2273,7 +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, }) }) @@ -2286,7 +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}} { @@ -2334,7 +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 { @@ -2367,7 +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() @@ -2404,8 +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 { @@ -2546,12 +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. @@ -2591,7 +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 { @@ -2637,26 +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. @@ -2674,7 +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): @@ -2735,7 +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. @@ -3479,7 +3548,7 @@ func TestReceivedInvalidSegmentCountIncrement(t *testing.T) { AckNum: c.IRS.Add(1), RcvWnd: 30000, }) - tcpbuf := vv.First()[header.IPv4MinimumSize:] + tcpbuf := vv.ToView()[header.IPv4MinimumSize:] tcpbuf[header.TCPDataOffset] = ((header.TCPMinimumSize - 1) / 4) << 4 c.SendSegment(vv) @@ -3506,7 +3575,7 @@ func TestReceivedIncorrectChecksumIncrement(t *testing.T) { AckNum: c.IRS.Add(1), RcvWnd: 30000, }) - tcpbuf := vv.First()[header.IPv4MinimumSize:] + tcpbuf := vv.ToView()[header.IPv4MinimumSize:] // Overwrite a byte in the payload which should cause checksum // verification to fail. tcpbuf[(tcpbuf[header.TCPDataOffset]>>4)*4] = 0x4 @@ -3852,26 +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) @@ -4117,11 +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) @@ -4444,11 +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. @@ -4539,7 +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 @@ -5074,25 +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. @@ -5100,7 +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) @@ -5579,7 +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) } }, )) @@ -5740,14 +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() @@ -5758,7 +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 @@ -5773,12 +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) } } @@ -6587,14 +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. @@ -6602,28 +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 8cea20fb5..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, "") @@ -217,7 +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) } @@ -235,7 +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") @@ -307,7 +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(), }) } @@ -363,7 +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, }) } @@ -371,7 +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), }) } @@ -380,7 +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), }) } @@ -415,6 +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) } @@ -423,6 +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), @@ -445,6 +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 @@ -486,7 +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") @@ -548,7 +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(), }) } @@ -567,6 +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 0af4514e1..756ab913a 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" @@ -144,6 +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 @@ -234,7 +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 } @@ -485,7 +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 @@ -499,11 +501,20 @@ 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. @@ -514,7 +525,18 @@ func (e *endpoint) SetSockOptBool(opt tcpip.SockOptBool, v bool) *tcpip.Error { e.mu.Lock() e.receiveTClass = v e.mu.Unlock() - return nil + + 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. @@ -531,13 +553,6 @@ func (e *endpoint) SetSockOptBool(opt tcpip.SockOptBool, v bool) *tcpip.Error { } e.v6only = v - return nil - - case tcpip.ReceiveIPPacketInfoOption: - e.mu.Lock() - e.receiveIPPacketInfo = v - e.mu.Unlock() - return nil } return nil @@ -545,22 +560,38 @@ 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() @@ -684,16 +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) { @@ -702,26 +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 } @@ -729,6 +730,21 @@ 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 @@ -746,6 +762,22 @@ func (e *endpoint) GetSockOptBool(opt tcpip.SockOptBool) (bool, *tcpip.Error) { 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 { @@ -758,19 +790,32 @@ func (e *endpoint) GetSockOptBool(opt tcpip.SockOptBool) (bool, *tcpip.Error) { return v, nil - case tcpip.ReceiveIPPacketInfoOption: - e.mu.RLock() - v := e.receiveIPPacketInfo - e.mu.RUnlock() - return v, nil + default: + return false, tcpip.ErrUnknownProtocolOption } - - 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() @@ -792,29 +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{ @@ -822,72 +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())) @@ -913,10 +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 @@ -1260,10 +1248,10 @@ 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() { + hdr, ok := pkt.Data.PullUp(header.UDPMinimumSize) + if !ok || int(header.UDP(hdr).Length()) > pkt.Data.Size() { // Malformed packet. e.stack.Stats().UDP.MalformedPacketsReceived.Increment() e.stats.ReceiveErrors.MalformedPacketsReceived.Increment() @@ -1298,7 +1286,7 @@ func (e *endpoint) HandlePacket(r *stack.Route, id stack.TransportEndpointID, pk senderAddress: tcpip.FullAddress{ NIC: r.NICID(), Addr: id.RemoteAddress, - Port: hdr.SourcePort(), + Port: header.UDP(hdr).SourcePort(), }, } packet.data = pkt.Data @@ -1327,7 +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. @@ -1357,3 +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/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 8df089d22..52af6de22 100644 --- a/pkg/tcpip/transport/udp/protocol.go +++ b/pkg/tcpip/transport/udp/protocol.go @@ -66,10 +66,15 @@ 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() { + h, ok := pkt.Data.PullUp(header.UDPMinimumSize) + if !ok { + // Malformed packet. + r.Stack().Stats().UDP.MalformedPacketsReceived.Increment() + return true + } + if int(header.UDP(h).Length()) > pkt.Data.Size() { // Malformed packet. r.Stack().Stats().UDP.MalformedPacketsReceived.Increment() return true @@ -135,7 +140,7 @@ 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 +177,7 @@ 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, }) diff --git a/pkg/tcpip/transport/udp/udp_test.go b/pkg/tcpip/transport/udp/udp_test.go index 34b7c2360..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") @@ -439,7 +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), @@ -486,7 +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), @@ -607,7 +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. @@ -1271,8 +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 @@ -1311,8 +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 @@ -1346,25 +1347,26 @@ func TestSetTOS(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(%T) failed: %s", v, 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(%T) = 0x%x, want = 0x%x", 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(%T, 0x%x) failed: %s", v, 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(%T) failed: %s", v, 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(%T) = 0x%x, want = 0x%x", 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)) @@ -1381,25 +1383,26 @@ func TestSetTClass(t *testing.T) { c.createEndpointForFlow(flow) const tClass = testTOS - var v tcpip.IPv6TrafficClassOption - if err := c.ep.GetSockOpt(&v); err != nil { - c.t.Errorf("GetSockopt(%T) failed: %s", v, err) + 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(%T) = 0x%x, want = 0x%x", v, v, 0) + c.t.Errorf("got GetSockOptInt(IPv6TrafficClassOption) = 0x%x, want = 0x%x", v, 0) } - if err := c.ep.SetSockOpt(tcpip.IPv6TrafficClassOption(tClass)); err != nil { - c.t.Errorf("SetSockOpt(%T, 0x%x) failed: %s", v, tcpip.IPv6TrafficClassOption(tClass), 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(%T) failed: %s", v, err) + v, err = c.ep.GetSockOptInt(tcpip.IPv6TrafficClassOption) + if err != nil { + c.t.Errorf("GetSockOptInt(IPv6TrafficClassOption) failed: %s", err) } - if want := tcpip.IPv6TrafficClassOption(tClass); v != want { - c.t.Errorf("got GetSockOpt(%T) = 0x%x, want = 0x%x", v, v, want) + if v != tClass { + c.t.Errorf("got GetSockOptInt(IPv6TrafficClassOption) = 0x%x, want = 0x%x", v, tClass) } // The header getter for TClass is called TOS, so use that checker. @@ -1430,7 +1433,7 @@ func TestReceiveTosTClass(t *testing.T) { // 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) + c.t.Errorf("GetSockOptBool(%s) failed: %s", name, err) } // Test for expected default value. if v != false { @@ -1444,7 +1447,7 @@ func TestReceiveTosTClass(t *testing.T) { got, err := c.ep.GetSockOptBool(option) if err != nil { - c.t.Errorf("GetSockoptBool(%s) failed: %s", name, err) + c.t.Errorf("GetSockOptBool(%s) failed: %s", name, err) } if got != want { @@ -1563,7 +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) } @@ -1571,7 +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") @@ -1639,7 +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) } @@ -1647,7 +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/test/criutil/BUILD b/pkg/test/criutil/BUILD new file mode 100644 index 000000000..a7b082cee --- /dev/null +++ b/pkg/test/criutil/BUILD @@ -0,0 +1,14 @@ +load("//tools:defs.bzl", "go_library") + +package(licenses = ["notice"]) + +go_library( + name = "criutil", + testonly = 1, + srcs = ["criutil.go"], + visibility = ["//:sandbox"], + deps = [ + "//pkg/test/dockerutil", + "//pkg/test/testutil", + ], +) diff --git a/pkg/test/criutil/criutil.go b/pkg/test/criutil/criutil.go new file mode 100644 index 000000000..bebebb48e --- /dev/null +++ b/pkg/test/criutil/criutil.go @@ -0,0 +1,306 @@ +// 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 criutil contains utility functions for interacting with the +// Container Runtime Interface (CRI), principally via the crictl command line +// tool. This requires critools to be installed on the local system. +package criutil + +import ( + "encoding/json" + "fmt" + "os" + "os/exec" + "strings" + "time" + + "gvisor.dev/gvisor/pkg/test/dockerutil" + "gvisor.dev/gvisor/pkg/test/testutil" +) + +// Crictl contains information required to run the crictl utility. +type Crictl struct { + logger testutil.Logger + endpoint string + cleanup []func() +} + +// resolvePath attempts to find binary paths. It may set the path to invalid, +// which will cause the execution to fail with a sensible error. +func resolvePath(executable string) string { + guess, err := exec.LookPath(executable) + if err != nil { + guess = fmt.Sprintf("/usr/local/bin/%s", executable) + } + return guess +} + +// NewCrictl returns a Crictl configured with a timeout and an endpoint over +// which it will talk to containerd. +func NewCrictl(logger testutil.Logger, endpoint string) *Crictl { + // Attempt to find the executable, but don't bother propagating the + // error at this point. The first command executed will return with a + // binary not found error. + return &Crictl{ + logger: logger, + endpoint: endpoint, + } +} + +// CleanUp executes cleanup functions. +func (cc *Crictl) CleanUp() { + for _, c := range cc.cleanup { + c() + } + cc.cleanup = nil +} + +// RunPod creates a sandbox. It corresponds to `crictl runp`. +func (cc *Crictl) RunPod(sbSpecFile string) (string, error) { + podID, err := cc.run("runp", sbSpecFile) + if err != nil { + return "", fmt.Errorf("runp failed: %v", err) + } + // Strip the trailing newline from crictl output. + return strings.TrimSpace(podID), nil +} + +// Create creates a container within a sandbox. It corresponds to `crictl +// create`. +func (cc *Crictl) Create(podID, contSpecFile, sbSpecFile string) (string, error) { + podID, err := cc.run("create", podID, contSpecFile, sbSpecFile) + if err != nil { + return "", fmt.Errorf("create failed: %v", err) + } + // Strip the trailing newline from crictl output. + return strings.TrimSpace(podID), nil +} + +// Start starts a container. It corresponds to `crictl start`. +func (cc *Crictl) Start(contID string) (string, error) { + output, err := cc.run("start", contID) + if err != nil { + return "", fmt.Errorf("start failed: %v", err) + } + return output, nil +} + +// Stop stops a container. It corresponds to `crictl stop`. +func (cc *Crictl) Stop(contID string) error { + _, err := cc.run("stop", contID) + return err +} + +// Exec execs a program inside a container. It corresponds to `crictl exec`. +func (cc *Crictl) Exec(contID string, args ...string) (string, error) { + a := []string{"exec", contID} + a = append(a, args...) + output, err := cc.run(a...) + if err != nil { + return "", fmt.Errorf("exec failed: %v", err) + } + return output, nil +} + +// Rm removes a container. It corresponds to `crictl rm`. +func (cc *Crictl) Rm(contID string) error { + _, err := cc.run("rm", contID) + return err +} + +// StopPod stops a pod. It corresponds to `crictl stopp`. +func (cc *Crictl) StopPod(podID string) error { + _, err := cc.run("stopp", podID) + return err +} + +// containsConfig is a minimal copy of +// https://github.com/kubernetes/kubernetes/blob/master/pkg/kubelet/apis/cri/runtime/v1alpha2/api.proto +// It only contains fields needed for testing. +type containerConfig struct { + Status containerStatus +} + +type containerStatus struct { + Network containerNetwork +} + +type containerNetwork struct { + IP string +} + +// PodIP returns a pod's IP address. +func (cc *Crictl) PodIP(podID string) (string, error) { + output, err := cc.run("inspectp", podID) + if err != nil { + return "", err + } + conf := &containerConfig{} + if err := json.Unmarshal([]byte(output), conf); err != nil { + return "", fmt.Errorf("failed to unmarshal JSON: %v, %s", err, output) + } + if conf.Status.Network.IP == "" { + return "", fmt.Errorf("no IP found in config: %s", output) + } + return conf.Status.Network.IP, nil +} + +// RmPod removes a container. It corresponds to `crictl rmp`. +func (cc *Crictl) RmPod(podID string) error { + _, err := cc.run("rmp", podID) + return err +} + +// Import imports the given container from the local Docker instance. +func (cc *Crictl) Import(image string) error { + // Note that we provide a 10 minute timeout after connect because we may + // be pushing a lot of bytes in order to import the image. The connect + // timeout stays the same and is inherited from the Crictl instance. + cmd := testutil.Command(cc.logger, + resolvePath("ctr"), + fmt.Sprintf("--connect-timeout=%s", 30*time.Second), + fmt.Sprintf("--address=%s", cc.endpoint), + "-n", "k8s.io", "images", "import", "-") + cmd.Stderr = os.Stderr // Pass through errors. + + // Create a pipe and start the program. + w, err := cmd.StdinPipe() + if err != nil { + return err + } + if err := cmd.Start(); err != nil { + return err + } + + // Save the image on the other end. + if err := dockerutil.Save(cc.logger, image, w); err != nil { + cmd.Wait() + return err + } + + // Close our pipe reference & see if it was loaded. + if err := w.Close(); err != nil { + return w.Close() + } + + return cmd.Wait() +} + +// StartContainer pulls the given image ands starts the container in the +// sandbox with the given podID. +// +// Note that the image will always be imported from the local docker daemon. +func (cc *Crictl) StartContainer(podID, image, sbSpec, contSpec string) (string, error) { + if err := cc.Import(image); err != nil { + return "", err + } + + // Write the specs to files that can be read by crictl. + sbSpecFile, cleanup, err := testutil.WriteTmpFile("sbSpec", sbSpec) + if err != nil { + return "", fmt.Errorf("failed to write sandbox spec: %v", err) + } + cc.cleanup = append(cc.cleanup, cleanup) + contSpecFile, cleanup, err := testutil.WriteTmpFile("contSpec", contSpec) + if err != nil { + return "", fmt.Errorf("failed to write container spec: %v", err) + } + cc.cleanup = append(cc.cleanup, cleanup) + + return cc.startContainer(podID, image, sbSpecFile, contSpecFile) +} + +func (cc *Crictl) startContainer(podID, image, sbSpecFile, contSpecFile string) (string, error) { + contID, err := cc.Create(podID, contSpecFile, sbSpecFile) + if err != nil { + return "", fmt.Errorf("failed to create container in pod %q: %v", podID, err) + } + + if _, err := cc.Start(contID); err != nil { + return "", fmt.Errorf("failed to start container %q in pod %q: %v", contID, podID, err) + } + + return contID, nil +} + +// StopContainer stops and deletes the container with the given container ID. +func (cc *Crictl) StopContainer(contID string) error { + if err := cc.Stop(contID); err != nil { + return fmt.Errorf("failed to stop container %q: %v", contID, err) + } + + if err := cc.Rm(contID); err != nil { + return fmt.Errorf("failed to remove container %q: %v", contID, err) + } + + return nil +} + +// StartPodAndContainer starts a sandbox and container in that sandbox. It +// returns the pod ID and container ID. +func (cc *Crictl) StartPodAndContainer(image, sbSpec, contSpec string) (string, string, error) { + if err := cc.Import(image); err != nil { + return "", "", err + } + + // Write the specs to files that can be read by crictl. + sbSpecFile, cleanup, err := testutil.WriteTmpFile("sbSpec", sbSpec) + if err != nil { + return "", "", fmt.Errorf("failed to write sandbox spec: %v", err) + } + cc.cleanup = append(cc.cleanup, cleanup) + contSpecFile, cleanup, err := testutil.WriteTmpFile("contSpec", contSpec) + if err != nil { + return "", "", fmt.Errorf("failed to write container spec: %v", err) + } + cc.cleanup = append(cc.cleanup, cleanup) + + podID, err := cc.RunPod(sbSpecFile) + if err != nil { + return "", "", err + } + + contID, err := cc.startContainer(podID, image, sbSpecFile, contSpecFile) + + return podID, contID, err +} + +// StopPodAndContainer stops a container and pod. +func (cc *Crictl) StopPodAndContainer(podID, contID string) error { + if err := cc.StopContainer(contID); err != nil { + return fmt.Errorf("failed to stop container %q in pod %q: %v", contID, podID, err) + } + + if err := cc.StopPod(podID); err != nil { + return fmt.Errorf("failed to stop pod %q: %v", podID, err) + } + + if err := cc.RmPod(podID); err != nil { + return fmt.Errorf("failed to remove pod %q: %v", podID, err) + } + + return nil +} + +// run runs crictl with the given args. +func (cc *Crictl) run(args ...string) (string, error) { + defaultArgs := []string{ + resolvePath("crictl"), + "--image-endpoint", fmt.Sprintf("unix://%s", cc.endpoint), + "--runtime-endpoint", fmt.Sprintf("unix://%s", cc.endpoint), + } + fullArgs := append(defaultArgs, args...) + out, err := testutil.Command(cc.logger, fullArgs...).CombinedOutput() + return string(out), err +} diff --git a/pkg/test/dockerutil/BUILD b/pkg/test/dockerutil/BUILD new file mode 100644 index 000000000..7c8758e35 --- /dev/null +++ b/pkg/test/dockerutil/BUILD @@ -0,0 +1,14 @@ +load("//tools:defs.bzl", "go_library") + +package(licenses = ["notice"]) + +go_library( + name = "dockerutil", + testonly = 1, + srcs = ["dockerutil.go"], + visibility = ["//:sandbox"], + deps = [ + "//pkg/test/testutil", + "@com_github_kr_pty//:go_default_library", + ], +) diff --git a/pkg/test/dockerutil/dockerutil.go b/pkg/test/dockerutil/dockerutil.go new file mode 100644 index 000000000..baa8fc2f2 --- /dev/null +++ b/pkg/test/dockerutil/dockerutil.go @@ -0,0 +1,581 @@ +// 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 dockerutil is a collection of utility functions. +package dockerutil + +import ( + "encoding/json" + "flag" + "fmt" + "io" + "io/ioutil" + "log" + "net" + "os" + "os/exec" + "path" + "regexp" + "strconv" + "strings" + "syscall" + "time" + + "github.com/kr/pty" + "gvisor.dev/gvisor/pkg/test/testutil" +) + +var ( + // runtime is the runtime to use for tests. This will be applied to all + // containers. Note that the default here ("runsc") corresponds to the + // default used by the installations. This is important, because the + // default installer for vm_tests (in tools/installers:head, invoked + // via tools/vm:defs.bzl) will install with this name. So without + // changing anything, tests should have a runsc runtime available to + // them. Otherwise installers should update the existing runtime + // instead of installing a new one. + runtime = flag.String("runtime", "runsc", "specify which runtime to use") + + // config is the default Docker daemon configuration path. + config = flag.String("config_path", "/etc/docker/daemon.json", "configuration file for reading paths") +) + +// EnsureSupportedDockerVersion checks if correct docker is installed. +// +// This logs directly to stderr, as it is typically called from a Main wrapper. +func EnsureSupportedDockerVersion() { + cmd := exec.Command("docker", "version") + out, err := cmd.CombinedOutput() + if err != nil { + log.Fatalf("error running %q: %v", "docker version", err) + } + re := regexp.MustCompile(`Version:\s+(\d+)\.(\d+)\.\d.*`) + matches := re.FindStringSubmatch(string(out)) + if len(matches) != 3 { + log.Fatalf("Invalid docker output: %s", out) + } + major, _ := strconv.Atoi(matches[1]) + minor, _ := strconv.Atoi(matches[2]) + if major < 17 || (major == 17 && minor < 9) { + log.Fatalf("Docker version 17.09.0 or greater is required, found: %02d.%02d", major, minor) + } +} + +// RuntimePath returns the binary path for the current runtime. +func RuntimePath() (string, error) { + // Read the configuration data; the file must exist. + configBytes, err := ioutil.ReadFile(*config) + if err != nil { + return "", err + } + + // Unmarshal the configuration. + c := make(map[string]interface{}) + if err := json.Unmarshal(configBytes, &c); err != nil { + return "", err + } + + // Decode the expected configuration. + r, ok := c["runtimes"] + if !ok { + return "", fmt.Errorf("no runtimes declared: %v", c) + } + rs, ok := r.(map[string]interface{}) + if !ok { + // The runtimes are not a map. + return "", fmt.Errorf("unexpected format: %v", c) + } + r, ok = rs[*runtime] + if !ok { + // The expected runtime is not declared. + return "", fmt.Errorf("runtime %q not found: %v", *runtime, c) + } + rs, ok = r.(map[string]interface{}) + if !ok { + // The runtime is not a map. + return "", fmt.Errorf("unexpected format: %v", c) + } + p, ok := rs["path"].(string) + if !ok { + // The runtime does not declare a path. + return "", fmt.Errorf("unexpected format: %v", c) + } + return p, nil +} + +// Save exports a container image to the given Writer. +// +// Note that the writer should be actively consuming the output, otherwise it +// is not guaranteed that the Save will make any progress and the call may +// stall indefinitely. +// +// This is called by criutil in order to import imports. +func Save(logger testutil.Logger, image string, w io.Writer) error { + cmd := testutil.Command(logger, "docker", "save", testutil.ImageByName(image)) + cmd.Stdout = w // Send directly to the writer. + return cmd.Run() +} + +// MountMode describes if the mount should be ro or rw. +type MountMode int + +const ( + // ReadOnly is what the name says. + ReadOnly MountMode = iota + // ReadWrite is what the name says. + ReadWrite +) + +// String returns the mount mode argument for this MountMode. +func (m MountMode) String() string { + switch m { + case ReadOnly: + return "ro" + case ReadWrite: + return "rw" + } + panic(fmt.Sprintf("invalid mode: %d", m)) +} + +// Docker contains the name and the runtime of a docker container. +type Docker struct { + logger testutil.Logger + Runtime string + Name string + copyErr error + mounts []string + cleanups []func() +} + +// MakeDocker sets up the struct for a Docker container. +// +// Names of containers will be unique. +func MakeDocker(logger testutil.Logger) *Docker { + return &Docker{ + logger: logger, + Name: testutil.RandomID(logger.Name()), + Runtime: *runtime, + } +} + +// Mount mounts the given source and makes it available in the container. +func (d *Docker) Mount(target, source string, mode MountMode) { + d.mounts = append(d.mounts, fmt.Sprintf("-v=%s:%s:%v", source, target, mode)) +} + +// CopyFiles copies in and mounts the given files. They are always ReadOnly. +func (d *Docker) CopyFiles(target string, sources ...string) { + dir, err := ioutil.TempDir("", d.Name) + if err != nil { + d.copyErr = fmt.Errorf("ioutil.TempDir failed: %v", err) + return + } + d.cleanups = append(d.cleanups, func() { os.RemoveAll(dir) }) + if err := os.Chmod(dir, 0755); err != nil { + d.copyErr = fmt.Errorf("os.Chmod(%q, 0755) failed: %v", dir, err) + return + } + for _, name := range sources { + src, err := testutil.FindFile(name) + if err != nil { + d.copyErr = fmt.Errorf("testutil.FindFile(%q) failed: %v", name, err) + return + } + dst := path.Join(dir, path.Base(name)) + if err := testutil.Copy(src, dst); err != nil { + d.copyErr = fmt.Errorf("testutil.Copy(%q, %q) failed: %v", src, dst, err) + return + } + d.logger.Logf("copy: %s -> %s", src, dst) + } + d.Mount(target, dir, ReadOnly) +} + +// Link links the given target. +func (d *Docker) Link(target string, source *Docker) { + d.mounts = append(d.mounts, fmt.Sprintf("--link=%s:%s", source.Name, target)) +} + +// RunOpts are options for running a container. +type RunOpts struct { + // Image is the image relative to images/. This will be mangled + // appropriately, to ensure that only first-party images are used. + Image string + + // Memory is the memory limit in kB. + Memory int + + // Ports are the ports to be allocated. + Ports []int + + // WorkDir sets the working directory. + WorkDir string + + // ReadOnly sets the read-only flag. + ReadOnly bool + + // Env are additional environment variables. + Env []string + + // User is the user to use. + User string + + // Privileged enables privileged mode. + Privileged bool + + // CapAdd are the extra set of capabilities to add. + CapAdd []string + + // CapDrop are the extra set of capabilities to drop. + CapDrop []string + + // Pty indicates that a pty will be allocated. If this is non-nil, then + // this will run after start-up with the *exec.Command and Pty file + // passed in to the function. + Pty func(*exec.Cmd, *os.File) + + // Foreground indicates that the container should be run in the + // foreground. If this is true, then the output will be available as a + // return value from the Run function. + Foreground bool + + // Extra are extra arguments that may be passed. + Extra []string +} + +// args returns common arguments. +// +// Note that this does not define the complete behavior. +func (d *Docker) argsFor(r *RunOpts, command string, p []string) (rv []string) { + isExec := command == "exec" + isRun := command == "run" + + if isRun || isExec { + rv = append(rv, "-i") + } + if r.Pty != nil { + rv = append(rv, "-t") + } + if r.User != "" { + rv = append(rv, fmt.Sprintf("--user=%s", r.User)) + } + if r.Privileged { + rv = append(rv, "--privileged") + } + for _, c := range r.CapAdd { + rv = append(rv, fmt.Sprintf("--cap-add=%s", c)) + } + for _, c := range r.CapDrop { + rv = append(rv, fmt.Sprintf("--cap-drop=%s", c)) + } + for _, e := range r.Env { + rv = append(rv, fmt.Sprintf("--env=%s", e)) + } + if r.WorkDir != "" { + rv = append(rv, fmt.Sprintf("--workdir=%s", r.WorkDir)) + } + if !isExec { + if r.Memory != 0 { + rv = append(rv, fmt.Sprintf("--memory=%dk", r.Memory)) + } + for _, p := range r.Ports { + rv = append(rv, fmt.Sprintf("--publish=%d", p)) + } + if r.ReadOnly { + rv = append(rv, fmt.Sprintf("--read-only")) + } + if len(p) > 0 { + rv = append(rv, "--entrypoint=") + } + } + + // Always attach the test environment & Extra. + rv = append(rv, fmt.Sprintf("--env=RUNSC_TEST_NAME=%s", d.Name)) + rv = append(rv, r.Extra...) + + // Attach necessary bits. + if isExec { + rv = append(rv, d.Name) + } else { + rv = append(rv, d.mounts...) + rv = append(rv, fmt.Sprintf("--runtime=%s", d.Runtime)) + rv = append(rv, fmt.Sprintf("--name=%s", d.Name)) + rv = append(rv, testutil.ImageByName(r.Image)) + } + + // Attach other arguments. + rv = append(rv, p...) + return rv +} + +// run runs a complete command. +func (d *Docker) run(r RunOpts, command string, p ...string) (string, error) { + if d.copyErr != nil { + return "", d.copyErr + } + basicArgs := []string{"docker"} + if command == "spawn" { + command = "run" + basicArgs = append(basicArgs, command) + basicArgs = append(basicArgs, "-d") + } else { + basicArgs = append(basicArgs, command) + } + customArgs := d.argsFor(&r, command, p) + cmd := testutil.Command(d.logger, append(basicArgs, customArgs...)...) + if r.Pty != nil { + // If allocating a terminal, then we just ignore the output + // from the command. + ptmx, err := pty.Start(cmd.Cmd) + if err != nil { + return "", err + } + defer cmd.Wait() // Best effort. + r.Pty(cmd.Cmd, ptmx) + } else { + // Can't support PTY or streaming. + out, err := cmd.CombinedOutput() + return string(out), err + } + return "", nil +} + +// Create calls 'docker create' with the arguments provided. +func (d *Docker) Create(r RunOpts, args ...string) error { + _, err := d.run(r, "create", args...) + return err +} + +// Start calls 'docker start'. +func (d *Docker) Start() error { + return testutil.Command(d.logger, "docker", "start", d.Name).Run() +} + +// Stop calls 'docker stop'. +func (d *Docker) Stop() error { + return testutil.Command(d.logger, "docker", "stop", d.Name).Run() +} + +// Run calls 'docker run' with the arguments provided. +func (d *Docker) Run(r RunOpts, args ...string) (string, error) { + return d.run(r, "run", args...) +} + +// Spawn starts the container and detaches. +func (d *Docker) Spawn(r RunOpts, args ...string) error { + _, err := d.run(r, "spawn", args...) + return err +} + +// Logs calls 'docker logs'. +func (d *Docker) Logs() (string, error) { + // Don't capture the output; since it will swamp the logs. + out, err := exec.Command("docker", "logs", d.Name).CombinedOutput() + return string(out), err +} + +// Exec calls 'docker exec' with the arguments provided. +func (d *Docker) Exec(r RunOpts, args ...string) (string, error) { + return d.run(r, "exec", args...) +} + +// Pause calls 'docker pause'. +func (d *Docker) Pause() error { + return testutil.Command(d.logger, "docker", "pause", d.Name).Run() +} + +// Unpause calls 'docker pause'. +func (d *Docker) Unpause() error { + return testutil.Command(d.logger, "docker", "unpause", d.Name).Run() +} + +// Checkpoint calls 'docker checkpoint'. +func (d *Docker) Checkpoint(name string) error { + return testutil.Command(d.logger, "docker", "checkpoint", "create", d.Name, name).Run() +} + +// Restore calls 'docker start --checkname [name]'. +func (d *Docker) Restore(name string) error { + return testutil.Command(d.logger, "docker", "start", fmt.Sprintf("--checkpoint=%s", name), d.Name).Run() +} + +// Kill calls 'docker kill'. +func (d *Docker) Kill() error { + // Skip logging this command, it will likely be an error. + out, err := exec.Command("docker", "kill", d.Name).CombinedOutput() + if err != nil && !strings.Contains(string(out), "is not running") { + return err + } + return nil +} + +// Remove calls 'docker rm'. +func (d *Docker) Remove() error { + return testutil.Command(d.logger, "docker", "rm", d.Name).Run() +} + +// CleanUp kills and deletes the container (best effort). +func (d *Docker) CleanUp() { + // Kill the container. + if err := d.Kill(); err != nil { + // Just log; can't do anything here. + d.logger.Logf("error killing container %q: %v", d.Name, err) + } + // Remove the image. + if err := d.Remove(); err != nil { + d.logger.Logf("error removing container %q: %v", d.Name, err) + } + // Forget all mounts. + d.mounts = nil + // Execute all cleanups. + for _, c := range d.cleanups { + c() + } + d.cleanups = nil +} + +// FindPort returns the host port that is mapped to 'sandboxPort'. This calls +// docker to allocate a free port in the host and prevent conflicts. +func (d *Docker) FindPort(sandboxPort int) (int, error) { + format := fmt.Sprintf(`{{ (index (index .NetworkSettings.Ports "%d/tcp") 0).HostPort }}`, sandboxPort) + out, err := testutil.Command(d.logger, "docker", "inspect", "-f", format, d.Name).CombinedOutput() + if err != nil { + return -1, fmt.Errorf("error retrieving port: %v", err) + } + port, err := strconv.Atoi(strings.TrimSuffix(string(out), "\n")) + if err != nil { + return -1, fmt.Errorf("error parsing port %q: %v", out, err) + } + return port, nil +} + +// FindIP returns the IP address of the container. +func (d *Docker) FindIP() (net.IP, error) { + const format = `{{range .NetworkSettings.Networks}}{{.IPAddress}}{{end}}` + out, err := testutil.Command(d.logger, "docker", "inspect", "-f", format, d.Name).CombinedOutput() + if err != nil { + return net.IP{}, fmt.Errorf("error retrieving IP: %v", err) + } + ip := net.ParseIP(strings.TrimSpace(string(out))) + if ip == nil { + return net.IP{}, fmt.Errorf("invalid IP: %q", string(out)) + } + return ip, nil +} + +// SandboxPid returns the PID to the sandbox process. +func (d *Docker) SandboxPid() (int, error) { + out, err := testutil.Command(d.logger, "docker", "inspect", "-f={{.State.Pid}}", d.Name).CombinedOutput() + if err != nil { + return -1, fmt.Errorf("error retrieving pid: %v", err) + } + pid, err := strconv.Atoi(strings.TrimSuffix(string(out), "\n")) + if err != nil { + return -1, fmt.Errorf("error parsing pid %q: %v", out, err) + } + return pid, nil +} + +// ID returns the container ID. +func (d *Docker) ID() (string, error) { + out, err := testutil.Command(d.logger, "docker", "inspect", "-f={{.Id}}", d.Name).CombinedOutput() + if err != nil { + return "", fmt.Errorf("error retrieving ID: %v", err) + } + return strings.TrimSpace(string(out)), nil +} + +// Wait waits for container to exit, up to the given timeout. Returns error if +// wait fails or timeout is hit. Returns the application return code otherwise. +// Note that the application may have failed even if err == nil, always check +// the exit code. +func (d *Docker) Wait(timeout time.Duration) (syscall.WaitStatus, error) { + timeoutChan := time.After(timeout) + waitChan := make(chan (syscall.WaitStatus)) + errChan := make(chan (error)) + + go func() { + out, err := testutil.Command(d.logger, "docker", "wait", d.Name).CombinedOutput() + if err != nil { + errChan <- fmt.Errorf("error waiting for container %q: %v", d.Name, err) + } + exit, err := strconv.Atoi(strings.TrimSuffix(string(out), "\n")) + if err != nil { + errChan <- fmt.Errorf("error parsing exit code %q: %v", out, err) + } + waitChan <- syscall.WaitStatus(uint32(exit)) + }() + + select { + case ws := <-waitChan: + return ws, nil + case err := <-errChan: + return syscall.WaitStatus(1), err + case <-timeoutChan: + return syscall.WaitStatus(1), fmt.Errorf("timeout waiting for container %q", d.Name) + } +} + +// WaitForOutput calls 'docker logs' to retrieve containers output and searches +// for the given pattern. +func (d *Docker) WaitForOutput(pattern string, timeout time.Duration) (string, error) { + matches, err := d.WaitForOutputSubmatch(pattern, timeout) + if err != nil { + return "", err + } + if len(matches) == 0 { + return "", nil + } + return matches[0], nil +} + +// WaitForOutputSubmatch calls 'docker logs' to retrieve containers output and +// searches for the given pattern. It returns any regexp submatches as well. +func (d *Docker) WaitForOutputSubmatch(pattern string, timeout time.Duration) ([]string, error) { + re := regexp.MustCompile(pattern) + var ( + lastOut string + stopped bool + ) + for exp := time.Now().Add(timeout); time.Now().Before(exp); { + out, err := d.Logs() + if err != nil { + return nil, err + } + if out != lastOut { + if lastOut == "" { + d.logger.Logf("output (start): %s", out) + } else if strings.HasPrefix(out, lastOut) { + d.logger.Logf("output (contn): %s", out[len(lastOut):]) + } else { + d.logger.Logf("output (trunc): %s", out) + } + lastOut = out // Save for future. + if matches := re.FindStringSubmatch(lastOut); matches != nil { + return matches, nil // Success! + } + } else if stopped { + // The sandbox stopped and we looked at the + // logs at least once since determining that. + return nil, fmt.Errorf("no longer running: %v", err) + } else if pid, err := d.SandboxPid(); pid == 0 || err != nil { + // The sandbox may have stopped, but it's + // possible that it has emitted the terminal + // line between the last call to Logs and here. + stopped = true + } + time.Sleep(100 * time.Millisecond) + } + return nil, fmt.Errorf("timeout waiting for output %q: %s", re.String(), lastOut) +} diff --git a/pkg/test/testutil/BUILD b/pkg/test/testutil/BUILD new file mode 100644 index 000000000..03b1b4677 --- /dev/null +++ b/pkg/test/testutil/BUILD @@ -0,0 +1,20 @@ +load("//tools:defs.bzl", "go_library") + +package(licenses = ["notice"]) + +go_library( + name = "testutil", + testonly = 1, + srcs = [ + "testutil.go", + "testutil_runfiles.go", + ], + visibility = ["//:sandbox"], + deps = [ + "//pkg/sync", + "//runsc/boot", + "//runsc/specutils", + "@com_github_cenkalti_backoff//:go_default_library", + "@com_github_opencontainers_runtime-spec//specs-go:go_default_library", + ], +) diff --git a/pkg/test/testutil/testutil.go b/pkg/test/testutil/testutil.go new file mode 100644 index 000000000..ee8c78014 --- /dev/null +++ b/pkg/test/testutil/testutil.go @@ -0,0 +1,518 @@ +// 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 testutil contains utility functions for runsc tests. +package testutil + +import ( + "bufio" + "context" + "debug/elf" + "encoding/base32" + "encoding/json" + "flag" + "fmt" + "io" + "io/ioutil" + "log" + "math" + "math/rand" + "net/http" + "os" + "os/exec" + "os/signal" + "path" + "path/filepath" + "strconv" + "strings" + "sync/atomic" + "syscall" + "testing" + "time" + + "github.com/cenkalti/backoff" + specs "github.com/opencontainers/runtime-spec/specs-go" + "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/runsc/boot" + "gvisor.dev/gvisor/runsc/specutils" +) + +var ( + checkpoint = flag.Bool("checkpoint", true, "control checkpoint/restore support") +) + +// IsCheckpointSupported returns the relevant command line flag. +func IsCheckpointSupported() bool { + return *checkpoint +} + +// ImageByName mangles the image name used locally. This depends on the image +// build infrastructure in images/ and tools/vm. +func ImageByName(name string) string { + return fmt.Sprintf("gvisor.dev/images/%s", name) +} + +// ConfigureExePath configures the executable for runsc in the test environment. +func ConfigureExePath() error { + path, err := FindFile("runsc/runsc") + if err != nil { + return err + } + specutils.ExePath = path + return nil +} + +// TmpDir returns the absolute path to a writable directory that can be used as +// scratch by the test. +func TmpDir() string { + dir := os.Getenv("TEST_TMPDIR") + if dir == "" { + dir = "/tmp" + } + return dir +} + +// Logger is a simple logging wrapper. +// +// This is designed to be implemented by *testing.T. +type Logger interface { + Name() string + Logf(fmt string, args ...interface{}) +} + +// DefaultLogger logs using the log package. +type DefaultLogger string + +// Name implements Logger.Name. +func (d DefaultLogger) Name() string { + return string(d) +} + +// Logf implements Logger.Logf. +func (d DefaultLogger) Logf(fmt string, args ...interface{}) { + log.Printf(fmt, args...) +} + +// Cmd is a simple wrapper. +type Cmd struct { + logger Logger + *exec.Cmd +} + +// CombinedOutput returns the output and logs. +func (c *Cmd) CombinedOutput() ([]byte, error) { + out, err := c.Cmd.CombinedOutput() + if len(out) > 0 { + c.logger.Logf("output: %s", string(out)) + } + if err != nil { + c.logger.Logf("error: %v", err) + } + return out, err +} + +// Command is a simple wrapper around exec.Command, that logs. +func Command(logger Logger, args ...string) *Cmd { + logger.Logf("command: %s", strings.Join(args, " ")) + return &Cmd{ + logger: logger, + Cmd: exec.Command(args[0], args[1:]...), + } +} + +// TestConfig returns the default configuration to use in tests. Note that +// 'RootDir' must be set by caller if required. +func TestConfig(t *testing.T) *boot.Config { + logDir := os.TempDir() + if dir, ok := os.LookupEnv("TEST_UNDECLARED_OUTPUTS_DIR"); ok { + logDir = dir + "/" + } + return &boot.Config{ + Debug: true, + DebugLog: path.Join(logDir, "runsc.log."+t.Name()+".%TIMESTAMP%.%COMMAND%"), + LogFormat: "text", + DebugLogFormat: "text", + LogPackets: true, + Network: boot.NetworkNone, + Strace: true, + Platform: "ptrace", + FileAccess: boot.FileAccessExclusive, + NumNetworkChannels: 1, + + TestOnlyAllowRunAsCurrentUserWithoutChroot: true, + } +} + +// NewSpecWithArgs creates a simple spec with the given args suitable for use +// in tests. +func NewSpecWithArgs(args ...string) *specs.Spec { + return &specs.Spec{ + // The host filesystem root is the container root. + Root: &specs.Root{ + Path: "/", + Readonly: true, + }, + Process: &specs.Process{ + Args: args, + Env: []string{ + "PATH=" + os.Getenv("PATH"), + }, + Capabilities: specutils.AllCapabilities(), + }, + Mounts: []specs.Mount{ + // Hide the host /etc to avoid any side-effects. + // For example, bash reads /etc/passwd and if it is + // very big, tests can fail by timeout. + { + Type: "tmpfs", + Destination: "/etc", + }, + // Root is readonly, but many tests want to write to tmpdir. + // This creates a writable mount inside the root. Also, when tmpdir points + // to "/tmp", it makes the the actual /tmp to be mounted and not a tmpfs + // inside the sentry. + { + Type: "bind", + Destination: TmpDir(), + Source: TmpDir(), + }, + }, + Hostname: "runsc-test-hostname", + } +} + +// SetupRootDir creates a root directory for containers. +func SetupRootDir() (string, func(), error) { + rootDir, err := ioutil.TempDir(TmpDir(), "containers") + if err != nil { + return "", nil, fmt.Errorf("error creating root dir: %v", err) + } + return rootDir, func() { os.RemoveAll(rootDir) }, nil +} + +// SetupContainer creates a bundle and root dir for the container, generates a +// test config, and writes the spec to config.json in the bundle dir. +func SetupContainer(spec *specs.Spec, conf *boot.Config) (rootDir, bundleDir string, cleanup func(), err error) { + rootDir, rootCleanup, err := SetupRootDir() + if err != nil { + return "", "", nil, err + } + conf.RootDir = rootDir + bundleDir, bundleCleanup, err := SetupBundleDir(spec) + if err != nil { + rootCleanup() + return "", "", nil, err + } + return rootDir, bundleDir, func() { + bundleCleanup() + rootCleanup() + }, err +} + +// SetupBundleDir creates a bundle dir and writes the spec to config.json. +func SetupBundleDir(spec *specs.Spec) (string, func(), error) { + bundleDir, err := ioutil.TempDir(TmpDir(), "bundle") + if err != nil { + return "", nil, fmt.Errorf("error creating bundle dir: %v", err) + } + cleanup := func() { os.RemoveAll(bundleDir) } + if err := writeSpec(bundleDir, spec); err != nil { + cleanup() + return "", nil, fmt.Errorf("error writing spec: %v", err) + } + return bundleDir, cleanup, nil +} + +// writeSpec writes the spec to disk in the given directory. +func writeSpec(dir string, spec *specs.Spec) error { + b, err := json.Marshal(spec) + if err != nil { + return err + } + return ioutil.WriteFile(filepath.Join(dir, "config.json"), b, 0755) +} + +// RandomID returns 20 random bytes following the given prefix. +func RandomID(prefix string) string { + // Read 20 random bytes. + b := make([]byte, 20) + // "[Read] always returns len(p) and a nil error." --godoc + if _, err := rand.Read(b); err != nil { + panic("rand.Read failed: " + err.Error()) + } + return fmt.Sprintf("%s-%s", prefix, base32.StdEncoding.EncodeToString(b)) +} + +// RandomContainerID generates a random container id for each test. +// +// The container id is used to create an abstract unix domain socket, which +// must be unique. While the container forbids creating two containers with the +// same name, sometimes between test runs the socket does not get cleaned up +// quickly enough, causing container creation to fail. +func RandomContainerID() string { + return RandomID("test-container-") +} + +// Copy copies file from src to dst. +func Copy(src, dst string) error { + in, err := os.Open(src) + if err != nil { + return err + } + defer in.Close() + + st, err := in.Stat() + if err != nil { + return err + } + + out, err := os.OpenFile(dst, os.O_CREATE|os.O_TRUNC|os.O_WRONLY, st.Mode().Perm()) + if err != nil { + return err + } + defer out.Close() + + // Mirror the local user's permissions across all users. This is + // because as we inject things into the container, the UID/GID will + // change. Also, the build system may generate artifacts with different + // modes. At the top-level (volume mapping) we have a big read-only + // knob that can be applied to prevent modifications. + // + // Note that this must be done via a separate Chmod call, otherwise the + // current process's umask will get in the way. + var mode os.FileMode + if st.Mode()&0100 != 0 { + mode |= 0111 + } + if st.Mode()&0200 != 0 { + mode |= 0222 + } + if st.Mode()&0400 != 0 { + mode |= 0444 + } + if err := os.Chmod(dst, mode); err != nil { + return err + } + + _, err = io.Copy(out, in) + return err +} + +// Poll is a shorthand function to poll for something with given timeout. +func Poll(cb func() error, timeout time.Duration) error { + ctx, cancel := context.WithTimeout(context.Background(), timeout) + defer cancel() + b := backoff.WithContext(backoff.NewConstantBackOff(100*time.Millisecond), ctx) + return backoff.Retry(cb, b) +} + +// WaitForHTTP tries GET requests on a port until the call succeeds or timeout. +func WaitForHTTP(port int, timeout time.Duration) error { + cb := func() error { + c := &http.Client{ + // Calculate timeout to be able to do minimum 5 attempts. + Timeout: timeout / 5, + } + url := fmt.Sprintf("http://localhost:%d/", port) + resp, err := c.Get(url) + if err != nil { + log.Printf("Waiting %s: %v", url, err) + return err + } + resp.Body.Close() + return nil + } + return Poll(cb, timeout) +} + +// Reaper reaps child processes. +type Reaper struct { + // mu protects ch, which will be nil if the reaper is not running. + mu sync.Mutex + ch chan os.Signal +} + +// Start starts reaping child processes. +func (r *Reaper) Start() { + r.mu.Lock() + defer r.mu.Unlock() + + if r.ch != nil { + panic("reaper.Start called on a running reaper") + } + + r.ch = make(chan os.Signal, 1) + signal.Notify(r.ch, syscall.SIGCHLD) + + go func() { + for { + r.mu.Lock() + ch := r.ch + r.mu.Unlock() + if ch == nil { + return + } + + _, ok := <-ch + if !ok { + // Channel closed. + return + } + for { + cpid, _ := syscall.Wait4(-1, nil, syscall.WNOHANG, nil) + if cpid < 1 { + break + } + } + } + }() +} + +// Stop stops reaping child processes. +func (r *Reaper) Stop() { + r.mu.Lock() + defer r.mu.Unlock() + + if r.ch == nil { + panic("reaper.Stop called on a stopped reaper") + } + + signal.Stop(r.ch) + close(r.ch) + r.ch = nil +} + +// StartReaper is a helper that starts a new Reaper and returns a function to +// stop it. +func StartReaper() func() { + r := &Reaper{} + r.Start() + return r.Stop +} + +// WaitUntilRead reads from the given reader until the wanted string is found +// or until timeout. +func WaitUntilRead(r io.Reader, want string, split bufio.SplitFunc, timeout time.Duration) error { + sc := bufio.NewScanner(r) + if split != nil { + sc.Split(split) + } + // done must be accessed atomically. A value greater than 0 indicates + // that the read loop can exit. + var done uint32 + doneCh := make(chan struct{}) + go func() { + for sc.Scan() { + t := sc.Text() + if strings.Contains(t, want) { + atomic.StoreUint32(&done, 1) + close(doneCh) + break + } + if atomic.LoadUint32(&done) > 0 { + break + } + } + }() + select { + case <-time.After(timeout): + atomic.StoreUint32(&done, 1) + return fmt.Errorf("timeout waiting to read %q", want) + case <-doneCh: + return nil + } +} + +// KillCommand kills the process running cmd unless it hasn't been started. It +// returns an error if it cannot kill the process unless the reason is that the +// process has already exited. +// +// KillCommand will also reap the process. +func KillCommand(cmd *exec.Cmd) error { + if cmd.Process == nil { + return nil + } + if err := cmd.Process.Kill(); err != nil { + if !strings.Contains(err.Error(), "process already finished") { + return fmt.Errorf("failed to kill process %v: %v", cmd, err) + } + } + return cmd.Wait() +} + +// WriteTmpFile writes text to a temporary file, closes the file, and returns +// the name of the file. A cleanup function is also returned. +func WriteTmpFile(pattern, text string) (string, func(), error) { + file, err := ioutil.TempFile(TmpDir(), pattern) + if err != nil { + return "", nil, err + } + defer file.Close() + if _, err := file.Write([]byte(text)); err != nil { + return "", nil, err + } + return file.Name(), func() { os.RemoveAll(file.Name()) }, nil +} + +// IsStatic returns true iff the given file is a static binary. +func IsStatic(filename string) (bool, error) { + f, err := elf.Open(filename) + if err != nil { + return false, err + } + for _, prog := range f.Progs { + if prog.Type == elf.PT_INTERP { + return false, nil // Has interpreter. + } + } + return true, nil +} + +// TestIndicesForShard returns indices for this test shard based on the +// TEST_SHARD_INDEX and TEST_TOTAL_SHARDS environment vars. +// +// If either of the env vars are not present, then the function will return all +// tests. If there are more shards than there are tests, then the returned list +// may be empty. +func TestIndicesForShard(numTests int) ([]int, error) { + var ( + shardIndex = 0 + shardTotal = 1 + ) + + indexStr, totalStr := os.Getenv("TEST_SHARD_INDEX"), os.Getenv("TEST_TOTAL_SHARDS") + if indexStr != "" && totalStr != "" { + // Parse index and total to ints. + var err error + shardIndex, err = strconv.Atoi(indexStr) + if err != nil { + return nil, fmt.Errorf("invalid TEST_SHARD_INDEX %q: %v", indexStr, err) + } + shardTotal, err = strconv.Atoi(totalStr) + if err != nil { + return nil, fmt.Errorf("invalid TEST_TOTAL_SHARDS %q: %v", totalStr, err) + } + } + + // Calculate! + var indices []int + numBlocks := int(math.Ceil(float64(numTests) / float64(shardTotal))) + for i := 0; i < numBlocks; i++ { + pick := i*shardTotal + shardIndex + if pick < numTests { + indices = append(indices, pick) + } + } + return indices, nil +} diff --git a/pkg/test/testutil/testutil_runfiles.go b/pkg/test/testutil/testutil_runfiles.go new file mode 100644 index 000000000..ece9ea9a1 --- /dev/null +++ b/pkg/test/testutil/testutil_runfiles.go @@ -0,0 +1,75 @@ +// 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 testutil + +import ( + "fmt" + "os" + "path/filepath" +) + +// FindFile searchs for a file inside the test run environment. It returns the +// full path to the file. It fails if none or more than one file is found. +func FindFile(path string) (string, error) { + wd, err := os.Getwd() + if err != nil { + return "", err + } + + // The test root is demarcated by a path element called "__main__". Search for + // it backwards from the working directory. + root := wd + for { + dir, name := filepath.Split(root) + if name == "__main__" { + break + } + if len(dir) == 0 { + return "", fmt.Errorf("directory __main__ not found in %q", wd) + } + // Remove ending slash to loop around. + root = dir[:len(dir)-1] + } + + // Annoyingly, bazel adds the build type to the directory path for go + // binaries, but not for c++ binaries. We use two different patterns to + // to find our file. + patterns := []string{ + // Try the obvious path first. + filepath.Join(root, path), + // If it was a go binary, use a wildcard to match the build + // type. The pattern is: /test-path/__main__/directories/*/file. + filepath.Join(root, filepath.Dir(path), "*", filepath.Base(path)), + } + + for _, p := range patterns { + matches, err := filepath.Glob(p) + if err != nil { + // "The only possible returned error is ErrBadPattern, + // when pattern is malformed." -godoc + return "", fmt.Errorf("error globbing %q: %v", p, err) + } + switch len(matches) { + case 0: + // Try the next pattern. + case 1: + // We found it. + return matches[0], nil + default: + return "", fmt.Errorf("more than one match found for %q: %s", path, matches) + } + } + return "", fmt.Errorf("file %q not found", path) +} diff --git a/pkg/usermem/usermem.go b/pkg/usermem/usermem.go index d2f4403b0..cd6a0ea6b 100644 --- a/pkg/usermem/usermem.go +++ b/pkg/usermem/usermem.go @@ -29,9 +29,6 @@ import ( ) // 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 < 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 |