diff options
Diffstat (limited to 'pkg/sentry')
232 files changed, 7022 insertions, 2131 deletions
diff --git a/pkg/sentry/arch/BUILD b/pkg/sentry/arch/BUILD index 99e2b3389..4af4d6e84 100644 --- a/pkg/sentry/arch/BUILD +++ b/pkg/sentry/arch/BUILD @@ -22,6 +22,7 @@ go_library( "signal_info.go", "signal_stack.go", "stack.go", + "stack_unsafe.go", "syscalls_amd64.go", "syscalls_arm64.go", ], diff --git a/pkg/sentry/arch/arch_aarch64.go b/pkg/sentry/arch/arch_aarch64.go index 0f433ee79..fd73751e7 100644 --- a/pkg/sentry/arch/arch_aarch64.go +++ b/pkg/sentry/arch/arch_aarch64.go @@ -154,6 +154,7 @@ func (s State) Proto() *rpb.Registers { Sp: s.Regs.Sp, Pc: s.Regs.Pc, Pstate: s.Regs.Pstate, + Tls: s.Regs.TPIDR_EL0, } return &rpb.Registers{Arch: &rpb.Registers_Arm64{Arm64: regs}} } @@ -232,6 +233,7 @@ func (s *State) RegisterMap() (map[string]uintptr, error) { "Sp": uintptr(s.Regs.Sp), "Pc": uintptr(s.Regs.Pc), "Pstate": uintptr(s.Regs.Pstate), + "Tls": uintptr(s.Regs.TPIDR_EL0), }, nil } diff --git a/pkg/sentry/arch/registers.proto b/pkg/sentry/arch/registers.proto index 60c027aab..2727ba08a 100644 --- a/pkg/sentry/arch/registers.proto +++ b/pkg/sentry/arch/registers.proto @@ -83,6 +83,7 @@ message ARM64Registers { uint64 sp = 32; uint64 pc = 33; uint64 pstate = 34; + uint64 tls = 35; } message Registers { oneof arch { diff --git a/pkg/sentry/arch/signal_amd64.go b/pkg/sentry/arch/signal_amd64.go index 6fb756f0e..72e07a988 100644 --- a/pkg/sentry/arch/signal_amd64.go +++ b/pkg/sentry/arch/signal_amd64.go @@ -17,17 +17,19 @@ package arch import ( - "encoding/binary" "math" "syscall" "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/log" + "gvisor.dev/gvisor/pkg/marshal/primitive" "gvisor.dev/gvisor/pkg/usermem" ) // SignalContext64 is equivalent to struct sigcontext, the type passed as the // second argument to signal handlers set by signal(2). +// +// +marshal type SignalContext64 struct { R8 uint64 R9 uint64 @@ -68,6 +70,8 @@ const ( ) // UContext64 is equivalent to ucontext_t on 64-bit x86. +// +// +marshal type UContext64 struct { Flags uint64 Link uint64 @@ -172,12 +176,7 @@ func (c *context64) SignalSetup(st *Stack, act *SignalAct, info *SignalInfo, alt // "... the value (%rsp+8) is always a multiple of 16 (...) when // control is transferred to the function entry point." - AMD64 ABI - ucSize := binary.Size(uc) - if ucSize < 0 { - // This can only happen if we've screwed up the definition of - // UContext64. - panic("can't get size of UContext64") - } + ucSize := uc.SizeBytes() // 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 @@ -195,18 +194,18 @@ func (c *context64) SignalSetup(st *Stack, act *SignalAct, info *SignalInfo, alt info.FixSignalCodeForUser() // Set up the stack frame. - infoAddr, err := st.Push(info) - if err != nil { + if _, err := info.CopyOut(st, StackBottomMagic); err != nil { return err } - ucAddr, err := st.Push(uc) - if err != nil { + infoAddr := st.Bottom + if _, err := uc.CopyOut(st, StackBottomMagic); err != nil { return err } + ucAddr := st.Bottom if act.HasRestorer() { // Push the restorer return address. // Note that this doesn't need to be popped. - if _, err := st.Push(usermem.Addr(act.Restorer)); err != nil { + if _, err := primitive.CopyUint64Out(st, StackBottomMagic, act.Restorer); err != nil { return err } } else { @@ -240,11 +239,11 @@ func (c *context64) SignalSetup(st *Stack, act *SignalAct, info *SignalInfo, alt func (c *context64) SignalRestore(st *Stack, rt bool) (linux.SignalSet, SignalStack, error) { // Copy out the stack frame. var uc UContext64 - if _, err := st.Pop(&uc); err != nil { + if _, err := uc.CopyIn(st, StackBottomMagic); err != nil { return 0, SignalStack{}, err } var info SignalInfo - if _, err := st.Pop(&info); err != nil { + if _, err := info.CopyIn(st, StackBottomMagic); err != nil { return 0, SignalStack{}, err } diff --git a/pkg/sentry/arch/signal_arm64.go b/pkg/sentry/arch/signal_arm64.go index 642c79dda..7fde5d34e 100644 --- a/pkg/sentry/arch/signal_arm64.go +++ b/pkg/sentry/arch/signal_arm64.go @@ -12,10 +12,11 @@ // See the License for the specific language governing permissions and // limitations under the License. +// +build arm64 + package arch import ( - "encoding/binary" "syscall" "gvisor.dev/gvisor/pkg/abi/linux" @@ -25,6 +26,8 @@ import ( // SignalContext64 is equivalent to struct sigcontext, the type passed as the // second argument to signal handlers set by signal(2). +// +// +marshal type SignalContext64 struct { FaultAddr uint64 Regs [31]uint64 @@ -36,6 +39,7 @@ type SignalContext64 struct { Reserved [3568]uint8 } +// +marshal type aarch64Ctx struct { Magic uint32 Size uint32 @@ -43,6 +47,8 @@ type aarch64Ctx struct { // FpsimdContext is equivalent to struct fpsimd_context on arm64 // (arch/arm64/include/uapi/asm/sigcontext.h). +// +// +marshal type FpsimdContext struct { Head aarch64Ctx Fpsr uint32 @@ -51,13 +57,15 @@ type FpsimdContext struct { } // UContext64 is equivalent to ucontext on arm64(arch/arm64/include/uapi/asm/ucontext.h). +// +// +marshal type UContext64 struct { Flags uint64 Link uint64 Stack SignalStack Sigset linux.SignalSet // glibc uses a 1024-bit sigset_t - _pad [(1024 - 64) / 8]byte + _pad [120]byte // (1024 - 64) / 8 = 120 // sigcontext must be aligned to 16-byte _pad2 [8]byte // last for future expansion @@ -94,11 +102,7 @@ func (c *context64) SignalSetup(st *Stack, act *SignalAct, info *SignalInfo, alt }, Sigset: sigset, } - - ucSize := binary.Size(uc) - if ucSize < 0 { - panic("can't get size of UContext64") - } + ucSize := uc.SizeBytes() // frameSize = ucSize + sizeof(siginfo). // sizeof(siginfo) == 128. @@ -119,14 +123,14 @@ func (c *context64) SignalSetup(st *Stack, act *SignalAct, info *SignalInfo, alt info.FixSignalCodeForUser() // Set up the stack frame. - infoAddr, err := st.Push(info) - if err != nil { + if _, err := info.CopyOut(st, StackBottomMagic); err != nil { return err } - ucAddr, err := st.Push(uc) - if err != nil { + infoAddr := st.Bottom + if _, err := uc.CopyOut(st, StackBottomMagic); err != nil { return err } + ucAddr := st.Bottom // Set up registers. c.Regs.Sp = uint64(st.Bottom) @@ -147,11 +151,11 @@ func (c *context64) SignalSetup(st *Stack, act *SignalAct, info *SignalInfo, alt func (c *context64) SignalRestore(st *Stack, rt bool) (linux.SignalSet, SignalStack, error) { // Copy out the stack frame. var uc UContext64 - if _, err := st.Pop(&uc); err != nil { + if _, err := uc.CopyIn(st, StackBottomMagic); err != nil { return 0, SignalStack{}, err } var info SignalInfo - if _, err := st.Pop(&info); err != nil { + if _, err := info.CopyIn(st, StackBottomMagic); err != nil { return 0, SignalStack{}, err } diff --git a/pkg/sentry/arch/stack.go b/pkg/sentry/arch/stack.go index 1108fa0bd..5f06c751d 100644 --- a/pkg/sentry/arch/stack.go +++ b/pkg/sentry/arch/stack.go @@ -15,14 +15,16 @@ package arch import ( - "encoding/binary" - "fmt" - "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/marshal/primitive" "gvisor.dev/gvisor/pkg/usermem" ) -// Stack is a simple wrapper around a usermem.IO and an address. +// Stack is a simple wrapper around a usermem.IO and an address. Stack +// implements marshal.CopyContext, and marshallable values can be pushed or +// popped from the stack through the marshal.Marshallable interface. +// +// Stack is not thread-safe. type Stack struct { // Our arch info. // We use this for automatic Native conversion of usermem.Addrs during @@ -34,105 +36,60 @@ type Stack struct { // Our current stack bottom. Bottom usermem.Addr -} -// Push pushes the given values on to the stack. -// -// (This method supports Addrs and treats them as native types.) -func (s *Stack) Push(vals ...interface{}) (usermem.Addr, error) { - for _, v := range vals { - - // We convert some types to well-known serializable quanities. - var norm interface{} - - // For array types, we will automatically add an appropriate - // terminal value. This is done simply to make the interface - // easier to use. - var term interface{} - - switch v.(type) { - case string: - norm = []byte(v.(string)) - term = byte(0) - case []int8, []uint8: - norm = v - term = byte(0) - case []int16, []uint16: - norm = v - term = uint16(0) - case []int32, []uint32: - norm = v - term = uint32(0) - case []int64, []uint64: - norm = v - term = uint64(0) - case []usermem.Addr: - // Special case: simply push recursively. - _, err := s.Push(s.Arch.Native(uintptr(0))) - if err != nil { - return 0, err - } - varr := v.([]usermem.Addr) - for i := len(varr) - 1; i >= 0; i-- { - _, err := s.Push(varr[i]) - if err != nil { - return 0, err - } - } - continue - case usermem.Addr: - norm = s.Arch.Native(uintptr(v.(usermem.Addr))) - default: - norm = v - } + // Scratch buffer used for marshalling to avoid having to repeatedly + // allocate scratch memory. + scratchBuf []byte +} - if term != nil { - _, err := s.Push(term) - if err != nil { - return 0, err - } - } +// scratchBufLen is the default length of Stack.scratchBuf. The +// largest structs the stack regularly serializes are arch.SignalInfo +// and arch.UContext64. We'll set the default size as the larger of +// the two, arch.UContext64. +var scratchBufLen = (*UContext64)(nil).SizeBytes() - c := binary.Size(norm) - if c < 0 { - return 0, fmt.Errorf("bad binary.Size for %T", v) - } - n, err := usermem.CopyObjectOut(context.Background(), s.IO, s.Bottom-usermem.Addr(c), norm, usermem.IOOpts{}) - if err != nil || c != n { - return 0, err - } +// CopyScratchBuffer implements marshal.CopyContext.CopyScratchBuffer. +func (s *Stack) CopyScratchBuffer(size int) []byte { + if len(s.scratchBuf) < size { + s.scratchBuf = make([]byte, size) + } + return s.scratchBuf[:size] +} +// StackBottomMagic is the special address callers must past to all stack +// marshalling operations to cause the src/dst address to be computed based on +// the current end of the stack. +const StackBottomMagic = ^usermem.Addr(0) // usermem.Addr(-1) + +// CopyOutBytes implements marshal.CopyContext.CopyOutBytes. CopyOutBytes +// computes an appropriate address based on the current end of the +// stack. Callers use the sentinel address StackBottomMagic to marshal methods +// to indicate this. +func (s *Stack) CopyOutBytes(sentinel usermem.Addr, b []byte) (int, error) { + if sentinel != StackBottomMagic { + panic("Attempted to copy out to stack with absolute address") + } + c := len(b) + n, err := s.IO.CopyOut(context.Background(), s.Bottom-usermem.Addr(c), b, usermem.IOOpts{}) + if err == nil && n == c { s.Bottom -= usermem.Addr(n) } - - return s.Bottom, nil + return n, err } -// Pop pops the given values off the stack. -// -// (This method supports Addrs and treats them as native types.) -func (s *Stack) Pop(vals ...interface{}) (usermem.Addr, error) { - for _, v := range vals { - - vaddr, isVaddr := v.(*usermem.Addr) - - var n int - var err error - if isVaddr { - value := s.Arch.Native(uintptr(0)) - n, err = usermem.CopyObjectIn(context.Background(), s.IO, s.Bottom, value, usermem.IOOpts{}) - *vaddr = usermem.Addr(s.Arch.Value(value)) - } else { - n, err = usermem.CopyObjectIn(context.Background(), s.IO, s.Bottom, v, usermem.IOOpts{}) - } - if err != nil { - return 0, err - } - +// CopyInBytes implements marshal.CopyContext.CopyInBytes. CopyInBytes computes +// an appropriate address based on the current end of the stack. Callers must +// use the sentinel address StackBottomMagic to marshal methods to indicate +// this. +func (s *Stack) CopyInBytes(sentinel usermem.Addr, b []byte) (int, error) { + if sentinel != StackBottomMagic { + panic("Attempted to copy in from stack with absolute address") + } + n, err := s.IO.CopyIn(context.Background(), s.Bottom, b, usermem.IOOpts{}) + if err == nil { s.Bottom += usermem.Addr(n) } - - return s.Bottom, nil + return n, err } // Align aligns the stack to the given offset. @@ -142,6 +99,22 @@ func (s *Stack) Align(offset int) { } } +// PushNullTerminatedByteSlice writes bs to the stack, followed by an extra null +// byte at the end. On error, the contents of the stack and the bottom cursor +// are undefined. +func (s *Stack) PushNullTerminatedByteSlice(bs []byte) (int, error) { + // Note: Stack grows up, so write the terminal null byte first. + nNull, err := primitive.CopyUint8Out(s, StackBottomMagic, 0) + if err != nil { + return 0, err + } + n, err := primitive.CopyByteSliceOut(s, StackBottomMagic, bs) + if err != nil { + return 0, err + } + return n + nNull, nil +} + // StackLayout describes the location of the arguments and environment on the // stack. type StackLayout struct { @@ -177,11 +150,10 @@ func (s *Stack) Load(args []string, env []string, aux Auxv) (StackLayout, error) l.EnvvEnd = s.Bottom envAddrs := make([]usermem.Addr, len(env)) for i := len(env) - 1; i >= 0; i-- { - addr, err := s.Push(env[i]) - if err != nil { + if _, err := s.PushNullTerminatedByteSlice([]byte(env[i])); err != nil { return StackLayout{}, err } - envAddrs[i] = addr + envAddrs[i] = s.Bottom } l.EnvvStart = s.Bottom @@ -189,11 +161,10 @@ func (s *Stack) Load(args []string, env []string, aux Auxv) (StackLayout, error) l.ArgvEnd = s.Bottom argAddrs := make([]usermem.Addr, len(args)) for i := len(args) - 1; i >= 0; i-- { - addr, err := s.Push(args[i]) - if err != nil { + if _, err := s.PushNullTerminatedByteSlice([]byte(args[i])); err != nil { return StackLayout{}, err } - argAddrs[i] = addr + argAddrs[i] = s.Bottom } l.ArgvStart = s.Bottom @@ -222,26 +193,26 @@ func (s *Stack) Load(args []string, env []string, aux Auxv) (StackLayout, error) auxv = append(auxv, usermem.Addr(a.Key), a.Value) } auxv = append(auxv, usermem.Addr(0)) - _, err := s.Push(auxv) + _, err := s.pushAddrSliceAndTerminator(auxv) if err != nil { return StackLayout{}, err } // Push environment. - _, err = s.Push(envAddrs) + _, err = s.pushAddrSliceAndTerminator(envAddrs) if err != nil { return StackLayout{}, err } // Push args. - _, err = s.Push(argAddrs) + _, err = s.pushAddrSliceAndTerminator(argAddrs) if err != nil { return StackLayout{}, err } // Push arg count. - _, err = s.Push(usermem.Addr(len(args))) - if err != nil { + lenP := s.Arch.Native(uintptr(len(args))) + if _, err = lenP.CopyOut(s, StackBottomMagic); err != nil { return StackLayout{}, err } diff --git a/pkg/sentry/arch/stack_unsafe.go b/pkg/sentry/arch/stack_unsafe.go new file mode 100644 index 000000000..a90d297ee --- /dev/null +++ b/pkg/sentry/arch/stack_unsafe.go @@ -0,0 +1,69 @@ +// 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 arch + +import ( + "reflect" + "runtime" + "unsafe" + + "gvisor.dev/gvisor/pkg/marshal/primitive" + "gvisor.dev/gvisor/pkg/usermem" +) + +// pushAddrSliceAndTerminator copies a slices of addresses to the stack, and +// also pushes an extra null address element at the end of the slice. +// +// Internally, we unsafely transmute the slice type from the arch-dependent +// []usermem.Addr type, to a slice of fixed-sized ints so that we can pass it to +// go-marshal. +// +// On error, the contents of the stack and the bottom cursor are undefined. +func (s *Stack) pushAddrSliceAndTerminator(src []usermem.Addr) (int, error) { + // Note: Stack grows upwards, so push the terminator first. + srcHdr := (*reflect.SliceHeader)(unsafe.Pointer(&src)) + switch s.Arch.Width() { + case 8: + nNull, err := primitive.CopyUint64Out(s, StackBottomMagic, 0) + if err != nil { + return 0, err + } + var dst []uint64 + dstHdr := (*reflect.SliceHeader)(unsafe.Pointer(&dst)) + dstHdr.Data = srcHdr.Data + dstHdr.Len = srcHdr.Len + dstHdr.Cap = srcHdr.Cap + n, err := primitive.CopyUint64SliceOut(s, StackBottomMagic, dst) + // Ensures src doesn't get GCed until we're done using it through dst. + runtime.KeepAlive(src) + return n + nNull, err + case 4: + nNull, err := primitive.CopyUint32Out(s, StackBottomMagic, 0) + if err != nil { + return 0, err + } + var dst []uint32 + dstHdr := (*reflect.SliceHeader)(unsafe.Pointer(&dst)) + dstHdr.Data = srcHdr.Data + dstHdr.Len = srcHdr.Len + dstHdr.Cap = srcHdr.Cap + n, err := primitive.CopyUint32SliceOut(s, StackBottomMagic, dst) + // Ensure src doesn't get GCed until we're done using it through dst. + runtime.KeepAlive(src) + return n + nNull, err + default: + panic("Unsupported arch width") + } +} diff --git a/pkg/sentry/devices/memdev/BUILD b/pkg/sentry/devices/memdev/BUILD index abe58f818..4c8604d58 100644 --- a/pkg/sentry/devices/memdev/BUILD +++ b/pkg/sentry/devices/memdev/BUILD @@ -18,9 +18,10 @@ go_library( "//pkg/rand", "//pkg/safemem", "//pkg/sentry/fsimpl/devtmpfs", + "//pkg/sentry/fsimpl/tmpfs", + "//pkg/sentry/kernel", + "//pkg/sentry/kernel/auth", "//pkg/sentry/memmap", - "//pkg/sentry/mm", - "//pkg/sentry/pgalloc", "//pkg/sentry/vfs", "//pkg/syserror", "//pkg/usermem", diff --git a/pkg/sentry/devices/memdev/full.go b/pkg/sentry/devices/memdev/full.go index 511179e31..fece3e762 100644 --- a/pkg/sentry/devices/memdev/full.go +++ b/pkg/sentry/devices/memdev/full.go @@ -24,6 +24,8 @@ import ( const fullDevMinor = 7 // fullDevice implements vfs.Device for /dev/full. +// +// +stateify savable type fullDevice struct{} // Open implements vfs.Device.Open. @@ -38,6 +40,8 @@ func (fullDevice) Open(ctx context.Context, mnt *vfs.Mount, vfsd *vfs.Dentry, op } // fullFD implements vfs.FileDescriptionImpl for /dev/full. +// +// +stateify savable type fullFD struct { vfsfd vfs.FileDescription vfs.FileDescriptionDefaultImpl diff --git a/pkg/sentry/devices/memdev/null.go b/pkg/sentry/devices/memdev/null.go index 4918dbeeb..ff5837747 100644 --- a/pkg/sentry/devices/memdev/null.go +++ b/pkg/sentry/devices/memdev/null.go @@ -25,6 +25,8 @@ import ( const nullDevMinor = 3 // nullDevice implements vfs.Device for /dev/null. +// +// +stateify savable type nullDevice struct{} // Open implements vfs.Device.Open. @@ -39,6 +41,8 @@ func (nullDevice) Open(ctx context.Context, mnt *vfs.Mount, vfsd *vfs.Dentry, op } // nullFD implements vfs.FileDescriptionImpl for /dev/null. +// +// +stateify savable type nullFD struct { vfsfd vfs.FileDescription vfs.FileDescriptionDefaultImpl diff --git a/pkg/sentry/devices/memdev/random.go b/pkg/sentry/devices/memdev/random.go index 5e7fe0280..ac943e3ba 100644 --- a/pkg/sentry/devices/memdev/random.go +++ b/pkg/sentry/devices/memdev/random.go @@ -30,6 +30,8 @@ const ( ) // randomDevice implements vfs.Device for /dev/random and /dev/urandom. +// +// +stateify savable type randomDevice struct{} // Open implements vfs.Device.Open. @@ -44,6 +46,8 @@ func (randomDevice) Open(ctx context.Context, mnt *vfs.Mount, vfsd *vfs.Dentry, } // randomFD implements vfs.FileDescriptionImpl for /dev/random. +// +// +stateify savable type randomFD struct { vfsfd vfs.FileDescription vfs.FileDescriptionDefaultImpl diff --git a/pkg/sentry/devices/memdev/zero.go b/pkg/sentry/devices/memdev/zero.go index 2e631a252..1929e41cd 100644 --- a/pkg/sentry/devices/memdev/zero.go +++ b/pkg/sentry/devices/memdev/zero.go @@ -16,9 +16,10 @@ package memdev import ( "gvisor.dev/gvisor/pkg/context" + "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/memmap" - "gvisor.dev/gvisor/pkg/sentry/mm" - "gvisor.dev/gvisor/pkg/sentry/pgalloc" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/usermem" ) @@ -26,6 +27,8 @@ import ( const zeroDevMinor = 5 // zeroDevice implements vfs.Device for /dev/zero. +// +// +stateify savable type zeroDevice struct{} // Open implements vfs.Device.Open. @@ -40,6 +43,8 @@ func (zeroDevice) Open(ctx context.Context, mnt *vfs.Mount, vfsd *vfs.Dentry, op } // zeroFD implements vfs.FileDescriptionImpl for /dev/zero. +// +// +stateify savable type zeroFD struct { vfsfd vfs.FileDescription vfs.FileDescriptionDefaultImpl @@ -79,11 +84,22 @@ func (fd *zeroFD) Seek(ctx context.Context, offset int64, whence int32) (int64, // ConfigureMMap implements vfs.FileDescriptionImpl.ConfigureMMap. func (fd *zeroFD) ConfigureMMap(ctx context.Context, opts *memmap.MMapOpts) error { - m, err := mm.NewSharedAnonMappable(opts.Length, pgalloc.MemoryFileProviderFromContext(ctx)) + if opts.Private || !opts.MaxPerms.Write { + // This mapping will never permit writing to the "underlying file" (in + // Linux terms, it isn't VM_SHARED), so implement it as an anonymous + // mapping, but back it with fd; this is what Linux does, and is + // actually application-visible because the resulting VMA will show up + // in /proc/[pid]/maps with fd.vfsfd.VirtualDentry()'s path rather than + // "/dev/zero (deleted)". + opts.Offset = 0 + opts.MappingIdentity = &fd.vfsfd + opts.MappingIdentity.IncRef() + return nil + } + tmpfsFD, err := tmpfs.NewZeroFile(ctx, auth.CredentialsFromContext(ctx), kernel.KernelFromContext(ctx).ShmMount(), opts.Length) if err != nil { return err } - opts.MappingIdentity = m - opts.Mappable = m - return nil + defer tmpfsFD.DecRef(ctx) + return tmpfsFD.ConfigureMMap(ctx, opts) } diff --git a/pkg/sentry/devices/ttydev/ttydev.go b/pkg/sentry/devices/ttydev/ttydev.go index 664e54498..a287c65ca 100644 --- a/pkg/sentry/devices/ttydev/ttydev.go +++ b/pkg/sentry/devices/ttydev/ttydev.go @@ -30,6 +30,8 @@ const ( ) // ttyDevice implements vfs.Device for /dev/tty. +// +// +stateify savable type ttyDevice struct{} // Open implements vfs.Device.Open. diff --git a/pkg/sentry/fs/g3doc/fuse.md b/pkg/sentry/fs/g3doc/fuse.md index 2ca84dd74..05e043583 100644 --- a/pkg/sentry/fs/g3doc/fuse.md +++ b/pkg/sentry/fs/g3doc/fuse.md @@ -79,7 +79,7 @@ ops can be implemented in parallel. - Implement `/dev/fuse` - a character device used to establish an FD for communication between the sentry and the server daemon. -- Implement basic FUSE ops like `FUSE_INIT`, `FUSE_DESTROY`. +- Implement basic FUSE ops like `FUSE_INIT`. #### Read-only mount with basic file operations @@ -95,6 +95,103 @@ ops can be implemented in parallel. - Implement the remaining FUSE ops and decide if we can omit rarely used operations like ioctl. +### Design Details + +#### Lifecycle for a FUSE Request + +- User invokes a syscall +- Sentry prepares corresponding request + - If FUSE device is available + - Write the request in binary + - If FUSE device is full + - Kernel task blocked until available +- Sentry notifies the readers of fuse device that it's ready for read +- FUSE daemon reads the request and processes it +- Sentry waits until a reply is written to the FUSE device + - but returns directly for async requests +- FUSE daemon writes to the fuse device +- Sentry processes the reply + - For sync requests, unblock blocked kernel task + - For async requests, execute pre-specified callback if any +- Sentry returns the syscall to the user + +#### Channels and Queues for Requests in Different Stages + +`connection.initializedChan` + +- a channel that the requests issued before connection initialization blocks + on. + +`fd.queue` + +- a queue of requests that haven’t been read by the FUSE daemon yet. + +`fd.completions` + +- a map of the requests that have been prepared but not yet received a + response, including the ones on the `fd.queue`. + +`fd.waitQueue` + +- a queue of waiters that is waiting for the fuse device fd to be available, + such as the FUSE daemon. + +`fd.fullQueueCh` + +- a channel that the kernel task will be blocked on when the fd is not + available. + +#### Basic I/O Implementation + +Currently we have implemented basic functionalities of read and write for our +FUSE. We describe the design and ways to improve it here: + +##### Basic FUSE Read + +The vfs2 expects implementations of `vfs.FileDescriptionImpl.Read()` and +`vfs.FileDescriptionImpl.PRead()`. When a syscall is made, it will eventually +reach our implementation of those interface functions located at +`pkg/sentry/fsimpl/fuse/regular_file.go` for regular files. + +After validation checks of the input, sentry sends `FUSE_READ` requests to the +FUSE daemon. The FUSE daemon returns data after the `fuse_out_header` as the +responses. For the first version, we create a copy in kernel memory of those +data. They are represented as a byte slice in the marshalled struct. This +happens as a common process for all the FUSE responses at this moment at +`pkg/sentry/fsimpl/fuse/dev.go:writeLocked()`. We then directly copy from this +intermediate buffer to the input buffer provided by the read syscall. + +There is an extra requirement for FUSE: When mounting the FUSE fs, the mounter +or the FUSE daemon can specify a `max_read` or a `max_pages` parameter. They are +the upperbound of the bytes to read in each `FUSE_READ` request. We implemented +the code to handle the fragmented reads. + +To improve the performance: ideally we should have buffer cache to copy those +data from the responses of FUSE daemon into, as is also the design of several +other existing file system implementations for sentry, instead of a single-use +temporary buffer. Directly mapping the memory of one process to another could +also boost the performance, but to keep them isolated, we did not choose to do +so. + +##### Basic FUSE Write + +The vfs2 invokes implementations of `vfs.FileDescriptionImpl.Write()` and +`vfs.FileDescriptionImpl.PWrite()` on the regular file descriptor of FUSE when a +user makes write(2) and pwrite(2) syscall. + +For valid writes, sentry sends the bytes to write after a `FUSE_WRITE` header +(can be regarded as a request with 2 payloads) to the FUSE daemon. For the first +version, we allocate a buffer inside kernel memory to store the bytes from the +user, and copy directly from that buffer to the memory of FUSE daemon. This +happens at `pkg/sentry/fsimpl/fuse/dev.go:readLocked()` + +The parameters `max_write` and `max_pages` restrict the number of bytes in one +`FUSE_WRITE`. There are code handling fragmented writes in current +implementation. + +To have better performance: the extra copy created to store the bytes to write +can be replaced by the buffer cache as well. + # Appendix ## FUSE Protocol diff --git a/pkg/sentry/fs/proc/BUILD b/pkg/sentry/fs/proc/BUILD index 77c2c5c0e..b8b2281a8 100644 --- a/pkg/sentry/fs/proc/BUILD +++ b/pkg/sentry/fs/proc/BUILD @@ -50,6 +50,7 @@ go_library( "//pkg/sync", "//pkg/syserror", "//pkg/tcpip/header", + "//pkg/tcpip/network/ipv4", "//pkg/usermem", "//pkg/waiter", ], diff --git a/pkg/sentry/fs/proc/sys_net.go b/pkg/sentry/fs/proc/sys_net.go index 8615b60f0..e555672ad 100644 --- a/pkg/sentry/fs/proc/sys_net.go +++ b/pkg/sentry/fs/proc/sys_net.go @@ -26,6 +26,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/fs/ramfs" "gvisor.dev/gvisor/pkg/sentry/inet" "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/pkg/tcpip/network/ipv4" "gvisor.dev/gvisor/pkg/usermem" "gvisor.dev/gvisor/pkg/waiter" ) @@ -54,7 +55,7 @@ type tcpMemInode struct { // size stores the tcp buffer size during save, and sets the buffer // size in netstack in restore. We must save/restore this here, since - // netstack itself is stateless. + // a netstack instance is created on restore. size inet.TCPBufferSize // mu protects against concurrent reads/writes to files based on this @@ -258,6 +259,9 @@ func (f *tcpSackFile) Write(ctx context.Context, _ *fs.File, src usermem.IOSeque if src.NumBytes() == 0 { return 0, nil } + + // Only consider size of one memory page for input for performance reasons. + // We are only reading if it's zero or not anyway. src = src.TakeFirst(usermem.PageSize - 1) var v int32 @@ -383,11 +387,125 @@ func (p *proc) newSysNetCore(ctx context.Context, msrc *fs.MountSource, s inet.S return newProcInode(ctx, d, msrc, fs.SpecialDirectory, nil) } +// ipForwarding implements fs.InodeOperations. +// +// ipForwarding is used to enable/disable packet forwarding of netstack. +// +// +stateify savable +type ipForwarding struct { + fsutil.SimpleFileInode + + stack inet.Stack `state:"wait"` + + // enabled stores the IPv4 forwarding state on save. + // We must save/restore this here, since a netstack instance + // is created on restore. + enabled *bool +} + +func newIPForwardingInode(ctx context.Context, msrc *fs.MountSource, s inet.Stack) *fs.Inode { + ipf := &ipForwarding{ + SimpleFileInode: *fsutil.NewSimpleFileInode(ctx, fs.RootOwner, fs.FilePermsFromMode(0444), linux.PROC_SUPER_MAGIC), + stack: s, + } + sattr := fs.StableAttr{ + DeviceID: device.ProcDevice.DeviceID(), + InodeID: device.ProcDevice.NextIno(), + BlockSize: usermem.PageSize, + Type: fs.SpecialFile, + } + return fs.NewInode(ctx, ipf, msrc, sattr) +} + +// Truncate implements fs.InodeOperations.Truncate. Truncate is called when +// O_TRUNC is specified for any kind of existing Dirent but is not called via +// (f)truncate for proc files. +func (*ipForwarding) Truncate(context.Context, *fs.Inode, int64) error { + return nil +} + +// +stateify savable +type ipForwardingFile struct { + fsutil.FileGenericSeek `state:"nosave"` + fsutil.FileNoIoctl `state:"nosave"` + fsutil.FileNoMMap `state:"nosave"` + fsutil.FileNoSplice `state:"nosave"` + fsutil.FileNoopFlush `state:"nosave"` + fsutil.FileNoopFsync `state:"nosave"` + fsutil.FileNoopRelease `state:"nosave"` + fsutil.FileNotDirReaddir `state:"nosave"` + fsutil.FileUseInodeUnstableAttr `state:"nosave"` + waiter.AlwaysReady `state:"nosave"` + + ipf *ipForwarding + + stack inet.Stack `state:"wait"` +} + +// GetFile implements fs.InodeOperations.GetFile. +func (ipf *ipForwarding) GetFile(ctx context.Context, dirent *fs.Dirent, flags fs.FileFlags) (*fs.File, error) { + flags.Pread = true + flags.Pwrite = true + return fs.NewFile(ctx, dirent, flags, &ipForwardingFile{ + stack: ipf.stack, + ipf: ipf, + }), nil +} + +// Read implements fs.FileOperations.Read. +func (f *ipForwardingFile) Read(ctx context.Context, _ *fs.File, dst usermem.IOSequence, offset int64) (int64, error) { + if offset != 0 { + return 0, io.EOF + } + + if f.ipf.enabled == nil { + enabled := f.stack.Forwarding(ipv4.ProtocolNumber) + f.ipf.enabled = &enabled + } + + val := "0\n" + if *f.ipf.enabled { + // Technically, this is not quite compatible with Linux. Linux + // stores these as an integer, so if you write "2" into + // ip_forward, you should get 2 back. + val = "1\n" + } + n, err := dst.CopyOut(ctx, []byte(val)) + return int64(n), err +} + +// Write implements fs.FileOperations.Write. +// +// Offset is ignored, multiple writes are not supported. +func (f *ipForwardingFile) Write(ctx context.Context, _ *fs.File, src usermem.IOSequence, offset int64) (int64, error) { + if src.NumBytes() == 0 { + return 0, nil + } + + // Only consider size of one memory page for input for performance reasons. + // We are only reading if it's zero or not anyway. + src = src.TakeFirst(usermem.PageSize - 1) + + var v int32 + n, err := usermem.CopyInt32StringInVec(ctx, src.IO, src.Addrs, &v, src.Opts) + if err != nil { + return n, err + } + if f.ipf.enabled == nil { + f.ipf.enabled = new(bool) + } + *f.ipf.enabled = v != 0 + return n, f.stack.SetForwarding(ipv4.ProtocolNumber, *f.ipf.enabled) +} + func (p *proc) newSysNetIPv4Dir(ctx context.Context, msrc *fs.MountSource, s inet.Stack) *fs.Inode { contents := map[string]*fs.Inode{ // Add tcp_sack. "tcp_sack": newTCPSackInode(ctx, msrc, s), + // Add ip_forward. + "ip_forward": newIPForwardingInode(ctx, msrc, s), + // The following files are simple stubs until they are // implemented in netstack, most of these files are // configuration related. We use the value closest to the diff --git a/pkg/sentry/fs/proc/sys_net_state.go b/pkg/sentry/fs/proc/sys_net_state.go index 6eba709c6..4cb4741af 100644 --- a/pkg/sentry/fs/proc/sys_net_state.go +++ b/pkg/sentry/fs/proc/sys_net_state.go @@ -14,7 +14,11 @@ package proc -import "fmt" +import ( + "fmt" + + "gvisor.dev/gvisor/pkg/tcpip/network/ipv4" +) // beforeSave is invoked by stateify. func (t *tcpMemInode) beforeSave() { @@ -40,3 +44,12 @@ func (s *tcpSack) afterLoad() { } } } + +// afterLoad is invoked by stateify. +func (ipf *ipForwarding) afterLoad() { + if ipf.enabled != nil { + if err := ipf.stack.SetForwarding(ipv4.ProtocolNumber, *ipf.enabled); err != nil { + panic(fmt.Sprintf("failed to set IPv4 forwarding [%v]: %v", *ipf.enabled, err)) + } + } +} diff --git a/pkg/sentry/fs/proc/sys_net_test.go b/pkg/sentry/fs/proc/sys_net_test.go index 355e83d47..6ef5738e7 100644 --- a/pkg/sentry/fs/proc/sys_net_test.go +++ b/pkg/sentry/fs/proc/sys_net_test.go @@ -123,3 +123,76 @@ func TestConfigureRecvBufferSize(t *testing.T) { } } } + +// TestIPForwarding tests the implementation of +// /proc/sys/net/ipv4/ip_forwarding +func TestIPForwarding(t *testing.T) { + ctx := context.Background() + s := inet.NewTestStack() + + var cases = []struct { + comment string + initial bool + str string + final bool + }{ + { + comment: `Forwarding is disabled; write 1 and enable forwarding`, + initial: false, + str: "1", + final: true, + }, + { + comment: `Forwarding is disabled; write 0 and disable forwarding`, + initial: false, + str: "0", + final: false, + }, + { + comment: `Forwarding is enabled; write 1 and enable forwarding`, + initial: true, + str: "1", + final: true, + }, + { + comment: `Forwarding is enabled; write 0 and disable forwarding`, + initial: true, + str: "0", + final: false, + }, + { + comment: `Forwarding is disabled; write 2404 and enable forwarding`, + initial: false, + str: "2404", + final: true, + }, + { + comment: `Forwarding is enabled; write 2404 and enable forwarding`, + initial: true, + str: "2404", + final: true, + }, + } + for _, c := range cases { + t.Run(c.comment, func(t *testing.T) { + s.IPForwarding = c.initial + ipf := &ipForwarding{stack: s} + file := &ipForwardingFile{ + stack: s, + ipf: ipf, + } + + // Write the values. + src := usermem.BytesIOSequence([]byte(c.str)) + if n, err := file.Write(ctx, nil, src, 0); n != int64(len(c.str)) || err != nil { + t.Errorf("file.Write(ctx, nil, %q, 0) = (%d, %v); want (%d, nil)", c.str, n, err, len(c.str)) + } + + // Read the values from the stack and check them. + if got, want := s.IPForwarding, c.final; got != want { + t.Errorf("s.IPForwarding incorrect; got: %v, want: %v", got, want) + } + + }) + } +} diff --git a/pkg/sentry/fs/proc/task.go b/pkg/sentry/fs/proc/task.go index 103bfc600..22d658acf 100644 --- a/pkg/sentry/fs/proc/task.go +++ b/pkg/sentry/fs/proc/task.go @@ -84,6 +84,7 @@ func (p *proc) newTaskDir(t *kernel.Task, msrc *fs.MountSource, isThreadGroup bo "auxv": newAuxvec(t, msrc), "cmdline": newExecArgInode(t, msrc, cmdlineExecArg), "comm": newComm(t, msrc), + "cwd": newCwd(t, msrc), "environ": newExecArgInode(t, msrc, environExecArg), "exe": newExe(t, msrc), "fd": newFdDir(t, msrc), @@ -300,6 +301,49 @@ func (e *exe) Readlink(ctx context.Context, inode *fs.Inode) (string, error) { return exec.PathnameWithDeleted(ctx), nil } +// cwd is an fs.InodeOperations symlink for the /proc/PID/cwd file. +// +// +stateify savable +type cwd struct { + ramfs.Symlink + + t *kernel.Task +} + +func newCwd(t *kernel.Task, msrc *fs.MountSource) *fs.Inode { + cwdSymlink := &cwd{ + Symlink: *ramfs.NewSymlink(t, fs.RootOwner, ""), + t: t, + } + return newProcInode(t, cwdSymlink, msrc, fs.Symlink, t) +} + +// Readlink implements fs.InodeOperations. +func (e *cwd) Readlink(ctx context.Context, inode *fs.Inode) (string, error) { + if !kernel.ContextCanTrace(ctx, e.t, false) { + return "", syserror.EACCES + } + if err := checkTaskState(e.t); err != nil { + return "", err + } + cwd := e.t.FSContext().WorkingDirectory() + if cwd == nil { + // It could have raced with process deletion. + return "", syserror.ESRCH + } + defer cwd.DecRef(ctx) + + root := fs.RootFromContext(ctx) + if root == nil { + // It could have raced with process deletion. + return "", syserror.ESRCH + } + defer root.DecRef(ctx) + + name, _ := cwd.FullName(root) + return name, nil +} + // namespaceSymlink represents a symlink in the namespacefs, such as the files // in /proc/<pid>/ns. // diff --git a/pkg/sentry/fsbridge/vfs.go b/pkg/sentry/fsbridge/vfs.go index 323506d33..be0900030 100644 --- a/pkg/sentry/fsbridge/vfs.go +++ b/pkg/sentry/fsbridge/vfs.go @@ -122,7 +122,7 @@ func NewVFSLookup(mntns *vfs.MountNamespace, root, workingDir vfs.VirtualDentry) // remainingTraversals is not configurable in VFS2, all callers are using the // default anyways. func (l *vfsLookup) OpenPath(ctx context.Context, pathname string, opts vfs.OpenOptions, _ *uint, resolveFinal bool) (File, error) { - vfsObj := l.mntns.Root().Mount().Filesystem().VirtualFilesystem() + vfsObj := l.root.Mount().Filesystem().VirtualFilesystem() creds := auth.CredentialsFromContext(ctx) path := fspath.Parse(pathname) pop := &vfs.PathOperation{ diff --git a/pkg/sentry/fsimpl/devpts/devpts.go b/pkg/sentry/fsimpl/devpts/devpts.go index dcf1ee25b..903135fae 100644 --- a/pkg/sentry/fsimpl/devpts/devpts.go +++ b/pkg/sentry/fsimpl/devpts/devpts.go @@ -35,6 +35,8 @@ import ( const Name = "devpts" // FilesystemType implements vfs.FilesystemType. +// +// +stateify savable type FilesystemType struct{} // Name implements vfs.FilesystemType.Name. @@ -58,6 +60,7 @@ func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virt return fs.Filesystem.VFSFilesystem(), root.VFSDentry(), nil } +// +stateify savable type filesystem struct { kernfs.Filesystem @@ -110,6 +113,8 @@ func (fs *filesystem) Release(ctx context.Context) { } // rootInode is the root directory inode for the devpts mounts. +// +// +stateify savable type rootInode struct { implStatFS kernfs.AlwaysValid @@ -131,7 +136,7 @@ type rootInode struct { root *rootInode // mu protects the fields below. - mu sync.Mutex + mu sync.Mutex `state:"nosave"` // replicas maps pty ids to replica inodes. replicas map[uint32]*replicaInode @@ -187,8 +192,8 @@ func (i *rootInode) masterClose(t *Terminal) { } // 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, &i.locks, &opts, kernfs.GenericDirectoryFDOptions{ +func (i *rootInode) Open(ctx context.Context, rp *vfs.ResolvingPath, d *kernfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), d, &i.OrderedChildren, &i.locks, &opts, kernfs.GenericDirectoryFDOptions{ SeekEnd: kernfs.SeekEndStaticEntries, }) if err != nil { @@ -198,7 +203,7 @@ func (i *rootInode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.D } // Lookup implements kernfs.Inode.Lookup. -func (i *rootInode) Lookup(ctx context.Context, name string) (*vfs.Dentry, error) { +func (i *rootInode) Lookup(ctx context.Context, name string) (*kernfs.Dentry, error) { idx, err := strconv.ParseUint(name, 10, 32) if err != nil { return nil, syserror.ENOENT @@ -207,7 +212,7 @@ func (i *rootInode) Lookup(ctx context.Context, name string) (*vfs.Dentry, error defer i.mu.Unlock() if si, ok := i.replicas[uint32(idx)]; ok { si.dentry.IncRef() - return si.dentry.VFSDentry(), nil + return &si.dentry, nil } return nil, syserror.ENOENT @@ -237,11 +242,12 @@ func (i *rootInode) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback, return offset, nil } -// DecRef implements kernfs.Inode. +// DecRef implements kernfs.Inode.DecRef. func (i *rootInode) DecRef(context.Context) { i.rootInodeRefs.DecRef(i.Destroy) } +// +stateify savable type implStatFS struct{} // StatFS implements kernfs.Inode.StatFS. diff --git a/pkg/sentry/fsimpl/devpts/master.go b/pkg/sentry/fsimpl/devpts/master.go index 83d790b38..69c2fe951 100644 --- a/pkg/sentry/fsimpl/devpts/master.go +++ b/pkg/sentry/fsimpl/devpts/master.go @@ -31,6 +31,8 @@ import ( ) // masterInode is the inode for the master end of the Terminal. +// +// +stateify savable type masterInode struct { implStatFS kernfs.InodeAttrs @@ -50,20 +52,18 @@ type masterInode struct { 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) { +func (mi *masterInode) Open(ctx context.Context, rp *vfs.ResolvingPath, d *kernfs.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, } fd.LockFD.Init(&mi.locks) - if err := fd.vfsfd.Init(fd, opts.Flags, rp.Mount(), vfsd, &vfs.FileDescriptionOptions{}); err != nil { - mi.DecRef(ctx) + if err := fd.vfsfd.Init(fd, opts.Flags, rp.Mount(), d.VFSDentry(), &vfs.FileDescriptionOptions{}); err != nil { return nil, err } return &fd.vfsfd, nil @@ -89,6 +89,7 @@ func (mi *masterInode) SetStat(ctx context.Context, vfsfs *vfs.Filesystem, creds return mi.InodeAttrs.SetStat(ctx, vfsfs, creds, opts) } +// +stateify savable type masterFileDescription struct { vfsfd vfs.FileDescription vfs.FileDescriptionDefaultImpl @@ -103,7 +104,6 @@ var _ vfs.FileDescriptionImpl = (*masterFileDescription)(nil) // Release implements vfs.FileDescriptionImpl.Release. func (mfd *masterFileDescription) Release(ctx context.Context) { mfd.inode.root.masterClose(mfd.t) - mfd.inode.DecRef(ctx) } // EventRegister implements waiter.Waitable.EventRegister. diff --git a/pkg/sentry/fsimpl/devpts/replica.go b/pkg/sentry/fsimpl/devpts/replica.go index 58f6c1d3a..6515c5536 100644 --- a/pkg/sentry/fsimpl/devpts/replica.go +++ b/pkg/sentry/fsimpl/devpts/replica.go @@ -30,6 +30,8 @@ import ( ) // replicaInode is the inode for the replica end of the Terminal. +// +// +stateify savable type replicaInode struct { implStatFS kernfs.InodeAttrs @@ -52,14 +54,12 @@ type replicaInode struct { var _ kernfs.Inode = (*replicaInode)(nil) // Open implements kernfs.Inode.Open. -func (si *replicaInode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { - si.IncRef() +func (ri *replicaInode) Open(ctx context.Context, rp *vfs.ResolvingPath, d *kernfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { fd := &replicaFileDescription{ - inode: si, + inode: ri, } - fd.LockFD.Init(&si.locks) - if err := fd.vfsfd.Init(fd, opts.Flags, rp.Mount(), vfsd, &vfs.FileDescriptionOptions{}); err != nil { - si.DecRef(ctx) + fd.LockFD.Init(&ri.locks) + if err := fd.vfsfd.Init(fd, opts.Flags, rp.Mount(), d.VFSDentry(), &vfs.FileDescriptionOptions{}); err != nil { return nil, err } return &fd.vfsfd, nil @@ -67,34 +67,35 @@ func (si *replicaInode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *v } // Valid implements kernfs.Inode.Valid. -func (si *replicaInode) Valid(context.Context) bool { +func (ri *replicaInode) Valid(context.Context) bool { // Return valid if the replica still exists. - si.root.mu.Lock() - defer si.root.mu.Unlock() - _, ok := si.root.replicas[si.t.n] + ri.root.mu.Lock() + defer ri.root.mu.Unlock() + _, ok := ri.root.replicas[ri.t.n] return ok } // Stat implements kernfs.Inode.Stat. -func (si *replicaInode) Stat(ctx context.Context, vfsfs *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) { - statx, err := si.InodeAttrs.Stat(ctx, vfsfs, opts) +func (ri *replicaInode) Stat(ctx context.Context, vfsfs *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) { + statx, err := ri.InodeAttrs.Stat(ctx, vfsfs, opts) if err != nil { return linux.Statx{}, err } statx.Blksize = 1024 statx.RdevMajor = linux.UNIX98_PTY_REPLICA_MAJOR - statx.RdevMinor = si.t.n + statx.RdevMinor = ri.t.n return statx, nil } // SetStat implements kernfs.Inode.SetStat -func (si *replicaInode) SetStat(ctx context.Context, vfsfs *vfs.Filesystem, creds *auth.Credentials, opts vfs.SetStatOptions) error { +func (ri *replicaInode) 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) + return ri.InodeAttrs.SetStat(ctx, vfsfs, creds, opts) } +// +stateify savable type replicaFileDescription struct { vfsfd vfs.FileDescription vfs.FileDescriptionDefaultImpl @@ -106,37 +107,35 @@ type replicaFileDescription struct { var _ vfs.FileDescriptionImpl = (*replicaFileDescription)(nil) // Release implements fs.FileOperations.Release. -func (sfd *replicaFileDescription) Release(ctx context.Context) { - sfd.inode.DecRef(ctx) -} +func (rfd *replicaFileDescription) Release(ctx context.Context) {} // EventRegister implements waiter.Waitable.EventRegister. -func (sfd *replicaFileDescription) EventRegister(e *waiter.Entry, mask waiter.EventMask) { - sfd.inode.t.ld.replicaWaiter.EventRegister(e, mask) +func (rfd *replicaFileDescription) EventRegister(e *waiter.Entry, mask waiter.EventMask) { + rfd.inode.t.ld.replicaWaiter.EventRegister(e, mask) } // EventUnregister implements waiter.Waitable.EventUnregister. -func (sfd *replicaFileDescription) EventUnregister(e *waiter.Entry) { - sfd.inode.t.ld.replicaWaiter.EventUnregister(e) +func (rfd *replicaFileDescription) EventUnregister(e *waiter.Entry) { + rfd.inode.t.ld.replicaWaiter.EventUnregister(e) } // Readiness implements waiter.Waitable.Readiness. -func (sfd *replicaFileDescription) Readiness(mask waiter.EventMask) waiter.EventMask { - return sfd.inode.t.ld.replicaReadiness() +func (rfd *replicaFileDescription) Readiness(mask waiter.EventMask) waiter.EventMask { + return rfd.inode.t.ld.replicaReadiness() } // Read implements vfs.FileDescriptionImpl.Read. -func (sfd *replicaFileDescription) Read(ctx context.Context, dst usermem.IOSequence, _ vfs.ReadOptions) (int64, error) { - return sfd.inode.t.ld.inputQueueRead(ctx, dst) +func (rfd *replicaFileDescription) Read(ctx context.Context, dst usermem.IOSequence, _ vfs.ReadOptions) (int64, error) { + return rfd.inode.t.ld.inputQueueRead(ctx, dst) } // Write implements vfs.FileDescriptionImpl.Write. -func (sfd *replicaFileDescription) Write(ctx context.Context, src usermem.IOSequence, _ vfs.WriteOptions) (int64, error) { - return sfd.inode.t.ld.outputQueueWrite(ctx, src) +func (rfd *replicaFileDescription) Write(ctx context.Context, src usermem.IOSequence, _ vfs.WriteOptions) (int64, error) { + return rfd.inode.t.ld.outputQueueWrite(ctx, src) } // Ioctl implements vfs.FileDescriptionImpl.Ioctl. -func (sfd *replicaFileDescription) Ioctl(ctx context.Context, io usermem.IO, args arch.SyscallArguments) (uintptr, error) { +func (rfd *replicaFileDescription) Ioctl(ctx context.Context, io usermem.IO, args arch.SyscallArguments) (uintptr, error) { t := kernel.TaskFromContext(ctx) if t == nil { // ioctl(2) may only be called from a task goroutine. @@ -146,35 +145,35 @@ func (sfd *replicaFileDescription) Ioctl(ctx context.Context, io usermem.IO, arg 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(t, io, args) + return 0, rfd.inode.t.ld.inputQueueReadSize(t, io, args) case linux.TCGETS: - return sfd.inode.t.ld.getTermios(t, args) + return rfd.inode.t.ld.getTermios(t, args) case linux.TCSETS: - return sfd.inode.t.ld.setTermios(t, args) + return rfd.inode.t.ld.setTermios(t, args) case linux.TCSETSW: // TODO(b/29356795): This should drain the output queue first. - return sfd.inode.t.ld.setTermios(t, args) + return rfd.inode.t.ld.setTermios(t, args) case linux.TIOCGPTN: - nP := primitive.Uint32(sfd.inode.t.n) + nP := primitive.Uint32(rfd.inode.t.n) _, err := nP.CopyOut(t, args[2].Pointer()) return 0, err case linux.TIOCGWINSZ: - return 0, sfd.inode.t.ld.windowSize(t, args) + return 0, rfd.inode.t.ld.windowSize(t, args) case linux.TIOCSWINSZ: - return 0, sfd.inode.t.ld.setWindowSize(t, args) + return 0, rfd.inode.t.ld.setWindowSize(t, args) case linux.TIOCSCTTY: // Make the given terminal the controlling terminal of the // calling process. - return 0, sfd.inode.t.setControllingTTY(ctx, args, false /* isMaster */) + return 0, rfd.inode.t.setControllingTTY(ctx, args, false /* isMaster */) case linux.TIOCNOTTY: // Release this process's controlling terminal. - return 0, sfd.inode.t.releaseControllingTTY(ctx, args, false /* isMaster */) + return 0, rfd.inode.t.releaseControllingTTY(ctx, args, false /* isMaster */) case linux.TIOCGPGRP: // Get the foreground process group. - return sfd.inode.t.foregroundProcessGroup(ctx, args, false /* isMaster */) + return rfd.inode.t.foregroundProcessGroup(ctx, args, false /* isMaster */) case linux.TIOCSPGRP: // Set the foreground process group. - return sfd.inode.t.setForegroundProcessGroup(ctx, args, false /* isMaster */) + return rfd.inode.t.setForegroundProcessGroup(ctx, args, false /* isMaster */) default: maybeEmitUnimplementedEvent(ctx, cmd) return 0, syserror.ENOTTY @@ -182,24 +181,24 @@ func (sfd *replicaFileDescription) Ioctl(ctx context.Context, io usermem.IO, arg } // SetStat implements vfs.FileDescriptionImpl.SetStat. -func (sfd *replicaFileDescription) SetStat(ctx context.Context, opts vfs.SetStatOptions) error { +func (rfd *replicaFileDescription) 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) + fs := rfd.vfsfd.VirtualDentry().Mount().Filesystem() + return rfd.inode.SetStat(ctx, fs, creds, opts) } // Stat implements vfs.FileDescriptionImpl.Stat. -func (sfd *replicaFileDescription) Stat(ctx context.Context, opts vfs.StatOptions) (linux.Statx, error) { - fs := sfd.vfsfd.VirtualDentry().Mount().Filesystem() - return sfd.inode.Stat(ctx, fs, opts) +func (rfd *replicaFileDescription) Stat(ctx context.Context, opts vfs.StatOptions) (linux.Statx, error) { + fs := rfd.vfsfd.VirtualDentry().Mount().Filesystem() + return rfd.inode.Stat(ctx, fs, opts) } // LockPOSIX implements vfs.FileDescriptionImpl.LockPOSIX. -func (sfd *replicaFileDescription) LockPOSIX(ctx context.Context, uid fslock.UniqueID, t fslock.LockType, start, length uint64, whence int16, block fslock.Blocker) error { - return sfd.Locks().LockPOSIX(ctx, &sfd.vfsfd, uid, t, start, length, whence, block) +func (rfd *replicaFileDescription) LockPOSIX(ctx context.Context, uid fslock.UniqueID, t fslock.LockType, start, length uint64, whence int16, block fslock.Blocker) error { + return rfd.Locks().LockPOSIX(ctx, &rfd.vfsfd, uid, t, start, length, whence, block) } // UnlockPOSIX implements vfs.FileDescriptionImpl.UnlockPOSIX. -func (sfd *replicaFileDescription) UnlockPOSIX(ctx context.Context, uid fslock.UniqueID, start, length uint64, whence int16) error { - return sfd.Locks().UnlockPOSIX(ctx, &sfd.vfsfd, uid, start, length, whence) +func (rfd *replicaFileDescription) UnlockPOSIX(ctx context.Context, uid fslock.UniqueID, start, length uint64, whence int16) error { + return rfd.Locks().UnlockPOSIX(ctx, &rfd.vfsfd, uid, start, length, whence) } diff --git a/pkg/sentry/fsimpl/devtmpfs/devtmpfs.go b/pkg/sentry/fsimpl/devtmpfs/devtmpfs.go index a23094e54..6d1753080 100644 --- a/pkg/sentry/fsimpl/devtmpfs/devtmpfs.go +++ b/pkg/sentry/fsimpl/devtmpfs/devtmpfs.go @@ -33,8 +33,10 @@ import ( const Name = "devtmpfs" // FilesystemType implements vfs.FilesystemType. +// +// +stateify savable type FilesystemType struct { - initOnce sync.Once + initOnce sync.Once `state:"nosave"` // FIXME(gvisor.dev/issue/1664): not yet supported. initErr error // fs is the tmpfs filesystem that backs all mounts of this FilesystemType. diff --git a/pkg/sentry/fsimpl/eventfd/eventfd.go b/pkg/sentry/fsimpl/eventfd/eventfd.go index 812171fa3..1c27ad700 100644 --- a/pkg/sentry/fsimpl/eventfd/eventfd.go +++ b/pkg/sentry/fsimpl/eventfd/eventfd.go @@ -30,9 +30,11 @@ import ( "gvisor.dev/gvisor/pkg/waiter" ) -// EventFileDescription implements FileDescriptionImpl for file-based event +// EventFileDescription implements vfs.FileDescriptionImpl for file-based event // notification (eventfd). Eventfds are usually internal to the Sentry but in // certain situations they may be converted into a host-backed eventfd. +// +// +stateify savable type EventFileDescription struct { vfsfd vfs.FileDescription vfs.FileDescriptionDefaultImpl @@ -106,7 +108,7 @@ func (efd *EventFileDescription) HostFD() (int, error) { return efd.hostfd, nil } -// Release implements FileDescriptionImpl.Release() +// Release implements vfs.FileDescriptionImpl.Release. func (efd *EventFileDescription) Release(context.Context) { efd.mu.Lock() defer efd.mu.Unlock() @@ -119,7 +121,7 @@ func (efd *EventFileDescription) Release(context.Context) { } } -// Read implements FileDescriptionImpl.Read. +// Read implements vfs.FileDescriptionImpl.Read. func (efd *EventFileDescription) Read(ctx context.Context, dst usermem.IOSequence, _ vfs.ReadOptions) (int64, error) { if dst.NumBytes() < 8 { return 0, syscall.EINVAL @@ -130,7 +132,7 @@ func (efd *EventFileDescription) Read(ctx context.Context, dst usermem.IOSequenc return 8, nil } -// Write implements FileDescriptionImpl.Write. +// Write implements vfs.FileDescriptionImpl.Write. func (efd *EventFileDescription) Write(ctx context.Context, src usermem.IOSequence, _ vfs.WriteOptions) (int64, error) { if src.NumBytes() < 8 { return 0, syscall.EINVAL diff --git a/pkg/sentry/fsimpl/ext/BUILD b/pkg/sentry/fsimpl/ext/BUILD index abc610ef3..7b1eec3da 100644 --- a/pkg/sentry/fsimpl/ext/BUILD +++ b/pkg/sentry/fsimpl/ext/BUILD @@ -51,6 +51,8 @@ go_library( "//pkg/fd", "//pkg/fspath", "//pkg/log", + "//pkg/marshal", + "//pkg/marshal/primitive", "//pkg/safemem", "//pkg/sentry/arch", "//pkg/sentry/fs", @@ -86,9 +88,9 @@ go_test( library = ":ext", deps = [ "//pkg/abi/linux", - "//pkg/binary", "//pkg/context", "//pkg/fspath", + "//pkg/marshal/primitive", "//pkg/sentry/contexttest", "//pkg/sentry/fsimpl/ext/disklayout", "//pkg/sentry/kernel/auth", diff --git a/pkg/sentry/fsimpl/ext/block_map_file.go b/pkg/sentry/fsimpl/ext/block_map_file.go index 8bb104ff0..1165234f9 100644 --- a/pkg/sentry/fsimpl/ext/block_map_file.go +++ b/pkg/sentry/fsimpl/ext/block_map_file.go @@ -18,7 +18,7 @@ import ( "io" "math" - "gvisor.dev/gvisor/pkg/binary" + "gvisor.dev/gvisor/pkg/marshal/primitive" "gvisor.dev/gvisor/pkg/syserror" ) @@ -34,19 +34,19 @@ type blockMapFile struct { // directBlks are the direct blocks numbers. The physical blocks pointed by // these holds file data. Contains file blocks 0 to 11. - directBlks [numDirectBlks]uint32 + directBlks [numDirectBlks]primitive.Uint32 // indirectBlk is the physical block which contains (blkSize/4) direct block // numbers (as uint32 integers). - indirectBlk uint32 + indirectBlk primitive.Uint32 // doubleIndirectBlk is the physical block which contains (blkSize/4) indirect // block numbers (as uint32 integers). - doubleIndirectBlk uint32 + doubleIndirectBlk primitive.Uint32 // tripleIndirectBlk is the physical block which contains (blkSize/4) doubly // indirect block numbers (as uint32 integers). - tripleIndirectBlk uint32 + tripleIndirectBlk primitive.Uint32 // coverage at (i)th index indicates the amount of file data a node at // height (i) covers. Height 0 is the direct block. @@ -68,10 +68,12 @@ func newBlockMapFile(args inodeArgs) (*blockMapFile, error) { } blkMap := file.regFile.inode.diskInode.Data() - binary.Unmarshal(blkMap[:numDirectBlks*4], binary.LittleEndian, &file.directBlks) - binary.Unmarshal(blkMap[numDirectBlks*4:(numDirectBlks+1)*4], binary.LittleEndian, &file.indirectBlk) - binary.Unmarshal(blkMap[(numDirectBlks+1)*4:(numDirectBlks+2)*4], binary.LittleEndian, &file.doubleIndirectBlk) - binary.Unmarshal(blkMap[(numDirectBlks+2)*4:(numDirectBlks+3)*4], binary.LittleEndian, &file.tripleIndirectBlk) + for i := 0; i < numDirectBlks; i++ { + file.directBlks[i].UnmarshalBytes(blkMap[i*4 : (i+1)*4]) + } + file.indirectBlk.UnmarshalBytes(blkMap[numDirectBlks*4 : (numDirectBlks+1)*4]) + file.doubleIndirectBlk.UnmarshalBytes(blkMap[(numDirectBlks+1)*4 : (numDirectBlks+2)*4]) + file.tripleIndirectBlk.UnmarshalBytes(blkMap[(numDirectBlks+2)*4 : (numDirectBlks+3)*4]) return file, nil } @@ -117,16 +119,16 @@ func (f *blockMapFile) ReadAt(dst []byte, off int64) (int, error) { switch { case offset < dirBlksEnd: // Direct block. - curR, err = f.read(f.directBlks[offset/f.regFile.inode.blkSize], offset%f.regFile.inode.blkSize, 0, dst[read:]) + curR, err = f.read(uint32(f.directBlks[offset/f.regFile.inode.blkSize]), offset%f.regFile.inode.blkSize, 0, dst[read:]) case offset < indirBlkEnd: // Indirect block. - curR, err = f.read(f.indirectBlk, offset-dirBlksEnd, 1, dst[read:]) + curR, err = f.read(uint32(f.indirectBlk), offset-dirBlksEnd, 1, dst[read:]) case offset < doubIndirBlkEnd: // Doubly indirect block. - curR, err = f.read(f.doubleIndirectBlk, offset-indirBlkEnd, 2, dst[read:]) + curR, err = f.read(uint32(f.doubleIndirectBlk), offset-indirBlkEnd, 2, dst[read:]) default: // Triply indirect block. - curR, err = f.read(f.tripleIndirectBlk, offset-doubIndirBlkEnd, 3, dst[read:]) + curR, err = f.read(uint32(f.tripleIndirectBlk), offset-doubIndirBlkEnd, 3, dst[read:]) } read += curR @@ -174,13 +176,13 @@ func (f *blockMapFile) read(curPhyBlk uint32, relFileOff uint64, height uint, ds read := 0 curChildOff := relFileOff % childCov for i := startIdx; i < endIdx; i++ { - var childPhyBlk uint32 + var childPhyBlk primitive.Uint32 err := readFromDisk(f.regFile.inode.fs.dev, curPhyBlkOff+int64(i*4), &childPhyBlk) if err != nil { return read, err } - n, err := f.read(childPhyBlk, curChildOff, height-1, dst[read:]) + n, err := f.read(uint32(childPhyBlk), curChildOff, height-1, dst[read:]) read += n if err != nil { return read, err diff --git a/pkg/sentry/fsimpl/ext/block_map_test.go b/pkg/sentry/fsimpl/ext/block_map_test.go index 6fa84e7aa..ed98b482e 100644 --- a/pkg/sentry/fsimpl/ext/block_map_test.go +++ b/pkg/sentry/fsimpl/ext/block_map_test.go @@ -20,7 +20,7 @@ import ( "testing" "github.com/google/go-cmp/cmp" - "gvisor.dev/gvisor/pkg/binary" + "gvisor.dev/gvisor/pkg/marshal/primitive" "gvisor.dev/gvisor/pkg/sentry/fsimpl/ext/disklayout" ) @@ -87,29 +87,33 @@ func blockMapSetUp(t *testing.T) (*blockMapFile, []byte) { mockDisk := make([]byte, mockBMDiskSize) var fileData []byte blkNums := newBlkNumGen() - var data []byte + off := 0 + data := make([]byte, (numDirectBlks+3)*(*primitive.Uint32)(nil).SizeBytes()) // Write the direct blocks. for i := 0; i < numDirectBlks; i++ { - curBlkNum := blkNums.next() - data = binary.Marshal(data, binary.LittleEndian, curBlkNum) - fileData = append(fileData, writeFileDataToBlock(mockDisk, curBlkNum, 0, blkNums)...) + curBlkNum := primitive.Uint32(blkNums.next()) + curBlkNum.MarshalBytes(data[off:]) + off += curBlkNum.SizeBytes() + fileData = append(fileData, writeFileDataToBlock(mockDisk, uint32(curBlkNum), 0, blkNums)...) } // Write to indirect block. - indirectBlk := blkNums.next() - data = binary.Marshal(data, binary.LittleEndian, indirectBlk) - fileData = append(fileData, writeFileDataToBlock(mockDisk, indirectBlk, 1, blkNums)...) - - // Write to indirect block. - doublyIndirectBlk := blkNums.next() - data = binary.Marshal(data, binary.LittleEndian, doublyIndirectBlk) - fileData = append(fileData, writeFileDataToBlock(mockDisk, doublyIndirectBlk, 2, blkNums)...) - - // Write to indirect block. - triplyIndirectBlk := blkNums.next() - data = binary.Marshal(data, binary.LittleEndian, triplyIndirectBlk) - fileData = append(fileData, writeFileDataToBlock(mockDisk, triplyIndirectBlk, 3, blkNums)...) + indirectBlk := primitive.Uint32(blkNums.next()) + indirectBlk.MarshalBytes(data[off:]) + off += indirectBlk.SizeBytes() + fileData = append(fileData, writeFileDataToBlock(mockDisk, uint32(indirectBlk), 1, blkNums)...) + + // Write to double indirect block. + doublyIndirectBlk := primitive.Uint32(blkNums.next()) + doublyIndirectBlk.MarshalBytes(data[off:]) + off += doublyIndirectBlk.SizeBytes() + fileData = append(fileData, writeFileDataToBlock(mockDisk, uint32(doublyIndirectBlk), 2, blkNums)...) + + // Write to triple indirect block. + triplyIndirectBlk := primitive.Uint32(blkNums.next()) + triplyIndirectBlk.MarshalBytes(data[off:]) + fileData = append(fileData, writeFileDataToBlock(mockDisk, uint32(triplyIndirectBlk), 3, blkNums)...) args := inodeArgs{ fs: &filesystem{ @@ -142,9 +146,9 @@ func writeFileDataToBlock(disk []byte, blkNum uint32, height uint, blkNums *blkN var fileData []byte for off := blkNum * mockBMBlkSize; off < (blkNum+1)*mockBMBlkSize; off += 4 { - curBlkNum := blkNums.next() - copy(disk[off:off+4], binary.Marshal(nil, binary.LittleEndian, curBlkNum)) - fileData = append(fileData, writeFileDataToBlock(disk, curBlkNum, height-1, blkNums)...) + curBlkNum := primitive.Uint32(blkNums.next()) + curBlkNum.MarshalBytes(disk[off : off+4]) + fileData = append(fileData, writeFileDataToBlock(disk, uint32(curBlkNum), height-1, blkNums)...) } return fileData } diff --git a/pkg/sentry/fsimpl/ext/dentry.go b/pkg/sentry/fsimpl/ext/dentry.go index 7a1b4219f..9bfed883a 100644 --- a/pkg/sentry/fsimpl/ext/dentry.go +++ b/pkg/sentry/fsimpl/ext/dentry.go @@ -20,6 +20,8 @@ import ( ) // dentry implements vfs.DentryImpl. +// +// +stateify savable type dentry struct { vfsd vfs.Dentry diff --git a/pkg/sentry/fsimpl/ext/directory.go b/pkg/sentry/fsimpl/ext/directory.go index 0fc01668d..0ad79b381 100644 --- a/pkg/sentry/fsimpl/ext/directory.go +++ b/pkg/sentry/fsimpl/ext/directory.go @@ -16,7 +16,6 @@ package ext import ( "gvisor.dev/gvisor/pkg/abi/linux" - "gvisor.dev/gvisor/pkg/binary" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/log" "gvisor.dev/gvisor/pkg/sentry/fs" @@ -28,6 +27,8 @@ import ( ) // directory represents a directory inode. It holds the childList in memory. +// +// +stateify savable type directory struct { inode inode @@ -39,7 +40,7 @@ type directory struct { // Lock Order (outermost locks must be taken first): // directory.mu // filesystem.mu - mu sync.Mutex + mu sync.Mutex `state:"nosave"` // childList is a list containing (1) child dirents and (2) fake dirents // (with diskDirent == nil) that represent the iteration position of @@ -98,7 +99,7 @@ func newDirectory(args inodeArgs, newDirent bool) (*directory, error) { } else { curDirent.diskDirent = &disklayout.DirentOld{} } - binary.Unmarshal(buf, binary.LittleEndian, curDirent.diskDirent) + curDirent.diskDirent.UnmarshalBytes(buf) if curDirent.diskDirent.Inode() != 0 && len(curDirent.diskDirent.FileName()) != 0 { // Inode number and name length fields being set to 0 is used to indicate @@ -120,6 +121,8 @@ func (i *inode) isDir() bool { } // dirent is the directory.childList node. +// +// +stateify savable type dirent struct { diskDirent disklayout.Dirent @@ -129,6 +132,8 @@ type dirent struct { // directoryFD represents a directory file description. It implements // vfs.FileDescriptionImpl. +// +// +stateify savable type directoryFD struct { fileDescription vfs.DirectoryFileDescriptionDefaultImpl diff --git a/pkg/sentry/fsimpl/ext/disklayout/BUILD b/pkg/sentry/fsimpl/ext/disklayout/BUILD index 9bd9c76c0..d98a05dd8 100644 --- a/pkg/sentry/fsimpl/ext/disklayout/BUILD +++ b/pkg/sentry/fsimpl/ext/disklayout/BUILD @@ -22,10 +22,11 @@ go_library( "superblock_old.go", "test_utils.go", ], + marshal = True, visibility = ["//pkg/sentry:internal"], deps = [ "//pkg/abi/linux", - "//pkg/binary", + "//pkg/marshal", "//pkg/sentry/fs", "//pkg/sentry/kernel/auth", "//pkg/sentry/kernel/time", diff --git a/pkg/sentry/fsimpl/ext/disklayout/block_group.go b/pkg/sentry/fsimpl/ext/disklayout/block_group.go index ad6f4fef8..0d56ae9da 100644 --- a/pkg/sentry/fsimpl/ext/disklayout/block_group.go +++ b/pkg/sentry/fsimpl/ext/disklayout/block_group.go @@ -14,6 +14,10 @@ package disklayout +import ( + "gvisor.dev/gvisor/pkg/marshal" +) + // BlockGroup represents a Linux ext block group descriptor. An ext file system // is split into a series of block groups. This provides an access layer to // information needed to access and use a block group. @@ -30,6 +34,8 @@ package disklayout // // See https://www.kernel.org/doc/html/latest/filesystems/ext4/globals.html#block-group-descriptors. type BlockGroup interface { + marshal.Marshallable + // InodeTable returns the absolute block number of the block containing the // inode table. This points to an array of Inode structs. Inode tables are // statically allocated at mkfs time. The superblock records the number of diff --git a/pkg/sentry/fsimpl/ext/disklayout/block_group_32.go b/pkg/sentry/fsimpl/ext/disklayout/block_group_32.go index 3e16c76db..a35fa22a0 100644 --- a/pkg/sentry/fsimpl/ext/disklayout/block_group_32.go +++ b/pkg/sentry/fsimpl/ext/disklayout/block_group_32.go @@ -17,6 +17,8 @@ package disklayout // BlockGroup32Bit emulates the first half of struct ext4_group_desc in // fs/ext4/ext4.h. It is the block group descriptor struct for ext2, ext3 and // 32-bit ext4 filesystems. It implements BlockGroup interface. +// +// +marshal type BlockGroup32Bit struct { BlockBitmapLo uint32 InodeBitmapLo uint32 diff --git a/pkg/sentry/fsimpl/ext/disklayout/block_group_64.go b/pkg/sentry/fsimpl/ext/disklayout/block_group_64.go index 9a809197a..d54d1d345 100644 --- a/pkg/sentry/fsimpl/ext/disklayout/block_group_64.go +++ b/pkg/sentry/fsimpl/ext/disklayout/block_group_64.go @@ -18,6 +18,8 @@ package disklayout // It is the block group descriptor struct for 64-bit ext4 filesystems. // It implements BlockGroup interface. It is an extension of the 32-bit // version of BlockGroup. +// +// +marshal type BlockGroup64Bit struct { // We embed the 32-bit struct here because 64-bit version is just an extension // of the 32-bit version. diff --git a/pkg/sentry/fsimpl/ext/disklayout/block_group_test.go b/pkg/sentry/fsimpl/ext/disklayout/block_group_test.go index 0ef4294c0..e4ce484e4 100644 --- a/pkg/sentry/fsimpl/ext/disklayout/block_group_test.go +++ b/pkg/sentry/fsimpl/ext/disklayout/block_group_test.go @@ -21,6 +21,8 @@ import ( // TestBlockGroupSize tests that the block group descriptor structs are of the // correct size. func TestBlockGroupSize(t *testing.T) { - assertSize(t, BlockGroup32Bit{}, 32) - assertSize(t, BlockGroup64Bit{}, 64) + var bgSmall BlockGroup32Bit + assertSize(t, &bgSmall, 32) + var bgBig BlockGroup64Bit + assertSize(t, &bgBig, 64) } diff --git a/pkg/sentry/fsimpl/ext/disklayout/dirent.go b/pkg/sentry/fsimpl/ext/disklayout/dirent.go index 417b6cf65..568c8cb4c 100644 --- a/pkg/sentry/fsimpl/ext/disklayout/dirent.go +++ b/pkg/sentry/fsimpl/ext/disklayout/dirent.go @@ -15,6 +15,7 @@ package disklayout import ( + "gvisor.dev/gvisor/pkg/marshal" "gvisor.dev/gvisor/pkg/sentry/fs" ) @@ -51,6 +52,8 @@ var ( // // See https://www.kernel.org/doc/html/latest/filesystems/ext4/dynamic.html#linear-classic-directories. type Dirent interface { + marshal.Marshallable + // Inode returns the absolute inode number of the underlying inode. // Inode number 0 signifies an unused dirent. Inode() uint32 diff --git a/pkg/sentry/fsimpl/ext/disklayout/dirent_new.go b/pkg/sentry/fsimpl/ext/disklayout/dirent_new.go index 29ae4a5c2..51f9c2946 100644 --- a/pkg/sentry/fsimpl/ext/disklayout/dirent_new.go +++ b/pkg/sentry/fsimpl/ext/disklayout/dirent_new.go @@ -29,12 +29,14 @@ import ( // Note: This struct can be of variable size on disk. The one described below // is of maximum size and the FileName beyond NameLength bytes might contain // garbage. +// +// +marshal type DirentNew struct { InodeNumber uint32 RecordLength uint16 NameLength uint8 FileTypeRaw uint8 - FileNameRaw [MaxFileName]byte + FileNameRaw [MaxFileName]byte `marshal:"unaligned"` } // Compiles only if DirentNew implements Dirent. diff --git a/pkg/sentry/fsimpl/ext/disklayout/dirent_old.go b/pkg/sentry/fsimpl/ext/disklayout/dirent_old.go index 6fff12a6e..d4b19e086 100644 --- a/pkg/sentry/fsimpl/ext/disklayout/dirent_old.go +++ b/pkg/sentry/fsimpl/ext/disklayout/dirent_old.go @@ -22,11 +22,13 @@ import "gvisor.dev/gvisor/pkg/sentry/fs" // Note: This struct can be of variable size on disk. The one described below // is of maximum size and the FileName beyond NameLength bytes might contain // garbage. +// +// +marshal type DirentOld struct { InodeNumber uint32 RecordLength uint16 NameLength uint16 - FileNameRaw [MaxFileName]byte + FileNameRaw [MaxFileName]byte `marshal:"unaligned"` } // Compiles only if DirentOld implements Dirent. diff --git a/pkg/sentry/fsimpl/ext/disklayout/dirent_test.go b/pkg/sentry/fsimpl/ext/disklayout/dirent_test.go index 934919f8a..3486864dc 100644 --- a/pkg/sentry/fsimpl/ext/disklayout/dirent_test.go +++ b/pkg/sentry/fsimpl/ext/disklayout/dirent_test.go @@ -21,6 +21,8 @@ import ( // TestDirentSize tests that the dirent structs are of the correct // size. func TestDirentSize(t *testing.T) { - assertSize(t, DirentOld{}, uintptr(DirentSize)) - assertSize(t, DirentNew{}, uintptr(DirentSize)) + var dOld DirentOld + assertSize(t, &dOld, DirentSize) + var dNew DirentNew + assertSize(t, &dNew, DirentSize) } diff --git a/pkg/sentry/fsimpl/ext/disklayout/disklayout.go b/pkg/sentry/fsimpl/ext/disklayout/disklayout.go index bdf4e2132..0834e9ba8 100644 --- a/pkg/sentry/fsimpl/ext/disklayout/disklayout.go +++ b/pkg/sentry/fsimpl/ext/disklayout/disklayout.go @@ -36,8 +36,6 @@ // escape analysis on an unknown implementation at compile time. // // Notes: -// - All fields in these structs are exported because binary.Read would -// panic otherwise. // - All structures on disk are in little-endian order. Only jbd2 (journal) // structures are in big-endian order. // - All OS dependent fields in these structures will be interpretted using diff --git a/pkg/sentry/fsimpl/ext/disklayout/extent.go b/pkg/sentry/fsimpl/ext/disklayout/extent.go index 4110649ab..b13999bfc 100644 --- a/pkg/sentry/fsimpl/ext/disklayout/extent.go +++ b/pkg/sentry/fsimpl/ext/disklayout/extent.go @@ -14,6 +14,10 @@ package disklayout +import ( + "gvisor.dev/gvisor/pkg/marshal" +) + // Extents were introduced in ext4 and provide huge performance gains in terms // data locality and reduced metadata block usage. Extents are organized in // extent trees. The root node is contained in inode.BlocksRaw. @@ -64,6 +68,8 @@ type ExtentNode struct { // ExtentEntry represents an extent tree node entry. The entry can either be // an ExtentIdx or Extent itself. This exists to simplify navigation logic. type ExtentEntry interface { + marshal.Marshallable + // FileBlock returns the first file block number covered by this entry. FileBlock() uint32 @@ -75,6 +81,8 @@ type ExtentEntry interface { // tree node begins with this and is followed by `NumEntries` number of: // - Extent if `Depth` == 0 // - ExtentIdx otherwise +// +// +marshal type ExtentHeader struct { // Magic in the extent magic number, must be 0xf30a. Magic uint16 @@ -96,6 +104,8 @@ type ExtentHeader struct { // internal nodes. Sorted in ascending order based on FirstFileBlock since // Linux does a binary search on this. This points to a block containing the // child node. +// +// +marshal type ExtentIdx struct { FirstFileBlock uint32 ChildBlockLo uint32 @@ -121,6 +131,8 @@ func (ei *ExtentIdx) PhysicalBlock() uint64 { // nodes. Sorted in ascending order based on FirstFileBlock since Linux does a // binary search on this. This points to an array of data blocks containing the // file data. It covers `Length` data blocks starting from `StartBlock`. +// +// +marshal type Extent struct { FirstFileBlock uint32 Length uint16 diff --git a/pkg/sentry/fsimpl/ext/disklayout/extent_test.go b/pkg/sentry/fsimpl/ext/disklayout/extent_test.go index 8762b90db..c96002e19 100644 --- a/pkg/sentry/fsimpl/ext/disklayout/extent_test.go +++ b/pkg/sentry/fsimpl/ext/disklayout/extent_test.go @@ -21,7 +21,10 @@ import ( // TestExtentSize tests that the extent structs are of the correct // size. func TestExtentSize(t *testing.T) { - assertSize(t, ExtentHeader{}, ExtentHeaderSize) - assertSize(t, ExtentIdx{}, ExtentEntrySize) - assertSize(t, Extent{}, ExtentEntrySize) + var h ExtentHeader + assertSize(t, &h, ExtentHeaderSize) + var i ExtentIdx + assertSize(t, &i, ExtentEntrySize) + var e Extent + assertSize(t, &e, ExtentEntrySize) } diff --git a/pkg/sentry/fsimpl/ext/disklayout/inode.go b/pkg/sentry/fsimpl/ext/disklayout/inode.go index 88ae913f5..ef25040a9 100644 --- a/pkg/sentry/fsimpl/ext/disklayout/inode.go +++ b/pkg/sentry/fsimpl/ext/disklayout/inode.go @@ -16,6 +16,7 @@ package disklayout import ( "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/marshal" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/kernel/time" ) @@ -38,6 +39,8 @@ const ( // // See https://www.kernel.org/doc/html/latest/filesystems/ext4/dynamic.html#index-nodes. type Inode interface { + marshal.Marshallable + // Mode returns the linux file mode which is majorly used to extract // information like: // - File permissions (read/write/execute by user/group/others). diff --git a/pkg/sentry/fsimpl/ext/disklayout/inode_new.go b/pkg/sentry/fsimpl/ext/disklayout/inode_new.go index 8f9f574ce..a4503f5cf 100644 --- a/pkg/sentry/fsimpl/ext/disklayout/inode_new.go +++ b/pkg/sentry/fsimpl/ext/disklayout/inode_new.go @@ -27,6 +27,8 @@ import "gvisor.dev/gvisor/pkg/sentry/kernel/time" // are used to provide nanoscond precision. Hence, these timestamps will now // overflow in May 2446. // See https://www.kernel.org/doc/html/latest/filesystems/ext4/dynamic.html#inode-timestamps. +// +// +marshal type InodeNew struct { InodeOld diff --git a/pkg/sentry/fsimpl/ext/disklayout/inode_old.go b/pkg/sentry/fsimpl/ext/disklayout/inode_old.go index db25b11b6..e6b28babf 100644 --- a/pkg/sentry/fsimpl/ext/disklayout/inode_old.go +++ b/pkg/sentry/fsimpl/ext/disklayout/inode_old.go @@ -30,6 +30,8 @@ const ( // // All fields representing time are in seconds since the epoch. Which means that // they will overflow in January 2038. +// +// +marshal type InodeOld struct { ModeRaw uint16 UIDLo uint16 diff --git a/pkg/sentry/fsimpl/ext/disklayout/inode_test.go b/pkg/sentry/fsimpl/ext/disklayout/inode_test.go index dd03ee50e..90744e956 100644 --- a/pkg/sentry/fsimpl/ext/disklayout/inode_test.go +++ b/pkg/sentry/fsimpl/ext/disklayout/inode_test.go @@ -24,10 +24,12 @@ import ( // TestInodeSize tests that the inode structs are of the correct size. func TestInodeSize(t *testing.T) { - assertSize(t, InodeOld{}, OldInodeSize) + var iOld InodeOld + assertSize(t, &iOld, OldInodeSize) // This was updated from 156 bytes to 160 bytes in Oct 2015. - assertSize(t, InodeNew{}, 160) + var iNew InodeNew + assertSize(t, &iNew, 160) } // TestTimestampSeconds tests that the seconds part of [a/c/m] timestamps in diff --git a/pkg/sentry/fsimpl/ext/disklayout/superblock.go b/pkg/sentry/fsimpl/ext/disklayout/superblock.go index 8bb327006..70948ebe9 100644 --- a/pkg/sentry/fsimpl/ext/disklayout/superblock.go +++ b/pkg/sentry/fsimpl/ext/disklayout/superblock.go @@ -14,6 +14,10 @@ package disklayout +import ( + "gvisor.dev/gvisor/pkg/marshal" +) + const ( // SbOffset is the absolute offset at which the superblock is placed. SbOffset = 1024 @@ -38,6 +42,8 @@ const ( // // See https://www.kernel.org/doc/html/latest/filesystems/ext4/globals.html#super-block. type SuperBlock interface { + marshal.Marshallable + // InodesCount returns the total number of inodes in this filesystem. InodesCount() uint32 diff --git a/pkg/sentry/fsimpl/ext/disklayout/superblock_32.go b/pkg/sentry/fsimpl/ext/disklayout/superblock_32.go index 53e515fd3..4dc6080fb 100644 --- a/pkg/sentry/fsimpl/ext/disklayout/superblock_32.go +++ b/pkg/sentry/fsimpl/ext/disklayout/superblock_32.go @@ -17,6 +17,8 @@ package disklayout // SuperBlock32Bit implements SuperBlock and represents the 32-bit version of // the ext4_super_block struct in fs/ext4/ext4.h. Should be used only if // RevLevel = DynamicRev and 64-bit feature is disabled. +// +// +marshal type SuperBlock32Bit struct { // We embed the old superblock struct here because the 32-bit version is just // an extension of the old version. diff --git a/pkg/sentry/fsimpl/ext/disklayout/superblock_64.go b/pkg/sentry/fsimpl/ext/disklayout/superblock_64.go index 7c1053fb4..2c9039327 100644 --- a/pkg/sentry/fsimpl/ext/disklayout/superblock_64.go +++ b/pkg/sentry/fsimpl/ext/disklayout/superblock_64.go @@ -19,6 +19,8 @@ package disklayout // 1024 bytes (smallest possible block size) and hence the superblock always // fits in no more than one data block. Should only be used when the 64-bit // feature is set. +// +// +marshal type SuperBlock64Bit struct { // We embed the 32-bit struct here because 64-bit version is just an extension // of the 32-bit version. diff --git a/pkg/sentry/fsimpl/ext/disklayout/superblock_old.go b/pkg/sentry/fsimpl/ext/disklayout/superblock_old.go index 9221e0251..e4709f23c 100644 --- a/pkg/sentry/fsimpl/ext/disklayout/superblock_old.go +++ b/pkg/sentry/fsimpl/ext/disklayout/superblock_old.go @@ -16,6 +16,8 @@ package disklayout // SuperBlockOld implements SuperBlock and represents the old version of the // superblock struct. Should be used only if RevLevel = OldRev. +// +// +marshal type SuperBlockOld struct { InodesCountRaw uint32 BlocksCountLo uint32 diff --git a/pkg/sentry/fsimpl/ext/disklayout/superblock_test.go b/pkg/sentry/fsimpl/ext/disklayout/superblock_test.go index 463b5ba21..b734b6987 100644 --- a/pkg/sentry/fsimpl/ext/disklayout/superblock_test.go +++ b/pkg/sentry/fsimpl/ext/disklayout/superblock_test.go @@ -21,7 +21,10 @@ import ( // TestSuperBlockSize tests that the superblock structs are of the correct // size. func TestSuperBlockSize(t *testing.T) { - assertSize(t, SuperBlockOld{}, 84) - assertSize(t, SuperBlock32Bit{}, 336) - assertSize(t, SuperBlock64Bit{}, 1024) + var sbOld SuperBlockOld + assertSize(t, &sbOld, 84) + var sb32 SuperBlock32Bit + assertSize(t, &sb32, 336) + var sb64 SuperBlock64Bit + assertSize(t, &sb64, 1024) } diff --git a/pkg/sentry/fsimpl/ext/disklayout/test_utils.go b/pkg/sentry/fsimpl/ext/disklayout/test_utils.go index 9c63f04c0..a4bc08411 100644 --- a/pkg/sentry/fsimpl/ext/disklayout/test_utils.go +++ b/pkg/sentry/fsimpl/ext/disklayout/test_utils.go @@ -18,13 +18,13 @@ import ( "reflect" "testing" - "gvisor.dev/gvisor/pkg/binary" + "gvisor.dev/gvisor/pkg/marshal" ) -func assertSize(t *testing.T, v interface{}, want uintptr) { +func assertSize(t *testing.T, v marshal.Marshallable, want int) { t.Helper() - if got := binary.Size(v); got != want { + if got := v.SizeBytes(); got != want { t.Errorf("struct %s should be exactly %d bytes but is %d bytes", reflect.TypeOf(v).Name(), want, got) } } diff --git a/pkg/sentry/fsimpl/ext/ext.go b/pkg/sentry/fsimpl/ext/ext.go index 08ffc2834..aca258d40 100644 --- a/pkg/sentry/fsimpl/ext/ext.go +++ b/pkg/sentry/fsimpl/ext/ext.go @@ -34,6 +34,8 @@ import ( const Name = "ext" // FilesystemType implements vfs.FilesystemType. +// +// +stateify savable type FilesystemType struct{} // Compiles only if FilesystemType implements vfs.FilesystemType. diff --git a/pkg/sentry/fsimpl/ext/extent_file.go b/pkg/sentry/fsimpl/ext/extent_file.go index c36225a7c..778460107 100644 --- a/pkg/sentry/fsimpl/ext/extent_file.go +++ b/pkg/sentry/fsimpl/ext/extent_file.go @@ -18,12 +18,13 @@ import ( "io" "sort" - "gvisor.dev/gvisor/pkg/binary" "gvisor.dev/gvisor/pkg/sentry/fsimpl/ext/disklayout" "gvisor.dev/gvisor/pkg/syserror" ) // extentFile is a type of regular file which uses extents to store file data. +// +// +stateify savable type extentFile struct { regFile regularFile @@ -58,7 +59,7 @@ func newExtentFile(args inodeArgs) (*extentFile, error) { func (f *extentFile) buildExtTree() error { rootNodeData := f.regFile.inode.diskInode.Data() - binary.Unmarshal(rootNodeData[:disklayout.ExtentHeaderSize], binary.LittleEndian, &f.root.Header) + f.root.Header.UnmarshalBytes(rootNodeData[:disklayout.ExtentHeaderSize]) // Root node can not have more than 4 entries: 60 bytes = 1 header + 4 entries. if f.root.Header.NumEntries > 4 { @@ -77,7 +78,7 @@ func (f *extentFile) buildExtTree() error { // Internal node. curEntry = &disklayout.ExtentIdx{} } - binary.Unmarshal(rootNodeData[off:off+disklayout.ExtentEntrySize], binary.LittleEndian, curEntry) + curEntry.UnmarshalBytes(rootNodeData[off : off+disklayout.ExtentEntrySize]) f.root.Entries[i].Entry = curEntry } diff --git a/pkg/sentry/fsimpl/ext/extent_test.go b/pkg/sentry/fsimpl/ext/extent_test.go index cd10d46ee..985f76ac0 100644 --- a/pkg/sentry/fsimpl/ext/extent_test.go +++ b/pkg/sentry/fsimpl/ext/extent_test.go @@ -21,7 +21,6 @@ import ( "github.com/google/go-cmp/cmp" "github.com/google/go-cmp/cmp/cmpopts" - "gvisor.dev/gvisor/pkg/binary" "gvisor.dev/gvisor/pkg/sentry/fsimpl/ext/disklayout" ) @@ -202,13 +201,14 @@ func extentTreeSetUp(t *testing.T, root *disklayout.ExtentNode) (*extentFile, [] // writeTree writes the tree represented by `root` to the inode and disk. It // also writes random file data on disk. func writeTree(in *inode, disk []byte, root *disklayout.ExtentNode, mockExtentBlkSize uint64) []byte { - rootData := binary.Marshal(nil, binary.LittleEndian, root.Header) + rootData := in.diskInode.Data() + root.Header.MarshalBytes(rootData) + off := root.Header.SizeBytes() for _, ep := range root.Entries { - rootData = binary.Marshal(rootData, binary.LittleEndian, ep.Entry) + ep.Entry.MarshalBytes(rootData[off:]) + off += ep.Entry.SizeBytes() } - copy(in.diskInode.Data(), rootData) - var fileData []byte for _, ep := range root.Entries { if root.Header.Height == 0 { @@ -223,13 +223,14 @@ func writeTree(in *inode, disk []byte, root *disklayout.ExtentNode, mockExtentBl // writeTreeToDisk is the recursive step for writeTree which writes the tree // on the disk only. Also writes random file data on disk. func writeTreeToDisk(disk []byte, curNode disklayout.ExtentEntryPair) []byte { - nodeData := binary.Marshal(nil, binary.LittleEndian, curNode.Node.Header) + nodeData := disk[curNode.Entry.PhysicalBlock()*mockExtentBlkSize:] + curNode.Node.Header.MarshalBytes(nodeData) + off := curNode.Node.Header.SizeBytes() for _, ep := range curNode.Node.Entries { - nodeData = binary.Marshal(nodeData, binary.LittleEndian, ep.Entry) + ep.Entry.MarshalBytes(nodeData[off:]) + off += ep.Entry.SizeBytes() } - copy(disk[curNode.Entry.PhysicalBlock()*mockExtentBlkSize:], nodeData) - var fileData []byte for _, ep := range curNode.Node.Entries { if curNode.Node.Header.Height == 0 { diff --git a/pkg/sentry/fsimpl/ext/filesystem.go b/pkg/sentry/fsimpl/ext/filesystem.go index 075de0e22..917f1873d 100644 --- a/pkg/sentry/fsimpl/ext/filesystem.go +++ b/pkg/sentry/fsimpl/ext/filesystem.go @@ -38,11 +38,13 @@ var ( ) // filesystem implements vfs.FilesystemImpl. +// +// +stateify savable type filesystem struct { vfsfs vfs.Filesystem // mu serializes changes to the Dentry tree. - mu sync.RWMutex + mu sync.RWMutex `state:"nosave"` // dev represents the underlying fs device. It does not require protection // because io.ReaderAt permits concurrent read calls to it. It translates to @@ -490,7 +492,7 @@ func (fs *filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error return syserror.EROFS } -// BoundEndpointAt implements FilesystemImpl.BoundEndpointAt. +// BoundEndpointAt implements vfs.FilesystemImpl.BoundEndpointAt. func (fs *filesystem) BoundEndpointAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.BoundEndpointOptions) (transport.BoundEndpoint, error) { _, inode, err := fs.walk(ctx, rp, false) if err != nil { diff --git a/pkg/sentry/fsimpl/ext/inode.go b/pkg/sentry/fsimpl/ext/inode.go index 30636cf66..9009ba3c7 100644 --- a/pkg/sentry/fsimpl/ext/inode.go +++ b/pkg/sentry/fsimpl/ext/inode.go @@ -37,6 +37,8 @@ import ( // |-- regular-- // |-- extent file // |-- block map file +// +// +stateify savable type inode struct { // refs is a reference count. refs is accessed using atomic memory operations. refs int64 diff --git a/pkg/sentry/fsimpl/ext/regular_file.go b/pkg/sentry/fsimpl/ext/regular_file.go index e73e740d6..4a5539b37 100644 --- a/pkg/sentry/fsimpl/ext/regular_file.go +++ b/pkg/sentry/fsimpl/ext/regular_file.go @@ -31,6 +31,8 @@ import ( // regularFile represents a regular file's inode. This too follows the // inheritance pattern prevelant in the vfs layer described in // pkg/sentry/vfs/README.md. +// +// +stateify savable type regularFile struct { inode inode @@ -67,6 +69,8 @@ func (in *inode) isRegular() bool { // directoryFD represents a directory file description. It implements // vfs.FileDescriptionImpl. +// +// +stateify savable type regularFileFD struct { fileDescription vfs.LockFD @@ -75,7 +79,7 @@ type regularFileFD struct { off int64 // offMu serializes operations that may mutate off. - offMu sync.Mutex + offMu sync.Mutex `state:"nosave"` } // Release implements vfs.FileDescriptionImpl.Release. diff --git a/pkg/sentry/fsimpl/ext/symlink.go b/pkg/sentry/fsimpl/ext/symlink.go index 2fd0d1fa8..5e2bcc837 100644 --- a/pkg/sentry/fsimpl/ext/symlink.go +++ b/pkg/sentry/fsimpl/ext/symlink.go @@ -23,6 +23,8 @@ import ( ) // symlink represents a symlink inode. +// +// +stateify savable type symlink struct { inode inode target string // immutable @@ -61,9 +63,11 @@ func (in *inode) isSymlink() bool { return ok } -// symlinkFD represents a symlink file description and implements implements +// symlinkFD represents a symlink file description and implements // vfs.FileDescriptionImpl. which may only be used if open options contains // O_PATH. For this reason most of the functions return EBADF. +// +// +stateify savable type symlinkFD struct { fileDescription vfs.NoLockFD diff --git a/pkg/sentry/fsimpl/ext/utils.go b/pkg/sentry/fsimpl/ext/utils.go index d8b728f8c..58ef7b9b8 100644 --- a/pkg/sentry/fsimpl/ext/utils.go +++ b/pkg/sentry/fsimpl/ext/utils.go @@ -17,21 +17,21 @@ package ext import ( "io" - "gvisor.dev/gvisor/pkg/binary" + "gvisor.dev/gvisor/pkg/marshal" "gvisor.dev/gvisor/pkg/sentry/fsimpl/ext/disklayout" "gvisor.dev/gvisor/pkg/syserror" ) // readFromDisk performs a binary read from disk into the given struct from // the absolute offset provided. -func readFromDisk(dev io.ReaderAt, abOff int64, v interface{}) error { - n := binary.Size(v) +func readFromDisk(dev io.ReaderAt, abOff int64, v marshal.Marshallable) error { + n := v.SizeBytes() buf := make([]byte, n) if read, _ := dev.ReadAt(buf, abOff); read < int(n) { return syserror.EIO } - binary.Unmarshal(buf, binary.LittleEndian, v) + v.UnmarshalBytes(buf) return nil } diff --git a/pkg/sentry/fsimpl/fuse/BUILD b/pkg/sentry/fsimpl/fuse/BUILD index 999c16bfd..045d7ab08 100644 --- a/pkg/sentry/fsimpl/fuse/BUILD +++ b/pkg/sentry/fsimpl/fuse/BUILD @@ -30,12 +30,17 @@ go_library( name = "fuse", srcs = [ "connection.go", + "connection_control.go", "dev.go", + "directory.go", + "file.go", "fusefs.go", - "init.go", "inode_refs.go", + "read_write.go", "register.go", + "regular_file.go", "request_list.go", + "request_response.go", ], visibility = ["//pkg/sentry:internal"], deps = [ @@ -44,6 +49,7 @@ go_library( "//pkg/log", "//pkg/marshal", "//pkg/refs", + "//pkg/safemem", "//pkg/sentry/fsimpl/devtmpfs", "//pkg/sentry/fsimpl/kernfs", "//pkg/sentry/kernel", @@ -60,7 +66,11 @@ go_library( go_test( name = "fuse_test", size = "small", - srcs = ["dev_test.go"], + srcs = [ + "connection_test.go", + "dev_test.go", + "utils_test.go", + ], library = ":fuse", deps = [ "//pkg/abi/linux", diff --git a/pkg/sentry/fsimpl/fuse/connection.go b/pkg/sentry/fsimpl/fuse/connection.go index eb1d1a2b7..8ccda1264 100644 --- a/pkg/sentry/fsimpl/fuse/connection.go +++ b/pkg/sentry/fsimpl/fuse/connection.go @@ -15,31 +15,17 @@ package fuse import ( - "errors" - "fmt" "sync" - "sync/atomic" - "syscall" "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/log" - "gvisor.dev/gvisor/pkg/marshal" "gvisor.dev/gvisor/pkg/sentry/kernel" - "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/waiter" ) -// maxActiveRequestsDefault is the default setting controlling the upper bound -// on the number of active requests at any given time. -const maxActiveRequestsDefault = 10000 - -// Ordinary requests have even IDs, while interrupts IDs are odd. -// Used to increment the unique ID for each FUSE request. -var reqIDStep uint64 = 2 - const ( // fuseDefaultMaxBackground is the default value for MaxBackground. fuseDefaultMaxBackground = 12 @@ -52,43 +38,39 @@ const ( fuseDefaultMaxPagesPerReq = 32 ) -// Request represents a FUSE operation request that hasn't been sent to the -// server yet. +// connection is the struct by which the sentry communicates with the FUSE server daemon. // -// +stateify savable -type Request struct { - requestEntry - - id linux.FUSEOpID - hdr *linux.FUSEHeaderIn - data []byte -} - -// Response represents an actual response from the server, including the -// response payload. +// Lock order: +// - conn.fd.mu +// - conn.mu +// - conn.asyncMu // // +stateify savable -type Response struct { - opcode linux.FUSEOpcode - hdr linux.FUSEHeaderOut - data []byte -} - -// connection is the struct by which the sentry communicates with the FUSE server daemon. type connection struct { fd *DeviceFD + // mu protects access to struct memebers. + mu sync.Mutex `state:"nosave"` + + // attributeVersion is the version of connection's attributes. + attributeVersion uint64 + + // We target FUSE 7.23. // The following FUSE_INIT flags are currently unsupported by this implementation: - // - FUSE_ATOMIC_O_TRUNC: requires open(..., O_TRUNC) // - FUSE_EXPORT_SUPPORT - // - FUSE_HANDLE_KILLPRIV // - FUSE_POSIX_LOCKS: requires POSIX locks // - FUSE_FLOCK_LOCKS: requires POSIX locks // - FUSE_AUTO_INVAL_DATA: requires page caching eviction - // - FUSE_EXPLICIT_INVAL_DATA: requires page caching eviction // - FUSE_DO_READDIRPLUS/FUSE_READDIRPLUS_AUTO: requires FUSE_READDIRPLUS implementation // - FUSE_ASYNC_DIO - // - FUSE_POSIX_ACL: affects defaultPermissions, posixACL, xattr handler + // - FUSE_PARALLEL_DIROPS (7.25) + // - FUSE_HANDLE_KILLPRIV (7.26) + // - FUSE_POSIX_ACL: affects defaultPermissions, posixACL, xattr handler (7.26) + // - FUSE_ABORT_ERROR (7.27) + // - FUSE_CACHE_SYMLINKS (7.28) + // - FUSE_NO_OPENDIR_SUPPORT (7.29) + // - FUSE_EXPLICIT_INVAL_DATA: requires page caching eviction (7.30) + // - FUSE_MAP_ALIGNMENT (7.31) // initialized after receiving FUSE_INIT reply. // Until it's set, suspend sending FUSE requests. @@ -96,11 +78,7 @@ type connection struct { initialized int32 // initializedChan is used to block requests before initialization. - initializedChan chan struct{} - - // blocked when there are too many outstading backgrounds requests (NumBackground == MaxBackground). - // TODO(gvisor.dev/issue/3185): update the numBackground accordingly; use a channel to block. - blocked bool + initializedChan chan struct{} `state:".(bool)"` // connected (connection established) when a new FUSE file system is created. // Set to false when: @@ -109,48 +87,55 @@ type connection struct { // device release. connected bool - // aborted via sysfs. - // TODO(gvisor.dev/issue/3185): abort all queued requests. - aborted bool - // connInitError if FUSE_INIT encountered error (major version mismatch). // Only set in INIT. connInitError bool // connInitSuccess if FUSE_INIT is successful. // Only set in INIT. - // Used for destory. + // Used for destory (not yet implemented). connInitSuccess bool - // TODO(gvisor.dev/issue/3185): All the queue logic are working in progress. - - // NumberBackground is the number of requests in the background. - numBackground uint16 + // aborted via sysfs, and will send ECONNABORTED to read after disconnection (instead of ENODEV). + // Set only if abortErr is true and via fuse control fs (not yet implemented). + // TODO(gvisor.dev/issue/3525): set this to true when user aborts. + aborted bool - // congestionThreshold for NumBackground. - // Negotiated in FUSE_INIT. - congestionThreshold uint16 + // numWating is the number of requests waiting to be + // sent to FUSE device or being processed by FUSE daemon. + numWaiting uint32 - // maxBackground is the maximum number of NumBackground. - // Block connection when it is reached. - // Negotiated in FUSE_INIT. - maxBackground uint16 + // Terminology note: + // + // - `asyncNumMax` is the `MaxBackground` in the FUSE_INIT_IN struct. + // + // - `asyncCongestionThreshold` is the `CongestionThreshold` in the FUSE_INIT_IN struct. + // + // We call the "background" requests in unix term as async requests. + // The "async requests" in unix term is our async requests that expect a reply, + // i.e. `!request.noReply` - // numActiveBackground is the number of requests in background and has being marked as active. - numActiveBackground uint16 + // asyncMu protects the async request fields. + asyncMu sync.Mutex `state:"nosave"` - // numWating is the number of requests waiting for completion. - numWaiting uint32 + // asyncNum is the number of async requests. + // Protected by asyncMu. + asyncNum uint16 - // TODO(gvisor.dev/issue/3185): BgQueue - // some queue for background queued requests. + // asyncCongestionThreshold the number of async requests. + // Negotiated in FUSE_INIT as "CongestionThreshold". + // TODO(gvisor.dev/issue/3529): add congestion control. + // Protected by asyncMu. + asyncCongestionThreshold uint16 - // bgLock protects: - // MaxBackground, CongestionThreshold, NumBackground, - // NumActiveBackground, BgQueue, Blocked. - bgLock sync.Mutex + // asyncNumMax is the maximum number of asyncNum. + // Connection blocks the async requests when it is reached. + // Negotiated in FUSE_INIT as "MaxBackground". + // Protected by asyncMu. + asyncNumMax uint16 // maxRead is the maximum size of a read buffer in in bytes. + // Initialized from a fuse fs parameter. maxRead uint32 // maxWrite is the maximum size of a write buffer in bytes. @@ -165,23 +150,20 @@ type connection struct { // Negotiated and only set in INIT. minor uint32 - // asyncRead if read pages asynchronously. + // atomicOTrunc is true when FUSE does not send a separate SETATTR request + // before open with O_TRUNC flag. // Negotiated and only set in INIT. - asyncRead bool + atomicOTrunc bool - // abortErr is true if kernel need to return an unique read error after abort. + // asyncRead if read pages asynchronously. // Negotiated and only set in INIT. - abortErr bool + asyncRead bool // writebackCache is true for write-back cache policy, // false for write-through policy. // Negotiated and only set in INIT. writebackCache bool - // cacheSymlinks if filesystem needs to cache READLINK responses in page cache. - // Negotiated and only set in INIT. - cacheSymlinks bool - // bigWrites if doing multi-page cached writes. // Negotiated and only set in INIT. bigWrites bool @@ -189,116 +171,86 @@ type connection struct { // dontMask if filestestem does not apply umask to creation modes. // Negotiated in INIT. dontMask bool + + // noOpen if FUSE server doesn't support open operation. + // This flag only influence performance, not correctness of the program. + noOpen bool +} + +func (conn *connection) saveInitializedChan() bool { + select { + case <-conn.initializedChan: + return true // Closed. + default: + return false // Not closed. + } +} + +func (conn *connection) loadInitializedChan(closed bool) { + conn.initializedChan = make(chan struct{}, 1) + if closed { + close(conn.initializedChan) + } } // newFUSEConnection creates a FUSE connection to fd. -func newFUSEConnection(_ context.Context, fd *vfs.FileDescription, maxInFlightRequests uint64) (*connection, error) { +func newFUSEConnection(_ context.Context, fd *vfs.FileDescription, opts *filesystemOptions) (*connection, error) { // Mark the device as ready so it can be used. /dev/fuse can only be used if the FD was used to // mount a FUSE filesystem. fuseFD := fd.Impl().(*DeviceFD) - fuseFD.mounted = true // Create the writeBuf for the header to be stored in. hdrLen := uint32((*linux.FUSEHeaderOut)(nil).SizeBytes()) fuseFD.writeBuf = make([]byte, hdrLen) fuseFD.completions = make(map[linux.FUSEOpID]*futureResponse) - fuseFD.fullQueueCh = make(chan struct{}, maxInFlightRequests) + fuseFD.fullQueueCh = make(chan struct{}, opts.maxActiveRequests) fuseFD.writeCursor = 0 return &connection{ - fd: fuseFD, - maxBackground: fuseDefaultMaxBackground, - congestionThreshold: fuseDefaultCongestionThreshold, - maxPages: fuseDefaultMaxPagesPerReq, - initializedChan: make(chan struct{}), - connected: true, - }, nil -} - -// SetInitialized atomically sets the connection as initialized. -func (conn *connection) SetInitialized() { - // Unblock the requests sent before INIT. - close(conn.initializedChan) - - // Close the channel first to avoid the non-atomic situation - // where conn.initialized is true but there are - // tasks being blocked on the channel. - // And it prevents the newer tasks from gaining - // unnecessary higher chance to be issued before the blocked one. - - atomic.StoreInt32(&(conn.initialized), int32(1)) -} - -// IsInitialized atomically check if the connection is initialized. -// pairs with SetInitialized(). -func (conn *connection) Initialized() bool { - return atomic.LoadInt32(&(conn.initialized)) != 0 -} - -// NewRequest creates a new request that can be sent to the FUSE server. -func (conn *connection) NewRequest(creds *auth.Credentials, pid uint32, ino uint64, opcode linux.FUSEOpcode, payload marshal.Marshallable) (*Request, error) { - conn.fd.mu.Lock() - defer conn.fd.mu.Unlock() - conn.fd.nextOpID += linux.FUSEOpID(reqIDStep) - - hdrLen := (*linux.FUSEHeaderIn)(nil).SizeBytes() - hdr := linux.FUSEHeaderIn{ - Len: uint32(hdrLen + payload.SizeBytes()), - Opcode: opcode, - Unique: conn.fd.nextOpID, - NodeID: ino, - UID: uint32(creds.EffectiveKUID), - GID: uint32(creds.EffectiveKGID), - PID: pid, - } - - buf := make([]byte, hdr.Len) - hdr.MarshalUnsafe(buf[:hdrLen]) - payload.MarshalUnsafe(buf[hdrLen:]) - - return &Request{ - id: hdr.Unique, - hdr: &hdr, - data: buf, + fd: fuseFD, + asyncNumMax: fuseDefaultMaxBackground, + asyncCongestionThreshold: fuseDefaultCongestionThreshold, + maxRead: opts.maxRead, + maxPages: fuseDefaultMaxPagesPerReq, + initializedChan: make(chan struct{}), + connected: true, }, nil } -// Call makes a request to the server and blocks the invoking task until a -// server responds with a response. Task should never be nil. -// Requests will not be sent before the connection is initialized. -// For async tasks, use CallAsync(). -func (conn *connection) Call(t *kernel.Task, r *Request) (*Response, error) { - // Block requests sent before connection is initalized. - if !conn.Initialized() { - if err := t.Block(conn.initializedChan); err != nil { - return nil, err - } - } - - return conn.call(t, r) +// CallAsync makes an async (aka background) request. +// It's a simple wrapper around Call(). +func (conn *connection) CallAsync(t *kernel.Task, r *Request) error { + r.async = true + _, err := conn.Call(t, r) + return err } -// CallAsync makes an async (aka background) request. -// Those requests either do not expect a response (e.g. release) or -// the response should be handled by others (e.g. init). -// Return immediately unless the connection is blocked (before initialization). -// Async call example: init, release, forget, aio, interrupt. +// Call makes a request to the server. +// Block before the connection is initialized. // When the Request is FUSE_INIT, it will not be blocked before initialization. -func (conn *connection) CallAsync(t *kernel.Task, r *Request) error { +// Task should never be nil. +// +// For a sync request, it blocks the invoking task until +// a server responds with a response. +// +// For an async request (that do not expect a response immediately), +// it returns directly unless being blocked either before initialization +// or when there are too many async requests ongoing. +// +// Example for async request: +// init, readahead, write, async read/write, fuse_notify_reply, +// non-sync release, interrupt, forget. +// +// The forget request does not have a reply, +// as documented in include/uapi/linux/fuse.h:FUSE_FORGET. +func (conn *connection) Call(t *kernel.Task, r *Request) (*Response, error) { // Block requests sent before connection is initalized. if !conn.Initialized() && r.hdr.Opcode != linux.FUSE_INIT { if err := t.Block(conn.initializedChan); err != nil { - return err + return nil, err } } - // This should be the only place that invokes call() with a nil task. - _, err := conn.call(nil, r) - return err -} - -// call makes a call without blocking checks. -func (conn *connection) call(t *kernel.Task, r *Request) (*Response, error) { if !conn.connected { return nil, syserror.ENOTCONN } @@ -315,31 +267,6 @@ func (conn *connection) call(t *kernel.Task, r *Request) (*Response, error) { return fut.resolve(t) } -// Error returns the error of the FUSE call. -func (r *Response) Error() error { - errno := r.hdr.Error - if errno >= 0 { - return nil - } - - sysErrNo := syscall.Errno(-errno) - return error(sysErrNo) -} - -// UnmarshalPayload unmarshals the response data into m. -func (r *Response) UnmarshalPayload(m marshal.Marshallable) error { - hdrLen := r.hdr.SizeBytes() - haveDataLen := r.hdr.Len - uint32(hdrLen) - wantDataLen := uint32(m.SizeBytes()) - - if haveDataLen < wantDataLen { - return fmt.Errorf("payload too small. Minimum data lenth required: %d, but got data length %d", wantDataLen, haveDataLen) - } - - m.UnmarshalUnsafe(r.data[hdrLen:]) - return nil -} - // callFuture makes a request to the server and returns a future response. // Call resolve() when the response needs to be fulfilled. func (conn *connection) callFuture(t *kernel.Task, r *Request) (*futureResponse, error) { @@ -358,11 +285,6 @@ func (conn *connection) callFuture(t *kernel.Task, r *Request) (*futureResponse, // if there are always too many ongoing requests all the time. The // supported maxActiveRequests setting should be really high to avoid this. for conn.fd.numActiveRequests == conn.fd.fs.opts.maxActiveRequests { - if t == nil { - // Since there is no task that is waiting. We must error out. - return nil, errors.New("FUSE request queue full") - } - log.Infof("Blocking request %v from being queued. Too many active requests: %v", r.id, conn.fd.numActiveRequests) conn.fd.mu.Unlock() @@ -378,9 +300,19 @@ func (conn *connection) callFuture(t *kernel.Task, r *Request) (*futureResponse, // callFutureLocked makes a request to the server and returns a future response. func (conn *connection) callFutureLocked(t *kernel.Task, r *Request) (*futureResponse, error) { + // Check connected again holding conn.mu. + conn.mu.Lock() + if !conn.connected { + conn.mu.Unlock() + // we checked connected before, + // this must be due to aborted connection. + return nil, syserror.ECONNABORTED + } + conn.mu.Unlock() + conn.fd.queue.PushBack(r) - conn.fd.numActiveRequests += 1 - fut := newFutureResponse(r.hdr.Opcode) + conn.fd.numActiveRequests++ + fut := newFutureResponse(r) conn.fd.completions[r.id] = fut // Signal the readers that there is something to read. @@ -388,50 +320,3 @@ func (conn *connection) callFutureLocked(t *kernel.Task, r *Request) (*futureRes return fut, nil } - -// futureResponse represents an in-flight request, that may or may not have -// completed yet. Convert it to a resolved Response by calling Resolve, but note -// that this may block. -// -// +stateify savable -type futureResponse struct { - opcode linux.FUSEOpcode - ch chan struct{} - hdr *linux.FUSEHeaderOut - data []byte -} - -// newFutureResponse creates a future response to a FUSE request. -func newFutureResponse(opcode linux.FUSEOpcode) *futureResponse { - return &futureResponse{ - opcode: opcode, - ch: make(chan struct{}), - } -} - -// resolve blocks the task until the server responds to its corresponding request, -// then returns a resolved response. -func (f *futureResponse) resolve(t *kernel.Task) (*Response, error) { - // If there is no Task associated with this request - then we don't try to resolve - // the response. Instead, the task writing the response (proxy to the server) will - // process the response on our behalf. - if t == nil { - log.Infof("fuse.Response.resolve: Not waiting on a response from server.") - return nil, nil - } - - if err := t.Block(f.ch); err != nil { - return nil, err - } - - return f.getResponse(), nil -} - -// getResponse creates a Response from the data the futureResponse has. -func (f *futureResponse) getResponse() *Response { - return &Response{ - opcode: f.opcode, - hdr: *f.hdr, - data: f.data, - } -} diff --git a/pkg/sentry/fsimpl/fuse/init.go b/pkg/sentry/fsimpl/fuse/connection_control.go index 779c2bd3f..bfde78559 100644 --- a/pkg/sentry/fsimpl/fuse/init.go +++ b/pkg/sentry/fsimpl/fuse/connection_control.go @@ -15,7 +15,11 @@ package fuse import ( + "sync/atomic" + "syscall" + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" ) @@ -29,9 +33,10 @@ const ( // Follow the same behavior as unix fuse implementation. fuseMaxTimeGranNs = 1000000000 - // Minimum value for MaxWrite. + // Minimum value for MaxWrite and MaxRead. // Follow the same behavior as unix fuse implementation. fuseMinMaxWrite = 4096 + fuseMinMaxRead = 4096 // Temporary default value for max readahead, 128kb. fuseDefaultMaxReadahead = 131072 @@ -49,6 +54,26 @@ var ( MaxUserCongestionThreshold uint16 = fuseDefaultCongestionThreshold ) +// SetInitialized atomically sets the connection as initialized. +func (conn *connection) SetInitialized() { + // Unblock the requests sent before INIT. + close(conn.initializedChan) + + // Close the channel first to avoid the non-atomic situation + // where conn.initialized is true but there are + // tasks being blocked on the channel. + // And it prevents the newer tasks from gaining + // unnecessary higher chance to be issued before the blocked one. + + atomic.StoreInt32(&(conn.initialized), int32(1)) +} + +// IsInitialized atomically check if the connection is initialized. +// pairs with SetInitialized(). +func (conn *connection) Initialized() bool { + return atomic.LoadInt32(&(conn.initialized)) != 0 +} + // InitSend sends a FUSE_INIT request. func (conn *connection) InitSend(creds *auth.Credentials, pid uint32) error { in := linux.FUSEInitIn{ @@ -70,29 +95,31 @@ func (conn *connection) InitSend(creds *auth.Credentials, pid uint32) error { } // InitRecv receives a FUSE_INIT reply and process it. +// +// Preconditions: conn.asyncMu must not be held if minor verion is newer than 13. func (conn *connection) InitRecv(res *Response, hasSysAdminCap bool) error { if err := res.Error(); err != nil { return err } - var out linux.FUSEInitOut - if err := res.UnmarshalPayload(&out); err != nil { + initRes := fuseInitRes{initLen: res.DataLen()} + if err := res.UnmarshalPayload(&initRes); err != nil { return err } - return conn.initProcessReply(&out, hasSysAdminCap) + return conn.initProcessReply(&initRes.initOut, hasSysAdminCap) } // Process the FUSE_INIT reply from the FUSE server. +// It tries to acquire the conn.asyncMu lock if minor version is newer than 13. func (conn *connection) initProcessReply(out *linux.FUSEInitOut, hasSysAdminCap bool) error { + // No matter error or not, always set initialzied. + // to unblock the blocked requests. + defer conn.SetInitialized() + // No support for old major fuse versions. if out.Major != linux.FUSE_KERNEL_VERSION { conn.connInitError = true - - // Set the connection as initialized and unblock the blocked requests - // (i.e. return error for them). - conn.SetInitialized() - return nil } @@ -100,29 +127,14 @@ func (conn *connection) initProcessReply(out *linux.FUSEInitOut, hasSysAdminCap conn.connInitSuccess = true conn.minor = out.Minor - // No support for limits before minor version 13. - if out.Minor >= 13 { - conn.bgLock.Lock() - - if out.MaxBackground > 0 { - conn.maxBackground = out.MaxBackground - - if !hasSysAdminCap && - conn.maxBackground > MaxUserBackgroundRequest { - conn.maxBackground = MaxUserBackgroundRequest - } - } - - if out.CongestionThreshold > 0 { - conn.congestionThreshold = out.CongestionThreshold - - if !hasSysAdminCap && - conn.congestionThreshold > MaxUserCongestionThreshold { - conn.congestionThreshold = MaxUserCongestionThreshold - } - } - - conn.bgLock.Unlock() + // No support for negotiating MaxWrite before minor version 5. + if out.Minor >= 5 { + conn.maxWrite = out.MaxWrite + } else { + conn.maxWrite = fuseMinMaxWrite + } + if conn.maxWrite < fuseMinMaxWrite { + conn.maxWrite = fuseMinMaxWrite } // No support for the following flags before minor version 6. @@ -131,8 +143,6 @@ func (conn *connection) initProcessReply(out *linux.FUSEInitOut, hasSysAdminCap conn.bigWrites = out.Flags&linux.FUSE_BIG_WRITES != 0 conn.dontMask = out.Flags&linux.FUSE_DONT_MASK != 0 conn.writebackCache = out.Flags&linux.FUSE_WRITEBACK_CACHE != 0 - conn.cacheSymlinks = out.Flags&linux.FUSE_CACHE_SYMLINKS != 0 - conn.abortErr = out.Flags&linux.FUSE_ABORT_ERROR != 0 // TODO(gvisor.dev/issue/3195): figure out how to use TimeGran (0 < TimeGran <= fuseMaxTimeGranNs). @@ -148,19 +158,90 @@ func (conn *connection) initProcessReply(out *linux.FUSEInitOut, hasSysAdminCap } } - // No support for negotiating MaxWrite before minor version 5. - if out.Minor >= 5 { - conn.maxWrite = out.MaxWrite - } else { - conn.maxWrite = fuseMinMaxWrite + // No support for limits before minor version 13. + if out.Minor >= 13 { + conn.asyncMu.Lock() + + if out.MaxBackground > 0 { + conn.asyncNumMax = out.MaxBackground + + if !hasSysAdminCap && + conn.asyncNumMax > MaxUserBackgroundRequest { + conn.asyncNumMax = MaxUserBackgroundRequest + } + } + + if out.CongestionThreshold > 0 { + conn.asyncCongestionThreshold = out.CongestionThreshold + + if !hasSysAdminCap && + conn.asyncCongestionThreshold > MaxUserCongestionThreshold { + conn.asyncCongestionThreshold = MaxUserCongestionThreshold + } + } + + conn.asyncMu.Unlock() } - if conn.maxWrite < fuseMinMaxWrite { - conn.maxWrite = fuseMinMaxWrite + + return nil +} + +// Abort this FUSE connection. +// It tries to acquire conn.fd.mu, conn.lock, conn.bgLock in order. +// All possible requests waiting or blocking will be aborted. +// +// Preconditions: conn.fd.mu is locked. +func (conn *connection) Abort(ctx context.Context) { + conn.mu.Lock() + conn.asyncMu.Lock() + + if !conn.connected { + conn.asyncMu.Unlock() + conn.mu.Unlock() + conn.fd.mu.Unlock() + return } - // Set connection as initialized and unblock the requests - // issued before init. - conn.SetInitialized() + conn.connected = false - return nil + // Empty the `fd.queue` that holds the requests + // not yet read by the FUSE daemon yet. + // These are a subset of the requests in `fuse.completion` map. + for !conn.fd.queue.Empty() { + req := conn.fd.queue.Front() + conn.fd.queue.Remove(req) + } + + var terminate []linux.FUSEOpID + + // 2. Collect the requests have not been sent to FUSE daemon, + // or have not received a reply. + for unique := range conn.fd.completions { + terminate = append(terminate, unique) + } + + // Release locks to avoid deadlock. + conn.asyncMu.Unlock() + conn.mu.Unlock() + + // 1. The requets blocked before initialization. + // Will reach call() `connected` check and return. + if !conn.Initialized() { + conn.SetInitialized() + } + + // 2. Terminate the requests collected above. + // Set ECONNABORTED error. + // sendError() will remove them from `fd.completion` map. + // Will enter the path of a normally received error. + for _, toTerminate := range terminate { + conn.fd.sendError(ctx, -int32(syscall.ECONNABORTED), toTerminate) + } + + // 3. The requests not yet written to FUSE device. + // Early terminate. + // Will reach callFutureLocked() `connected` check and return. + close(conn.fd.fullQueueCh) + + // TODO(gvisor.dev/issue/3528): Forget all pending forget reqs. } diff --git a/pkg/sentry/fsimpl/fuse/connection_test.go b/pkg/sentry/fsimpl/fuse/connection_test.go new file mode 100644 index 000000000..91d16c1cf --- /dev/null +++ b/pkg/sentry/fsimpl/fuse/connection_test.go @@ -0,0 +1,117 @@ +// 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 fuse + +import ( + "math/rand" + "syscall" + "testing" + + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/syserror" +) + +// TestConnectionInitBlock tests if initialization +// correctly blocks and unblocks the connection. +// Since it's unfeasible to test kernelTask.Block() in unit test, +// the code in Call() are not tested here. +func TestConnectionInitBlock(t *testing.T) { + s := setup(t) + defer s.Destroy() + + k := kernel.KernelFromContext(s.Ctx) + + conn, _, err := newTestConnection(s, k, maxActiveRequestsDefault) + if err != nil { + t.Fatalf("newTestConnection: %v", err) + } + + select { + case <-conn.initializedChan: + t.Fatalf("initializedChan should be blocking before SetInitialized") + default: + } + + conn.SetInitialized() + + select { + case <-conn.initializedChan: + default: + t.Fatalf("initializedChan should not be blocking after SetInitialized") + } +} + +func TestConnectionAbort(t *testing.T) { + s := setup(t) + defer s.Destroy() + + k := kernel.KernelFromContext(s.Ctx) + creds := auth.CredentialsFromContext(s.Ctx) + task := kernel.TaskFromContext(s.Ctx) + + const numRequests uint64 = 256 + + conn, _, err := newTestConnection(s, k, numRequests) + if err != nil { + t.Fatalf("newTestConnection: %v", err) + } + + testObj := &testPayload{ + data: rand.Uint32(), + } + + var futNormal []*futureResponse + + for i := 0; i < int(numRequests); i++ { + req, err := conn.NewRequest(creds, uint32(i), uint64(i), 0, testObj) + if err != nil { + t.Fatalf("NewRequest creation failed: %v", err) + } + fut, err := conn.callFutureLocked(task, req) + if err != nil { + t.Fatalf("callFutureLocked failed: %v", err) + } + futNormal = append(futNormal, fut) + } + + conn.Abort(s.Ctx) + + // Abort should unblock the initialization channel. + // Note: no test requests are actually blocked on `conn.initializedChan`. + select { + case <-conn.initializedChan: + default: + t.Fatalf("initializedChan should not be blocking after SetInitialized") + } + + // Abort will return ECONNABORTED error to unblocked requests. + for _, fut := range futNormal { + if fut.getResponse().hdr.Error != -int32(syscall.ECONNABORTED) { + t.Fatalf("Incorrect error code received for aborted connection: %v", fut.getResponse().hdr.Error) + } + } + + // After abort, Call() should return directly with ENOTCONN. + req, err := conn.NewRequest(creds, 0, 0, 0, testObj) + if err != nil { + t.Fatalf("NewRequest creation failed: %v", err) + } + _, err = conn.Call(task, req) + if err != syserror.ENOTCONN { + t.Fatalf("Incorrect error code received for Call() after connection aborted") + } + +} diff --git a/pkg/sentry/fsimpl/fuse/dev.go b/pkg/sentry/fsimpl/fuse/dev.go index e522ff9a0..1b86a4b4c 100644 --- a/pkg/sentry/fsimpl/fuse/dev.go +++ b/pkg/sentry/fsimpl/fuse/dev.go @@ -19,7 +19,6 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" - "gvisor.dev/gvisor/pkg/log" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/vfs" @@ -32,6 +31,8 @@ import ( const fuseDevMinor = 229 // fuseDevice implements vfs.Device for /dev/fuse. +// +// +stateify savable type fuseDevice struct{} // Open implements vfs.Device.Open. @@ -50,15 +51,14 @@ func (fuseDevice) Open(ctx context.Context, mnt *vfs.Mount, vfsd *vfs.Dentry, op } // DeviceFD implements vfs.FileDescriptionImpl for /dev/fuse. +// +// +stateify savable type DeviceFD struct { vfsfd vfs.FileDescription vfs.FileDescriptionDefaultImpl vfs.DentryMetadataFileDescriptionImpl vfs.NoLockFD - // mounted specifies whether a FUSE filesystem was mounted using the DeviceFD. - mounted bool - // nextOpID is used to create new requests. nextOpID linux.FUSEOpID @@ -83,7 +83,7 @@ type DeviceFD struct { writeCursorFR *futureResponse // mu protects all the queues, maps, buffers and cursors and nextOpID. - mu sync.Mutex + mu sync.Mutex `state:"nosave"` // waitQueue is used to notify interested parties when the device becomes // readable or writable. @@ -92,21 +92,36 @@ type DeviceFD struct { // fullQueueCh is a channel used to synchronize the readers with the writers. // Writers (inbound requests to the filesystem) block if there are too many // unprocessed in-flight requests. - fullQueueCh chan struct{} + fullQueueCh chan struct{} `state:".(int)"` // fs is the FUSE filesystem that this FD is being used for. fs *filesystem } +func (fd *DeviceFD) saveFullQueueCh() int { + return cap(fd.fullQueueCh) +} + +func (fd *DeviceFD) loadFullQueueCh(capacity int) { + fd.fullQueueCh = make(chan struct{}, capacity) +} + // Release implements vfs.FileDescriptionImpl.Release. -func (fd *DeviceFD) Release(context.Context) { - fd.fs.conn.connected = false +func (fd *DeviceFD) Release(ctx context.Context) { + if fd.fs != nil { + fd.fs.conn.mu.Lock() + fd.fs.conn.connected = false + fd.fs.conn.mu.Unlock() + + fd.fs.VFSFilesystem().DecRef(ctx) + fd.fs = nil + } } // PRead implements vfs.FileDescriptionImpl.PRead. func (fd *DeviceFD) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { // Operations on /dev/fuse don't make sense until a FUSE filesystem is mounted. - if !fd.mounted { + if fd.fs == nil { return 0, syserror.EPERM } @@ -116,10 +131,16 @@ func (fd *DeviceFD) PRead(ctx context.Context, dst usermem.IOSequence, offset in // Read implements vfs.FileDescriptionImpl.Read. func (fd *DeviceFD) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { // Operations on /dev/fuse don't make sense until a FUSE filesystem is mounted. - if !fd.mounted { + if fd.fs == nil { return 0, syserror.EPERM } + // Return ENODEV if the filesystem is umounted. + if fd.fs.umounted { + // TODO(gvisor.dev/issue/3525): return ECONNABORTED if aborted via fuse control fs. + return 0, syserror.ENODEV + } + // We require that any Read done on this filesystem have a sane minimum // read buffer. It must have the capacity for the fixed parts of any request // header (Linux uses the request header and the FUSEWriteIn header for this @@ -143,58 +164,82 @@ func (fd *DeviceFD) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.R } // readLocked implements the reading of the fuse device while locked with DeviceFD.mu. +// +// Preconditions: dst is large enough for any reasonable request. func (fd *DeviceFD) readLocked(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { - if fd.queue.Empty() { - return 0, syserror.ErrWouldBlock - } + var req *Request - var readCursor uint32 - var bytesRead int64 - for { - req := fd.queue.Front() - if dst.NumBytes() < int64(req.hdr.Len) { - // The request is too large. Cannot process it. All requests must be smaller than the - // negotiated size as specified by Connection.MaxWrite set as part of the FUSE_INIT - // handshake. - errno := -int32(syscall.EIO) - if req.hdr.Opcode == linux.FUSE_SETXATTR { - errno = -int32(syscall.E2BIG) - } + // Find the first valid request. + // For the normal case this loop only execute once. + for !fd.queue.Empty() { + req = fd.queue.Front() - // Return the error to the calling task. - if err := fd.sendError(ctx, errno, req); err != nil { - return 0, err - } + if int64(req.hdr.Len)+int64(len(req.payload)) <= dst.NumBytes() { + break + } - // We're done with this request. - fd.queue.Remove(req) + // The request is too large. Cannot process it. All requests must be smaller than the + // negotiated size as specified by Connection.MaxWrite set as part of the FUSE_INIT + // handshake. + errno := -int32(syscall.EIO) + if req.hdr.Opcode == linux.FUSE_SETXATTR { + errno = -int32(syscall.E2BIG) + } - // Restart the read as this request was invalid. - log.Warningf("fuse.DeviceFD.Read: request found was too large. Restarting read.") - return fd.readLocked(ctx, dst, opts) + // Return the error to the calling task. + if err := fd.sendError(ctx, errno, req.hdr.Unique); err != nil { + return 0, err } - n, err := dst.CopyOut(ctx, req.data[readCursor:]) + // We're done with this request. + fd.queue.Remove(req) + req = nil + } + + if req == nil { + return 0, syserror.ErrWouldBlock + } + + // We already checked the size: dst must be able to fit the whole request. + // Now we write the marshalled header, the payload, + // and the potential additional payload + // to the user memory IOSequence. + + n, err := dst.CopyOut(ctx, req.data) + if err != nil { + return 0, err + } + if n != len(req.data) { + return 0, syserror.EIO + } + + if req.hdr.Opcode == linux.FUSE_WRITE { + written, err := dst.DropFirst(n).CopyOut(ctx, req.payload) if err != nil { return 0, err } - readCursor += uint32(n) - bytesRead += int64(n) - - if readCursor >= req.hdr.Len { - // Fully done with this req, remove it from the queue. - fd.queue.Remove(req) - break + if written != len(req.payload) { + return 0, syserror.EIO } + n += int(written) } - return bytesRead, nil + // Fully done with this req, remove it from the queue. + fd.queue.Remove(req) + + // Remove noReply ones from map of requests expecting a reply. + if req.noReply { + fd.numActiveRequests -= 1 + delete(fd.completions, req.hdr.Unique) + } + + return int64(n), nil } // PWrite implements vfs.FileDescriptionImpl.PWrite. func (fd *DeviceFD) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) { // Operations on /dev/fuse don't make sense until a FUSE filesystem is mounted. - if !fd.mounted { + if fd.fs == nil { return 0, syserror.EPERM } @@ -211,10 +256,15 @@ func (fd *DeviceFD) Write(ctx context.Context, src usermem.IOSequence, opts vfs. // writeLocked implements writing to the fuse device while locked with DeviceFD.mu. func (fd *DeviceFD) writeLocked(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) { // Operations on /dev/fuse don't make sense until a FUSE filesystem is mounted. - if !fd.mounted { + if fd.fs == nil { return 0, syserror.EPERM } + // Return ENODEV if the filesystem is umounted. + if fd.fs.umounted { + return 0, syserror.ENODEV + } + var cn, n int64 hdrLen := uint32((*linux.FUSEHeaderOut)(nil).SizeBytes()) @@ -276,7 +326,8 @@ func (fd *DeviceFD) writeLocked(ctx context.Context, src usermem.IOSequence, opt fut, ok := fd.completions[hdr.Unique] if !ok { - // Server sent us a response for a request we never sent? + // Server sent us a response for a request we never sent, + // or for which we already received a reply (e.g. aborted), an unlikely event. return 0, syserror.EINVAL } @@ -307,8 +358,23 @@ func (fd *DeviceFD) writeLocked(ctx context.Context, src usermem.IOSequence, opt // Readiness implements vfs.FileDescriptionImpl.Readiness. func (fd *DeviceFD) Readiness(mask waiter.EventMask) waiter.EventMask { + fd.mu.Lock() + defer fd.mu.Unlock() + return fd.readinessLocked(mask) +} + +// readinessLocked implements checking the readiness of the fuse device while +// locked with DeviceFD.mu. +func (fd *DeviceFD) readinessLocked(mask waiter.EventMask) waiter.EventMask { var ready waiter.EventMask - ready |= waiter.EventOut // FD is always writable + + if fd.fs.umounted { + ready |= waiter.EventErr + return ready & mask + } + + // FD is always writable. + ready |= waiter.EventOut if !fd.queue.Empty() { // Have reqs available, FD is readable. ready |= waiter.EventIn @@ -330,7 +396,7 @@ func (fd *DeviceFD) EventUnregister(e *waiter.Entry) { // Seek implements vfs.FileDescriptionImpl.Seek. func (fd *DeviceFD) Seek(ctx context.Context, offset int64, whence int32) (int64, error) { // Operations on /dev/fuse don't make sense until a FUSE filesystem is mounted. - if !fd.mounted { + if fd.fs == nil { return 0, syserror.EPERM } @@ -338,59 +404,59 @@ func (fd *DeviceFD) Seek(ctx context.Context, offset int64, whence int32) (int64 } // sendResponse sends a response to the waiting task (if any). +// +// Preconditions: fd.mu must be held. func (fd *DeviceFD) sendResponse(ctx context.Context, fut *futureResponse) error { - // See if the running task need to perform some action before returning. - // Since we just finished writing the future, we can be sure that - // getResponse generates a populated response. - if err := fd.noReceiverAction(ctx, fut.getResponse()); err != nil { - return err - } + // Signal the task waiting on a response if any. + defer close(fut.ch) // Signal that the queue is no longer full. select { case fd.fullQueueCh <- struct{}{}: default: } - fd.numActiveRequests -= 1 + fd.numActiveRequests-- + + if fut.async { + return fd.asyncCallBack(ctx, fut.getResponse()) + } - // Signal the task waiting on a response. - close(fut.ch) return nil } -// sendError sends an error response to the waiting task (if any). -func (fd *DeviceFD) sendError(ctx context.Context, errno int32, req *Request) error { +// sendError sends an error response to the waiting task (if any) by calling sendResponse(). +// +// Preconditions: fd.mu must be held. +func (fd *DeviceFD) sendError(ctx context.Context, errno int32, unique linux.FUSEOpID) error { // Return the error to the calling task. outHdrLen := uint32((*linux.FUSEHeaderOut)(nil).SizeBytes()) respHdr := linux.FUSEHeaderOut{ Len: outHdrLen, Error: errno, - Unique: req.hdr.Unique, + Unique: unique, } fut, ok := fd.completions[respHdr.Unique] if !ok { - // Server sent us a response for a request we never sent? + // A response for a request we never sent, + // or for which we already received a reply (e.g. aborted). return syserror.EINVAL } delete(fd.completions, respHdr.Unique) fut.hdr = &respHdr - if err := fd.sendResponse(ctx, fut); err != nil { - return err - } - - return nil + return fd.sendResponse(ctx, fut) } -// noReceiverAction has the calling kernel.Task do some action if its known that no -// receiver is going to be waiting on the future channel. This is to be used by: -// FUSE_INIT. -func (fd *DeviceFD) noReceiverAction(ctx context.Context, r *Response) error { - if r.opcode == linux.FUSE_INIT { +// asyncCallBack executes pre-defined callback function for async requests. +// Currently used by: FUSE_INIT. +func (fd *DeviceFD) asyncCallBack(ctx context.Context, r *Response) error { + switch r.opcode { + case linux.FUSE_INIT: creds := auth.CredentialsFromContext(ctx) rootUserNs := kernel.KernelFromContext(ctx).RootUserNamespace() return fd.fs.conn.InitRecv(r, creds.HasCapabilityIn(linux.CAP_SYS_ADMIN, rootUserNs)) + // TODO(gvisor.dev/issue/3247): support async read: correctly process the response. } return nil diff --git a/pkg/sentry/fsimpl/fuse/dev_test.go b/pkg/sentry/fsimpl/fuse/dev_test.go index 6baf56520..5986133e9 100644 --- a/pkg/sentry/fsimpl/fuse/dev_test.go +++ b/pkg/sentry/fsimpl/fuse/dev_test.go @@ -16,12 +16,10 @@ package fuse import ( "fmt" - "io" "math/rand" "testing" "gvisor.dev/gvisor/pkg/abi/linux" - "gvisor.dev/gvisor/pkg/marshal" "gvisor.dev/gvisor/pkg/sentry/fsimpl/testutil" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" @@ -35,10 +33,6 @@ import ( // will simply echo the payload back with the appropriate headers. const echoTestOpcode linux.FUSEOpcode = 1000 -type testPayload struct { - data uint32 -} - // TestFUSECommunication tests that the communication layer between the Sentry and the // FUSE server daemon works as expected. func TestFUSECommunication(t *testing.T) { @@ -327,102 +321,3 @@ func fuseServerRun(t *testing.T, s *testutil.System, k *kernel.Kernel, fd *vfs.F } } } - -func setup(t *testing.T) *testutil.System { - k, err := testutil.Boot() - if err != nil { - t.Fatalf("Error creating kernel: %v", err) - } - - ctx := k.SupervisorContext() - creds := auth.CredentialsFromContext(ctx) - - k.VFS().MustRegisterFilesystemType(Name, &FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ - AllowUserList: true, - AllowUserMount: true, - }) - - mntns, err := k.VFS().NewMountNamespace(ctx, creds, "", "tmpfs", &vfs.MountOptions{}) - if err != nil { - t.Fatalf("NewMountNamespace(): %v", err) - } - - return testutil.NewSystem(ctx, t, k.VFS(), mntns) -} - -// newTestConnection creates a fuse connection that the sentry can communicate with -// and the FD for the server to communicate with. -func newTestConnection(system *testutil.System, k *kernel.Kernel, maxActiveRequests uint64) (*connection, *vfs.FileDescription, error) { - vfsObj := &vfs.VirtualFilesystem{} - fuseDev := &DeviceFD{} - - if err := vfsObj.Init(system.Ctx); err != nil { - return nil, nil, err - } - - vd := vfsObj.NewAnonVirtualDentry("genCountFD") - defer vd.DecRef(system.Ctx) - if err := fuseDev.vfsfd.Init(fuseDev, linux.O_RDWR|linux.O_CREAT, vd.Mount(), vd.Dentry(), &vfs.FileDescriptionOptions{}); err != nil { - return nil, nil, err - } - - fsopts := filesystemOptions{ - maxActiveRequests: maxActiveRequests, - } - fs, err := NewFUSEFilesystem(system.Ctx, 0, &fsopts, &fuseDev.vfsfd) - if err != nil { - return nil, nil, err - } - - return fs.conn, &fuseDev.vfsfd, nil -} - -// SizeBytes implements marshal.Marshallable.SizeBytes. -func (t *testPayload) SizeBytes() int { - return 4 -} - -// MarshalBytes implements marshal.Marshallable.MarshalBytes. -func (t *testPayload) MarshalBytes(dst []byte) { - usermem.ByteOrder.PutUint32(dst[:4], t.data) -} - -// UnmarshalBytes implements marshal.Marshallable.UnmarshalBytes. -func (t *testPayload) UnmarshalBytes(src []byte) { - *t = testPayload{data: usermem.ByteOrder.Uint32(src[:4])} -} - -// Packed implements marshal.Marshallable.Packed. -func (t *testPayload) Packed() bool { - return true -} - -// MarshalUnsafe implements marshal.Marshallable.MarshalUnsafe. -func (t *testPayload) MarshalUnsafe(dst []byte) { - t.MarshalBytes(dst) -} - -// UnmarshalUnsafe implements marshal.Marshallable.UnmarshalUnsafe. -func (t *testPayload) UnmarshalUnsafe(src []byte) { - t.UnmarshalBytes(src) -} - -// CopyOutN implements marshal.Marshallable.CopyOutN. -func (t *testPayload) CopyOutN(task marshal.Task, addr usermem.Addr, limit int) (int, error) { - panic("not implemented") -} - -// CopyOut implements marshal.Marshallable.CopyOut. -func (t *testPayload) CopyOut(task marshal.Task, addr usermem.Addr) (int, error) { - panic("not implemented") -} - -// CopyIn implements marshal.Marshallable.CopyIn. -func (t *testPayload) CopyIn(task marshal.Task, addr usermem.Addr) (int, error) { - panic("not implemented") -} - -// WriteTo implements io.WriterTo.WriteTo. -func (t *testPayload) WriteTo(w io.Writer) (int64, error) { - panic("not implemented") -} diff --git a/pkg/sentry/fsimpl/fuse/directory.go b/pkg/sentry/fsimpl/fuse/directory.go new file mode 100644 index 000000000..8f220a04b --- /dev/null +++ b/pkg/sentry/fsimpl/fuse/directory.go @@ -0,0 +1,105 @@ +// 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 fuse + +import ( + "sync/atomic" + + "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/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +type directoryFD struct { + fileDescription +} + +// Allocate implements directoryFD.Allocate. +func (*directoryFD) Allocate(ctx context.Context, mode, offset, length uint64) error { + return syserror.EISDIR +} + +// PRead implements vfs.FileDescriptionImpl.PRead. +func (*directoryFD) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { + return 0, syserror.EISDIR +} + +// Read implements vfs.FileDescriptionImpl.Read. +func (*directoryFD) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { + return 0, syserror.EISDIR +} + +// PWrite implements vfs.FileDescriptionImpl.PWrite. +func (*directoryFD) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) { + return 0, syserror.EISDIR +} + +// Write implements vfs.FileDescriptionImpl.Write. +func (*directoryFD) Write(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) { + return 0, syserror.EISDIR +} + +// IterDirents implements vfs.FileDescriptionImpl.IterDirents. +func (dir *directoryFD) IterDirents(ctx context.Context, callback vfs.IterDirentsCallback) error { + fusefs := dir.inode().fs + task, creds := kernel.TaskFromContext(ctx), auth.CredentialsFromContext(ctx) + + in := linux.FUSEReadIn{ + Fh: dir.Fh, + Offset: uint64(atomic.LoadInt64(&dir.off)), + Size: linux.FUSE_PAGE_SIZE, + Flags: dir.statusFlags(), + } + + // TODO(gVisor.dev/issue/3404): Support FUSE_READDIRPLUS. + req, err := fusefs.conn.NewRequest(creds, uint32(task.ThreadID()), dir.inode().nodeID, linux.FUSE_READDIR, &in) + if err != nil { + return err + } + + res, err := fusefs.conn.Call(task, req) + if err != nil { + return err + } + if err := res.Error(); err != nil { + return err + } + + var out linux.FUSEDirents + if err := res.UnmarshalPayload(&out); err != nil { + return err + } + + for _, fuseDirent := range out.Dirents { + nextOff := int64(fuseDirent.Meta.Off) + dirent := vfs.Dirent{ + Name: fuseDirent.Name, + Type: uint8(fuseDirent.Meta.Type), + Ino: fuseDirent.Meta.Ino, + NextOff: nextOff, + } + + if err := callback.Handle(dirent); err != nil { + return err + } + atomic.StoreInt64(&dir.off, nextOff) + } + + return nil +} diff --git a/pkg/sentry/fsimpl/fuse/file.go b/pkg/sentry/fsimpl/fuse/file.go new file mode 100644 index 000000000..83f2816b7 --- /dev/null +++ b/pkg/sentry/fsimpl/fuse/file.go @@ -0,0 +1,133 @@ +// 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 fuse + +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" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/usermem" +) + +// fileDescription implements vfs.FileDescriptionImpl for fuse. +type fileDescription struct { + vfsfd vfs.FileDescription + vfs.FileDescriptionDefaultImpl + vfs.DentryMetadataFileDescriptionImpl + vfs.NoLockFD + + // the file handle used in userspace. + Fh uint64 + + // Nonseekable is indicate cannot perform seek on a file. + Nonseekable bool + + // DirectIO suggest fuse to use direct io operation. + DirectIO bool + + // OpenFlag is the flag returned by open. + OpenFlag uint32 + + // off is the file offset. + off int64 +} + +func (fd *fileDescription) dentry() *kernfs.Dentry { + return fd.vfsfd.Dentry().Impl().(*kernfs.Dentry) +} + +func (fd *fileDescription) inode() *inode { + return fd.dentry().Inode().(*inode) +} + +func (fd *fileDescription) filesystem() *vfs.Filesystem { + return fd.vfsfd.VirtualDentry().Mount().Filesystem() +} + +func (fd *fileDescription) statusFlags() uint32 { + return fd.vfsfd.StatusFlags() +} + +// Release implements vfs.FileDescriptionImpl.Release. +func (fd *fileDescription) Release(ctx context.Context) { + // no need to release if FUSE server doesn't implement Open. + conn := fd.inode().fs.conn + if conn.noOpen { + return + } + + in := linux.FUSEReleaseIn{ + Fh: fd.Fh, + Flags: fd.statusFlags(), + } + // TODO(gvisor.dev/issue/3245): add logic when we support file lock owner. + var opcode linux.FUSEOpcode + if fd.inode().Mode().IsDir() { + opcode = linux.FUSE_RELEASEDIR + } else { + opcode = linux.FUSE_RELEASE + } + kernelTask := kernel.TaskFromContext(ctx) + // ignoring errors and FUSE server reply is analogous to Linux's behavior. + req, err := conn.NewRequest(auth.CredentialsFromContext(ctx), uint32(kernelTask.ThreadID()), fd.inode().nodeID, opcode, &in) + if err != nil { + // No way to invoke Call() with an errored request. + return + } + // The reply will be ignored since no callback is defined in asyncCallBack(). + conn.CallAsync(kernelTask, req) +} + +// PRead implements vfs.FileDescriptionImpl.PRead. +func (fd *fileDescription) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { + return 0, nil +} + +// Read implements vfs.FileDescriptionImpl.Read. +func (fd *fileDescription) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { + return 0, nil +} + +// PWrite implements vfs.FileDescriptionImpl.PWrite. +func (fd *fileDescription) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) { + return 0, nil +} + +// Write implements vfs.FileDescriptionImpl.Write. +func (fd *fileDescription) Write(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) { + return 0, nil +} + +// Seek implements vfs.FileDescriptionImpl.Seek. +func (fd *fileDescription) Seek(ctx context.Context, offset int64, whence int32) (int64, error) { + return 0, nil +} + +// Stat implements vfs.FileDescriptionImpl.Stat. +func (fd *fileDescription) Stat(ctx context.Context, opts vfs.StatOptions) (linux.Statx, error) { + fs := fd.filesystem() + inode := fd.inode() + return inode.Stat(ctx, fs, opts) +} + +// SetStat implements vfs.FileDescriptionImpl.SetStat. +func (fd *fileDescription) SetStat(ctx context.Context, opts vfs.SetStatOptions) error { + fs := fd.filesystem() + creds := auth.CredentialsFromContext(ctx) + return fd.inode().setAttr(ctx, fs, creds, opts, true, fd.Fh) +} diff --git a/pkg/sentry/fsimpl/fuse/fusefs.go b/pkg/sentry/fsimpl/fuse/fusefs.go index 810819ae4..65786e42a 100644 --- a/pkg/sentry/fsimpl/fuse/fusefs.go +++ b/pkg/sentry/fsimpl/fuse/fusefs.go @@ -16,24 +16,36 @@ package fuse import ( + "math" "strconv" + "sync" + "sync/atomic" "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/log" + "gvisor.dev/gvisor/pkg/marshal" "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" + "gvisor.dev/gvisor/pkg/waiter" ) // Name is the default filesystem name. const Name = "fuse" +// maxActiveRequestsDefault is the default setting controlling the upper bound +// on the number of active requests at any given time. +const maxActiveRequestsDefault = 10000 + // FilesystemType implements vfs.FilesystemType. +// +// +stateify savable type FilesystemType struct{} +// +stateify savable type filesystemOptions struct { // userID specifies the numeric uid of the mount owner. // This option should not be specified by the filesystem owner. @@ -56,9 +68,16 @@ type filesystemOptions struct { // exist at any time. Any further requests will block when trying to // Call the server. maxActiveRequests uint64 + + // maxRead is the max number of bytes to read, + // specified as "max_read" in fs parameters. + // If not specified by user, use math.MaxUint32 as default value. + maxRead uint32 } // filesystem implements vfs.FilesystemImpl. +// +// +stateify savable type filesystem struct { kernfs.Filesystem devMinor uint32 @@ -69,6 +88,9 @@ type filesystem struct { // opts is the options the fusefs is initialized with. opts *filesystemOptions + + // umounted is true if filesystem.Release() has been called. + umounted bool } // Name implements vfs.FilesystemType.Name. @@ -142,14 +164,29 @@ func (fsType FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virt // Set the maxInFlightRequests option. fsopts.maxActiveRequests = maxActiveRequestsDefault + if maxReadStr, ok := mopts["max_read"]; ok { + delete(mopts, "max_read") + maxRead, err := strconv.ParseUint(maxReadStr, 10, 32) + if err != nil { + log.Warningf("%s.GetFilesystem: invalid max_read: max_read=%s", fsType.Name(), maxReadStr) + return nil, nil, syserror.EINVAL + } + if maxRead < fuseMinMaxRead { + maxRead = fuseMinMaxRead + } + fsopts.maxRead = uint32(maxRead) + } else { + fsopts.maxRead = math.MaxUint32 + } + // Check for unparsed options. if len(mopts) != 0 { - log.Warningf("%s.GetFilesystem: unknown options: %v", fsType.Name(), mopts) + log.Warningf("%s.GetFilesystem: unsupported or unknown options: %v", fsType.Name(), mopts) return nil, nil, syserror.EINVAL } // Create a new FUSE filesystem. - fs, err := NewFUSEFilesystem(ctx, devMinor, &fsopts, fuseFd) + fs, err := newFUSEFilesystem(ctx, devMinor, &fsopts, fuseFd) if err != nil { log.Warningf("%s.NewFUSEFilesystem: failed with error: %v", fsType.Name(), err) return nil, nil, err @@ -165,26 +202,28 @@ func (fsType FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virt } // root is the fusefs root directory. - root := fs.newInode(creds, fsopts.rootMode) + root := fs.newRootInode(creds, fsopts.rootMode) return fs.VFSFilesystem(), root.VFSDentry(), nil } -// NewFUSEFilesystem creates a new FUSE filesystem. -func NewFUSEFilesystem(ctx context.Context, devMinor uint32, opts *filesystemOptions, device *vfs.FileDescription) (*filesystem, error) { - fs := &filesystem{ - devMinor: devMinor, - opts: opts, - } - - conn, err := newFUSEConnection(ctx, device, opts.maxActiveRequests) +// newFUSEFilesystem creates a new FUSE filesystem. +func newFUSEFilesystem(ctx context.Context, devMinor uint32, opts *filesystemOptions, device *vfs.FileDescription) (*filesystem, error) { + conn, err := newFUSEConnection(ctx, device, opts) if err != nil { log.Warningf("fuse.NewFUSEFilesystem: NewFUSEConnection failed with error: %v", err) return nil, syserror.EINVAL } - fs.conn = conn fuseFD := device.Impl().(*DeviceFD) + + fs := &filesystem{ + devMinor: devMinor, + opts: opts, + conn: conn, + } + + fs.VFSFilesystem().IncRef() fuseFD.fs = fs return fs, nil @@ -192,11 +231,22 @@ func NewFUSEFilesystem(ctx context.Context, devMinor uint32, opts *filesystemOpt // Release implements vfs.FilesystemImpl.Release. func (fs *filesystem) Release(ctx context.Context) { + fs.conn.fd.mu.Lock() + + fs.umounted = true + fs.conn.Abort(ctx) + // Notify all the waiters on this fd. + fs.conn.fd.waitQueue.Notify(waiter.EventIn) + + fs.conn.fd.mu.Unlock() + fs.Filesystem.VFSFilesystem().VirtualFilesystem().PutAnonBlockDevMinor(fs.devMinor) fs.Filesystem.Release(ctx) } // inode implements kernfs.Inode. +// +// +stateify savable type inode struct { inodeRefs kernfs.InodeAttrs @@ -205,14 +255,50 @@ type inode struct { kernfs.InodeNotSymlink kernfs.OrderedChildren + dentry kernfs.Dentry + + // the owning filesystem. fs is immutable. + fs *filesystem + + // metaDataMu protects the metadata of this inode. + metadataMu sync.Mutex + + nodeID uint64 + locks vfs.FileLocks - dentry kernfs.Dentry + // size of the file. + size uint64 + + // attributeVersion is the version of inode's attributes. + attributeVersion uint64 + + // attributeTime is the remaining vaild time of attributes. + attributeTime uint64 + + // version of the inode. + version uint64 + + // link is result of following a symbolic link. + link string +} + +func (fs *filesystem) newRootInode(creds *auth.Credentials, mode linux.FileMode) *kernfs.Dentry { + i := &inode{fs: fs} + i.InodeAttrs.Init(creds, linux.UNNAMED_MAJOR, fs.devMinor, 1, linux.ModeDirectory|0755) + i.OrderedChildren.Init(kernfs.OrderedChildrenOptions{}) + i.EnableLeakCheck() + i.dentry.Init(i) + i.nodeID = 1 + + return &i.dentry } -func (fs *filesystem) newInode(creds *auth.Credentials, mode linux.FileMode) *kernfs.Dentry { - i := &inode{} - i.InodeAttrs.Init(creds, linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), linux.ModeDirectory|0755) +func (fs *filesystem) newInode(nodeID uint64, attr linux.FUSEAttr) *kernfs.Dentry { + i := &inode{fs: fs, nodeID: nodeID} + creds := auth.Credentials{EffectiveKGID: auth.KGID(attr.UID), EffectiveKUID: auth.KUID(attr.UID)} + i.InodeAttrs.Init(&creds, linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), linux.FileMode(attr.Mode)) + atomic.StoreUint64(&i.size, attr.Size) i.OrderedChildren.Init(kernfs.OrderedChildrenOptions{}) i.EnableLeakCheck() i.dentry.Init(i) @@ -221,14 +307,292 @@ func (fs *filesystem) newInode(creds *auth.Credentials, mode linux.FileMode) *ke } // Open implements kernfs.Inode.Open. -func (i *inode) 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, &i.locks, &opts, kernfs.GenericDirectoryFDOptions{ - SeekEnd: kernfs.SeekEndStaticEntries, - }) +func (i *inode) Open(ctx context.Context, rp *vfs.ResolvingPath, d *kernfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + isDir := i.InodeAttrs.Mode().IsDir() + // return error if specified to open directory but inode is not a directory. + if !isDir && opts.Mode.IsDir() { + return nil, syserror.ENOTDIR + } + if opts.Flags&linux.O_LARGEFILE == 0 && atomic.LoadUint64(&i.size) > linux.MAX_NON_LFS { + return nil, syserror.EOVERFLOW + } + + var fd *fileDescription + var fdImpl vfs.FileDescriptionImpl + if isDir { + directoryFD := &directoryFD{} + fd = &(directoryFD.fileDescription) + fdImpl = directoryFD + } else { + regularFD := ®ularFileFD{} + fd = &(regularFD.fileDescription) + fdImpl = regularFD + } + // FOPEN_KEEP_CACHE is the defualt flag for noOpen. + fd.OpenFlag = linux.FOPEN_KEEP_CACHE + + // Only send open request when FUSE server support open or is opening a directory. + if !i.fs.conn.noOpen || isDir { + kernelTask := kernel.TaskFromContext(ctx) + if kernelTask == nil { + log.Warningf("fusefs.Inode.Open: couldn't get kernel task from context") + return nil, syserror.EINVAL + } + + // Build the request. + var opcode linux.FUSEOpcode + if isDir { + opcode = linux.FUSE_OPENDIR + } else { + opcode = linux.FUSE_OPEN + } + + in := linux.FUSEOpenIn{Flags: opts.Flags & ^uint32(linux.O_CREAT|linux.O_EXCL|linux.O_NOCTTY)} + if !i.fs.conn.atomicOTrunc { + in.Flags &= ^uint32(linux.O_TRUNC) + } + + req, err := i.fs.conn.NewRequest(auth.CredentialsFromContext(ctx), uint32(kernelTask.ThreadID()), i.nodeID, opcode, &in) + if err != nil { + return nil, err + } + + // Send the request and receive the reply. + res, err := i.fs.conn.Call(kernelTask, req) + if err != nil { + return nil, err + } + if err := res.Error(); err == syserror.ENOSYS && !isDir { + i.fs.conn.noOpen = true + } else if err != nil { + return nil, err + } else { + out := linux.FUSEOpenOut{} + if err := res.UnmarshalPayload(&out); err != nil { + return nil, err + } + + // Process the reply. + fd.OpenFlag = out.OpenFlag + if isDir { + fd.OpenFlag &= ^uint32(linux.FOPEN_DIRECT_IO) + } + + fd.Fh = out.Fh + } + } + + // TODO(gvisor.dev/issue/3234): invalidate mmap after implemented it for FUSE Inode + fd.DirectIO = fd.OpenFlag&linux.FOPEN_DIRECT_IO != 0 + fdOptions := &vfs.FileDescriptionOptions{} + if fd.OpenFlag&linux.FOPEN_NONSEEKABLE != 0 { + fdOptions.DenyPRead = true + fdOptions.DenyPWrite = true + fd.Nonseekable = true + } + + // If we don't send SETATTR before open (which is indicated by atomicOTrunc) + // and O_TRUNC is set, update the inode's version number and clean existing data + // by setting the file size to 0. + if i.fs.conn.atomicOTrunc && opts.Flags&linux.O_TRUNC != 0 { + i.fs.conn.mu.Lock() + i.fs.conn.attributeVersion++ + i.attributeVersion = i.fs.conn.attributeVersion + atomic.StoreUint64(&i.size, 0) + i.fs.conn.mu.Unlock() + i.attributeTime = 0 + } + + if err := fd.vfsfd.Init(fdImpl, opts.Flags, rp.Mount(), d.VFSDentry(), fdOptions); err != nil { + return nil, err + } + return &fd.vfsfd, nil +} + +// Lookup implements kernfs.Inode.Lookup. +func (i *inode) Lookup(ctx context.Context, name string) (*kernfs.Dentry, error) { + in := linux.FUSELookupIn{Name: name} + return i.newEntry(ctx, name, 0, linux.FUSE_LOOKUP, &in) +} + +// IterDirents implements kernfs.Inode.IterDirents. +func (*inode) IterDirents(ctx context.Context, callback vfs.IterDirentsCallback, offset, relOffset int64) (int64, error) { + return offset, nil +} + +// Valid implements kernfs.Inode.Valid. +func (*inode) Valid(ctx context.Context) bool { + return true +} + +// NewFile implements kernfs.Inode.NewFile. +func (i *inode) NewFile(ctx context.Context, name string, opts vfs.OpenOptions) (*kernfs.Dentry, error) { + kernelTask := kernel.TaskFromContext(ctx) + if kernelTask == nil { + log.Warningf("fusefs.Inode.NewFile: couldn't get kernel task from context", i.nodeID) + return nil, syserror.EINVAL + } + in := linux.FUSECreateIn{ + CreateMeta: linux.FUSECreateMeta{ + Flags: opts.Flags, + Mode: uint32(opts.Mode) | linux.S_IFREG, + Umask: uint32(kernelTask.FSContext().Umask()), + }, + Name: name, + } + return i.newEntry(ctx, name, linux.S_IFREG, linux.FUSE_CREATE, &in) +} + +// NewNode implements kernfs.Inode.NewNode. +func (i *inode) NewNode(ctx context.Context, name string, opts vfs.MknodOptions) (*kernfs.Dentry, error) { + in := linux.FUSEMknodIn{ + MknodMeta: linux.FUSEMknodMeta{ + Mode: uint32(opts.Mode), + Rdev: linux.MakeDeviceID(uint16(opts.DevMajor), opts.DevMinor), + Umask: uint32(kernel.TaskFromContext(ctx).FSContext().Umask()), + }, + Name: name, + } + return i.newEntry(ctx, name, opts.Mode.FileType(), linux.FUSE_MKNOD, &in) +} + +// NewSymlink implements kernfs.Inode.NewSymlink. +func (i *inode) NewSymlink(ctx context.Context, name, target string) (*kernfs.Dentry, error) { + in := linux.FUSESymLinkIn{ + Name: name, + Target: target, + } + return i.newEntry(ctx, name, linux.S_IFLNK, linux.FUSE_SYMLINK, &in) +} + +// Unlink implements kernfs.Inode.Unlink. +func (i *inode) Unlink(ctx context.Context, name string, child *kernfs.Dentry) error { + kernelTask := kernel.TaskFromContext(ctx) + if kernelTask == nil { + log.Warningf("fusefs.Inode.newEntry: couldn't get kernel task from context", i.nodeID) + return syserror.EINVAL + } + in := linux.FUSEUnlinkIn{Name: name} + req, err := i.fs.conn.NewRequest(auth.CredentialsFromContext(ctx), uint32(kernelTask.ThreadID()), i.nodeID, linux.FUSE_UNLINK, &in) + if err != nil { + return err + } + res, err := i.fs.conn.Call(kernelTask, req) + if err != nil { + return err + } + // only return error, discard res. + if err := res.Error(); err != nil { + return err + } + return i.dentry.RemoveChildLocked(name, child) +} + +// NewDir implements kernfs.Inode.NewDir. +func (i *inode) NewDir(ctx context.Context, name string, opts vfs.MkdirOptions) (*kernfs.Dentry, error) { + in := linux.FUSEMkdirIn{ + MkdirMeta: linux.FUSEMkdirMeta{ + Mode: uint32(opts.Mode), + Umask: uint32(kernel.TaskFromContext(ctx).FSContext().Umask()), + }, + Name: name, + } + return i.newEntry(ctx, name, linux.S_IFDIR, linux.FUSE_MKDIR, &in) +} + +// RmDir implements kernfs.Inode.RmDir. +func (i *inode) RmDir(ctx context.Context, name string, child *kernfs.Dentry) error { + fusefs := i.fs + task, creds := kernel.TaskFromContext(ctx), auth.CredentialsFromContext(ctx) + + in := linux.FUSERmDirIn{Name: name} + req, err := fusefs.conn.NewRequest(creds, uint32(task.ThreadID()), i.nodeID, linux.FUSE_RMDIR, &in) + if err != nil { + return err + } + + res, err := i.fs.conn.Call(task, req) + if err != nil { + return err + } + if err := res.Error(); err != nil { + return err + } + + return i.dentry.RemoveChildLocked(name, child) +} + +// newEntry calls FUSE server for entry creation and allocates corresponding entry according to response. +// Shared by FUSE_MKNOD, FUSE_MKDIR, FUSE_SYMLINK, FUSE_LINK and FUSE_LOOKUP. +func (i *inode) newEntry(ctx context.Context, name string, fileType linux.FileMode, opcode linux.FUSEOpcode, payload marshal.Marshallable) (*kernfs.Dentry, error) { + kernelTask := kernel.TaskFromContext(ctx) + if kernelTask == nil { + log.Warningf("fusefs.Inode.newEntry: couldn't get kernel task from context", i.nodeID) + return nil, syserror.EINVAL + } + req, err := i.fs.conn.NewRequest(auth.CredentialsFromContext(ctx), uint32(kernelTask.ThreadID()), i.nodeID, opcode, payload) + if err != nil { + return nil, err + } + res, err := i.fs.conn.Call(kernelTask, req) if err != nil { return nil, err } - return fd.VFSFileDescription(), nil + if err := res.Error(); err != nil { + return nil, err + } + out := linux.FUSEEntryOut{} + if err := res.UnmarshalPayload(&out); err != nil { + return nil, err + } + if opcode != linux.FUSE_LOOKUP && ((out.Attr.Mode&linux.S_IFMT)^uint32(fileType) != 0 || out.NodeID == 0 || out.NodeID == linux.FUSE_ROOT_ID) { + return nil, syserror.EIO + } + child := i.fs.newInode(out.NodeID, out.Attr) + return child, nil +} + +// Getlink implements kernfs.Inode.Getlink. +func (i *inode) Getlink(ctx context.Context, mnt *vfs.Mount) (vfs.VirtualDentry, string, error) { + path, err := i.Readlink(ctx, mnt) + return vfs.VirtualDentry{}, path, err +} + +// Readlink implements kernfs.Inode.Readlink. +func (i *inode) Readlink(ctx context.Context, mnt *vfs.Mount) (string, error) { + if i.Mode().FileType()&linux.S_IFLNK == 0 { + return "", syserror.EINVAL + } + if len(i.link) == 0 { + kernelTask := kernel.TaskFromContext(ctx) + if kernelTask == nil { + log.Warningf("fusefs.Inode.Readlink: couldn't get kernel task from context") + return "", syserror.EINVAL + } + req, err := i.fs.conn.NewRequest(auth.CredentialsFromContext(ctx), uint32(kernelTask.ThreadID()), i.nodeID, linux.FUSE_READLINK, &linux.FUSEEmptyIn{}) + if err != nil { + return "", err + } + res, err := i.fs.conn.Call(kernelTask, req) + if err != nil { + return "", err + } + i.link = string(res.data[res.hdr.SizeBytes():]) + if !mnt.Options().ReadOnly { + i.attributeTime = 0 + } + } + return i.link, nil +} + +// getFUSEAttr returns a linux.FUSEAttr of this inode stored in local cache. +// TODO(gvisor.dev/issue/3679): Add support for other fields. +func (i *inode) getFUSEAttr() linux.FUSEAttr { + return linux.FUSEAttr{ + Ino: i.Ino(), + Size: atomic.LoadUint64(&i.size), + Mode: uint32(i.Mode()), + } } // statFromFUSEAttr makes attributes from linux.FUSEAttr to linux.Statx. The @@ -284,50 +648,92 @@ func statFromFUSEAttr(attr linux.FUSEAttr, mask, devMinor uint32) linux.Statx { return stat } -// Stat implements kernfs.Inode.Stat. -func (i *inode) Stat(ctx context.Context, fs *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) { - fusefs := fs.Impl().(*filesystem) - conn := fusefs.conn - task, creds := kernel.TaskFromContext(ctx), auth.CredentialsFromContext(ctx) +// getAttr gets the attribute of this inode by issuing a FUSE_GETATTR request +// or read from local cache. It updates the corresponding attributes if +// necessary. +func (i *inode) getAttr(ctx context.Context, fs *vfs.Filesystem, opts vfs.StatOptions, flags uint32, fh uint64) (linux.FUSEAttr, error) { + attributeVersion := atomic.LoadUint64(&i.fs.conn.attributeVersion) + + // TODO(gvisor.dev/issue/3679): send the request only if + // - invalid local cache for fields specified in the opts.Mask + // - forced update + // - i.attributeTime expired + // If local cache is still valid, return local cache. + // Currently we always send a request, + // and we always set the metadata with the new result, + // unless attributeVersion has changed. + + task := kernel.TaskFromContext(ctx) if task == nil { log.Warningf("couldn't get kernel task from context") - return linux.Statx{}, syserror.EINVAL + return linux.FUSEAttr{}, syserror.EINVAL } - var in linux.FUSEGetAttrIn - // We don't set any attribute in the request, because in VFS2 fstat(2) will - // finally be translated into vfs.FilesystemImpl.StatAt() (see - // pkg/sentry/syscalls/linux/vfs2/stat.go), resulting in the same flow - // as stat(2). Thus GetAttrFlags and Fh variable will never be used in VFS2. - req, err := conn.NewRequest(creds, uint32(task.ThreadID()), i.Ino(), linux.FUSE_GETATTR, &in) + creds := auth.CredentialsFromContext(ctx) + + in := linux.FUSEGetAttrIn{ + GetAttrFlags: flags, + Fh: fh, + } + req, err := i.fs.conn.NewRequest(creds, uint32(task.ThreadID()), i.nodeID, linux.FUSE_GETATTR, &in) if err != nil { - return linux.Statx{}, err + return linux.FUSEAttr{}, err } - res, err := conn.Call(task, req) + res, err := i.fs.conn.Call(task, req) if err != nil { - return linux.Statx{}, err + return linux.FUSEAttr{}, err } if err := res.Error(); err != nil { - return linux.Statx{}, err + return linux.FUSEAttr{}, err } var out linux.FUSEGetAttrOut if err := res.UnmarshalPayload(&out); err != nil { - return linux.Statx{}, err + return linux.FUSEAttr{}, err + } + + // Local version is newer, return the local one. + // Skip the update. + if attributeVersion != 0 && atomic.LoadUint64(&i.attributeVersion) > attributeVersion { + return i.getFUSEAttr(), nil } - // Set all metadata into kernfs.InodeAttrs. - if err := i.SetStat(ctx, fs, creds, vfs.SetStatOptions{ - Stat: statFromFUSEAttr(out.Attr, linux.STATX_ALL, fusefs.devMinor), + // Set the metadata of kernfs.InodeAttrs. + if err := i.SetInodeStat(ctx, fs, creds, vfs.SetStatOptions{ + Stat: statFromFUSEAttr(out.Attr, linux.STATX_ALL, i.fs.devMinor), }); err != nil { + return linux.FUSEAttr{}, err + } + + // Set the size if no error (after SetStat() check). + atomic.StoreUint64(&i.size, out.Attr.Size) + + return out.Attr, nil +} + +// reviseAttr attempts to update the attributes for internal purposes +// by calling getAttr with a pre-specified mask. +// Used by read, write, lseek. +func (i *inode) reviseAttr(ctx context.Context, flags uint32, fh uint64) error { + // Never need atime for internal purposes. + _, err := i.getAttr(ctx, i.fs.VFSFilesystem(), vfs.StatOptions{ + Mask: linux.STATX_BASIC_STATS &^ linux.STATX_ATIME, + }, flags, fh) + return err +} + +// Stat implements kernfs.Inode.Stat. +func (i *inode) Stat(ctx context.Context, fs *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) { + attr, err := i.getAttr(ctx, fs, opts, 0, 0) + if err != nil { return linux.Statx{}, err } - return statFromFUSEAttr(out.Attr, opts.Mask, fusefs.devMinor), nil + return statFromFUSEAttr(attr, opts.Mask, i.fs.devMinor), nil } -// DecRef implements kernfs.Inode. +// DecRef implements kernfs.Inode.DecRef. func (i *inode) DecRef(context.Context) { i.inodeRefs.DecRef(i.Destroy) } @@ -337,3 +743,84 @@ func (i *inode) StatFS(ctx context.Context, fs *vfs.Filesystem) (linux.Statfs, e // TODO(gvisor.dev/issues/3413): Complete the implementation of statfs. return vfs.GenericStatFS(linux.FUSE_SUPER_MAGIC), nil } + +// fattrMaskFromStats converts vfs.SetStatOptions.Stat.Mask to linux stats mask +// aligned with the attribute mask defined in include/linux/fs.h. +func fattrMaskFromStats(mask uint32) uint32 { + var fuseAttrMask uint32 + maskMap := map[uint32]uint32{ + linux.STATX_MODE: linux.FATTR_MODE, + linux.STATX_UID: linux.FATTR_UID, + linux.STATX_GID: linux.FATTR_GID, + linux.STATX_SIZE: linux.FATTR_SIZE, + linux.STATX_ATIME: linux.FATTR_ATIME, + linux.STATX_MTIME: linux.FATTR_MTIME, + linux.STATX_CTIME: linux.FATTR_CTIME, + } + for statxMask, fattrMask := range maskMap { + if mask&statxMask != 0 { + fuseAttrMask |= fattrMask + } + } + return fuseAttrMask +} + +// SetStat implements kernfs.Inode.SetStat. +func (i *inode) SetStat(ctx context.Context, fs *vfs.Filesystem, creds *auth.Credentials, opts vfs.SetStatOptions) error { + return i.setAttr(ctx, fs, creds, opts, false, 0) +} + +func (i *inode) setAttr(ctx context.Context, fs *vfs.Filesystem, creds *auth.Credentials, opts vfs.SetStatOptions, useFh bool, fh uint64) error { + conn := i.fs.conn + task := kernel.TaskFromContext(ctx) + if task == nil { + log.Warningf("couldn't get kernel task from context") + return syserror.EINVAL + } + + // We should retain the original file type when assigning new mode. + fileType := uint16(i.Mode()) & linux.S_IFMT + fattrMask := fattrMaskFromStats(opts.Stat.Mask) + if useFh { + fattrMask |= linux.FATTR_FH + } + in := linux.FUSESetAttrIn{ + Valid: fattrMask, + Fh: fh, + Size: opts.Stat.Size, + Atime: uint64(opts.Stat.Atime.Sec), + Mtime: uint64(opts.Stat.Mtime.Sec), + Ctime: uint64(opts.Stat.Ctime.Sec), + AtimeNsec: opts.Stat.Atime.Nsec, + MtimeNsec: opts.Stat.Mtime.Nsec, + CtimeNsec: opts.Stat.Ctime.Nsec, + Mode: uint32(fileType | opts.Stat.Mode), + UID: opts.Stat.UID, + GID: opts.Stat.GID, + } + req, err := conn.NewRequest(creds, uint32(task.ThreadID()), i.nodeID, linux.FUSE_SETATTR, &in) + if err != nil { + return err + } + + res, err := conn.Call(task, req) + if err != nil { + return err + } + if err := res.Error(); err != nil { + return err + } + out := linux.FUSEGetAttrOut{} + if err := res.UnmarshalPayload(&out); err != nil { + return err + } + + // Set the metadata of kernfs.InodeAttrs. + if err := i.SetInodeStat(ctx, fs, creds, vfs.SetStatOptions{ + Stat: statFromFUSEAttr(out.Attr, linux.STATX_ALL, i.fs.devMinor), + }); err != nil { + return err + } + + return nil +} diff --git a/pkg/sentry/fsimpl/fuse/read_write.go b/pkg/sentry/fsimpl/fuse/read_write.go new file mode 100644 index 000000000..625d1547f --- /dev/null +++ b/pkg/sentry/fsimpl/fuse/read_write.go @@ -0,0 +1,242 @@ +// 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 fuse + +import ( + "io" + "sync/atomic" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/log" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +// ReadInPages sends FUSE_READ requests for the size after round it up to +// a multiple of page size, blocks on it for reply, processes the reply +// and returns the payload (or joined payloads) as a byte slice. +// This is used for the general purpose reading. +// We do not support direct IO (which read the exact number of bytes) +// at this moment. +func (fs *filesystem) ReadInPages(ctx context.Context, fd *regularFileFD, off uint64, size uint32) ([][]byte, uint32, error) { + attributeVersion := atomic.LoadUint64(&fs.conn.attributeVersion) + + t := kernel.TaskFromContext(ctx) + if t == nil { + log.Warningf("fusefs.Read: couldn't get kernel task from context") + return nil, 0, syserror.EINVAL + } + + // Round up to a multiple of page size. + readSize, _ := usermem.PageRoundUp(uint64(size)) + + // One request cannnot exceed either maxRead or maxPages. + maxPages := fs.conn.maxRead >> usermem.PageShift + if maxPages > uint32(fs.conn.maxPages) { + maxPages = uint32(fs.conn.maxPages) + } + + var outs [][]byte + var sizeRead uint32 + + // readSize is a multiple of usermem.PageSize. + // Always request bytes as a multiple of pages. + pagesRead, pagesToRead := uint32(0), uint32(readSize>>usermem.PageShift) + + // Reuse the same struct for unmarshalling to avoid unnecessary memory allocation. + in := linux.FUSEReadIn{ + Fh: fd.Fh, + LockOwner: 0, // TODO(gvisor.dev/issue/3245): file lock + ReadFlags: 0, // TODO(gvisor.dev/issue/3245): |= linux.FUSE_READ_LOCKOWNER + Flags: fd.statusFlags(), + } + + // This loop is intended for fragmented read where the bytes to read is + // larger than either the maxPages or maxRead. + // For the majority of reads with normal size, this loop should only + // execute once. + for pagesRead < pagesToRead { + pagesCanRead := pagesToRead - pagesRead + if pagesCanRead > maxPages { + pagesCanRead = maxPages + } + + in.Offset = off + (uint64(pagesRead) << usermem.PageShift) + in.Size = pagesCanRead << usermem.PageShift + + req, err := fs.conn.NewRequest(auth.CredentialsFromContext(ctx), uint32(t.ThreadID()), fd.inode().nodeID, linux.FUSE_READ, &in) + if err != nil { + return nil, 0, err + } + + // TODO(gvisor.dev/issue/3247): support async read. + + res, err := fs.conn.Call(t, req) + if err != nil { + return nil, 0, err + } + if err := res.Error(); err != nil { + return nil, 0, err + } + + // Not enough bytes in response, + // either we reached EOF, + // or the FUSE server sends back a response + // that cannot even fit the hdr. + if len(res.data) <= res.hdr.SizeBytes() { + // We treat both case as EOF here for now + // since there is no reliable way to detect + // the over-short hdr case. + break + } + + // Directly using the slice to avoid extra copy. + out := res.data[res.hdr.SizeBytes():] + + outs = append(outs, out) + sizeRead += uint32(len(out)) + + pagesRead += pagesCanRead + } + + defer fs.ReadCallback(ctx, fd, off, size, sizeRead, attributeVersion) + + // No bytes returned: offset >= EOF. + if len(outs) == 0 { + return nil, 0, io.EOF + } + + return outs, sizeRead, nil +} + +// ReadCallback updates several information after receiving a read response. +// Due to readahead, sizeRead can be larger than size. +func (fs *filesystem) ReadCallback(ctx context.Context, fd *regularFileFD, off uint64, size uint32, sizeRead uint32, attributeVersion uint64) { + // TODO(gvisor.dev/issue/3247): support async read. + // If this is called by an async read, correctly process it. + // May need to update the signature. + + i := fd.inode() + // TODO(gvisor.dev/issue/1193): Invalidate or update atime. + + // Reached EOF. + if sizeRead < size { + // TODO(gvisor.dev/issue/3630): If we have writeback cache, then we need to fill this hole. + // Might need to update the buf to be returned from the Read(). + + // Update existing size. + newSize := off + uint64(sizeRead) + fs.conn.mu.Lock() + if attributeVersion == i.attributeVersion && newSize < atomic.LoadUint64(&i.size) { + fs.conn.attributeVersion++ + i.attributeVersion = i.fs.conn.attributeVersion + atomic.StoreUint64(&i.size, newSize) + } + fs.conn.mu.Unlock() + } +} + +// Write sends FUSE_WRITE requests and return the bytes +// written according to the response. +// +// Preconditions: len(data) == size. +func (fs *filesystem) Write(ctx context.Context, fd *regularFileFD, off uint64, size uint32, data []byte) (uint32, error) { + t := kernel.TaskFromContext(ctx) + if t == nil { + log.Warningf("fusefs.Read: couldn't get kernel task from context") + return 0, syserror.EINVAL + } + + // One request cannnot exceed either maxWrite or maxPages. + maxWrite := uint32(fs.conn.maxPages) << usermem.PageShift + if maxWrite > fs.conn.maxWrite { + maxWrite = fs.conn.maxWrite + } + + // Reuse the same struct for unmarshalling to avoid unnecessary memory allocation. + in := linux.FUSEWriteIn{ + Fh: fd.Fh, + // TODO(gvisor.dev/issue/3245): file lock + LockOwner: 0, + // TODO(gvisor.dev/issue/3245): |= linux.FUSE_READ_LOCKOWNER + // TODO(gvisor.dev/issue/3237): |= linux.FUSE_WRITE_CACHE (not added yet) + WriteFlags: 0, + Flags: fd.statusFlags(), + } + + var written uint32 + + // This loop is intended for fragmented write where the bytes to write is + // larger than either the maxWrite or maxPages or when bigWrites is false. + // Unless a small value for max_write is explicitly used, this loop + // is expected to execute only once for the majority of the writes. + for written < size { + toWrite := size - written + + // Limit the write size to one page. + // Note that the bigWrites flag is obsolete, + // latest libfuse always sets it on. + if !fs.conn.bigWrites && toWrite > usermem.PageSize { + toWrite = usermem.PageSize + } + + // Limit the write size to maxWrite. + if toWrite > maxWrite { + toWrite = maxWrite + } + + in.Offset = off + uint64(written) + in.Size = toWrite + + req, err := fs.conn.NewRequest(auth.CredentialsFromContext(ctx), uint32(t.ThreadID()), fd.inode().nodeID, linux.FUSE_WRITE, &in) + if err != nil { + return 0, err + } + + req.payload = data[written : written+toWrite] + + // TODO(gvisor.dev/issue/3247): support async write. + + res, err := fs.conn.Call(t, req) + if err != nil { + return 0, err + } + if err := res.Error(); err != nil { + return 0, err + } + + out := linux.FUSEWriteOut{} + if err := res.UnmarshalPayload(&out); err != nil { + return 0, err + } + + // Write more than requested? EIO. + if out.Size > toWrite { + return 0, syserror.EIO + } + + written += out.Size + + // Break if short write. Not necessarily an error. + if out.Size != toWrite { + break + } + } + + return written, nil +} diff --git a/pkg/sentry/fsimpl/fuse/regular_file.go b/pkg/sentry/fsimpl/fuse/regular_file.go new file mode 100644 index 000000000..5bdd096c3 --- /dev/null +++ b/pkg/sentry/fsimpl/fuse/regular_file.go @@ -0,0 +1,230 @@ +// 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 fuse + +import ( + "io" + "math" + "sync" + "sync/atomic" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +type regularFileFD struct { + fileDescription + + // off is the file offset. + off int64 + // offMu protects off. + offMu sync.Mutex +} + +// PRead implements vfs.FileDescriptionImpl.PRead. +func (fd *regularFileFD) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { + if offset < 0 { + return 0, syserror.EINVAL + } + + // Check that flags are supported. + // + // TODO(gvisor.dev/issue/2601): Support select preadv2 flags. + if opts.Flags&^linux.RWF_HIPRI != 0 { + return 0, syserror.EOPNOTSUPP + } + + size := dst.NumBytes() + if size == 0 { + // Early return if count is 0. + return 0, nil + } else if size > math.MaxUint32 { + // FUSE only supports uint32 for size. + // Overflow. + return 0, syserror.EINVAL + } + + // TODO(gvisor.dev/issue/3678): Add direct IO support. + + inode := fd.inode() + + // Reading beyond EOF, update file size if outdated. + if uint64(offset+size) > atomic.LoadUint64(&inode.size) { + if err := inode.reviseAttr(ctx, linux.FUSE_GETATTR_FH, fd.Fh); err != nil { + return 0, err + } + // If the offset after update is still too large, return error. + if uint64(offset) >= atomic.LoadUint64(&inode.size) { + return 0, io.EOF + } + } + + // Truncate the read with updated file size. + fileSize := atomic.LoadUint64(&inode.size) + if uint64(offset+size) > fileSize { + size = int64(fileSize) - offset + } + + buffers, n, err := inode.fs.ReadInPages(ctx, fd, uint64(offset), uint32(size)) + if err != nil { + return 0, err + } + + // TODO(gvisor.dev/issue/3237): support indirect IO (e.g. caching), + // store the bytes that were read ahead. + + // Update the number of bytes to copy for short read. + if n < uint32(size) { + size = int64(n) + } + + // Copy the bytes read to the dst. + // This loop is intended for fragmented reads. + // For the majority of reads, this loop only execute once. + var copied int64 + for _, buffer := range buffers { + toCopy := int64(len(buffer)) + if copied+toCopy > size { + toCopy = size - copied + } + cp, err := dst.DropFirst64(copied).CopyOut(ctx, buffer[:toCopy]) + if err != nil { + return 0, err + } + if int64(cp) != toCopy { + return 0, syserror.EIO + } + copied += toCopy + } + + return copied, nil +} + +// Read implements vfs.FileDescriptionImpl.Read. +func (fd *regularFileFD) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { + fd.offMu.Lock() + n, err := fd.PRead(ctx, dst, fd.off, opts) + fd.off += n + fd.offMu.Unlock() + return n, err +} + +// PWrite implements vfs.FileDescriptionImpl.PWrite. +func (fd *regularFileFD) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) { + n, _, err := fd.pwrite(ctx, src, offset, opts) + return n, err +} + +// Write implements vfs.FileDescriptionImpl.Write. +func (fd *regularFileFD) Write(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) { + fd.offMu.Lock() + n, off, err := fd.pwrite(ctx, src, fd.off, opts) + fd.off = off + fd.offMu.Unlock() + return n, err +} + +// pwrite returns the number of bytes written, final offset and error. The +// final offset should be ignored by PWrite. +func (fd *regularFileFD) pwrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (written, finalOff int64, err error) { + if offset < 0 { + return 0, offset, syserror.EINVAL + } + + // Check that flags are supported. + // + // TODO(gvisor.dev/issue/2601): Support select preadv2 flags. + if opts.Flags&^linux.RWF_HIPRI != 0 { + return 0, offset, syserror.EOPNOTSUPP + } + + inode := fd.inode() + inode.metadataMu.Lock() + defer inode.metadataMu.Unlock() + + // If the file is opened with O_APPEND, update offset to file size. + // Note: since our Open() implements the interface of kernfs, + // and kernfs currently does not support O_APPEND, this will never + // be true before we switch out from kernfs. + if fd.vfsfd.StatusFlags()&linux.O_APPEND != 0 { + // Locking inode.metadataMu is sufficient for reading size + offset = int64(inode.size) + } + + srclen := src.NumBytes() + + if srclen > math.MaxUint32 { + // FUSE only supports uint32 for size. + // Overflow. + return 0, offset, syserror.EINVAL + } + if end := offset + srclen; end < offset { + // Overflow. + return 0, offset, syserror.EINVAL + } + + srclen, err = vfs.CheckLimit(ctx, offset, srclen) + if err != nil { + return 0, offset, err + } + + if srclen == 0 { + // Return before causing any side effects. + return 0, offset, nil + } + + src = src.TakeFirst64(srclen) + + // TODO(gvisor.dev/issue/3237): Add cache support: + // buffer cache. Ideally we write from src to our buffer cache first. + // The slice passed to fs.Write() should be a slice from buffer cache. + data := make([]byte, srclen) + // Reason for making a copy here: connection.Call() blocks on kerneltask, + // which in turn acquires mm.activeMu lock. Functions like CopyInTo() will + // attemp to acquire the mm.activeMu lock as well -> deadlock. + // We must finish reading from the userspace memory before + // t.Block() deactivates it. + cp, err := src.CopyIn(ctx, data) + if err != nil { + return 0, offset, err + } + if int64(cp) != srclen { + return 0, offset, syserror.EIO + } + + n, err := fd.inode().fs.Write(ctx, fd, uint64(offset), uint32(srclen), data) + if err != nil { + return 0, offset, err + } + + if n == 0 { + // We have checked srclen != 0 previously. + // If err == nil, then it's a short write and we return EIO. + return 0, offset, syserror.EIO + } + + written = int64(n) + finalOff = offset + written + + if finalOff > int64(inode.size) { + atomic.StoreUint64(&inode.size, uint64(finalOff)) + atomic.AddUint64(&inode.fs.conn.attributeVersion, 1) + } + + return +} diff --git a/pkg/sentry/fsimpl/fuse/request_response.go b/pkg/sentry/fsimpl/fuse/request_response.go new file mode 100644 index 000000000..7fa00569b --- /dev/null +++ b/pkg/sentry/fsimpl/fuse/request_response.go @@ -0,0 +1,229 @@ +// 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 fuse + +import ( + "fmt" + "syscall" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/marshal" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/usermem" +) + +// fuseInitRes is a variable-length wrapper of linux.FUSEInitOut. The FUSE +// server may implement an older version of FUSE protocol, which contains a +// linux.FUSEInitOut with less attributes. +// +// Dynamically-sized objects cannot be marshalled. +type fuseInitRes struct { + marshal.StubMarshallable + + // initOut contains the response from the FUSE server. + initOut linux.FUSEInitOut + + // initLen is the total length of bytes of the response. + initLen uint32 +} + +// UnmarshalBytes deserializes src to the initOut attribute in a fuseInitRes. +func (r *fuseInitRes) UnmarshalBytes(src []byte) { + out := &r.initOut + + // Introduced before FUSE kernel version 7.13. + out.Major = uint32(usermem.ByteOrder.Uint32(src[:4])) + src = src[4:] + out.Minor = uint32(usermem.ByteOrder.Uint32(src[:4])) + src = src[4:] + out.MaxReadahead = uint32(usermem.ByteOrder.Uint32(src[:4])) + src = src[4:] + out.Flags = uint32(usermem.ByteOrder.Uint32(src[:4])) + src = src[4:] + out.MaxBackground = uint16(usermem.ByteOrder.Uint16(src[:2])) + src = src[2:] + out.CongestionThreshold = uint16(usermem.ByteOrder.Uint16(src[:2])) + src = src[2:] + out.MaxWrite = uint32(usermem.ByteOrder.Uint32(src[:4])) + src = src[4:] + + // Introduced in FUSE kernel version 7.23. + if len(src) >= 4 { + out.TimeGran = uint32(usermem.ByteOrder.Uint32(src[:4])) + src = src[4:] + } + // Introduced in FUSE kernel version 7.28. + if len(src) >= 2 { + out.MaxPages = uint16(usermem.ByteOrder.Uint16(src[:2])) + src = src[2:] + } +} + +// SizeBytes is the size of the payload of the FUSE_INIT response. +func (r *fuseInitRes) SizeBytes() int { + return int(r.initLen) +} + +// Ordinary requests have even IDs, while interrupts IDs are odd. +// Used to increment the unique ID for each FUSE request. +var reqIDStep uint64 = 2 + +// Request represents a FUSE operation request that hasn't been sent to the +// server yet. +// +// +stateify savable +type Request struct { + requestEntry + + id linux.FUSEOpID + hdr *linux.FUSEHeaderIn + data []byte + + // payload for this request: extra bytes to write after + // the data slice. Used by FUSE_WRITE. + payload []byte + + // If this request is async. + async bool + // If we don't care its response. + // Manually set by the caller. + noReply bool +} + +// NewRequest creates a new request that can be sent to the FUSE server. +func (conn *connection) NewRequest(creds *auth.Credentials, pid uint32, ino uint64, opcode linux.FUSEOpcode, payload marshal.Marshallable) (*Request, error) { + conn.fd.mu.Lock() + defer conn.fd.mu.Unlock() + conn.fd.nextOpID += linux.FUSEOpID(reqIDStep) + + hdrLen := (*linux.FUSEHeaderIn)(nil).SizeBytes() + hdr := linux.FUSEHeaderIn{ + Len: uint32(hdrLen + payload.SizeBytes()), + Opcode: opcode, + Unique: conn.fd.nextOpID, + NodeID: ino, + UID: uint32(creds.EffectiveKUID), + GID: uint32(creds.EffectiveKGID), + PID: pid, + } + + buf := make([]byte, hdr.Len) + + // TODO(gVisor.dev/issue/3698): Use the unsafe version once go_marshal is safe to use again. + hdr.MarshalBytes(buf[:hdrLen]) + payload.MarshalBytes(buf[hdrLen:]) + + return &Request{ + id: hdr.Unique, + hdr: &hdr, + data: buf, + }, nil +} + +// futureResponse represents an in-flight request, that may or may not have +// completed yet. Convert it to a resolved Response by calling Resolve, but note +// that this may block. +// +// +stateify savable +type futureResponse struct { + opcode linux.FUSEOpcode + ch chan struct{} + hdr *linux.FUSEHeaderOut + data []byte + + // If this request is async. + async bool +} + +// newFutureResponse creates a future response to a FUSE request. +func newFutureResponse(req *Request) *futureResponse { + return &futureResponse{ + opcode: req.hdr.Opcode, + ch: make(chan struct{}), + async: req.async, + } +} + +// resolve blocks the task until the server responds to its corresponding request, +// then returns a resolved response. +func (f *futureResponse) resolve(t *kernel.Task) (*Response, error) { + // Return directly for async requests. + if f.async { + return nil, nil + } + + if err := t.Block(f.ch); err != nil { + return nil, err + } + + return f.getResponse(), nil +} + +// getResponse creates a Response from the data the futureResponse has. +func (f *futureResponse) getResponse() *Response { + return &Response{ + opcode: f.opcode, + hdr: *f.hdr, + data: f.data, + } +} + +// Response represents an actual response from the server, including the +// response payload. +// +// +stateify savable +type Response struct { + opcode linux.FUSEOpcode + hdr linux.FUSEHeaderOut + data []byte +} + +// Error returns the error of the FUSE call. +func (r *Response) Error() error { + errno := r.hdr.Error + if errno >= 0 { + return nil + } + + sysErrNo := syscall.Errno(-errno) + return error(sysErrNo) +} + +// DataLen returns the size of the response without the header. +func (r *Response) DataLen() uint32 { + return r.hdr.Len - uint32(r.hdr.SizeBytes()) +} + +// UnmarshalPayload unmarshals the response data into m. +func (r *Response) UnmarshalPayload(m marshal.Marshallable) error { + hdrLen := r.hdr.SizeBytes() + haveDataLen := r.hdr.Len - uint32(hdrLen) + wantDataLen := uint32(m.SizeBytes()) + + if haveDataLen < wantDataLen { + return fmt.Errorf("payload too small. Minimum data lenth required: %d, but got data length %d", wantDataLen, haveDataLen) + } + + // The response data is empty unless there is some payload. And so, doesn't + // need to be unmarshalled. + if r.data == nil { + return nil + } + + // TODO(gVisor.dev/issue/3698): Use the unsafe version once go_marshal is safe to use again. + m.UnmarshalBytes(r.data[hdrLen:]) + return nil +} diff --git a/pkg/sentry/fsimpl/fuse/utils_test.go b/pkg/sentry/fsimpl/fuse/utils_test.go new file mode 100644 index 000000000..e1d9e3365 --- /dev/null +++ b/pkg/sentry/fsimpl/fuse/utils_test.go @@ -0,0 +1,132 @@ +// 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 fuse + +import ( + "io" + "testing" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/marshal" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/testutil" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/usermem" +) + +func setup(t *testing.T) *testutil.System { + k, err := testutil.Boot() + if err != nil { + t.Fatalf("Error creating kernel: %v", err) + } + + ctx := k.SupervisorContext() + creds := auth.CredentialsFromContext(ctx) + + k.VFS().MustRegisterFilesystemType(Name, &FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ + AllowUserList: true, + AllowUserMount: true, + }) + + mntns, err := k.VFS().NewMountNamespace(ctx, creds, "", "tmpfs", &vfs.MountOptions{}) + if err != nil { + t.Fatalf("NewMountNamespace(): %v", err) + } + + return testutil.NewSystem(ctx, t, k.VFS(), mntns) +} + +// newTestConnection creates a fuse connection that the sentry can communicate with +// and the FD for the server to communicate with. +func newTestConnection(system *testutil.System, k *kernel.Kernel, maxActiveRequests uint64) (*connection, *vfs.FileDescription, error) { + vfsObj := &vfs.VirtualFilesystem{} + fuseDev := &DeviceFD{} + + if err := vfsObj.Init(system.Ctx); err != nil { + return nil, nil, err + } + + vd := vfsObj.NewAnonVirtualDentry("genCountFD") + defer vd.DecRef(system.Ctx) + if err := fuseDev.vfsfd.Init(fuseDev, linux.O_RDWR|linux.O_CREAT, vd.Mount(), vd.Dentry(), &vfs.FileDescriptionOptions{}); err != nil { + return nil, nil, err + } + + fsopts := filesystemOptions{ + maxActiveRequests: maxActiveRequests, + } + fs, err := newFUSEFilesystem(system.Ctx, 0, &fsopts, &fuseDev.vfsfd) + if err != nil { + return nil, nil, err + } + + return fs.conn, &fuseDev.vfsfd, nil +} + +type testPayload struct { + marshal.StubMarshallable + data uint32 +} + +// SizeBytes implements marshal.Marshallable.SizeBytes. +func (t *testPayload) SizeBytes() int { + return 4 +} + +// MarshalBytes implements marshal.Marshallable.MarshalBytes. +func (t *testPayload) MarshalBytes(dst []byte) { + usermem.ByteOrder.PutUint32(dst[:4], t.data) +} + +// UnmarshalBytes implements marshal.Marshallable.UnmarshalBytes. +func (t *testPayload) UnmarshalBytes(src []byte) { + *t = testPayload{data: usermem.ByteOrder.Uint32(src[:4])} +} + +// Packed implements marshal.Marshallable.Packed. +func (t *testPayload) Packed() bool { + return true +} + +// MarshalUnsafe implements marshal.Marshallable.MarshalUnsafe. +func (t *testPayload) MarshalUnsafe(dst []byte) { + t.MarshalBytes(dst) +} + +// UnmarshalUnsafe implements marshal.Marshallable.UnmarshalUnsafe. +func (t *testPayload) UnmarshalUnsafe(src []byte) { + t.UnmarshalBytes(src) +} + +// CopyOutN implements marshal.Marshallable.CopyOutN. +func (t *testPayload) CopyOutN(task marshal.CopyContext, addr usermem.Addr, limit int) (int, error) { + panic("not implemented") +} + +// CopyOut implements marshal.Marshallable.CopyOut. +func (t *testPayload) CopyOut(task marshal.CopyContext, addr usermem.Addr) (int, error) { + panic("not implemented") +} + +// CopyIn implements marshal.Marshallable.CopyIn. +func (t *testPayload) CopyIn(task marshal.CopyContext, addr usermem.Addr) (int, error) { + panic("not implemented") +} + +// WriteTo implements io.WriterTo.WriteTo. +func (t *testPayload) WriteTo(w io.Writer) (int64, error) { + panic("not implemented") +} diff --git a/pkg/sentry/fsimpl/gofer/directory.go b/pkg/sentry/fsimpl/gofer/directory.go index 91d2ae199..18c884b59 100644 --- a/pkg/sentry/fsimpl/gofer/directory.go +++ b/pkg/sentry/fsimpl/gofer/directory.go @@ -117,11 +117,12 @@ func (d *dentry) createSyntheticChildLocked(opts *createSyntheticOpts) { d.syntheticChildren++ } +// +stateify savable type directoryFD struct { fileDescription vfs.DirectoryFileDescriptionDefaultImpl - mu sync.Mutex + mu sync.Mutex `state:"nosave"` off int64 dirents []vfs.Dirent } diff --git a/pkg/sentry/fsimpl/gofer/filesystem.go b/pkg/sentry/fsimpl/gofer/filesystem.go index b01121f9e..94d96261b 100644 --- a/pkg/sentry/fsimpl/gofer/filesystem.go +++ b/pkg/sentry/fsimpl/gofer/filesystem.go @@ -1026,7 +1026,7 @@ func (d *dentry) open(ctx context.Context, rp *vfs.ResolvingPath, opts *vfs.Open // step is required even if !d.cachedMetadataAuthoritative() because // d.mappings has to be updated. // d.metadataMu has already been acquired if trunc == true. - d.updateFileSizeLocked(0) + d.updateSizeLocked(0) if d.cachedMetadataAuthoritative() { d.touchCMtimeLocked() @@ -1311,6 +1311,9 @@ func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa if !renamed.isDir() { return syserror.EISDIR } + if genericIsAncestorDentry(replaced, renamed) { + return syserror.ENOTEMPTY + } } else { if rp.MustBeDir() || renamed.isDir() { return syserror.ENOTDIR @@ -1361,14 +1364,15 @@ func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa // with reference counts and queue oldParent for checkCachingLocked if the // parent isn't actually changing. if oldParent != newParent { + oldParent.decRefLocked() ds = appendDentry(ds, oldParent) newParent.IncRef() if renamed.isSynthetic() { oldParent.syntheticChildren-- newParent.syntheticChildren++ } + renamed.parent = newParent } - renamed.parent = newParent renamed.name = newName if newParent.children == nil { newParent.children = make(map[string]*dentry) @@ -1412,11 +1416,11 @@ func (fs *filesystem) SetStatAt(ctx context.Context, rp *vfs.ResolvingPath, opts fs.renameMuRUnlockAndCheckCaching(ctx, &ds) return err } - if err := d.setStat(ctx, rp.Credentials(), &opts, rp.Mount()); err != nil { - fs.renameMuRUnlockAndCheckCaching(ctx, &ds) + err = d.setStat(ctx, rp.Credentials(), &opts, rp.Mount()) + fs.renameMuRUnlockAndCheckCaching(ctx, &ds) + if err != nil { return err } - fs.renameMuRUnlockAndCheckCaching(ctx, &ds) if ev := vfs.InotifyEventFromStatMask(opts.Stat.Mask); ev != 0 { d.InotifyWithParent(ctx, ev, 0, vfs.InodeEvent) @@ -1491,7 +1495,7 @@ func (fs *filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error return fs.unlinkAt(ctx, rp, false /* dir */) } -// BoundEndpointAt implements FilesystemImpl.BoundEndpointAt. +// BoundEndpointAt implements vfs.FilesystemImpl.BoundEndpointAt. func (fs *filesystem) BoundEndpointAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.BoundEndpointOptions) (transport.BoundEndpoint, error) { var ds *[]*dentry fs.renameMu.RLock() @@ -1552,11 +1556,11 @@ func (fs *filesystem) SetXattrAt(ctx context.Context, rp *vfs.ResolvingPath, opt fs.renameMuRUnlockAndCheckCaching(ctx, &ds) return err } - if err := d.setXattr(ctx, rp.Credentials(), &opts); err != nil { - fs.renameMuRUnlockAndCheckCaching(ctx, &ds) + err = d.setXattr(ctx, rp.Credentials(), &opts) + fs.renameMuRUnlockAndCheckCaching(ctx, &ds) + if err != nil { return err } - fs.renameMuRUnlockAndCheckCaching(ctx, &ds) d.InotifyWithParent(ctx, linux.IN_ATTRIB, 0, vfs.InodeEvent) return nil @@ -1571,11 +1575,11 @@ func (fs *filesystem) RemoveXattrAt(ctx context.Context, rp *vfs.ResolvingPath, fs.renameMuRUnlockAndCheckCaching(ctx, &ds) return err } - if err := d.removeXattr(ctx, rp.Credentials(), name); err != nil { - fs.renameMuRUnlockAndCheckCaching(ctx, &ds) + err = d.removeXattr(ctx, rp.Credentials(), name) + fs.renameMuRUnlockAndCheckCaching(ctx, &ds) + if err != nil { return err } - fs.renameMuRUnlockAndCheckCaching(ctx, &ds) d.InotifyWithParent(ctx, linux.IN_ATTRIB, 0, vfs.InodeEvent) return nil diff --git a/pkg/sentry/fsimpl/gofer/gofer.go b/pkg/sentry/fsimpl/gofer/gofer.go index 0e21c31a4..8608471f8 100644 --- a/pkg/sentry/fsimpl/gofer/gofer.go +++ b/pkg/sentry/fsimpl/gofer/gofer.go @@ -62,9 +62,13 @@ import ( const Name = "9p" // FilesystemType implements vfs.FilesystemType. +// +// +stateify savable type FilesystemType struct{} // filesystem implements vfs.FilesystemImpl. +// +// +stateify savable type filesystem struct { vfsfs vfs.Filesystem @@ -77,7 +81,7 @@ type filesystem struct { iopts InternalFilesystemOptions // client is the client used by this filesystem. client is immutable. - client *p9.Client + client *p9.Client `state:"nosave"` // FIXME(gvisor.dev/issue/1663): not yet supported. // clock is a realtime clock used to set timestamps in file operations. clock ktime.Clock @@ -95,7 +99,7 @@ type filesystem struct { // reference count (such that it is usable as vfs.ResolvingPath.Start() or // is reachable from its children), or if it is a child dentry (such that // it is reachable from its parent). - renameMu sync.RWMutex + renameMu sync.RWMutex `state:"nosave"` // cachedDentries contains all dentries with 0 references. (Due to race // conditions, it may also contain dentries with non-zero references.) @@ -107,7 +111,7 @@ type filesystem 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 + syncMu sync.Mutex `state:"nosave"` syncableDentries map[*dentry]struct{} specialFileFDs map[*specialFileFD]struct{} @@ -120,6 +124,8 @@ type filesystem struct { // dentries, it comes from QID.Path from the 9P server. Synthetic dentries // have have their inodeNumber generated sequentially, with the MSB reserved to // prevent conflicts with regular dentries. +// +// +stateify savable type inodeNumber uint64 // Reserve MSB for synthetic mounts. @@ -132,6 +138,7 @@ func inoFromPath(path uint64) inodeNumber { return inodeNumber(path &^ syntheticInoMask) } +// +stateify savable type filesystemOptions struct { // "Standard" 9P options. fd int @@ -177,6 +184,8 @@ type filesystemOptions struct { // InteropMode controls the client's interaction with other remote filesystem // users. +// +// +stateify savable type InteropMode uint32 const ( @@ -235,6 +244,8 @@ const ( // InternalFilesystemOptions may be passed as // vfs.GetFilesystemOptions.InternalData to FilesystemType.GetFilesystem. +// +// +stateify savable type InternalFilesystemOptions struct { // If LeakConnection is true, do not close the connection to the server // when the Filesystem is released. This is necessary for deployments in @@ -534,6 +545,8 @@ func (fs *filesystem) Release(ctx context.Context) { } // dentry implements vfs.DentryImpl. +// +// +stateify savable type dentry struct { vfsd vfs.Dentry @@ -563,7 +576,7 @@ type dentry struct { // 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, the // only files that can be synthetic are sockets, pipes, and directories. - file p9file + file p9file `state:"nosave"` // FIXME(gvisor.dev/issue/1663): not yet supported. // If deleted is non-zero, the file represented by this dentry has been // deleted. deleted is accessed using atomic memory operations. @@ -575,7 +588,7 @@ type dentry struct { cached bool dentryEntry - dirMu sync.Mutex + dirMu sync.Mutex `state:"nosave"` // If this dentry represents a directory, children contains: // @@ -607,7 +620,7 @@ type dentry struct { // To mutate: // - Lock metadataMu and use atomic operations to update because we might // have atomic readers that don't hold the lock. - metadataMu sync.Mutex + metadataMu sync.Mutex `state:"nosave"` ino inodeNumber // immutable mode uint32 // type is immutable, perms are mutable uid uint32 // auth.KUID, but stored as raw uint32 for sync/atomic @@ -638,7 +651,7 @@ type dentry struct { // other metadata fields. nlink uint32 - mapsMu sync.Mutex + mapsMu sync.Mutex `state:"nosave"` // If this dentry represents a regular file, mappings tracks mappings of // the file into memmap.MappingSpaces. mappings is protected by mapsMu. @@ -662,12 +675,12 @@ type dentry struct { // either p9.File transitions from closed (isNil() == true) to open // (isNil() == false), it may be mutated with handleMu locked, but cannot // be closed until the dentry is destroyed. - handleMu sync.RWMutex - readFile p9file - writeFile p9file + handleMu sync.RWMutex `state:"nosave"` + readFile p9file `state:"nosave"` // FIXME(gvisor.dev/issue/1663): not yet supported. + writeFile p9file `state:"nosave"` // FIXME(gvisor.dev/issue/1663): not yet supported. hostFD int32 - dataMu sync.RWMutex + dataMu sync.RWMutex `state:"nosave"` // If this dentry represents a regular file that is client-cached, cache // maps offsets into the cached file to offsets into @@ -833,7 +846,7 @@ func (d *dentry) updateFromP9AttrsLocked(mask p9.AttrMask, attr *p9.Attr) { atomic.StoreUint32(&d.nlink, uint32(attr.NLink)) } if mask.Size { - d.updateFileSizeLocked(attr.Size) + d.updateSizeLocked(attr.Size) } } @@ -987,7 +1000,7 @@ func (d *dentry) setStat(ctx context.Context, creds *auth.Credentials, opts *vfs // d.size should be kept up to date, and privatized // copy-on-write mappings of truncated pages need to be // invalidated, even if InteropModeShared is in effect. - d.updateFileSizeLocked(stat.Size) + d.updateSizeLocked(stat.Size) } } if d.fs.opts.interop == InteropModeShared { @@ -1024,8 +1037,31 @@ func (d *dentry) setStat(ctx context.Context, creds *auth.Credentials, opts *vfs return nil } +// doAllocate performs an allocate operation on d. Note that d.metadataMu will +// be held when allocate is called. +func (d *dentry) doAllocate(ctx context.Context, offset, length uint64, allocate func() error) error { + d.metadataMu.Lock() + defer d.metadataMu.Unlock() + + // Allocating a smaller size is a noop. + size := offset + length + if d.cachedMetadataAuthoritative() && size <= d.size { + return nil + } + + err := allocate() + if err != nil { + return err + } + d.updateSizeLocked(size) + if d.cachedMetadataAuthoritative() { + d.touchCMtimeLocked() + } + return nil +} + // Preconditions: d.metadataMu must be locked. -func (d *dentry) updateFileSizeLocked(newSize uint64) { +func (d *dentry) updateSizeLocked(newSize uint64) { d.dataMu.Lock() oldSize := d.size atomic.StoreUint64(&d.size, newSize) @@ -1604,12 +1640,14 @@ func (d *dentry) decLinks() { // fileDescription is embedded by gofer implementations of // vfs.FileDescriptionImpl. +// +// +stateify savable type fileDescription struct { vfsfd vfs.FileDescription vfs.FileDescriptionDefaultImpl vfs.LockFD - lockLogging sync.Once + lockLogging sync.Once `state:"nosave"` // FIXME(gvisor.dev/issue/1663): not yet supported. } func (fd *fileDescription) filesystem() *filesystem { diff --git a/pkg/sentry/fsimpl/gofer/handle.go b/pkg/sentry/fsimpl/gofer/handle.go index 104157512..a9ebe1206 100644 --- a/pkg/sentry/fsimpl/gofer/handle.go +++ b/pkg/sentry/fsimpl/gofer/handle.go @@ -25,6 +25,8 @@ import ( // handle represents a remote "open file descriptor", consisting of an opened // fid (p9.File) and optionally a host file descriptor. +// +// These are explicitly not savable. type handle struct { file p9file fd int32 // -1 if unavailable diff --git a/pkg/sentry/fsimpl/gofer/regular_file.go b/pkg/sentry/fsimpl/gofer/regular_file.go index a2e9342d5..eeaf6e444 100644 --- a/pkg/sentry/fsimpl/gofer/regular_file.go +++ b/pkg/sentry/fsimpl/gofer/regular_file.go @@ -39,11 +39,12 @@ func (d *dentry) isRegularFile() bool { return d.fileType() == linux.S_IFREG } +// +stateify savable type regularFileFD struct { fileDescription // off is the file offset. off is protected by mu. - mu sync.Mutex + mu sync.Mutex `state:"nosave"` off int64 } @@ -79,28 +80,11 @@ func (fd *regularFileFD) OnClose(ctx context.Context) error { // Allocate implements vfs.FileDescriptionImpl.Allocate. func (fd *regularFileFD) Allocate(ctx context.Context, mode, offset, length uint64) error { d := fd.dentry() - d.metadataMu.Lock() - defer d.metadataMu.Unlock() - - // Allocating a smaller size is a noop. - size := offset + length - if d.cachedMetadataAuthoritative() && size <= d.size { - return nil - } - - d.handleMu.RLock() - err := d.writeFile.allocate(ctx, p9.ToAllocateMode(mode), offset, length) - d.handleMu.RUnlock() - if err != nil { - return err - } - d.dataMu.Lock() - atomic.StoreUint64(&d.size, size) - d.dataMu.Unlock() - if d.cachedMetadataAuthoritative() { - d.touchCMtimeLocked() - } - return nil + return d.doAllocate(ctx, offset, length, func() error { + d.handleMu.RLock() + defer d.handleMu.RUnlock() + return d.writeFile.allocate(ctx, p9.ToAllocateMode(mode), offset, length) + }) } // PRead implements vfs.FileDescriptionImpl.PRead. @@ -915,6 +899,8 @@ func (d *dentry) Evict(ctx context.Context, er pgalloc.EvictableRange) { // dentryPlatformFile is only used when a host FD representing the remote file // is available (i.e. dentry.hostFD >= 0), and that FD is used for application // memory mappings (i.e. !filesystem.opts.forcePageCache). +// +// +stateify savable type dentryPlatformFile struct { *dentry @@ -927,7 +913,7 @@ type dentryPlatformFile struct { hostFileMapper fsutil.HostFileMapper // hostFileMapperInitOnce is used to lazily initialize hostFileMapper. - hostFileMapperInitOnce sync.Once + hostFileMapperInitOnce sync.Once `state:"nosave"` // FIXME(gvisor.dev/issue/1663): not yet supported. } // IncRef implements memmap.File.IncRef. diff --git a/pkg/sentry/fsimpl/gofer/socket.go b/pkg/sentry/fsimpl/gofer/socket.go index 85d2bee72..326b940a7 100644 --- a/pkg/sentry/fsimpl/gofer/socket.go +++ b/pkg/sentry/fsimpl/gofer/socket.go @@ -36,12 +36,14 @@ func (d *dentry) isSocket() bool { // An endpoint's lifetime is the time between when filesystem.BoundEndpointAt() // is called and either BoundEndpoint.BidirectionalConnect or // BoundEndpoint.UnidirectionalConnect is called. +// +// +stateify savable type endpoint struct { // dentry is the filesystem dentry which produced this endpoint. dentry *dentry // file is the p9 file that contains a single unopened fid. - file p9.File + file p9.File `state:"nosave"` // FIXME(gvisor.dev/issue/1663): not yet supported. // path is the sentry path where this endpoint is bound. path string diff --git a/pkg/sentry/fsimpl/gofer/special_file.go b/pkg/sentry/fsimpl/gofer/special_file.go index 3c39aa9b7..71581736c 100644 --- a/pkg/sentry/fsimpl/gofer/special_file.go +++ b/pkg/sentry/fsimpl/gofer/special_file.go @@ -22,6 +22,7 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/fdnotifier" + "gvisor.dev/gvisor/pkg/p9" "gvisor.dev/gvisor/pkg/safemem" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserror" @@ -33,11 +34,13 @@ import ( // special files, and (when filesystemOptions.regularFilesUseSpecialFileFD is // in effect) regular files. specialFileFD differs from regularFileFD by using // per-FD handles instead of shared per-dentry handles, and never buffering I/O. +// +// +stateify savable type specialFileFD struct { fileDescription // handle is used for file I/O. handle is immutable. - handle handle + handle handle `state:"nosave"` // FIXME(gvisor.dev/issue/1663): not yet supported. // isRegularFile is true if this FD represents a regular file which is only // possible when filesystemOptions.regularFilesUseSpecialFileFD is in @@ -55,7 +58,7 @@ type specialFileFD struct { queue waiter.Queue // If seekable is true, off is the file offset. off is protected by mu. - mu sync.Mutex + mu sync.Mutex `state:"nosave"` off int64 } @@ -135,6 +138,16 @@ func (fd *specialFileFD) EventUnregister(e *waiter.Entry) { fd.fileDescription.EventUnregister(e) } +func (fd *specialFileFD) Allocate(ctx context.Context, mode, offset, length uint64) error { + if fd.isRegularFile { + d := fd.dentry() + return d.doAllocate(ctx, offset, length, func() error { + return fd.handle.file.allocate(ctx, p9.ToAllocateMode(mode), offset, length) + }) + } + return fd.FileDescriptionDefaultImpl.Allocate(ctx, mode, offset, length) +} + // PRead implements vfs.FileDescriptionImpl.PRead. func (fd *specialFileFD) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { if fd.seekable && offset < 0 { @@ -235,11 +248,12 @@ func (fd *specialFileFD) pwrite(ctx context.Context, src usermem.IOSequence, off d.touchCMtime() } buf := make([]byte, src.NumBytes()) - // Don't do partial writes if we get a partial read from src. - if _, err := src.CopyIn(ctx, buf); err != nil { - return 0, offset, err + copied, copyErr := src.CopyIn(ctx, buf) + if copied == 0 && copyErr != nil { + // Only return the error if we didn't get any data. + return 0, offset, copyErr } - n, err := fd.handle.writeFromBlocksAt(ctx, safemem.BlockSeqOf(safemem.BlockFromSafeSlice(buf)), uint64(offset)) + n, err := fd.handle.writeFromBlocksAt(ctx, safemem.BlockSeqOf(safemem.BlockFromSafeSlice(buf[:copied])), uint64(offset)) if err == syserror.EAGAIN { err = syserror.ErrWouldBlock } @@ -256,7 +270,10 @@ func (fd *specialFileFD) pwrite(ctx context.Context, src usermem.IOSequence, off atomic.StoreUint64(&d.size, uint64(offset)) } } - return int64(n), offset, err + if err != nil { + return int64(n), offset, err + } + return int64(n), offset, copyErr } // Write implements vfs.FileDescriptionImpl.Write. diff --git a/pkg/sentry/fsimpl/host/host.go b/pkg/sentry/fsimpl/host/host.go index 1bd0e4ee8..ffe4ddb32 100644 --- a/pkg/sentry/fsimpl/host/host.go +++ b/pkg/sentry/fsimpl/host/host.go @@ -58,7 +58,7 @@ func newInode(fs *filesystem, hostFD int, fileType linux.FileMode, isTTY bool) ( canMap: fileType == linux.S_IFREG, } i.pf.inode = i - i.refs.EnableLeakCheck() + i.EnableLeakCheck() // Non-seekable files can't be memory mapped, assert this. if !i.seekable && i.canMap { @@ -126,7 +126,7 @@ func NewFD(ctx context.Context, mnt *vfs.Mount, hostFD int, opts *NewFDOptions) // For simplicity, fileDescription.offset is set 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 when importing. - return i.open(ctx, d.VFSDentry(), mnt, flags) + return i.open(ctx, d, mnt, flags) } // ImportFD sets up and returns a vfs.FileDescription from a donated fd. @@ -137,14 +137,16 @@ func ImportFD(ctx context.Context, mnt *vfs.Mount, hostFD int, isTTY bool) (*vfs } // filesystemType implements vfs.FilesystemType. +// +// +stateify savable type filesystemType struct{} -// GetFilesystem implements FilesystemType.GetFilesystem. +// GetFilesystem implements vfs.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. +// Name implements vfs.FilesystemType.Name. func (filesystemType) Name() string { return "none" } @@ -166,6 +168,8 @@ func NewFilesystem(vfsObj *vfs.VirtualFilesystem) (*vfs.Filesystem, error) { } // filesystem implements vfs.FilesystemImpl. +// +// +stateify savable type filesystem struct { kernfs.Filesystem @@ -185,6 +189,8 @@ func (fs *filesystem) PrependPath(ctx context.Context, vfsroot, vd vfs.VirtualDe } // inode implements kernfs.Inode. +// +// +stateify savable type inode struct { kernfs.InodeNoStatFS kernfs.InodeNotDirectory @@ -193,7 +199,7 @@ type inode struct { locks vfs.FileLocks // When the reference count reaches zero, the host fd is closed. - refs inodeRefs + inodeRefs // hostFD contains the host fd that this file was originally created from, // which must be available at time of restore. @@ -233,7 +239,7 @@ type inode struct { canMap bool // mapsMu protects mappings. - mapsMu sync.Mutex + mapsMu sync.Mutex `state:"nosave"` // If canMap is true, mappings tracks mappings of hostFD into // memmap.MappingSpaces. @@ -243,7 +249,7 @@ type inode struct { pf inodePlatformFile } -// CheckPermissions implements kernfs.Inode. +// CheckPermissions implements kernfs.Inode.CheckPermissions. 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 { @@ -252,7 +258,7 @@ func (i *inode) CheckPermissions(ctx context.Context, creds *auth.Credentials, a return vfs.GenericCheckPermissions(creds, ats, linux.FileMode(s.Mode), auth.KUID(s.Uid), auth.KGID(s.Gid)) } -// Mode implements kernfs.Inode. +// Mode implements kernfs.Inode.Mode. func (i *inode) Mode() linux.FileMode { var s syscall.Stat_t if err := syscall.Fstat(i.hostFD, &s); err != nil { @@ -263,7 +269,7 @@ func (i *inode) Mode() linux.FileMode { return linux.FileMode(s.Mode) } -// Stat implements kernfs.Inode. +// Stat implements kernfs.Inode.Stat. func (i *inode) Stat(ctx context.Context, vfsfs *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) { if opts.Mask&linux.STATX__RESERVED != 0 { return linux.Statx{}, syserror.EINVAL @@ -376,7 +382,7 @@ func (i *inode) fstat(fs *filesystem) (linux.Statx, error) { }, nil } -// SetStat implements kernfs.Inode. +// SetStat implements kernfs.Inode.SetStat. func (i *inode) SetStat(ctx context.Context, fs *vfs.Filesystem, creds *auth.Credentials, opts vfs.SetStatOptions) error { s := &opts.Stat @@ -435,19 +441,9 @@ func (i *inode) SetStat(ctx context.Context, fs *vfs.Filesystem, creds *auth.Cre return nil } -// IncRef implements kernfs.Inode. -func (i *inode) IncRef() { - i.refs.IncRef() -} - -// TryIncRef implements kernfs.Inode. -func (i *inode) TryIncRef() bool { - return i.refs.TryIncRef() -} - -// DecRef implements kernfs.Inode. +// DecRef implements kernfs.Inode.DecRef. func (i *inode) DecRef(ctx context.Context) { - i.refs.DecRef(func() { + i.inodeRefs.DecRef(func() { if i.wouldBlock { fdnotifier.RemoveFD(int32(i.hostFD)) } @@ -457,16 +453,16 @@ func (i *inode) DecRef(ctx context.Context) { }) } -// Open implements kernfs.Inode. -func (i *inode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { +// Open implements kernfs.Inode.Open. +func (i *inode) Open(ctx context.Context, rp *vfs.ResolvingPath, d *kernfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { // Once created, we cannot re-open a socket fd through /proc/[pid]/fd/. if i.Mode().FileType() == linux.S_IFSOCK { return nil, syserror.ENXIO } - return i.open(ctx, vfsd, rp.Mount(), opts.Flags) + return i.open(ctx, d, rp.Mount(), opts.Flags) } -func (i *inode) open(ctx context.Context, d *vfs.Dentry, mnt *vfs.Mount, flags uint32) (*vfs.FileDescription, error) { +func (i *inode) open(ctx context.Context, d *kernfs.Dentry, mnt *vfs.Mount, flags uint32) (*vfs.FileDescription, error) { var s syscall.Stat_t if err := syscall.Fstat(i.hostFD, &s); err != nil { return nil, err @@ -490,7 +486,7 @@ func (i *inode) open(ctx context.Context, d *vfs.Dentry, mnt *vfs.Mount, flags u return nil, err } // Currently, we only allow Unix sockets to be imported. - return unixsocket.NewFileDescription(ep, ep.Type(), flags, mnt, d, &i.locks) + return unixsocket.NewFileDescription(ep, ep.Type(), flags, mnt, d.VFSDentry(), &i.locks) case syscall.S_IFREG, syscall.S_IFIFO, syscall.S_IFCHR: if i.isTTY { @@ -500,7 +496,7 @@ func (i *inode) open(ctx context.Context, d *vfs.Dentry, mnt *vfs.Mount, flags u } fd.LockFD.Init(&i.locks) vfsfd := &fd.vfsfd - if err := vfsfd.Init(fd, flags, mnt, d, &vfs.FileDescriptionOptions{}); err != nil { + if err := vfsfd.Init(fd, flags, mnt, d.VFSDentry(), &vfs.FileDescriptionOptions{}); err != nil { return nil, err } return vfsfd, nil @@ -509,7 +505,7 @@ func (i *inode) open(ctx context.Context, d *vfs.Dentry, mnt *vfs.Mount, flags u fd := &fileDescription{inode: i} fd.LockFD.Init(&i.locks) vfsfd := &fd.vfsfd - if err := vfsfd.Init(fd, flags, mnt, d, &vfs.FileDescriptionOptions{}); err != nil { + if err := vfsfd.Init(fd, flags, mnt, d.VFSDentry(), &vfs.FileDescriptionOptions{}); err != nil { return nil, err } return vfsfd, nil @@ -521,6 +517,8 @@ func (i *inode) open(ctx context.Context, d *vfs.Dentry, mnt *vfs.Mount, flags u } // fileDescription is embedded by host fd implementations of FileDescriptionImpl. +// +// +stateify savable type fileDescription struct { vfsfd vfs.FileDescription vfs.FileDescriptionDefaultImpl @@ -535,40 +533,35 @@ type fileDescription struct { inode *inode // offsetMu protects offset. - offsetMu sync.Mutex + offsetMu sync.Mutex `state:"nosave"` // offset specifies the current file offset. It is only meaningful when // inode.seekable is true. offset int64 } -// SetStat implements vfs.FileDescriptionImpl. +// SetStat implements vfs.FileDescriptionImpl.SetStat. func (f *fileDescription) SetStat(ctx context.Context, opts vfs.SetStatOptions) error { creds := auth.CredentialsFromContext(ctx) return f.inode.SetStat(ctx, f.vfsfd.Mount().Filesystem(), creds, opts) } -// Stat implements vfs.FileDescriptionImpl. +// Stat implements vfs.FileDescriptionImpl.Stat. func (f *fileDescription) Stat(ctx context.Context, opts vfs.StatOptions) (linux.Statx, error) { return f.inode.Stat(ctx, f.vfsfd.Mount().Filesystem(), opts) } -// Release implements vfs.FileDescriptionImpl. +// Release implements vfs.FileDescriptionImpl.Release. func (f *fileDescription) Release(context.Context) { // noop } -// Allocate implements vfs.FileDescriptionImpl. +// Allocate implements vfs.FileDescriptionImpl.Allocate. func (f *fileDescription) Allocate(ctx context.Context, mode, offset, length uint64) error { - if !f.inode.seekable { - return syserror.ESPIPE - } - - // TODO(gvisor.dev/issue/3589): Implement Allocate for non-pipe hostfds. - return syserror.EOPNOTSUPP + return unix.Fallocate(f.inode.hostFD, uint32(mode), int64(offset), int64(length)) } -// PRead implements FileDescriptionImpl. +// PRead implements vfs.FileDescriptionImpl.PRead. func (f *fileDescription) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { i := f.inode if !i.seekable { @@ -578,7 +571,7 @@ func (f *fileDescription) PRead(ctx context.Context, dst usermem.IOSequence, off return readFromHostFD(ctx, i.hostFD, dst, offset, opts.Flags) } -// Read implements FileDescriptionImpl. +// Read implements vfs.FileDescriptionImpl.Read. func (f *fileDescription) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { i := f.inode if !i.seekable { @@ -615,7 +608,7 @@ func readFromHostFD(ctx context.Context, hostFD int, dst usermem.IOSequence, off return int64(n), err } -// PWrite implements FileDescriptionImpl. +// PWrite implements vfs.FileDescriptionImpl.PWrite. func (f *fileDescription) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) { if !f.inode.seekable { return 0, syserror.ESPIPE @@ -624,7 +617,7 @@ func (f *fileDescription) PWrite(ctx context.Context, src usermem.IOSequence, of return f.writeToHostFD(ctx, src, offset, opts.Flags) } -// Write implements FileDescriptionImpl. +// Write implements vfs.FileDescriptionImpl.Write. func (f *fileDescription) Write(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) { i := f.inode if !i.seekable { @@ -672,7 +665,7 @@ func (f *fileDescription) writeToHostFD(ctx context.Context, src usermem.IOSeque return int64(n), err } -// Seek implements FileDescriptionImpl. +// Seek implements vfs.FileDescriptionImpl.Seek. // // Note that we do not support seeking on directories, since we do not even // allow directory fds to be imported at all. @@ -737,13 +730,13 @@ func (f *fileDescription) Seek(_ context.Context, offset int64, whence int32) (i return f.offset, nil } -// Sync implements FileDescriptionImpl. +// Sync implements vfs.FileDescriptionImpl.Sync. func (f *fileDescription) Sync(context.Context) error { // TODO(gvisor.dev/issue/1897): Currently, we always sync everything. return unix.Fsync(f.inode.hostFD) } -// ConfigureMMap implements FileDescriptionImpl. +// ConfigureMMap implements vfs.FileDescriptionImpl.ConfigureMMap. func (f *fileDescription) ConfigureMMap(_ context.Context, opts *memmap.MMapOpts) error { if !f.inode.canMap { return syserror.ENODEV diff --git a/pkg/sentry/fsimpl/host/mmap.go b/pkg/sentry/fsimpl/host/mmap.go index 65d3af38c..b51a17bed 100644 --- a/pkg/sentry/fsimpl/host/mmap.go +++ b/pkg/sentry/fsimpl/host/mmap.go @@ -27,11 +27,13 @@ import ( // cannot implement both kernfs.Inode.IncRef and memmap.File.IncRef. // // inodePlatformFile should only be used if inode.canMap is true. +// +// +stateify savable type inodePlatformFile struct { *inode // fdRefsMu protects fdRefs. - fdRefsMu sync.Mutex + fdRefsMu sync.Mutex `state:"nosave"` // fdRefs counts references on memmap.File offsets. It is used solely for // memory accounting. @@ -41,7 +43,7 @@ type inodePlatformFile struct { fileMapper fsutil.HostFileMapper // fileMapperInitOnce is used to lazily initialize fileMapper. - fileMapperInitOnce sync.Once + fileMapperInitOnce sync.Once `state:"nosave"` // FIXME(gvisor.dev/issue/1663): not yet supported. } // IncRef implements memmap.File.IncRef. diff --git a/pkg/sentry/fsimpl/host/socket.go b/pkg/sentry/fsimpl/host/socket.go index 131145b85..8a447e29f 100644 --- a/pkg/sentry/fsimpl/host/socket.go +++ b/pkg/sentry/fsimpl/host/socket.go @@ -348,10 +348,10 @@ func (e *SCMConnectedEndpoint) Init() error { func (e *SCMConnectedEndpoint) Release(ctx context.Context) { e.DecRef(func() { e.mu.Lock() + fdnotifier.RemoveFD(int32(e.fd)) if err := syscall.Close(e.fd); err != nil { log.Warningf("Failed to close host fd %d: %v", err) } - fdnotifier.RemoveFD(int32(e.fd)) e.destroyLocked() e.mu.Unlock() }) diff --git a/pkg/sentry/fsimpl/host/tty.go b/pkg/sentry/fsimpl/host/tty.go index 97cefa350..f5c596fec 100644 --- a/pkg/sentry/fsimpl/host/tty.go +++ b/pkg/sentry/fsimpl/host/tty.go @@ -30,6 +30,8 @@ import ( // TTYFileDescription implements vfs.FileDescriptionImpl for a host file // descriptor that wraps a TTY FD. +// +// +stateify savable type TTYFileDescription struct { fileDescription @@ -76,7 +78,7 @@ func (t *TTYFileDescription) Release(ctx context.Context) { t.fileDescription.Release(ctx) } -// PRead implements vfs.FileDescriptionImpl. +// PRead implements vfs.FileDescriptionImpl.PRead. // // Reading from a TTY is only allowed for foreground process groups. Background // process groups will either get EIO or a SIGTTIN. @@ -94,7 +96,7 @@ func (t *TTYFileDescription) PRead(ctx context.Context, dst usermem.IOSequence, return t.fileDescription.PRead(ctx, dst, offset, opts) } -// Read implements vfs.FileDescriptionImpl. +// Read implements vfs.FileDescriptionImpl.Read. // // Reading from a TTY is only allowed for foreground process groups. Background // process groups will either get EIO or a SIGTTIN. @@ -112,7 +114,7 @@ func (t *TTYFileDescription) Read(ctx context.Context, dst usermem.IOSequence, o return t.fileDescription.Read(ctx, dst, opts) } -// PWrite implements vfs.FileDescriptionImpl. +// PWrite implements vfs.FileDescriptionImpl.PWrite. func (t *TTYFileDescription) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) { t.mu.Lock() defer t.mu.Unlock() @@ -127,7 +129,7 @@ func (t *TTYFileDescription) PWrite(ctx context.Context, src usermem.IOSequence, return t.fileDescription.PWrite(ctx, src, offset, opts) } -// Write implements vfs.FileDescriptionImpl. +// Write implements vfs.FileDescriptionImpl.Write. func (t *TTYFileDescription) Write(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) { t.mu.Lock() defer t.mu.Unlock() @@ -142,7 +144,7 @@ func (t *TTYFileDescription) Write(ctx context.Context, src usermem.IOSequence, return t.fileDescription.Write(ctx, src, opts) } -// Ioctl implements vfs.FileDescriptionImpl. +// Ioctl implements vfs.FileDescriptionImpl.Ioctl. func (t *TTYFileDescription) Ioctl(ctx context.Context, io usermem.IO, args arch.SyscallArguments) (uintptr, error) { task := kernel.TaskFromContext(ctx) if task == nil { diff --git a/pkg/sentry/fsimpl/kernfs/dynamic_bytes_file.go b/pkg/sentry/fsimpl/kernfs/dynamic_bytes_file.go index 1ee089620..b929118b1 100644 --- a/pkg/sentry/fsimpl/kernfs/dynamic_bytes_file.go +++ b/pkg/sentry/fsimpl/kernfs/dynamic_bytes_file.go @@ -56,9 +56,9 @@ func (f *DynamicBytesFile) Init(creds *auth.Credentials, devMajor, devMinor uint } // Open implements Inode.Open. -func (f *DynamicBytesFile) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { +func (f *DynamicBytesFile) Open(ctx context.Context, rp *vfs.ResolvingPath, d *Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { fd := &DynamicBytesFD{} - if err := fd.Init(rp.Mount(), vfsd, f.data, &f.locks, opts.Flags); err != nil { + if err := fd.Init(rp.Mount(), d, f.data, &f.locks, opts.Flags); err != nil { return nil, err } return &fd.vfsfd, nil @@ -87,12 +87,12 @@ type DynamicBytesFD struct { } // Init initializes a DynamicBytesFD. -func (fd *DynamicBytesFD) Init(m *vfs.Mount, d *vfs.Dentry, data vfs.DynamicBytesSource, locks *vfs.FileLocks, flags uint32) error { +func (fd *DynamicBytesFD) Init(m *vfs.Mount, d *Dentry, data vfs.DynamicBytesSource, locks *vfs.FileLocks, flags uint32) error { fd.LockFD.Init(locks) - if err := fd.vfsfd.Init(fd, flags, m, d, &vfs.FileDescriptionOptions{}); err != nil { + if err := fd.vfsfd.Init(fd, flags, m, d.VFSDentry(), &vfs.FileDescriptionOptions{}); err != nil { return err } - fd.inode = d.Impl().(*Dentry).inode + fd.inode = d.inode fd.SetDataSource(data) return nil } diff --git a/pkg/sentry/fsimpl/kernfs/fd_impl_util.go b/pkg/sentry/fsimpl/kernfs/fd_impl_util.go index 6518ff5cd..0a4cd4057 100644 --- a/pkg/sentry/fsimpl/kernfs/fd_impl_util.go +++ b/pkg/sentry/fsimpl/kernfs/fd_impl_util.go @@ -29,6 +29,8 @@ import ( ) // SeekEndConfig describes the SEEK_END behaviour for FDs. +// +// +stateify savable type SeekEndConfig int // Constants related to SEEK_END behaviour for FDs. @@ -41,6 +43,8 @@ const ( ) // GenericDirectoryFDOptions contains configuration for a GenericDirectoryFD. +// +// +stateify savable type GenericDirectoryFDOptions struct { SeekEnd SeekEndConfig } @@ -56,6 +60,8 @@ type GenericDirectoryFDOptions struct { // Must be initialize with Init before first use. // // Lock ordering: mu => children.mu. +// +// +stateify savable type GenericDirectoryFD struct { vfs.FileDescriptionDefaultImpl vfs.DirectoryFileDescriptionDefaultImpl @@ -68,7 +74,7 @@ type GenericDirectoryFD struct { children *OrderedChildren // mu protects the fields below. - mu sync.Mutex + mu sync.Mutex `state:"nosave"` // off is the current directory offset. Protected by "mu". off int64 @@ -76,12 +82,12 @@ type GenericDirectoryFD struct { // NewGenericDirectoryFD creates a new GenericDirectoryFD and returns its // dentry. -func NewGenericDirectoryFD(m *vfs.Mount, d *vfs.Dentry, children *OrderedChildren, locks *vfs.FileLocks, opts *vfs.OpenOptions, fdOpts GenericDirectoryFDOptions) (*GenericDirectoryFD, error) { +func NewGenericDirectoryFD(m *vfs.Mount, d *Dentry, children *OrderedChildren, locks *vfs.FileLocks, opts *vfs.OpenOptions, fdOpts GenericDirectoryFDOptions) (*GenericDirectoryFD, error) { fd := &GenericDirectoryFD{} if err := fd.Init(children, locks, opts, fdOpts); err != nil { return nil, err } - if err := fd.vfsfd.Init(fd, opts.Flags, m, d, &vfs.FileDescriptionOptions{}); err != nil { + if err := fd.vfsfd.Init(fd, opts.Flags, m, d.VFSDentry(), &vfs.FileDescriptionOptions{}); err != nil { return nil, err } return fd, nil @@ -195,8 +201,7 @@ func (fd *GenericDirectoryFD) IterDirents(ctx context.Context, cb vfs.IterDirent // these. childIdx := fd.off - 2 for it := fd.children.nthLocked(childIdx); it != nil; it = it.Next() { - inode := it.Dentry.Impl().(*Dentry).inode - stat, err := inode.Stat(ctx, fd.filesystem(), opts) + stat, err := it.Dentry.inode.Stat(ctx, fd.filesystem(), opts) if err != nil { return err } diff --git a/pkg/sentry/fsimpl/kernfs/filesystem.go b/pkg/sentry/fsimpl/kernfs/filesystem.go index d7d3e8f48..5cc1c4281 100644 --- a/pkg/sentry/fsimpl/kernfs/filesystem.go +++ b/pkg/sentry/fsimpl/kernfs/filesystem.go @@ -37,8 +37,7 @@ import ( // * !rp.Done(). // // Postcondition: Caller must call fs.processDeferredDecRefs*. -func (fs *Filesystem) stepExistingLocked(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, mayFollowSymlinks bool) (*vfs.Dentry, error) { - d := vfsd.Impl().(*Dentry) +func (fs *Filesystem) stepExistingLocked(ctx context.Context, rp *vfs.ResolvingPath, d *Dentry, mayFollowSymlinks bool) (*Dentry, error) { if !d.isDir() { return nil, syserror.ENOTDIR } @@ -55,20 +54,20 @@ afterSymlink: // calls d_revalidate(), but walk_component() => handle_dots() does not. if name == "." { rp.Advance() - return vfsd, nil + return d, nil } if name == ".." { - if isRoot, err := rp.CheckRoot(ctx, vfsd); err != nil { + if isRoot, err := rp.CheckRoot(ctx, d.VFSDentry()); err != nil { return nil, err } else if isRoot || d.parent == nil { rp.Advance() - return vfsd, nil + return d, nil } - if err := rp.CheckMount(ctx, &d.parent.vfsd); err != nil { + if err := rp.CheckMount(ctx, d.parent.VFSDentry()); err != nil { return nil, err } rp.Advance() - return &d.parent.vfsd, nil + return d.parent, nil } if len(name) > linux.NAME_MAX { return nil, syserror.ENAMETOOLONG @@ -79,7 +78,7 @@ afterSymlink: if err != nil { return nil, err } - if err := rp.CheckMount(ctx, &next.vfsd); err != nil { + if err := rp.CheckMount(ctx, next.VFSDentry()); err != nil { return nil, err } // Resolve any symlink at current path component. @@ -102,7 +101,7 @@ afterSymlink: goto afterSymlink } rp.Advance() - return &next.vfsd, nil + return next, nil } // revalidateChildLocked must be called after a call to parent.vfsd.Child(name) @@ -122,25 +121,21 @@ func (fs *Filesystem) revalidateChildLocked(ctx context.Context, vfsObj *vfs.Vir if !child.inode.Valid(ctx) { delete(parent.children, name) vfsObj.InvalidateDentry(ctx, &child.vfsd) - fs.deferDecRef(&child.vfsd) // Reference from Lookup. + fs.deferDecRef(child) // Reference from Lookup. child = nil } } if child == nil { - // Dentry isn't cached; it either doesn't exist or failed - // revalidation. Attempt to resolve it via Lookup. - // - // FIXME(gvisor.dev/issue/1193): Inode.Lookup() should return - // *(kernfs.)Dentry, not *vfs.Dentry, since (kernfs.)Filesystem assumes - // that all dentries in the filesystem are (kernfs.)Dentry and performs - // vfs.DentryImpl casts accordingly. - childVFSD, err := parent.inode.Lookup(ctx, name) + // Dentry isn't cached; it either doesn't exist or failed revalidation. + // Attempt to resolve it via Lookup. + c, err := parent.inode.Lookup(ctx, name) if err != nil { return nil, err } - // Reference on childVFSD dropped by a corresponding Valid. - child = childVFSD.Impl().(*Dentry) - parent.insertChildLocked(name, child) + // Reference on c (provided by Lookup) will be dropped when the dentry + // fails validation. + parent.InsertChildLocked(name, c) + child = c } return child, nil } @@ -153,20 +148,19 @@ func (fs *Filesystem) revalidateChildLocked(ctx context.Context, vfsObj *vfs.Vir // Preconditions: Filesystem.mu must be locked for at least reading. // // Postconditions: Caller must call fs.processDeferredDecRefs*. -func (fs *Filesystem) walkExistingLocked(ctx context.Context, rp *vfs.ResolvingPath) (*vfs.Dentry, Inode, error) { - vfsd := rp.Start() +func (fs *Filesystem) walkExistingLocked(ctx context.Context, rp *vfs.ResolvingPath) (*Dentry, error) { + d := rp.Start().Impl().(*Dentry) for !rp.Done() { var err error - vfsd, err = fs.stepExistingLocked(ctx, rp, vfsd, true /* mayFollowSymlinks */) + d, err = fs.stepExistingLocked(ctx, rp, d, true /* mayFollowSymlinks */) if err != nil { - return nil, nil, err + return nil, err } } - d := vfsd.Impl().(*Dentry) if rp.MustBeDir() && !d.isDir() { - return nil, nil, syserror.ENOTDIR + return nil, syserror.ENOTDIR } - return vfsd, d.inode, nil + return d, nil } // walkParentDirLocked resolves all but the last path component of rp to an @@ -181,20 +175,19 @@ func (fs *Filesystem) walkExistingLocked(ctx context.Context, rp *vfs.ResolvingP // * !rp.Done(). // // Postconditions: Caller must call fs.processDeferredDecRefs*. -func (fs *Filesystem) walkParentDirLocked(ctx context.Context, rp *vfs.ResolvingPath) (*vfs.Dentry, Inode, error) { - vfsd := rp.Start() +func (fs *Filesystem) walkParentDirLocked(ctx context.Context, rp *vfs.ResolvingPath) (*Dentry, error) { + d := rp.Start().Impl().(*Dentry) for !rp.Final() { var err error - vfsd, err = fs.stepExistingLocked(ctx, rp, vfsd, true /* mayFollowSymlinks */) + d, err = fs.stepExistingLocked(ctx, rp, d, true /* mayFollowSymlinks */) if err != nil { - return nil, nil, err + return nil, err } } - d := vfsd.Impl().(*Dentry) if !d.isDir() { - return nil, nil, syserror.ENOTDIR + return nil, syserror.ENOTDIR } - return vfsd, d.inode, nil + return d, nil } // checkCreateLocked checks that a file named rp.Component() may be created in @@ -202,10 +195,9 @@ func (fs *Filesystem) walkParentDirLocked(ctx context.Context, rp *vfs.Resolving // // Preconditions: // * Filesystem.mu must be locked for at least reading. -// * parentInode == parentVFSD.Impl().(*Dentry).Inode. // * isDir(parentInode) == true. -func checkCreateLocked(ctx context.Context, rp *vfs.ResolvingPath, parentVFSD *vfs.Dentry, parentInode Inode) (string, error) { - if err := parentInode.CheckPermissions(ctx, rp.Credentials(), vfs.MayWrite|vfs.MayExec); err != nil { +func checkCreateLocked(ctx context.Context, rp *vfs.ResolvingPath, parent *Dentry) (string, error) { + if err := parent.inode.CheckPermissions(ctx, rp.Credentials(), vfs.MayWrite|vfs.MayExec); err != nil { return "", err } pc := rp.Component() @@ -215,11 +207,10 @@ func checkCreateLocked(ctx context.Context, rp *vfs.ResolvingPath, parentVFSD *v if len(pc) > linux.NAME_MAX { return "", syserror.ENAMETOOLONG } - // FIXME(gvisor.dev/issue/1193): Data race due to not holding dirMu. - if _, ok := parentVFSD.Impl().(*Dentry).children[pc]; ok { + if _, ok := parent.children[pc]; ok { return "", syserror.EEXIST } - if parentVFSD.IsDead() { + if parent.VFSDentry().IsDead() { return "", syserror.ENOENT } return pc, nil @@ -228,8 +219,8 @@ func checkCreateLocked(ctx context.Context, rp *vfs.ResolvingPath, parentVFSD *v // checkDeleteLocked checks that the file represented by vfsd may be deleted. // // Preconditions: Filesystem.mu must be locked for at least reading. -func checkDeleteLocked(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry) error { - parent := vfsd.Impl().(*Dentry).parent +func checkDeleteLocked(ctx context.Context, rp *vfs.ResolvingPath, d *Dentry) error { + parent := d.parent if parent == nil { return syserror.EBUSY } @@ -258,11 +249,11 @@ func (fs *Filesystem) AccessAt(ctx context.Context, rp *vfs.ResolvingPath, creds defer fs.processDeferredDecRefs(ctx) defer fs.mu.RUnlock() - _, inode, err := fs.walkExistingLocked(ctx, rp) + d, err := fs.walkExistingLocked(ctx, rp) if err != nil { return err } - return inode.CheckPermissions(ctx, creds, ats) + return d.inode.CheckPermissions(ctx, creds, ats) } // GetDentryAt implements vfs.FilesystemImpl.GetDentryAt. @@ -270,20 +261,20 @@ func (fs *Filesystem) GetDentryAt(ctx context.Context, rp *vfs.ResolvingPath, op fs.mu.RLock() defer fs.processDeferredDecRefs(ctx) defer fs.mu.RUnlock() - vfsd, inode, err := fs.walkExistingLocked(ctx, rp) + d, err := fs.walkExistingLocked(ctx, rp) if err != nil { return nil, err } if opts.CheckSearchable { - d := vfsd.Impl().(*Dentry) if !d.isDir() { return nil, syserror.ENOTDIR } - if err := inode.CheckPermissions(ctx, rp.Credentials(), vfs.MayExec); err != nil { + if err := d.inode.CheckPermissions(ctx, rp.Credentials(), vfs.MayExec); err != nil { return nil, err } } + vfsd := d.VFSDentry() vfsd.IncRef() // Ownership transferred to caller. return vfsd, nil } @@ -293,12 +284,12 @@ func (fs *Filesystem) GetParentDentryAt(ctx context.Context, rp *vfs.ResolvingPa fs.mu.RLock() defer fs.processDeferredDecRefs(ctx) defer fs.mu.RUnlock() - vfsd, _, err := fs.walkParentDirLocked(ctx, rp) + d, err := fs.walkParentDirLocked(ctx, rp) if err != nil { return nil, err } - vfsd.IncRef() // Ownership transferred to caller. - return vfsd, nil + d.IncRef() // Ownership transferred to caller. + return d.VFSDentry(), nil } // LinkAt implements vfs.FilesystemImpl.LinkAt. @@ -308,12 +299,15 @@ func (fs *Filesystem) LinkAt(ctx context.Context, rp *vfs.ResolvingPath, vd vfs. } fs.mu.Lock() defer fs.mu.Unlock() - parentVFSD, parentInode, err := fs.walkParentDirLocked(ctx, rp) + parent, err := fs.walkParentDirLocked(ctx, rp) fs.processDeferredDecRefsLocked(ctx) if err != nil { return err } - pc, err := checkCreateLocked(ctx, rp, parentVFSD, parentInode) + + parent.dirMu.Lock() + defer parent.dirMu.Unlock() + pc, err := checkCreateLocked(ctx, rp, parent) if err != nil { return err } @@ -330,11 +324,11 @@ func (fs *Filesystem) LinkAt(ctx context.Context, rp *vfs.ResolvingPath, vd vfs. return syserror.EPERM } - childVFSD, err := parentInode.NewLink(ctx, pc, d.inode) + child, err := parent.inode.NewLink(ctx, pc, d.inode) if err != nil { return err } - parentVFSD.Impl().(*Dentry).InsertChild(pc, childVFSD.Impl().(*Dentry)) + parent.InsertChildLocked(pc, child) return nil } @@ -345,12 +339,15 @@ func (fs *Filesystem) MkdirAt(ctx context.Context, rp *vfs.ResolvingPath, opts v } fs.mu.Lock() defer fs.mu.Unlock() - parentVFSD, parentInode, err := fs.walkParentDirLocked(ctx, rp) + parent, err := fs.walkParentDirLocked(ctx, rp) fs.processDeferredDecRefsLocked(ctx) if err != nil { return err } - pc, err := checkCreateLocked(ctx, rp, parentVFSD, parentInode) + + parent.dirMu.Lock() + defer parent.dirMu.Unlock() + pc, err := checkCreateLocked(ctx, rp, parent) if err != nil { return err } @@ -358,14 +355,14 @@ func (fs *Filesystem) MkdirAt(ctx context.Context, rp *vfs.ResolvingPath, opts v return err } defer rp.Mount().EndWrite() - childVFSD, err := parentInode.NewDir(ctx, pc, opts) + child, err := parent.inode.NewDir(ctx, pc, opts) if err != nil { if !opts.ForSyntheticMountpoint || err == syserror.EEXIST { return err } - childVFSD = newSyntheticDirectory(rp.Credentials(), opts.Mode) + child = newSyntheticDirectory(rp.Credentials(), opts.Mode) } - parentVFSD.Impl().(*Dentry).InsertChild(pc, childVFSD.Impl().(*Dentry)) + parent.InsertChildLocked(pc, child) return nil } @@ -376,12 +373,15 @@ func (fs *Filesystem) MknodAt(ctx context.Context, rp *vfs.ResolvingPath, opts v } fs.mu.Lock() defer fs.mu.Unlock() - parentVFSD, parentInode, err := fs.walkParentDirLocked(ctx, rp) + parent, err := fs.walkParentDirLocked(ctx, rp) fs.processDeferredDecRefsLocked(ctx) if err != nil { return err } - pc, err := checkCreateLocked(ctx, rp, parentVFSD, parentInode) + + parent.dirMu.Lock() + defer parent.dirMu.Unlock() + pc, err := checkCreateLocked(ctx, rp, parent) if err != nil { return err } @@ -389,11 +389,11 @@ func (fs *Filesystem) MknodAt(ctx context.Context, rp *vfs.ResolvingPath, opts v return err } defer rp.Mount().EndWrite() - newVFSD, err := parentInode.NewNode(ctx, pc, opts) + newD, err := parent.inode.NewNode(ctx, pc, opts) if err != nil { return err } - parentVFSD.Impl().(*Dentry).InsertChild(pc, newVFSD.Impl().(*Dentry)) + parent.InsertChildLocked(pc, newD) return nil } @@ -409,28 +409,27 @@ func (fs *Filesystem) OpenAt(ctx context.Context, rp *vfs.ResolvingPath, opts vf // Do not create new file. if opts.Flags&linux.O_CREAT == 0 { fs.mu.RLock() - vfsd, inode, err := fs.walkExistingLocked(ctx, rp) + d, err := fs.walkExistingLocked(ctx, rp) if err != nil { fs.mu.RUnlock() fs.processDeferredDecRefs(ctx) return nil, err } - if err := inode.CheckPermissions(ctx, rp.Credentials(), ats); err != nil { + if err := d.inode.CheckPermissions(ctx, rp.Credentials(), ats); err != nil { fs.mu.RUnlock() fs.processDeferredDecRefs(ctx) return nil, err } - inode.IncRef() - defer inode.DecRef(ctx) + d.inode.IncRef() + defer d.inode.DecRef(ctx) fs.mu.RUnlock() fs.processDeferredDecRefs(ctx) - return inode.Open(ctx, rp, vfsd, opts) + return d.inode.Open(ctx, rp, d, opts) } // May create new file. mustCreate := opts.Flags&linux.O_EXCL != 0 - vfsd := rp.Start() - inode := vfsd.Impl().(*Dentry).inode + d := rp.Start().Impl().(*Dentry) fs.mu.Lock() unlocked := false unlock := func() { @@ -447,22 +446,22 @@ func (fs *Filesystem) OpenAt(ctx context.Context, rp *vfs.ResolvingPath, opts vf if mustCreate { return nil, syserror.EEXIST } - if err := inode.CheckPermissions(ctx, rp.Credentials(), ats); err != nil { + if err := d.inode.CheckPermissions(ctx, rp.Credentials(), ats); err != nil { return nil, err } - inode.IncRef() - defer inode.DecRef(ctx) + d.inode.IncRef() + defer d.inode.DecRef(ctx) unlock() - return inode.Open(ctx, rp, vfsd, opts) + return d.inode.Open(ctx, rp, d, opts) } afterTrailingSymlink: - parentVFSD, parentInode, err := fs.walkParentDirLocked(ctx, rp) + parent, err := fs.walkParentDirLocked(ctx, rp) fs.processDeferredDecRefsLocked(ctx) if err != nil { return nil, err } // Check for search permission in the parent directory. - if err := parentInode.CheckPermissions(ctx, rp.Credentials(), vfs.MayExec); err != nil { + if err := parent.inode.CheckPermissions(ctx, rp.Credentials(), vfs.MayExec); err != nil { return nil, err } // Reject attempts to open directories with O_CREAT. @@ -477,10 +476,10 @@ afterTrailingSymlink: return nil, syserror.ENAMETOOLONG } // Determine whether or not we need to create a file. - childVFSD, err := fs.stepExistingLocked(ctx, rp, parentVFSD, false /* mayFollowSymlinks */) + child, err := fs.stepExistingLocked(ctx, rp, parent, false /* mayFollowSymlinks */) if err == syserror.ENOENT { // Already checked for searchability above; now check for writability. - if err := parentInode.CheckPermissions(ctx, rp.Credentials(), vfs.MayWrite); err != nil { + if err := parent.inode.CheckPermissions(ctx, rp.Credentials(), vfs.MayWrite); err != nil { return nil, err } if err := rp.Mount().CheckBeginWrite(); err != nil { @@ -488,16 +487,18 @@ afterTrailingSymlink: } defer rp.Mount().EndWrite() // Create and open the child. - childVFSD, err = parentInode.NewFile(ctx, pc, opts) + child, err := parent.inode.NewFile(ctx, pc, opts) if err != nil { return nil, err } - child := childVFSD.Impl().(*Dentry) - parentVFSD.Impl().(*Dentry).InsertChild(pc, child) + // FIXME(gvisor.dev/issue/1193): Race between checking existence with + // fs.stepExistingLocked and parent.InsertChild. If possible, we should hold + // dirMu from one to the other. + parent.InsertChild(pc, child) child.inode.IncRef() defer child.inode.DecRef(ctx) unlock() - return child.inode.Open(ctx, rp, childVFSD, opts) + return child.inode.Open(ctx, rp, child, opts) } if err != nil { return nil, err @@ -506,7 +507,6 @@ afterTrailingSymlink: if mustCreate { return nil, syserror.EEXIST } - child := childVFSD.Impl().(*Dentry) if rp.ShouldFollowSymlink() && child.isSymlink() { targetVD, targetPathname, err := child.inode.Getlink(ctx, rp.Mount()) if err != nil { @@ -533,22 +533,22 @@ afterTrailingSymlink: child.inode.IncRef() defer child.inode.DecRef(ctx) unlock() - return child.inode.Open(ctx, rp, &child.vfsd, opts) + return child.inode.Open(ctx, rp, child, opts) } // ReadlinkAt implements vfs.FilesystemImpl.ReadlinkAt. func (fs *Filesystem) ReadlinkAt(ctx context.Context, rp *vfs.ResolvingPath) (string, error) { fs.mu.RLock() - d, inode, err := fs.walkExistingLocked(ctx, rp) + d, err := fs.walkExistingLocked(ctx, rp) fs.mu.RUnlock() fs.processDeferredDecRefs(ctx) if err != nil { return "", err } - if !d.Impl().(*Dentry).isSymlink() { + if !d.isSymlink() { return "", syserror.EINVAL } - return inode.Readlink(ctx) + return d.inode.Readlink(ctx, rp.Mount()) } // RenameAt implements vfs.FilesystemImpl.RenameAt. @@ -565,11 +565,10 @@ func (fs *Filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa // Resolve the destination directory first to verify that it's on this // Mount. - dstDirVFSD, dstDirInode, err := fs.walkParentDirLocked(ctx, rp) + dstDir, err := fs.walkParentDirLocked(ctx, rp) if err != nil { return err } - dstDir := dstDirVFSD.Impl().(*Dentry) mnt := rp.Mount() if mnt != oldParentVD.Mount() { return syserror.EXDEV @@ -587,16 +586,15 @@ func (fs *Filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa if err != nil { return err } - srcVFSD := &src.vfsd // Can we remove the src dentry? - if err := checkDeleteLocked(ctx, rp, srcVFSD); err != nil { + if err := checkDeleteLocked(ctx, rp, src); err != nil { return err } // Can we create the dst dentry? var dst *Dentry - pc, err := checkCreateLocked(ctx, rp, dstDirVFSD, dstDirInode) + pc, err := checkCreateLocked(ctx, rp, dstDir) switch err { case nil: // Ok, continue with rename as replacement. @@ -607,14 +605,14 @@ func (fs *Filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa } dst = dstDir.children[pc] if dst == nil { - panic(fmt.Sprintf("Child %q for parent Dentry %+v disappeared inside atomic section?", pc, dstDirVFSD)) + panic(fmt.Sprintf("Child %q for parent Dentry %+v disappeared inside atomic section?", pc, dstDir)) } default: return err } var dstVFSD *vfs.Dentry if dst != nil { - dstVFSD = &dst.vfsd + dstVFSD = dst.VFSDentry() } mntns := vfs.MountNamespaceFromContext(ctx) @@ -630,17 +628,18 @@ func (fs *Filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa defer dstDir.dirMu.Unlock() } + srcVFSD := src.VFSDentry() if err := virtfs.PrepareRenameDentry(mntns, srcVFSD, dstVFSD); err != nil { return err } - replaced, err := srcDir.inode.Rename(ctx, src.name, pc, srcVFSD, dstDirVFSD) + replaced, err := srcDir.inode.Rename(ctx, src.name, pc, src, dstDir) if err != nil { virtfs.AbortRenameDentry(srcVFSD, dstVFSD) return err } delete(srcDir.children, src.name) if srcDir != dstDir { - fs.deferDecRef(srcDirVFSD) + fs.deferDecRef(srcDir) dstDir.IncRef() } src.parent = dstDir @@ -649,7 +648,11 @@ func (fs *Filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa dstDir.children = make(map[string]*Dentry) } dstDir.children[pc] = src - virtfs.CommitRenameReplaceDentry(ctx, srcVFSD, replaced) + var replaceVFSD *vfs.Dentry + if replaced != nil { + replaceVFSD = replaced.VFSDentry() + } + virtfs.CommitRenameReplaceDentry(ctx, srcVFSD, replaceVFSD) return nil } @@ -657,7 +660,8 @@ 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() - vfsd, inode, err := fs.walkExistingLocked(ctx, rp) + + d, err := fs.walkExistingLocked(ctx, rp) fs.processDeferredDecRefsLocked(ctx) if err != nil { return err @@ -666,14 +670,13 @@ func (fs *Filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error return err } defer rp.Mount().EndWrite() - if err := checkDeleteLocked(ctx, rp, vfsd); err != nil { + if err := checkDeleteLocked(ctx, rp, d); err != nil { return err } - d := vfsd.Impl().(*Dentry) if !d.isDir() { return syserror.ENOTDIR } - if inode.HasChildren() { + if d.inode.HasChildren() { return syserror.ENOTEMPTY } virtfs := rp.VirtualFilesystem() @@ -683,10 +686,12 @@ func (fs *Filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error mntns := vfs.MountNamespaceFromContext(ctx) defer mntns.DecRef(ctx) + vfsd := d.VFSDentry() if err := virtfs.PrepareDeleteDentry(mntns, vfsd); err != nil { return err } - if err := parentDentry.inode.RmDir(ctx, rp.Component(), vfsd); err != nil { + + if err := parentDentry.inode.RmDir(ctx, d.name, d); err != nil { virtfs.AbortDeleteDentry(vfsd) return err } @@ -697,7 +702,7 @@ func (fs *Filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error // SetStatAt implements vfs.FilesystemImpl.SetStatAt. func (fs *Filesystem) SetStatAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetStatOptions) error { fs.mu.RLock() - _, inode, err := fs.walkExistingLocked(ctx, rp) + d, err := fs.walkExistingLocked(ctx, rp) fs.mu.RUnlock() fs.processDeferredDecRefs(ctx) if err != nil { @@ -706,31 +711,31 @@ func (fs *Filesystem) SetStatAt(ctx context.Context, rp *vfs.ResolvingPath, opts if opts.Stat.Mask == 0 { return nil } - return inode.SetStat(ctx, fs.VFSFilesystem(), rp.Credentials(), opts) + return d.inode.SetStat(ctx, fs.VFSFilesystem(), rp.Credentials(), opts) } // StatAt implements vfs.FilesystemImpl.StatAt. func (fs *Filesystem) StatAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.StatOptions) (linux.Statx, error) { fs.mu.RLock() - _, inode, err := fs.walkExistingLocked(ctx, rp) + d, err := fs.walkExistingLocked(ctx, rp) fs.mu.RUnlock() fs.processDeferredDecRefs(ctx) if err != nil { return linux.Statx{}, err } - return inode.Stat(ctx, fs.VFSFilesystem(), opts) + return d.inode.Stat(ctx, fs.VFSFilesystem(), opts) } // StatFSAt implements vfs.FilesystemImpl.StatFSAt. func (fs *Filesystem) StatFSAt(ctx context.Context, rp *vfs.ResolvingPath) (linux.Statfs, error) { fs.mu.RLock() - _, inode, err := fs.walkExistingLocked(ctx, rp) + d, err := fs.walkExistingLocked(ctx, rp) fs.mu.RUnlock() fs.processDeferredDecRefs(ctx) if err != nil { return linux.Statfs{}, err } - return inode.StatFS(ctx, fs.VFSFilesystem()) + return d.inode.StatFS(ctx, fs.VFSFilesystem()) } // SymlinkAt implements vfs.FilesystemImpl.SymlinkAt. @@ -740,12 +745,15 @@ func (fs *Filesystem) SymlinkAt(ctx context.Context, rp *vfs.ResolvingPath, targ } fs.mu.Lock() defer fs.mu.Unlock() - parentVFSD, parentInode, err := fs.walkParentDirLocked(ctx, rp) + parent, err := fs.walkParentDirLocked(ctx, rp) fs.processDeferredDecRefsLocked(ctx) if err != nil { return err } - pc, err := checkCreateLocked(ctx, rp, parentVFSD, parentInode) + parent.dirMu.Lock() + defer parent.dirMu.Unlock() + + pc, err := checkCreateLocked(ctx, rp, parent) if err != nil { return err } @@ -753,11 +761,11 @@ func (fs *Filesystem) SymlinkAt(ctx context.Context, rp *vfs.ResolvingPath, targ return err } defer rp.Mount().EndWrite() - childVFSD, err := parentInode.NewSymlink(ctx, pc, target) + child, err := parent.inode.NewSymlink(ctx, pc, target) if err != nil { return err } - parentVFSD.Impl().(*Dentry).InsertChild(pc, childVFSD.Impl().(*Dentry)) + parent.InsertChildLocked(pc, child) return nil } @@ -765,7 +773,8 @@ 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() - vfsd, _, err := fs.walkExistingLocked(ctx, rp) + + d, err := fs.walkExistingLocked(ctx, rp) fs.processDeferredDecRefsLocked(ctx) if err != nil { return err @@ -774,10 +783,9 @@ func (fs *Filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error return err } defer rp.Mount().EndWrite() - if err := checkDeleteLocked(ctx, rp, vfsd); err != nil { + if err := checkDeleteLocked(ctx, rp, d); err != nil { return err } - d := vfsd.Impl().(*Dentry) if d.isDir() { return syserror.EISDIR } @@ -787,10 +795,11 @@ func (fs *Filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error defer parentDentry.dirMu.Unlock() mntns := vfs.MountNamespaceFromContext(ctx) defer mntns.DecRef(ctx) + vfsd := d.VFSDentry() if err := virtfs.PrepareDeleteDentry(mntns, vfsd); err != nil { return err } - if err := parentDentry.inode.Unlink(ctx, rp.Component(), vfsd); err != nil { + if err := parentDentry.inode.Unlink(ctx, d.name, d); err != nil { virtfs.AbortDeleteDentry(vfsd) return err } @@ -798,16 +807,16 @@ func (fs *Filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error return nil } -// BoundEndpointAt implements FilesystemImpl.BoundEndpointAt. +// BoundEndpointAt implements vfs.FilesystemImpl.BoundEndpointAt. func (fs *Filesystem) BoundEndpointAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.BoundEndpointOptions) (transport.BoundEndpoint, error) { fs.mu.RLock() - _, inode, err := fs.walkExistingLocked(ctx, rp) + d, err := fs.walkExistingLocked(ctx, rp) fs.mu.RUnlock() fs.processDeferredDecRefs(ctx) if err != nil { return nil, err } - if err := inode.CheckPermissions(ctx, rp.Credentials(), vfs.MayWrite); err != nil { + if err := d.inode.CheckPermissions(ctx, rp.Credentials(), vfs.MayWrite); err != nil { return nil, err } return nil, syserror.ECONNREFUSED @@ -816,7 +825,7 @@ func (fs *Filesystem) BoundEndpointAt(ctx context.Context, rp *vfs.ResolvingPath // ListXattrAt implements vfs.FilesystemImpl.ListXattrAt. func (fs *Filesystem) ListXattrAt(ctx context.Context, rp *vfs.ResolvingPath, size uint64) ([]string, error) { fs.mu.RLock() - _, _, err := fs.walkExistingLocked(ctx, rp) + _, err := fs.walkExistingLocked(ctx, rp) fs.mu.RUnlock() fs.processDeferredDecRefs(ctx) if err != nil { @@ -829,7 +838,7 @@ func (fs *Filesystem) ListXattrAt(ctx context.Context, rp *vfs.ResolvingPath, si // GetXattrAt implements vfs.FilesystemImpl.GetXattrAt. func (fs *Filesystem) GetXattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.GetXattrOptions) (string, error) { fs.mu.RLock() - _, _, err := fs.walkExistingLocked(ctx, rp) + _, err := fs.walkExistingLocked(ctx, rp) fs.mu.RUnlock() fs.processDeferredDecRefs(ctx) if err != nil { @@ -842,7 +851,7 @@ func (fs *Filesystem) GetXattrAt(ctx context.Context, rp *vfs.ResolvingPath, opt // SetXattrAt implements vfs.FilesystemImpl.SetXattrAt. func (fs *Filesystem) SetXattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetXattrOptions) error { fs.mu.RLock() - _, _, err := fs.walkExistingLocked(ctx, rp) + _, err := fs.walkExistingLocked(ctx, rp) fs.mu.RUnlock() fs.processDeferredDecRefs(ctx) if err != nil { @@ -855,7 +864,7 @@ func (fs *Filesystem) SetXattrAt(ctx context.Context, rp *vfs.ResolvingPath, opt // RemoveXattrAt implements vfs.FilesystemImpl.RemoveXattrAt. func (fs *Filesystem) RemoveXattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) error { fs.mu.RLock() - _, _, err := fs.walkExistingLocked(ctx, rp) + _, err := fs.walkExistingLocked(ctx, rp) fs.mu.RUnlock() fs.processDeferredDecRefs(ctx) if err != nil { diff --git a/pkg/sentry/fsimpl/kernfs/inode_impl_util.go b/pkg/sentry/fsimpl/kernfs/inode_impl_util.go index 74408e322..49210e748 100644 --- a/pkg/sentry/fsimpl/kernfs/inode_impl_util.go +++ b/pkg/sentry/fsimpl/kernfs/inode_impl_util.go @@ -31,6 +31,8 @@ import ( // count for inodes, performing no extra actions when references are obtained or // released. This is suitable for simple file inodes that don't reference any // resources. +// +// +stateify savable type InodeNoopRefCount struct { } @@ -50,30 +52,32 @@ func (InodeNoopRefCount) TryIncRef() bool { // InodeDirectoryNoNewChildren partially implements the Inode interface. // InodeDirectoryNoNewChildren represents a directory inode which does not // support creation of new children. +// +// +stateify savable type InodeDirectoryNoNewChildren struct{} // NewFile implements Inode.NewFile. -func (InodeDirectoryNoNewChildren) NewFile(context.Context, string, vfs.OpenOptions) (*vfs.Dentry, error) { +func (InodeDirectoryNoNewChildren) NewFile(context.Context, string, vfs.OpenOptions) (*Dentry, error) { return nil, syserror.EPERM } // NewDir implements Inode.NewDir. -func (InodeDirectoryNoNewChildren) NewDir(context.Context, string, vfs.MkdirOptions) (*vfs.Dentry, error) { +func (InodeDirectoryNoNewChildren) NewDir(context.Context, string, vfs.MkdirOptions) (*Dentry, error) { return nil, syserror.EPERM } // NewLink implements Inode.NewLink. -func (InodeDirectoryNoNewChildren) NewLink(context.Context, string, Inode) (*vfs.Dentry, error) { +func (InodeDirectoryNoNewChildren) NewLink(context.Context, string, Inode) (*Dentry, error) { return nil, syserror.EPERM } // NewSymlink implements Inode.NewSymlink. -func (InodeDirectoryNoNewChildren) NewSymlink(context.Context, string, string) (*vfs.Dentry, error) { +func (InodeDirectoryNoNewChildren) NewSymlink(context.Context, string, string) (*Dentry, error) { return nil, syserror.EPERM } // NewNode implements Inode.NewNode. -func (InodeDirectoryNoNewChildren) NewNode(context.Context, string, vfs.MknodOptions) (*vfs.Dentry, error) { +func (InodeDirectoryNoNewChildren) NewNode(context.Context, string, vfs.MknodOptions) (*Dentry, error) { return nil, syserror.EPERM } @@ -81,6 +85,8 @@ func (InodeDirectoryNoNewChildren) NewNode(context.Context, string, vfs.MknodOpt // inodeDirectory and inodeDynamicDirectory sub interfaces. Inodes that do not // represent directories can embed this to provide no-op implementations for // directory-related functions. +// +// +stateify savable type InodeNotDirectory struct { } @@ -90,47 +96,47 @@ func (InodeNotDirectory) HasChildren() bool { } // NewFile implements Inode.NewFile. -func (InodeNotDirectory) NewFile(context.Context, string, vfs.OpenOptions) (*vfs.Dentry, error) { +func (InodeNotDirectory) NewFile(context.Context, string, vfs.OpenOptions) (*Dentry, error) { panic("NewFile called on non-directory inode") } // NewDir implements Inode.NewDir. -func (InodeNotDirectory) NewDir(context.Context, string, vfs.MkdirOptions) (*vfs.Dentry, error) { +func (InodeNotDirectory) NewDir(context.Context, string, vfs.MkdirOptions) (*Dentry, error) { panic("NewDir called on non-directory inode") } // NewLink implements Inode.NewLinkink. -func (InodeNotDirectory) NewLink(context.Context, string, Inode) (*vfs.Dentry, error) { +func (InodeNotDirectory) NewLink(context.Context, string, Inode) (*Dentry, error) { panic("NewLink called on non-directory inode") } // NewSymlink implements Inode.NewSymlink. -func (InodeNotDirectory) NewSymlink(context.Context, string, string) (*vfs.Dentry, error) { +func (InodeNotDirectory) NewSymlink(context.Context, string, string) (*Dentry, error) { panic("NewSymlink called on non-directory inode") } // NewNode implements Inode.NewNode. -func (InodeNotDirectory) NewNode(context.Context, string, vfs.MknodOptions) (*vfs.Dentry, error) { +func (InodeNotDirectory) NewNode(context.Context, string, vfs.MknodOptions) (*Dentry, error) { panic("NewNode called on non-directory inode") } // Unlink implements Inode.Unlink. -func (InodeNotDirectory) Unlink(context.Context, string, *vfs.Dentry) error { +func (InodeNotDirectory) Unlink(context.Context, string, *Dentry) error { panic("Unlink called on non-directory inode") } // RmDir implements Inode.RmDir. -func (InodeNotDirectory) RmDir(context.Context, string, *vfs.Dentry) error { +func (InodeNotDirectory) RmDir(context.Context, string, *Dentry) error { panic("RmDir called on non-directory inode") } // Rename implements Inode.Rename. -func (InodeNotDirectory) Rename(context.Context, string, string, *vfs.Dentry, *vfs.Dentry) (*vfs.Dentry, error) { +func (InodeNotDirectory) Rename(context.Context, string, string, *Dentry, *Dentry) (*Dentry, error) { panic("Rename called on non-directory inode") } // Lookup implements Inode.Lookup. -func (InodeNotDirectory) Lookup(ctx context.Context, name string) (*vfs.Dentry, error) { +func (InodeNotDirectory) Lookup(ctx context.Context, name string) (*Dentry, error) { panic("Lookup called on non-directory inode") } @@ -149,10 +155,12 @@ func (InodeNotDirectory) Valid(context.Context) bool { // dymanic entries (i.e. entries that are not "hashed" into the // vfs.Dentry.children) can embed this to provide no-op implementations for // functions related to dynamic entries. +// +// +stateify savable type InodeNoDynamicLookup struct{} // Lookup implements Inode.Lookup. -func (InodeNoDynamicLookup) Lookup(ctx context.Context, name string) (*vfs.Dentry, error) { +func (InodeNoDynamicLookup) Lookup(ctx context.Context, name string) (*Dentry, error) { return nil, syserror.ENOENT } @@ -169,10 +177,12 @@ func (InodeNoDynamicLookup) Valid(ctx context.Context) bool { // InodeNotSymlink partially implements the Inode interface, specifically the // inodeSymlink sub interface. All inodes that are not symlinks may embed this // to return the appropriate errors from symlink-related functions. +// +// +stateify savable type InodeNotSymlink struct{} // Readlink implements Inode.Readlink. -func (InodeNotSymlink) Readlink(context.Context) (string, error) { +func (InodeNotSymlink) Readlink(context.Context, *vfs.Mount) (string, error) { return "", syserror.EINVAL } @@ -186,6 +196,8 @@ func (InodeNotSymlink) Getlink(context.Context, *vfs.Mount) (vfs.VirtualDentry, // inode attributes. // // Must be initialized by Init prior to first use. +// +// +stateify savable type InodeAttrs struct { devMajor uint32 devMinor uint32 @@ -256,6 +268,13 @@ func (a *InodeAttrs) Stat(context.Context, *vfs.Filesystem, vfs.StatOptions) (li // SetStat implements Inode.SetStat. func (a *InodeAttrs) SetStat(ctx context.Context, fs *vfs.Filesystem, creds *auth.Credentials, opts vfs.SetStatOptions) error { + return a.SetInodeStat(ctx, fs, creds, opts) +} + +// SetInodeStat sets the corresponding attributes from opts to InodeAttrs. +// This function can be used by other kernfs-based filesystem implementation to +// sets the unexported attributes into kernfs.InodeAttrs. +func (a *InodeAttrs) SetInodeStat(ctx context.Context, fs *vfs.Filesystem, creds *auth.Credentials, opts vfs.SetStatOptions) error { if opts.Stat.Mask == 0 { return nil } @@ -323,13 +342,16 @@ func (a *InodeAttrs) DecLinks() { } } +// +stateify savable type slot struct { Name string - Dentry *vfs.Dentry + Dentry *Dentry slotEntry } // OrderedChildrenOptions contains initialization options for OrderedChildren. +// +// +stateify savable type OrderedChildrenOptions struct { // Writable indicates whether vfs.FilesystemImpl methods implemented by // OrderedChildren may modify the tracked children. This applies to @@ -345,12 +367,14 @@ type OrderedChildrenOptions struct { // directories. // // Must be initialize with Init before first use. +// +// +stateify savable type OrderedChildren struct { // Can children be modified by user syscalls? It set to false, interface // methods that would modify the children return EPERM. Immutable. writable bool - mu sync.RWMutex + mu sync.RWMutex `state:"nosave"` order slotList set map[string]*slot } @@ -383,7 +407,7 @@ func (o *OrderedChildren) Populate(d *Dentry, children map[string]*Dentry) uint3 if child.isDir() { links++ } - if err := o.Insert(name, child.VFSDentry()); err != nil { + if err := o.Insert(name, child); err != nil { panic(fmt.Sprintf("Collision when attempting to insert child %q (%+v) into %+v", name, child, d)) } d.InsertChild(name, child) @@ -400,7 +424,7 @@ func (o *OrderedChildren) HasChildren() bool { // Insert inserts child into o. This ignores the writability of o, as this is // not part of the vfs.FilesystemImpl interface, and is a lower-level operation. -func (o *OrderedChildren) Insert(name string, child *vfs.Dentry) error { +func (o *OrderedChildren) Insert(name string, child *Dentry) error { o.mu.Lock() defer o.mu.Unlock() if _, ok := o.set[name]; ok { @@ -424,10 +448,10 @@ func (o *OrderedChildren) removeLocked(name string) { } // Precondition: caller must hold o.mu for writing. -func (o *OrderedChildren) replaceChildLocked(name string, new *vfs.Dentry) *vfs.Dentry { +func (o *OrderedChildren) replaceChildLocked(name string, new *Dentry) *Dentry { if s, ok := o.set[name]; ok { // Existing slot with given name, simply replace the dentry. - var old *vfs.Dentry + var old *Dentry old, s.Dentry = s.Dentry, new return old } @@ -443,7 +467,7 @@ func (o *OrderedChildren) replaceChildLocked(name string, new *vfs.Dentry) *vfs. } // Precondition: caller must hold o.mu for reading or writing. -func (o *OrderedChildren) checkExistingLocked(name string, child *vfs.Dentry) error { +func (o *OrderedChildren) checkExistingLocked(name string, child *Dentry) error { s, ok := o.set[name] if !ok { return syserror.ENOENT @@ -455,7 +479,7 @@ func (o *OrderedChildren) checkExistingLocked(name string, child *vfs.Dentry) er } // Unlink implements Inode.Unlink. -func (o *OrderedChildren) Unlink(ctx context.Context, name string, child *vfs.Dentry) error { +func (o *OrderedChildren) Unlink(ctx context.Context, name string, child *Dentry) error { if !o.writable { return syserror.EPERM } @@ -471,12 +495,13 @@ func (o *OrderedChildren) Unlink(ctx context.Context, name string, child *vfs.De } // Rmdir implements Inode.Rmdir. -func (o *OrderedChildren) RmDir(ctx context.Context, name string, child *vfs.Dentry) error { +func (o *OrderedChildren) RmDir(ctx context.Context, name string, child *Dentry) error { // We're not responsible for checking that child is a directory, that it's // empty, or updating any link counts; so this is the same as unlink. return o.Unlink(ctx, name, child) } +// +stateify savable type renameAcrossDifferentImplementationsError struct{} func (renameAcrossDifferentImplementationsError) Error() string { @@ -492,8 +517,8 @@ func (renameAcrossDifferentImplementationsError) Error() string { // that will support Rename. // // Postcondition: reference on any replaced dentry transferred to caller. -func (o *OrderedChildren) Rename(ctx context.Context, oldname, newname string, child, dstDir *vfs.Dentry) (*vfs.Dentry, error) { - dst, ok := dstDir.Impl().(*Dentry).inode.(interface{}).(*OrderedChildren) +func (o *OrderedChildren) Rename(ctx context.Context, oldname, newname string, child, dstDir *Dentry) (*Dentry, error) { + dst, ok := dstDir.inode.(interface{}).(*OrderedChildren) if !ok { return nil, renameAcrossDifferentImplementationsError{} } @@ -535,12 +560,14 @@ func (o *OrderedChildren) nthLocked(i int64) *slot { } // InodeSymlink partially implements Inode interface for symlinks. +// +// +stateify savable type InodeSymlink struct { InodeNotDirectory } // Open implements Inode.Open. -func (InodeSymlink) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { +func (InodeSymlink) Open(ctx context.Context, rp *vfs.ResolvingPath, d *Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { return nil, syserror.ELOOP } @@ -567,6 +594,7 @@ var _ Inode = (*StaticDirectory)(nil) func NewStaticDir(creds *auth.Credentials, devMajor, devMinor uint32, ino uint64, perm linux.FileMode, children map[string]*Dentry, fdOpts GenericDirectoryFDOptions) *Dentry { inode := &StaticDirectory{} inode.Init(creds, devMajor, devMinor, ino, perm, fdOpts) + inode.EnableLeakCheck() dentry := &Dentry{} dentry.Init(inode) @@ -587,9 +615,9 @@ func (s *StaticDirectory) Init(creds *auth.Credentials, devMajor, devMinor uint3 s.InodeAttrs.Init(creds, devMajor, devMinor, ino, linux.ModeDirectory|perm) } -// Open implements kernfs.Inode. -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, &s.locks, &opts, s.fdOpts) +// Open implements kernfs.Inode.Open. +func (s *StaticDirectory) Open(ctx context.Context, rp *vfs.ResolvingPath, d *Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd, err := NewGenericDirectoryFD(rp.Mount(), d, &s.OrderedChildren, &s.locks, &opts, s.fdOpts) if err != nil { return nil, err } @@ -601,21 +629,25 @@ func (*StaticDirectory) SetStat(context.Context, *vfs.Filesystem, *auth.Credenti return syserror.EPERM } -// DecRef implements kernfs.Inode. +// DecRef implements kernfs.Inode.DecRef. func (s *StaticDirectory) DecRef(context.Context) { s.StaticDirectoryRefs.DecRef(s.Destroy) } // AlwaysValid partially implements kernfs.inodeDynamicLookup. +// +// +stateify savable type AlwaysValid struct{} -// Valid implements kernfs.inodeDynamicLookup. +// Valid implements kernfs.inodeDynamicLookup.Valid. func (*AlwaysValid) Valid(context.Context) bool { return true } // InodeNoStatFS partially implements the Inode interface, where the client // filesystem doesn't support statfs(2). +// +// +stateify savable type InodeNoStatFS struct{} // StatFS implements Inode.StatFS. diff --git a/pkg/sentry/fsimpl/kernfs/kernfs.go b/pkg/sentry/fsimpl/kernfs/kernfs.go index 88fcd54aa..6d3d79333 100644 --- a/pkg/sentry/fsimpl/kernfs/kernfs.go +++ b/pkg/sentry/fsimpl/kernfs/kernfs.go @@ -29,7 +29,7 @@ // // Reference Model: // -// Kernfs dentries represents named pointers to inodes. Dentries and inode have +// Kernfs dentries represents named pointers to inodes. Dentries and inodes have // independent lifetimes and reference counts. A child dentry unconditionally // holds a reference on its parent directory's dentry. A dentry also holds a // reference on the inode it points to. Multiple dentries can point to the same @@ -60,20 +60,23 @@ import ( "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/pkg/syserror" ) // Filesystem mostly implements vfs.FilesystemImpl for a generic in-memory // filesystem. Concrete implementations are expected to embed this in their own // Filesystem type. +// +// +stateify savable type Filesystem struct { vfsfs vfs.Filesystem - droppedDentriesMu sync.Mutex + droppedDentriesMu sync.Mutex `state:"nosave"` // droppedDentries is a list of dentries waiting to be DecRef()ed. This is // used to defer dentry destruction until mu can be acquired for // writing. Protected by droppedDentriesMu. - droppedDentries []*vfs.Dentry + droppedDentries []*Dentry // mu synchronizes the lifetime of Dentries on this filesystem. Holding it // for reading guarantees continued existence of any resolved dentries, but @@ -96,7 +99,7 @@ type Filesystem struct { // defer fs.mu.RUnlock() // ... // fs.deferDecRef(dentry) - mu sync.RWMutex + mu sync.RWMutex `state:"nosave"` // nextInoMinusOne is used to to allocate inode numbers on this // filesystem. Must be accessed by atomic operations. @@ -107,7 +110,7 @@ type Filesystem struct { // processDeferredDecRefs{,Locked}. See comment on Filesystem.mu. // // Precondition: d must not already be pending destruction. -func (fs *Filesystem) deferDecRef(d *vfs.Dentry) { +func (fs *Filesystem) deferDecRef(d *Dentry) { fs.droppedDentriesMu.Lock() fs.droppedDentries = append(fs.droppedDentries, d) fs.droppedDentriesMu.Unlock() @@ -159,6 +162,8 @@ const ( // to, and child dentries hold a reference on their parent. // // Must be initialized by Init prior to first use. +// +// +stateify savable type Dentry struct { DentryRefs @@ -172,7 +177,11 @@ type Dentry struct { name string // dirMu protects children and the names of child Dentries. - dirMu sync.Mutex + // + // Note that holding fs.mu for writing is not sufficient; + // revalidateChildLocked(), which is a very hot path, may modify children with + // fs.mu acquired for reading only. + dirMu sync.Mutex `state:"nosave"` children map[string]*Dentry inode Inode @@ -239,24 +248,25 @@ func (d *Dentry) Watches() *vfs.Watches { func (d *Dentry) OnZeroWatches(context.Context) {} // InsertChild inserts child into the vfs dentry cache with the given name under -// this dentry. This does not update the directory inode, so calling this on -// its own isn't sufficient to insert a child into a directory. InsertChild -// updates the link count on d if required. +// this dentry. This does not update the directory inode, so calling this on its +// own isn't sufficient to insert a child into a directory. // // Precondition: d must represent a directory inode. func (d *Dentry) InsertChild(name string, child *Dentry) { d.dirMu.Lock() - d.insertChildLocked(name, child) + d.InsertChildLocked(name, child) d.dirMu.Unlock() } -// insertChildLocked is equivalent to InsertChild, with additional +// InsertChildLocked is equivalent to InsertChild, with additional // preconditions. // -// Precondition: d.dirMu must be locked. -func (d *Dentry) insertChildLocked(name string, child *Dentry) { +// Preconditions: +// * d must represent a directory inode. +// * d.dirMu must be locked. +func (d *Dentry) InsertChildLocked(name string, child *Dentry) { if !d.isDir() { - panic(fmt.Sprintf("InsertChild called on non-directory Dentry: %+v.", d)) + panic(fmt.Sprintf("InsertChildLocked called on non-directory Dentry: %+v.", d)) } d.IncRef() // DecRef in child's Dentry.destroy. child.parent = d @@ -267,6 +277,36 @@ func (d *Dentry) insertChildLocked(name string, child *Dentry) { d.children[name] = child } +// RemoveChild removes child from the vfs dentry cache. This does not update the +// directory inode or modify the inode to be unlinked. So calling this on its own +// isn't sufficient to remove a child from a directory. +// +// Precondition: d must represent a directory inode. +func (d *Dentry) RemoveChild(name string, child *Dentry) error { + d.dirMu.Lock() + defer d.dirMu.Unlock() + return d.RemoveChildLocked(name, child) +} + +// RemoveChildLocked is equivalent to RemoveChild, with additional +// preconditions. +// +// Precondition: d.dirMu must be locked. +func (d *Dentry) RemoveChildLocked(name string, child *Dentry) error { + if !d.isDir() { + panic(fmt.Sprintf("RemoveChild called on non-directory Dentry: %+v.", d)) + } + c, ok := d.children[name] + if !ok { + return syserror.ENOENT + } + if c != child { + panic(fmt.Sprintf("Dentry hashed into inode doesn't match what vfs thinks! Child: %+v, vfs: %+v", c, child)) + } + delete(d.children, name) + return nil +} + // Inode returns the dentry's inode. func (d *Dentry) Inode() Inode { return d.inode @@ -287,7 +327,6 @@ func (d *Dentry) Inode() Inode { // // - Checking that dentries passed to methods are of the appropriate file type. // - Checking permissions. -// - Updating link and reference counts. // // Specific responsibilities of implementations are documented below. type Inode interface { @@ -297,7 +336,8 @@ type Inode interface { inodeRefs // Methods related to node metadata. A generic implementation is provided by - // InodeAttrs. + // InodeAttrs. Note that a concrete filesystem using kernfs is responsible for + // managing link counts. inodeMetadata // Method for inodes that represent symlink. InodeNotSymlink provides a @@ -315,11 +355,11 @@ type Inode interface { // Open creates a file description for the filesystem object represented by // this inode. The returned file description should hold a reference on the - // inode for its lifetime. + // dentry for its lifetime. // // Precondition: rp.Done(). vfsd.Impl() must be the kernfs Dentry containing // the inode on which Open() is being called. - Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) + Open(ctx context.Context, rp *vfs.ResolvingPath, d *Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) // StatFS returns filesystem statistics for the client filesystem. This // corresponds to vfs.FilesystemImpl.StatFSAt. If the client filesystem @@ -369,30 +409,30 @@ type inodeDirectory interface { HasChildren() bool // NewFile creates a new regular file inode. - NewFile(ctx context.Context, name string, opts vfs.OpenOptions) (*vfs.Dentry, error) + NewFile(ctx context.Context, name string, opts vfs.OpenOptions) (*Dentry, error) // NewDir creates a new directory inode. - NewDir(ctx context.Context, name string, opts vfs.MkdirOptions) (*vfs.Dentry, error) + NewDir(ctx context.Context, name string, opts vfs.MkdirOptions) (*Dentry, error) // NewLink creates a new hardlink to a specified inode in this // directory. Implementations should create a new kernfs Dentry pointing to // target, and update target's link count. - NewLink(ctx context.Context, name string, target Inode) (*vfs.Dentry, error) + NewLink(ctx context.Context, name string, target Inode) (*Dentry, error) // NewSymlink creates a new symbolic link inode. - NewSymlink(ctx context.Context, name, target string) (*vfs.Dentry, error) + NewSymlink(ctx context.Context, name, target string) (*Dentry, error) // NewNode creates a new filesystem node for a mknod syscall. - NewNode(ctx context.Context, name string, opts vfs.MknodOptions) (*vfs.Dentry, error) + NewNode(ctx context.Context, name string, opts vfs.MknodOptions) (*Dentry, error) // Unlink removes a child dentry from this directory inode. - Unlink(ctx context.Context, name string, child *vfs.Dentry) error + Unlink(ctx context.Context, name string, child *Dentry) error // RmDir removes an empty child directory from this directory // inode. Implementations must update the parent directory's link count, // if required. Implementations are not responsible for checking that child // is a directory, checking for an empty directory. - RmDir(ctx context.Context, name string, child *vfs.Dentry) error + RmDir(ctx context.Context, name string, child *Dentry) error // Rename is called on the source directory containing an inode being // renamed. child should point to the resolved child in the source @@ -400,7 +440,7 @@ type inodeDirectory interface { // should return the replaced dentry or nil otherwise. // // Precondition: Caller must serialize concurrent calls to Rename. - Rename(ctx context.Context, oldname, newname string, child, dstDir *vfs.Dentry) (replaced *vfs.Dentry, err error) + Rename(ctx context.Context, oldname, newname string, child, dstDir *Dentry) (replaced *Dentry, err error) } type inodeDynamicLookup interface { @@ -418,14 +458,14 @@ type inodeDynamicLookup interface { // // Lookup returns the child with an extra reference and the caller owns this // reference. - Lookup(ctx context.Context, name string) (*vfs.Dentry, error) + Lookup(ctx context.Context, name string) (*Dentry, error) // Valid should return true if this inode is still valid, or needs to // be resolved again by a call to Lookup. Valid(ctx context.Context) bool // IterDirents is used to iterate over dynamically created entries. It invokes - // cb on each entry in the directory represented by the FileDescription. + // cb on each entry in the directory represented by the Inode. // 'offset' is the offset for the entire IterDirents call, which may include // results from the caller (e.g. "." and ".."). 'relOffset' is the offset // inside the entries returned by this IterDirents invocation. In other words, @@ -437,7 +477,7 @@ type inodeDynamicLookup interface { type inodeSymlink interface { // 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) + Readlink(ctx context.Context, mnt *vfs.Mount) (string, error) // Getlink returns the target of a symbolic link, as used by path // resolution: diff --git a/pkg/sentry/fsimpl/kernfs/kernfs_test.go b/pkg/sentry/fsimpl/kernfs/kernfs_test.go index 09806a3f2..e413242dc 100644 --- a/pkg/sentry/fsimpl/kernfs/kernfs_test.go +++ b/pkg/sentry/fsimpl/kernfs/kernfs_test.go @@ -121,8 +121,8 @@ func (fs *filesystem) newReadonlyDir(creds *auth.Credentials, mode linux.FileMod return &dir.dentry } -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, &d.locks, &opts, kernfs.GenericDirectoryFDOptions{ +func (d *readonlyDir) Open(ctx context.Context, rp *vfs.ResolvingPath, kd *kernfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), kd, &d.OrderedChildren, &d.locks, &opts, kernfs.GenericDirectoryFDOptions{ SeekEnd: kernfs.SeekEndStaticEntries, }) if err != nil { @@ -162,8 +162,8 @@ func (fs *filesystem) newDir(creds *auth.Credentials, mode linux.FileMode, conte return &dir.dentry } -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, &d.locks, &opts, kernfs.GenericDirectoryFDOptions{ +func (d *dir) Open(ctx context.Context, rp *vfs.ResolvingPath, kd *kernfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), kd, &d.OrderedChildren, &d.locks, &opts, kernfs.GenericDirectoryFDOptions{ SeekEnd: kernfs.SeekEndStaticEntries, }) if err != nil { @@ -176,38 +176,36 @@ func (d *dir) DecRef(context.Context) { d.dirRefs.DecRef(d.Destroy) } -func (d *dir) NewDir(ctx context.Context, name string, opts vfs.MkdirOptions) (*vfs.Dentry, error) { +func (d *dir) NewDir(ctx context.Context, name string, opts vfs.MkdirOptions) (*kernfs.Dentry, error) { creds := auth.CredentialsFromContext(ctx) dir := d.fs.newDir(creds, opts.Mode, nil) - dirVFSD := dir.VFSDentry() - if err := d.OrderedChildren.Insert(name, dirVFSD); err != nil { + if err := d.OrderedChildren.Insert(name, dir); err != nil { dir.DecRef(ctx) return nil, err } d.IncLinks(1) - return dirVFSD, nil + return dir, nil } -func (d *dir) NewFile(ctx context.Context, name string, opts vfs.OpenOptions) (*vfs.Dentry, error) { +func (d *dir) NewFile(ctx context.Context, name string, opts vfs.OpenOptions) (*kernfs.Dentry, error) { creds := auth.CredentialsFromContext(ctx) f := d.fs.newFile(creds, "") - fVFSD := f.VFSDentry() - if err := d.OrderedChildren.Insert(name, fVFSD); err != nil { + if err := d.OrderedChildren.Insert(name, f); err != nil { f.DecRef(ctx) return nil, err } - return fVFSD, nil + return f, nil } -func (*dir) NewLink(context.Context, string, kernfs.Inode) (*vfs.Dentry, error) { +func (*dir) NewLink(context.Context, string, kernfs.Inode) (*kernfs.Dentry, error) { return nil, syserror.EPERM } -func (*dir) NewSymlink(context.Context, string, string) (*vfs.Dentry, error) { +func (*dir) NewSymlink(context.Context, string, string) (*kernfs.Dentry, error) { return nil, syserror.EPERM } -func (*dir) NewNode(context.Context, string, vfs.MknodOptions) (*vfs.Dentry, error) { +func (*dir) NewNode(context.Context, string, vfs.MknodOptions) (*kernfs.Dentry, error) { return nil, syserror.EPERM } diff --git a/pkg/sentry/fsimpl/kernfs/symlink.go b/pkg/sentry/fsimpl/kernfs/symlink.go index 64731a3e4..58a93eaac 100644 --- a/pkg/sentry/fsimpl/kernfs/symlink.go +++ b/pkg/sentry/fsimpl/kernfs/symlink.go @@ -24,6 +24,8 @@ import ( // StaticSymlink provides an Inode implementation for symlinks that point to // a immutable target. +// +// +stateify savable type StaticSymlink struct { InodeAttrs InodeNoopRefCount @@ -51,8 +53,8 @@ func (s *StaticSymlink) Init(creds *auth.Credentials, devMajor uint32, devMinor s.InodeAttrs.Init(creds, devMajor, devMinor, ino, linux.ModeSymlink|0777) } -// Readlink implements Inode. -func (s *StaticSymlink) Readlink(_ context.Context) (string, error) { +// Readlink implements Inode.Readlink. +func (s *StaticSymlink) Readlink(_ context.Context, _ *vfs.Mount) (string, error) { return s.target, nil } diff --git a/pkg/sentry/fsimpl/kernfs/synthetic_directory.go b/pkg/sentry/fsimpl/kernfs/synthetic_directory.go index 01ba72fa8..ea7f073eb 100644 --- a/pkg/sentry/fsimpl/kernfs/synthetic_directory.go +++ b/pkg/sentry/fsimpl/kernfs/synthetic_directory.go @@ -41,12 +41,12 @@ type syntheticDirectory struct { var _ Inode = (*syntheticDirectory)(nil) -func newSyntheticDirectory(creds *auth.Credentials, perm linux.FileMode) *vfs.Dentry { +func newSyntheticDirectory(creds *auth.Credentials, perm linux.FileMode) *Dentry { inode := &syntheticDirectory{} inode.Init(creds, 0 /* devMajor */, 0 /* devMinor */, 0 /* ino */, perm) d := &Dentry{} d.Init(inode) - return &d.vfsd + return d } func (dir *syntheticDirectory) Init(creds *auth.Credentials, devMajor, devMinor uint32, ino uint64, perm linux.FileMode) { @@ -60,8 +60,8 @@ func (dir *syntheticDirectory) Init(creds *auth.Credentials, devMajor, devMinor } // Open implements Inode.Open. -func (dir *syntheticDirectory) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { - fd, err := NewGenericDirectoryFD(rp.Mount(), vfsd, &dir.OrderedChildren, &dir.locks, &opts, GenericDirectoryFDOptions{}) +func (dir *syntheticDirectory) Open(ctx context.Context, rp *vfs.ResolvingPath, d *Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd, err := NewGenericDirectoryFD(rp.Mount(), d, &dir.OrderedChildren, &dir.locks, &opts, GenericDirectoryFDOptions{}) if err != nil { return nil, err } @@ -69,12 +69,12 @@ func (dir *syntheticDirectory) Open(ctx context.Context, rp *vfs.ResolvingPath, } // NewFile implements Inode.NewFile. -func (dir *syntheticDirectory) NewFile(ctx context.Context, name string, opts vfs.OpenOptions) (*vfs.Dentry, error) { +func (dir *syntheticDirectory) NewFile(ctx context.Context, name string, opts vfs.OpenOptions) (*Dentry, error) { return nil, syserror.EPERM } // NewDir implements Inode.NewDir. -func (dir *syntheticDirectory) NewDir(ctx context.Context, name string, opts vfs.MkdirOptions) (*vfs.Dentry, error) { +func (dir *syntheticDirectory) NewDir(ctx context.Context, name string, opts vfs.MkdirOptions) (*Dentry, error) { if !opts.ForSyntheticMountpoint { return nil, syserror.EPERM } @@ -87,16 +87,16 @@ func (dir *syntheticDirectory) NewDir(ctx context.Context, name string, opts vfs } // NewLink implements Inode.NewLink. -func (dir *syntheticDirectory) NewLink(ctx context.Context, name string, target Inode) (*vfs.Dentry, error) { +func (dir *syntheticDirectory) NewLink(ctx context.Context, name string, target Inode) (*Dentry, error) { return nil, syserror.EPERM } // NewSymlink implements Inode.NewSymlink. -func (dir *syntheticDirectory) NewSymlink(ctx context.Context, name, target string) (*vfs.Dentry, error) { +func (dir *syntheticDirectory) NewSymlink(ctx context.Context, name, target string) (*Dentry, error) { return nil, syserror.EPERM } // NewNode implements Inode.NewNode. -func (dir *syntheticDirectory) NewNode(ctx context.Context, name string, opts vfs.MknodOptions) (*vfs.Dentry, error) { +func (dir *syntheticDirectory) NewNode(ctx context.Context, name string, opts vfs.MknodOptions) (*Dentry, error) { return nil, syserror.EPERM } diff --git a/pkg/sentry/fsimpl/overlay/copy_up.go b/pkg/sentry/fsimpl/overlay/copy_up.go index 360b77ef6..73b126669 100644 --- a/pkg/sentry/fsimpl/overlay/copy_up.go +++ b/pkg/sentry/fsimpl/overlay/copy_up.go @@ -92,7 +92,7 @@ func (d *dentry) copyUpLocked(ctx context.Context) error { err = vfsObj.UnlinkAt(ctx, d.fs.creds, &newpop) } if err != nil { - ctx.Warningf("Unrecoverable overlayfs inconsistency: failed to delete upper layer file after copy-up error: %v", err) + panic(fmt.Sprintf("unrecoverable overlayfs inconsistency: failed to delete upper layer file after copy-up error: %v", err)) } if d.upperVD.Ok() { d.upperVD.DecRef(ctx) @@ -362,7 +362,7 @@ func (d *dentry) copyXattrsLocked(ctx context.Context) error { // There are no guarantees as to the contents of lowerXattrs. return nil } - ctx.Warningf("failed to copy up xattrs because ListXattrAt failed: %v", err) + ctx.Infof("failed to copy up xattrs because ListXattrAt failed: %v", err) return err } @@ -374,14 +374,48 @@ func (d *dentry) copyXattrsLocked(ctx context.Context) error { value, err := vfsObj.GetXattrAt(ctx, d.fs.creds, lowerPop, &vfs.GetXattrOptions{Name: name, Size: 0}) if err != nil { - ctx.Warningf("failed to copy up xattrs because GetXattrAt failed: %v", err) + ctx.Infof("failed to copy up xattrs because GetXattrAt failed: %v", err) return err } if err := vfsObj.SetXattrAt(ctx, d.fs.creds, upperPop, &vfs.SetXattrOptions{Name: name, Value: value}); err != nil { - ctx.Warningf("failed to copy up xattrs because SetXattrAt failed: %v", err) + ctx.Infof("failed to copy up xattrs because SetXattrAt failed: %v", err) return err } } return nil } + +// copyUpDescendantsLocked ensures that all descendants of d are copied up. +// +// Preconditions: +// * filesystem.renameMu must be locked. +// * d.dirMu must be locked. +// * d.isDir(). +func (d *dentry) copyUpDescendantsLocked(ctx context.Context, ds **[]*dentry) error { + dirents, err := d.getDirentsLocked(ctx) + if err != nil { + return err + } + for _, dirent := range dirents { + if dirent.Name == "." || dirent.Name == ".." { + continue + } + child, err := d.fs.getChildLocked(ctx, d, dirent.Name, ds) + if err != nil { + return err + } + if err := child.copyUpLocked(ctx); err != nil { + return err + } + if child.isDir() { + child.dirMu.Lock() + err := child.copyUpDescendantsLocked(ctx, ds) + child.dirMu.Unlock() + if err != nil { + return err + } + } + } + return nil +} diff --git a/pkg/sentry/fsimpl/overlay/directory.go b/pkg/sentry/fsimpl/overlay/directory.go index b1b292e83..df4492346 100644 --- a/pkg/sentry/fsimpl/overlay/directory.go +++ b/pkg/sentry/fsimpl/overlay/directory.go @@ -100,12 +100,13 @@ func (d *dentry) collectWhiteoutsForRmdirLocked(ctx context.Context) (map[string return whiteouts, readdirErr } +// +stateify savable type directoryFD struct { fileDescription vfs.DirectoryFileDescriptionDefaultImpl vfs.DentryMetadataFileDescriptionImpl - mu sync.Mutex + mu sync.Mutex `state:"nosave"` off int64 dirents []vfs.Dirent } @@ -116,10 +117,12 @@ func (fd *directoryFD) Release(ctx context.Context) { // IterDirents implements vfs.FileDescriptionImpl.IterDirents. func (fd *directoryFD) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback) error { + d := fd.dentry() + defer d.InotifyWithParent(ctx, linux.IN_ACCESS, 0, vfs.PathEvent) + fd.mu.Lock() defer fd.mu.Unlock() - d := fd.dentry() if fd.dirents == nil { ds, err := d.getDirents(ctx) if err != nil { @@ -143,7 +146,14 @@ func (d *dentry) getDirents(ctx context.Context) ([]vfs.Dirent, error) { defer d.fs.renameMu.RUnlock() d.dirMu.Lock() defer d.dirMu.Unlock() + return d.getDirentsLocked(ctx) +} +// Preconditions: +// * filesystem.renameMu must be locked. +// * d.dirMu must be locked. +// * d.isDir(). +func (d *dentry) getDirentsLocked(ctx context.Context) ([]vfs.Dirent, error) { if d.dirents != nil { return d.dirents, nil } diff --git a/pkg/sentry/fsimpl/overlay/filesystem.go b/pkg/sentry/fsimpl/overlay/filesystem.go index b530851c5..bd11372d5 100644 --- a/pkg/sentry/fsimpl/overlay/filesystem.go +++ b/pkg/sentry/fsimpl/overlay/filesystem.go @@ -15,6 +15,7 @@ package overlay import ( + "fmt" "strings" "sync/atomic" @@ -498,7 +499,13 @@ func (fs *filesystem) doCreateAt(ctx context.Context, rp *vfs.ResolvingPath, dir if err := create(parent, name, childLayer == lookupLayerUpperWhiteout); err != nil { return err } + parent.dirents = nil + ev := linux.IN_CREATE + if dir { + ev |= linux.IN_ISDIR + } + parent.watches.Notify(ctx, name, uint32(ev), 0 /* cookie */, vfs.InodeEvent, false /* unlinked */) return nil } @@ -512,7 +519,7 @@ func (fs *filesystem) createWhiteout(ctx context.Context, vfsObj *vfs.VirtualFil func (fs *filesystem) cleanupRecreateWhiteout(ctx context.Context, vfsObj *vfs.VirtualFilesystem, pop *vfs.PathOperation) { if err := fs.createWhiteout(ctx, vfsObj, pop); err != nil { - ctx.Warningf("Unrecoverable overlayfs inconsistency: failed to recreate whiteout after failed file creation: %v", err) + panic(fmt.Sprintf("unrecoverable overlayfs inconsistency: failed to recreate whiteout after failed file creation: %v", err)) } } @@ -624,12 +631,13 @@ func (fs *filesystem) LinkAt(ctx context.Context, rp *vfs.ResolvingPath, vd vfs. }, }); err != nil { if cleanupErr := vfsObj.UnlinkAt(ctx, fs.creds, &newpop); cleanupErr != nil { - ctx.Warningf("Unrecoverable overlayfs inconsistency: failed to delete upper layer file after LinkAt metadata update failure: %v", cleanupErr) + panic(fmt.Sprintf("unrecoverable overlayfs inconsistency: failed to delete upper layer file after LinkAt metadata update failure: %v", cleanupErr)) } else if haveUpperWhiteout { fs.cleanupRecreateWhiteout(ctx, vfsObj, &newpop) } return err } + old.watches.Notify(ctx, "", linux.IN_ATTRIB, 0 /* cookie */, vfs.InodeEvent, false /* unlinked */) return nil }) } @@ -663,7 +671,7 @@ func (fs *filesystem) MkdirAt(ctx context.Context, rp *vfs.ResolvingPath, opts v }, }); err != nil { if cleanupErr := vfsObj.RmdirAt(ctx, fs.creds, &pop); cleanupErr != nil { - ctx.Warningf("Unrecoverable overlayfs inconsistency: failed to delete upper layer directory after MkdirAt metadata update failure: %v", cleanupErr) + panic(fmt.Sprintf("unrecoverable overlayfs inconsistency: failed to delete upper layer directory after MkdirAt metadata update failure: %v", cleanupErr)) } else if haveUpperWhiteout { fs.cleanupRecreateWhiteout(ctx, vfsObj, &pop) } @@ -678,7 +686,7 @@ func (fs *filesystem) MkdirAt(ctx context.Context, rp *vfs.ResolvingPath, opts v Value: "y", }); err != nil { if cleanupErr := vfsObj.RmdirAt(ctx, fs.creds, &pop); cleanupErr != nil { - ctx.Warningf("Unrecoverable overlayfs inconsistency: failed to delete upper layer directory after MkdirAt set-opaque failure: %v", cleanupErr) + panic(fmt.Sprintf("unrecoverable overlayfs inconsistency: failed to delete upper layer directory after MkdirAt set-opaque failure: %v", cleanupErr)) } else { fs.cleanupRecreateWhiteout(ctx, vfsObj, &pop) } @@ -722,7 +730,7 @@ func (fs *filesystem) MknodAt(ctx context.Context, rp *vfs.ResolvingPath, opts v }, }); err != nil { if cleanupErr := vfsObj.UnlinkAt(ctx, fs.creds, &pop); cleanupErr != nil { - ctx.Warningf("Unrecoverable overlayfs inconsistency: failed to delete upper layer file after MknodAt metadata update failure: %v", cleanupErr) + panic(fmt.Sprintf("unrecoverable overlayfs inconsistency: failed to delete upper layer file after MknodAt metadata update failure: %v", cleanupErr)) } else if haveUpperWhiteout { fs.cleanupRecreateWhiteout(ctx, vfsObj, &pop) } @@ -942,7 +950,7 @@ func (fs *filesystem) createAndOpenLocked(ctx context.Context, rp *vfs.Resolving }, }); err != nil { if cleanupErr := vfsObj.UnlinkAt(ctx, fs.creds, &pop); cleanupErr != nil { - ctx.Warningf("Unrecoverable overlayfs inconsistency: failed to delete upper layer file after OpenAt(O_CREAT) metadata update failure: %v", cleanupErr) + panic(fmt.Sprintf("unrecoverable overlayfs inconsistency: failed to delete upper layer file after OpenAt(O_CREAT) metadata update failure: %v", cleanupErr)) } else if haveUpperWhiteout { fs.cleanupRecreateWhiteout(ctx, vfsObj, &pop) } @@ -953,7 +961,7 @@ func (fs *filesystem) createAndOpenLocked(ctx context.Context, rp *vfs.Resolving child, err := fs.getChildLocked(ctx, parent, childName, ds) if err != nil { if cleanupErr := vfsObj.UnlinkAt(ctx, fs.creds, &pop); cleanupErr != nil { - ctx.Warningf("Unrecoverable overlayfs inconsistency: failed to delete upper layer file after OpenAt(O_CREAT) dentry lookup failure: %v", cleanupErr) + panic(fmt.Sprintf("unrecoverable overlayfs inconsistency: failed to delete upper layer file after OpenAt(O_CREAT) dentry lookup failure: %v", cleanupErr)) } else if haveUpperWhiteout { fs.cleanupRecreateWhiteout(ctx, vfsObj, &pop) } @@ -974,6 +982,7 @@ func (fs *filesystem) createAndOpenLocked(ctx context.Context, rp *vfs.Resolving // just can't open it anymore for some reason. return nil, err } + parent.watches.Notify(ctx, childName, linux.IN_CREATE, 0 /* cookie */, vfs.PathEvent, false /* unlinked */) return &fd.vfsfd, nil } @@ -1019,9 +1028,224 @@ func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa } defer mnt.EndWrite() - // FIXME(gvisor.dev/issue/1199): Actually implement rename. - _ = newParent - return syserror.EXDEV + oldParent := oldParentVD.Dentry().Impl().(*dentry) + creds := rp.Credentials() + if err := oldParent.checkPermissions(creds, vfs.MayWrite|vfs.MayExec); err != nil { + return err + } + // We need a dentry representing the renamed file since, if it's a + // directory, we need to check for write permission on it. + oldParent.dirMu.Lock() + defer oldParent.dirMu.Unlock() + renamed, err := fs.getChildLocked(ctx, oldParent, oldName, &ds) + if err != nil { + return err + } + if err := vfs.CheckDeleteSticky(creds, linux.FileMode(atomic.LoadUint32(&oldParent.mode)), auth.KUID(atomic.LoadUint32(&renamed.uid))); err != nil { + return err + } + if renamed.isDir() { + if renamed == newParent || genericIsAncestorDentry(renamed, newParent) { + return syserror.EINVAL + } + if oldParent != newParent { + if err := renamed.checkPermissions(creds, vfs.MayWrite); err != nil { + return err + } + } + } else { + if opts.MustBeDir || rp.MustBeDir() { + return syserror.ENOTDIR + } + } + + if oldParent != newParent { + if err := newParent.checkPermissions(creds, vfs.MayWrite|vfs.MayExec); err != nil { + return err + } + newParent.dirMu.Lock() + defer newParent.dirMu.Unlock() + } + if newParent.vfsd.IsDead() { + return syserror.ENOENT + } + replacedLayer, err := fs.lookupLayerLocked(ctx, newParent, newName) + if err != nil { + return err + } + var ( + replaced *dentry + replacedVFSD *vfs.Dentry + whiteouts map[string]bool + ) + if replacedLayer.existsInOverlay() { + replaced, err = fs.getChildLocked(ctx, newParent, newName, &ds) + if err != nil { + return err + } + replacedVFSD = &replaced.vfsd + if replaced.isDir() { + if !renamed.isDir() { + return syserror.EISDIR + } + if genericIsAncestorDentry(replaced, renamed) { + return syserror.ENOTEMPTY + } + replaced.dirMu.Lock() + defer replaced.dirMu.Unlock() + whiteouts, err = replaced.collectWhiteoutsForRmdirLocked(ctx) + if err != nil { + return err + } + } else { + if rp.MustBeDir() || renamed.isDir() { + return syserror.ENOTDIR + } + } + } + + if oldParent == newParent && oldName == newName { + return nil + } + + // renamed and oldParent need to be copied-up before they're renamed on the + // upper layer. + if err := renamed.copyUpLocked(ctx); err != nil { + return err + } + // If renamed is a directory, all of its descendants need to be copied-up + // before they're renamed on the upper layer. + if renamed.isDir() { + if err := renamed.copyUpDescendantsLocked(ctx, &ds); err != nil { + return err + } + } + // newParent must be copied-up before it can contain renamed on the upper + // layer. + if err := newParent.copyUpLocked(ctx); err != nil { + return err + } + // If replaced exists, it doesn't need to be copied-up, but we do need to + // serialize with copy-up. Holding renameMu for writing should be + // sufficient, but out of an abundance of caution... + if replaced != nil { + replaced.copyMu.RLock() + defer replaced.copyMu.RUnlock() + } + + vfsObj := rp.VirtualFilesystem() + mntns := vfs.MountNamespaceFromContext(ctx) + defer mntns.DecRef(ctx) + if err := vfsObj.PrepareRenameDentry(mntns, &renamed.vfsd, replacedVFSD); err != nil { + return err + } + + newpop := vfs.PathOperation{ + Root: newParent.upperVD, + Start: newParent.upperVD, + Path: fspath.Parse(newName), + } + + needRecreateWhiteouts := false + cleanupRecreateWhiteouts := func() { + if !needRecreateWhiteouts { + return + } + for whiteoutName, whiteoutUpper := range whiteouts { + if !whiteoutUpper { + continue + } + if err := fs.createWhiteout(ctx, vfsObj, &vfs.PathOperation{ + Root: replaced.upperVD, + Start: replaced.upperVD, + Path: fspath.Parse(whiteoutName), + }); err != nil && err != syserror.EEXIST { + panic(fmt.Sprintf("unrecoverable overlayfs inconsistency: failed to recreate deleted whiteout after RenameAt failure: %v", err)) + } + } + } + if renamed.isDir() { + if replacedLayer == lookupLayerUpper { + // Remove whiteouts from the directory being replaced. + needRecreateWhiteouts = true + for whiteoutName, whiteoutUpper := range whiteouts { + if !whiteoutUpper { + continue + } + if err := vfsObj.UnlinkAt(ctx, fs.creds, &vfs.PathOperation{ + Root: replaced.upperVD, + Start: replaced.upperVD, + Path: fspath.Parse(whiteoutName), + }); err != nil { + cleanupRecreateWhiteouts() + vfsObj.AbortRenameDentry(&renamed.vfsd, replacedVFSD) + return err + } + } + } else if replacedLayer == lookupLayerUpperWhiteout { + // We need to explicitly remove the whiteout since otherwise rename + // on the upper layer will fail with ENOTDIR. + if err := vfsObj.UnlinkAt(ctx, fs.creds, &newpop); err != nil { + vfsObj.AbortRenameDentry(&renamed.vfsd, replacedVFSD) + return err + } + } + } + + // Essentially no gVisor filesystem supports RENAME_WHITEOUT, so just do a + // regular rename and create the whiteout at the origin manually. Unlike + // RENAME_WHITEOUT, this isn't atomic with respect to other users of the + // upper filesystem, but this is already the case for virtually all other + // overlay filesystem operations too. + oldpop := vfs.PathOperation{ + Root: oldParent.upperVD, + Start: oldParent.upperVD, + Path: fspath.Parse(oldName), + } + if err := vfsObj.RenameAt(ctx, creds, &oldpop, &newpop, &opts); err != nil { + cleanupRecreateWhiteouts() + vfsObj.AbortRenameDentry(&renamed.vfsd, replacedVFSD) + return err + } + + // Below this point, the renamed dentry is now at newpop, and anything we + // replaced is gone forever. Commit the rename, update the overlay + // filesystem tree, and abandon attempts to recover from errors. + vfsObj.CommitRenameReplaceDentry(ctx, &renamed.vfsd, replacedVFSD) + delete(oldParent.children, oldName) + if replaced != nil { + ds = appendDentry(ds, replaced) + } + if oldParent != newParent { + newParent.dirents = nil + // This can't drop the last reference on oldParent because one is held + // by oldParentVD, so lock recursion is impossible. + oldParent.DecRef(ctx) + ds = appendDentry(ds, oldParent) + newParent.IncRef() + renamed.parent = newParent + } + renamed.name = newName + if newParent.children == nil { + newParent.children = make(map[string]*dentry) + } + newParent.children[newName] = renamed + oldParent.dirents = nil + + if err := fs.createWhiteout(ctx, vfsObj, &oldpop); err != nil { + panic(fmt.Sprintf("unrecoverable overlayfs inconsistency: failed to create whiteout at origin after RenameAt: %v", err)) + } + if renamed.isDir() { + if err := vfsObj.SetXattrAt(ctx, fs.creds, &newpop, &vfs.SetXattrOptions{ + Name: _OVL_XATTR_OPAQUE, + Value: "y", + }); err != nil { + panic(fmt.Sprintf("unrecoverable overlayfs inconsistency: failed to make renamed directory opaque: %v", err)) + } + } + + vfs.InotifyRename(ctx, &renamed.watches, &oldParent.watches, &newParent.watches, oldName, newName, renamed.isDir()) + return nil } // RmdirAt implements vfs.FilesystemImpl.RmdirAt. @@ -1100,7 +1324,7 @@ func (fs *filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error Start: child.upperVD, Path: fspath.Parse(whiteoutName), }); err != nil && err != syserror.EEXIST { - ctx.Warningf("Unrecoverable overlayfs inconsistency: failed to recreate deleted whiteout after RmdirAt failure: %v", err) + panic(fmt.Sprintf("unrecoverable overlayfs inconsistency: failed to recreate deleted whiteout after RmdirAt failure: %v", err)) } } } @@ -1130,15 +1354,14 @@ func (fs *filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error // Don't attempt to recover from this: the original directory is // already gone, so any dentries representing it are invalid, and // creating a new directory won't undo that. - ctx.Warningf("Unrecoverable overlayfs inconsistency: failed to create whiteout during RmdirAt: %v", err) - vfsObj.AbortDeleteDentry(&child.vfsd) - return err + panic(fmt.Sprintf("unrecoverable overlayfs inconsistency: failed to create whiteout during RmdirAt: %v", err)) } vfsObj.CommitDeleteDentry(ctx, &child.vfsd) delete(parent.children, name) ds = appendDentry(ds, child) parent.dirents = nil + parent.watches.Notify(ctx, name, linux.IN_DELETE|linux.IN_ISDIR, 0 /* cookie */, vfs.InodeEvent, true /* unlinked */) return nil } @@ -1146,12 +1369,25 @@ func (fs *filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error func (fs *filesystem) SetStatAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetStatOptions) error { var ds *[]*dentry fs.renameMu.RLock() - defer fs.renameMuRUnlockAndCheckDrop(ctx, &ds) d, err := fs.resolveLocked(ctx, rp, &ds) if err != nil { + fs.renameMuRUnlockAndCheckDrop(ctx, &ds) + return err + } + err = d.setStatLocked(ctx, rp, opts) + fs.renameMuRUnlockAndCheckDrop(ctx, &ds) + if err != nil { return err } + if ev := vfs.InotifyEventFromStatMask(opts.Stat.Mask); ev != 0 { + d.InotifyWithParent(ctx, ev, 0 /* cookie */, vfs.InodeEvent) + } + return nil +} + +// Precondition: d.fs.renameMu must be held for reading. +func (d *dentry) setStatLocked(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetStatOptions) error { mode := linux.FileMode(atomic.LoadUint32(&d.mode)) if err := vfs.CheckSetStat(ctx, rp.Credentials(), &opts, mode, auth.KUID(atomic.LoadUint32(&d.uid)), auth.KGID(atomic.LoadUint32(&d.gid))); err != nil { return err @@ -1246,7 +1482,7 @@ func (fs *filesystem) SymlinkAt(ctx context.Context, rp *vfs.ResolvingPath, targ }, }); err != nil { if cleanupErr := vfsObj.UnlinkAt(ctx, fs.creds, &pop); cleanupErr != nil { - ctx.Warningf("Unrecoverable overlayfs inconsistency: failed to delete upper layer file after SymlinkAt metadata update failure: %v", cleanupErr) + panic(fmt.Sprintf("unrecoverable overlayfs inconsistency: failed to delete upper layer file after SymlinkAt metadata update failure: %v", cleanupErr)) } else if haveUpperWhiteout { fs.cleanupRecreateWhiteout(ctx, vfsObj, &pop) } @@ -1339,18 +1575,17 @@ func (fs *filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error } } if err := fs.createWhiteout(ctx, vfsObj, &pop); err != nil { - ctx.Warningf("Unrecoverable overlayfs inconsistency: failed to create whiteout during UnlinkAt: %v", err) - if child != nil { - vfsObj.AbortDeleteDentry(&child.vfsd) - } - return err + panic(fmt.Sprintf("unrecoverable overlayfs inconsistency: failed to create whiteout during UnlinkAt: %v", err)) } + var cw *vfs.Watches if child != nil { vfsObj.CommitDeleteDentry(ctx, &child.vfsd) delete(parent.children, name) ds = appendDentry(ds, child) + cw = &child.watches } + vfs.InotifyRemoveChild(ctx, cw, &parent.watches, name) parent.dirents = nil return nil } @@ -1427,13 +1662,20 @@ func (fs *filesystem) getXattr(ctx context.Context, d *dentry, creds *auth.Crede func (fs *filesystem) SetXattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetXattrOptions) error { var ds *[]*dentry fs.renameMu.RLock() - defer fs.renameMuRUnlockAndCheckDrop(ctx, &ds) d, err := fs.resolveLocked(ctx, rp, &ds) if err != nil { + fs.renameMuRUnlockAndCheckDrop(ctx, &ds) + return err + } + + err = fs.setXattrLocked(ctx, d, rp.Mount(), rp.Credentials(), &opts) + fs.renameMuRUnlockAndCheckDrop(ctx, &ds) + if err != nil { return err } - return fs.setXattrLocked(ctx, d, rp.Mount(), rp.Credentials(), &opts) + d.InotifyWithParent(ctx, linux.IN_ATTRIB, 0 /* cookie */, vfs.InodeEvent) + return nil } // Precondition: fs.renameMu must be locked. @@ -1464,13 +1706,20 @@ func (fs *filesystem) setXattrLocked(ctx context.Context, d *dentry, mnt *vfs.Mo func (fs *filesystem) RemoveXattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) error { var ds *[]*dentry fs.renameMu.RLock() - defer fs.renameMuRUnlockAndCheckDrop(ctx, &ds) d, err := fs.resolveLocked(ctx, rp, &ds) if err != nil { + fs.renameMuRUnlockAndCheckDrop(ctx, &ds) return err } - return fs.removeXattrLocked(ctx, d, rp.Mount(), rp.Credentials(), name) + err = fs.removeXattrLocked(ctx, d, rp.Mount(), rp.Credentials(), name) + fs.renameMuRUnlockAndCheckDrop(ctx, &ds) + if err != nil { + return err + } + + d.InotifyWithParent(ctx, linux.IN_ATTRIB, 0 /* cookie */, vfs.InodeEvent) + return nil } // Precondition: fs.renameMu must be locked. diff --git a/pkg/sentry/fsimpl/overlay/non_directory.go b/pkg/sentry/fsimpl/overlay/non_directory.go index 74cfd3799..853aee951 100644 --- a/pkg/sentry/fsimpl/overlay/non_directory.go +++ b/pkg/sentry/fsimpl/overlay/non_directory.go @@ -39,6 +39,7 @@ func (d *dentry) readlink(ctx context.Context) (string, error) { }) } +// +stateify savable type nonDirectoryFD struct { fileDescription @@ -47,7 +48,7 @@ type nonDirectoryFD struct { // fileDescription.dentry().upperVD. cachedFlags is the last known value of // cachedFD.StatusFlags(). copiedUp, cachedFD, and cachedFlags are // protected by mu. - mu sync.Mutex + mu sync.Mutex `state:"nosave"` copiedUp bool cachedFD *vfs.FileDescription cachedFlags uint32 @@ -147,6 +148,16 @@ func (fd *nonDirectoryFD) Stat(ctx context.Context, opts vfs.StatOptions) (linux return stat, nil } +// Allocate implements vfs.FileDescriptionImpl.Allocate. +func (fd *nonDirectoryFD) Allocate(ctx context.Context, mode, offset, length uint64) error { + wrappedFD, err := fd.getCurrentFD(ctx) + if err != nil { + return err + } + defer wrappedFD.DecRef(ctx) + return wrappedFD.Allocate(ctx, mode, offset, length) +} + // SetStat implements vfs.FileDescriptionImpl.SetStat. func (fd *nonDirectoryFD) SetStat(ctx context.Context, opts vfs.SetStatOptions) error { d := fd.dentry() @@ -173,6 +184,9 @@ func (fd *nonDirectoryFD) SetStat(ctx context.Context, opts vfs.SetStatOptions) return err } d.updateAfterSetStatLocked(&opts) + if ev := vfs.InotifyEventFromStatMask(opts.Stat.Mask); ev != 0 { + d.InotifyWithParent(ctx, ev, 0, vfs.InodeEvent) + } return nil } diff --git a/pkg/sentry/fsimpl/overlay/overlay.go b/pkg/sentry/fsimpl/overlay/overlay.go index b2efe5f80..dfbccd05f 100644 --- a/pkg/sentry/fsimpl/overlay/overlay.go +++ b/pkg/sentry/fsimpl/overlay/overlay.go @@ -51,6 +51,8 @@ import ( const Name = "overlay" // FilesystemType implements vfs.FilesystemType. +// +// +stateify savable type FilesystemType struct{} // Name implements vfs.FilesystemType.Name. @@ -60,6 +62,8 @@ func (FilesystemType) Name() string { // FilesystemOptions may be passed as vfs.GetFilesystemOptions.InternalData to // FilesystemType.GetFilesystem. +// +// +stateify savable type FilesystemOptions struct { // Callers passing FilesystemOptions to // overlay.FilesystemType.GetFilesystem() are responsible for ensuring that @@ -76,6 +80,8 @@ type FilesystemOptions struct { } // filesystem implements vfs.FilesystemImpl. +// +// +stateify savable type filesystem struct { vfsfs vfs.Filesystem @@ -98,7 +104,7 @@ type filesystem struct { // renameMu synchronizes renaming with non-renaming operations in order to // ensure consistent lock ordering between dentry.dirMu in different // dentries. - renameMu sync.RWMutex + renameMu sync.RWMutex `state:"nosave"` // lastDirIno is the last inode number assigned to a directory. lastDirIno // is accessed using atomic memory operations. @@ -111,16 +117,16 @@ func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virt fsoptsRaw := opts.InternalData fsopts, haveFSOpts := fsoptsRaw.(FilesystemOptions) if fsoptsRaw != nil && !haveFSOpts { - ctx.Warningf("overlay.FilesystemType.GetFilesystem: GetFilesystemOptions.InternalData has type %T, wanted overlay.FilesystemOptions or nil", fsoptsRaw) + ctx.Infof("overlay.FilesystemType.GetFilesystem: GetFilesystemOptions.InternalData has type %T, wanted overlay.FilesystemOptions or nil", fsoptsRaw) return nil, nil, syserror.EINVAL } if haveFSOpts { if len(fsopts.LowerRoots) == 0 { - ctx.Warningf("overlay.FilesystemType.GetFilesystem: LowerRoots must be non-empty") + ctx.Infof("overlay.FilesystemType.GetFilesystem: LowerRoots must be non-empty") return nil, nil, syserror.EINVAL } if len(fsopts.LowerRoots) < 2 && !fsopts.UpperRoot.Ok() { - ctx.Warningf("overlay.FilesystemType.GetFilesystem: at least two LowerRoots are required when UpperRoot is unspecified") + ctx.Infof("overlay.FilesystemType.GetFilesystem: at least two LowerRoots are required when UpperRoot is unspecified") return nil, nil, syserror.EINVAL } // We don't enforce a maximum number of lower layers when not @@ -137,7 +143,7 @@ func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virt delete(mopts, "workdir") upperPath := fspath.Parse(upperPathname) if !upperPath.Absolute { - ctx.Warningf("overlay.FilesystemType.GetFilesystem: upperdir %q must be absolute", upperPathname) + ctx.Infof("overlay.FilesystemType.GetFilesystem: upperdir %q must be absolute", upperPathname) return nil, nil, syserror.EINVAL } upperRoot, err := vfsObj.GetDentryAt(ctx, creds, &vfs.PathOperation{ @@ -149,13 +155,13 @@ func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virt CheckSearchable: true, }) if err != nil { - ctx.Warningf("overlay.FilesystemType.GetFilesystem: failed to resolve upperdir %q: %v", upperPathname, err) + ctx.Infof("overlay.FilesystemType.GetFilesystem: failed to resolve upperdir %q: %v", upperPathname, err) return nil, nil, err } defer upperRoot.DecRef(ctx) privateUpperRoot, err := clonePrivateMount(vfsObj, upperRoot, false /* forceReadOnly */) if err != nil { - ctx.Warningf("overlay.FilesystemType.GetFilesystem: failed to make private bind mount of upperdir %q: %v", upperPathname, err) + ctx.Infof("overlay.FilesystemType.GetFilesystem: failed to make private bind mount of upperdir %q: %v", upperPathname, err) return nil, nil, err } defer privateUpperRoot.DecRef(ctx) @@ -163,24 +169,24 @@ func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virt } lowerPathnamesStr, ok := mopts["lowerdir"] if !ok { - ctx.Warningf("overlay.FilesystemType.GetFilesystem: missing required option lowerdir") + ctx.Infof("overlay.FilesystemType.GetFilesystem: missing required option lowerdir") return nil, nil, syserror.EINVAL } delete(mopts, "lowerdir") lowerPathnames := strings.Split(lowerPathnamesStr, ":") const maxLowerLayers = 500 // Linux: fs/overlay/super.c:OVL_MAX_STACK if len(lowerPathnames) < 2 && !fsopts.UpperRoot.Ok() { - ctx.Warningf("overlay.FilesystemType.GetFilesystem: at least two lowerdirs are required when upperdir is unspecified") + ctx.Infof("overlay.FilesystemType.GetFilesystem: at least two lowerdirs are required when upperdir is unspecified") return nil, nil, syserror.EINVAL } if len(lowerPathnames) > maxLowerLayers { - ctx.Warningf("overlay.FilesystemType.GetFilesystem: %d lowerdirs specified, maximum %d", len(lowerPathnames), maxLowerLayers) + ctx.Infof("overlay.FilesystemType.GetFilesystem: %d lowerdirs specified, maximum %d", len(lowerPathnames), maxLowerLayers) return nil, nil, syserror.EINVAL } for _, lowerPathname := range lowerPathnames { lowerPath := fspath.Parse(lowerPathname) if !lowerPath.Absolute { - ctx.Warningf("overlay.FilesystemType.GetFilesystem: lowerdir %q must be absolute", lowerPathname) + ctx.Infof("overlay.FilesystemType.GetFilesystem: lowerdir %q must be absolute", lowerPathname) return nil, nil, syserror.EINVAL } lowerRoot, err := vfsObj.GetDentryAt(ctx, creds, &vfs.PathOperation{ @@ -192,13 +198,13 @@ func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virt CheckSearchable: true, }) if err != nil { - ctx.Warningf("overlay.FilesystemType.GetFilesystem: failed to resolve lowerdir %q: %v", lowerPathname, err) + ctx.Infof("overlay.FilesystemType.GetFilesystem: failed to resolve lowerdir %q: %v", lowerPathname, err) return nil, nil, err } defer lowerRoot.DecRef(ctx) privateLowerRoot, err := clonePrivateMount(vfsObj, lowerRoot, true /* forceReadOnly */) if err != nil { - ctx.Warningf("overlay.FilesystemType.GetFilesystem: failed to make private bind mount of lowerdir %q: %v", lowerPathname, err) + ctx.Infof("overlay.FilesystemType.GetFilesystem: failed to make private bind mount of lowerdir %q: %v", lowerPathname, err) return nil, nil, err } defer privateLowerRoot.DecRef(ctx) @@ -206,7 +212,7 @@ func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virt } } if len(mopts) != 0 { - ctx.Warningf("overlay.FilesystemType.GetFilesystem: unused options: %v", mopts) + ctx.Infof("overlay.FilesystemType.GetFilesystem: unused options: %v", mopts) return nil, nil, syserror.EINVAL } @@ -279,7 +285,7 @@ func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virt return nil, nil, syserror.EREMOTE } if isWhiteout(&rootStat) { - ctx.Warningf("overlay.FilesystemType.GetFilesystem: filesystem root is a whiteout") + ctx.Infof("overlay.FilesystemType.GetFilesystem: filesystem root is a whiteout") root.destroyLocked(ctx) fs.vfsfs.DecRef(ctx) return nil, nil, syserror.EINVAL @@ -367,6 +373,8 @@ func (fs *filesystem) newDirIno() uint64 { } // dentry implements vfs.DentryImpl. +// +// +stateify savable type dentry struct { vfsd vfs.Dentry @@ -399,7 +407,7 @@ type dentry struct { // and dirents (if not nil) is a cache of dirents as returned by // directoryFDs representing this directory. children is protected by // dirMu. - dirMu sync.Mutex + dirMu sync.Mutex `state:"nosave"` children map[string]*dentry dirents []vfs.Dirent @@ -409,7 +417,7 @@ type dentry struct { // If !upperVD.Ok(), it can transition to a valid vfs.VirtualDentry (i.e. // be copied up) with copyMu locked for writing; otherwise, it is // immutable. lowerVDs is always immutable. - copyMu sync.RWMutex + copyMu sync.RWMutex `state:"nosave"` upperVD vfs.VirtualDentry lowerVDs []vfs.VirtualDentry @@ -454,6 +462,13 @@ type dentry struct { isMappable uint32 locks vfs.FileLocks + + // watches is the set of inotify watches on the file repesented by this dentry. + // + // Note that hard links to the same file will not share the same set of + // watches, due to the fact that we do not have inode structures in this + // overlay implementation. + watches vfs.Watches } // newDentry creates a new dentry. The dentry initially has no references; it @@ -513,6 +528,14 @@ func (d *dentry) checkDropLocked(ctx context.Context) { if atomic.LoadInt64(&d.refs) != 0 { return } + + // Make sure that we do not lose watches on dentries that have not been + // deleted. Note that overlayfs never calls VFS.InvalidateDentry(), so + // d.vfsd.IsDead() indicates that d was deleted. + if !d.vfsd.IsDead() && d.watches.Size() > 0 { + return + } + // Refs is still zero; destroy it. d.destroyLocked(ctx) return @@ -541,6 +564,8 @@ func (d *dentry) destroyLocked(ctx context.Context) { lowerVD.DecRef(ctx) } + d.watches.HandleDeletion(ctx) + if d.parent != nil { d.parent.dirMu.Lock() if !d.vfsd.IsDead() { @@ -559,19 +584,36 @@ func (d *dentry) destroyLocked(ctx context.Context) { // InotifyWithParent implements vfs.DentryImpl.InotifyWithParent. func (d *dentry) InotifyWithParent(ctx context.Context, events uint32, cookie uint32, et vfs.EventType) { - // TODO(gvisor.dev/issue/1479): Implement inotify. + if d.isDir() { + events |= linux.IN_ISDIR + } + + // overlayfs never calls VFS.InvalidateDentry(), so d.vfsd.IsDead() indicates + // that d was deleted. + deleted := d.vfsd.IsDead() + + d.fs.renameMu.RLock() + // The ordering below is important, Linux always notifies the parent first. + if d.parent != nil { + d.parent.watches.Notify(ctx, d.name, events, cookie, et, deleted) + } + d.watches.Notify(ctx, "", events, cookie, et, deleted) + d.fs.renameMu.RUnlock() } // Watches implements vfs.DentryImpl.Watches. func (d *dentry) Watches() *vfs.Watches { - // TODO(gvisor.dev/issue/1479): Implement inotify. - return nil + return &d.watches } // OnZeroWatches implements vfs.DentryImpl.OnZeroWatches. -// -// TODO(gvisor.dev/issue/1479): Implement inotify. -func (d *dentry) OnZeroWatches(context.Context) {} +func (d *dentry) OnZeroWatches(ctx context.Context) { + if atomic.LoadInt64(&d.refs) == 0 { + d.fs.renameMu.Lock() + d.checkDropLocked(ctx) + d.fs.renameMu.Unlock() + } +} // iterLayers invokes yield on each layer comprising d, from top to bottom. If // any call to yield returns false, iterLayer stops iteration. @@ -652,6 +694,8 @@ func (d *dentry) updateAfterSetStatLocked(opts *vfs.SetStatOptions) { // fileDescription is embedded by overlay implementations of // vfs.FileDescriptionImpl. +// +// +stateify savable type fileDescription struct { vfsfd vfs.FileDescription vfs.FileDescriptionDefaultImpl @@ -679,17 +723,33 @@ func (fd *fileDescription) GetXattr(ctx context.Context, opts vfs.GetXattrOption // SetXattr implements vfs.FileDescriptionImpl.SetXattr. func (fd *fileDescription) SetXattr(ctx context.Context, opts vfs.SetXattrOptions) error { fs := fd.filesystem() + d := fd.dentry() + fs.renameMu.RLock() - defer fs.renameMu.RUnlock() - return fs.setXattrLocked(ctx, fd.dentry(), fd.vfsfd.Mount(), auth.CredentialsFromContext(ctx), &opts) + err := fs.setXattrLocked(ctx, d, fd.vfsfd.Mount(), auth.CredentialsFromContext(ctx), &opts) + fs.renameMu.RUnlock() + if err != nil { + return err + } + + d.InotifyWithParent(ctx, linux.IN_ATTRIB, 0, vfs.InodeEvent) + return nil } // RemoveXattr implements vfs.FileDescriptionImpl.RemoveXattr. func (fd *fileDescription) RemoveXattr(ctx context.Context, name string) error { fs := fd.filesystem() + d := fd.dentry() + fs.renameMu.RLock() - defer fs.renameMu.RUnlock() - return fs.removeXattrLocked(ctx, fd.dentry(), fd.vfsfd.Mount(), auth.CredentialsFromContext(ctx), name) + err := fs.removeXattrLocked(ctx, d, fd.vfsfd.Mount(), auth.CredentialsFromContext(ctx), name) + fs.renameMu.RUnlock() + if err != nil { + return err + } + + d.InotifyWithParent(ctx, linux.IN_ATTRIB, 0, vfs.InodeEvent) + return nil } // LockPOSIX implements vfs.FileDescriptionImpl.LockPOSIX. diff --git a/pkg/sentry/fsimpl/pipefs/pipefs.go b/pkg/sentry/fsimpl/pipefs/pipefs.go index 7053ad6db..4e2da4810 100644 --- a/pkg/sentry/fsimpl/pipefs/pipefs.go +++ b/pkg/sentry/fsimpl/pipefs/pipefs.go @@ -31,6 +31,7 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) +// +stateify savable type filesystemType struct{} // Name implements vfs.FilesystemType.Name. @@ -43,6 +44,7 @@ func (filesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFile panic("pipefs.filesystemType.GetFilesystem should never be called") } +// +stateify savable type filesystem struct { kernfs.Filesystem @@ -76,6 +78,8 @@ func (fs *filesystem) PrependPath(ctx context.Context, vfsroot, vd vfs.VirtualDe } // inode implements kernfs.Inode. +// +// +stateify savable type inode struct { kernfs.InodeNotDirectory kernfs.InodeNotSymlink @@ -144,8 +148,8 @@ func (i *inode) SetStat(ctx context.Context, vfsfs *vfs.Filesystem, creds *auth. } // 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, &i.locks) +func (i *inode) Open(ctx context.Context, rp *vfs.ResolvingPath, d *kernfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + return i.pipe.Open(ctx, rp.Mount(), d.VFSDentry(), opts.Flags, &i.locks) } // StatFS implements kernfs.Inode.StatFS. diff --git a/pkg/sentry/fsimpl/proc/BUILD b/pkg/sentry/fsimpl/proc/BUILD index a45b44440..2e086e34c 100644 --- a/pkg/sentry/fsimpl/proc/BUILD +++ b/pkg/sentry/fsimpl/proc/BUILD @@ -100,6 +100,7 @@ go_library( "//pkg/sync", "//pkg/syserror", "//pkg/tcpip/header", + "//pkg/tcpip/network/ipv4", "//pkg/usermem", ], ) diff --git a/pkg/sentry/fsimpl/proc/filesystem.go b/pkg/sentry/fsimpl/proc/filesystem.go index 03b5941b9..05d7948ea 100644 --- a/pkg/sentry/fsimpl/proc/filesystem.go +++ b/pkg/sentry/fsimpl/proc/filesystem.go @@ -41,6 +41,7 @@ func (FilesystemType) Name() string { return Name } +// +stateify savable type filesystem struct { kernfs.Filesystem @@ -84,6 +85,8 @@ func (fs *filesystem) Release(ctx context.Context) { // dynamicInode is an overfitted interface for common Inodes with // dynamicByteSource types used in procfs. +// +// +stateify savable type dynamicInode interface { kernfs.Inode vfs.DynamicBytesSource @@ -99,6 +102,7 @@ func (fs *filesystem) newDentry(creds *auth.Credentials, ino uint64, perm linux. return d } +// +stateify savable type staticFile struct { kernfs.DynamicBytesFile vfs.StaticData @@ -118,10 +122,13 @@ func newStaticDir(creds *auth.Credentials, devMajor, devMinor uint32, ino uint64 // InternalData contains internal data passed in to the procfs mount via // vfs.GetFilesystemOptions.InternalData. +// +// +stateify savable type InternalData struct { Cgroups map[string]string } +// +stateify savable type implStatFS struct{} // StatFS implements kernfs.Inode.StatFS. diff --git a/pkg/sentry/fsimpl/proc/subtasks.go b/pkg/sentry/fsimpl/proc/subtasks.go index d57d94dbc..47ecd941c 100644 --- a/pkg/sentry/fsimpl/proc/subtasks.go +++ b/pkg/sentry/fsimpl/proc/subtasks.go @@ -68,8 +68,8 @@ func (fs *filesystem) newSubtasks(task *kernel.Task, pidns *kernel.PIDNamespace, return dentry } -// Lookup implements kernfs.inodeDynamicLookup. -func (i *subtasksInode) Lookup(ctx context.Context, name string) (*vfs.Dentry, error) { +// Lookup implements kernfs.inodeDynamicLookup.Lookup. +func (i *subtasksInode) Lookup(ctx context.Context, name string) (*kernfs.Dentry, error) { tid, err := strconv.ParseUint(name, 10, 32) if err != nil { return nil, syserror.ENOENT @@ -82,12 +82,10 @@ func (i *subtasksInode) Lookup(ctx context.Context, name string) (*vfs.Dentry, e if subTask.ThreadGroup() != i.task.ThreadGroup() { return nil, syserror.ENOENT } - - subTaskDentry := i.fs.newTaskInode(subTask, i.pidns, false, i.cgroupControllers) - return subTaskDentry.VFSDentry(), nil + return i.fs.newTaskInode(subTask, i.pidns, false, i.cgroupControllers), nil } -// IterDirents implements kernfs.inodeDynamicLookup. +// IterDirents implements kernfs.inodeDynamicLookup.IterDirents. func (i *subtasksInode) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback, offset, relOffset int64) (int64, error) { tasks := i.task.ThreadGroup().MemberIDs(i.pidns) if len(tasks) == 0 { @@ -118,6 +116,7 @@ func (i *subtasksInode) IterDirents(ctx context.Context, cb vfs.IterDirentsCallb return offset, nil } +// +stateify savable type subtasksFD struct { kernfs.GenericDirectoryFD @@ -155,21 +154,21 @@ func (fd *subtasksFD) SetStat(ctx context.Context, opts vfs.SetStatOptions) erro return fd.GenericDirectoryFD.SetStat(ctx, opts) } -// Open implements kernfs.Inode. -func (i *subtasksInode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { +// Open implements kernfs.Inode.Open. +func (i *subtasksInode) Open(ctx context.Context, rp *vfs.ResolvingPath, d *kernfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { fd := &subtasksFD{task: i.task} if err := fd.Init(&i.OrderedChildren, &i.locks, &opts, kernfs.GenericDirectoryFDOptions{ SeekEnd: kernfs.SeekEndZero, }); err != nil { return nil, err } - if err := fd.VFSFileDescription().Init(fd, opts.Flags, rp.Mount(), vfsd, &vfs.FileDescriptionOptions{}); err != nil { + if err := fd.VFSFileDescription().Init(fd, opts.Flags, rp.Mount(), d.VFSDentry(), &vfs.FileDescriptionOptions{}); err != nil { return nil, err } return fd.VFSFileDescription(), nil } -// Stat implements kernfs.Inode. +// Stat implements kernfs.Inode.Stat. func (i *subtasksInode) Stat(ctx context.Context, vsfs *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) { stat, err := i.InodeAttrs.Stat(ctx, vsfs, opts) if err != nil { @@ -181,12 +180,12 @@ func (i *subtasksInode) Stat(ctx context.Context, vsfs *vfs.Filesystem, opts vfs return stat, nil } -// SetStat implements Inode.SetStat not allowing inode attributes to be changed. +// SetStat implements kernfs.Inode.SetStat not allowing inode attributes to be changed. func (*subtasksInode) SetStat(context.Context, *vfs.Filesystem, *auth.Credentials, vfs.SetStatOptions) error { return syserror.EPERM } -// DecRef implements kernfs.Inode. +// DecRef implements kernfs.Inode.DecRef. func (i *subtasksInode) DecRef(context.Context) { i.subtasksInodeRefs.DecRef(i.Destroy) } diff --git a/pkg/sentry/fsimpl/proc/task.go b/pkg/sentry/fsimpl/proc/task.go index dbdb5d929..a7cd6f57e 100644 --- a/pkg/sentry/fsimpl/proc/task.go +++ b/pkg/sentry/fsimpl/proc/task.go @@ -53,6 +53,7 @@ func (fs *filesystem) newTaskInode(task *kernel.Task, pidns *kernel.PIDNamespace "auxv": fs.newTaskOwnedFile(task, fs.NextIno(), 0444, &auxvData{task: task}), "cmdline": fs.newTaskOwnedFile(task, fs.NextIno(), 0444, &cmdlineData{task: task, arg: cmdlineDataArg}), "comm": fs.newComm(task, fs.NextIno(), 0444), + "cwd": fs.newCwdSymlink(task, fs.NextIno()), "environ": fs.newTaskOwnedFile(task, fs.NextIno(), 0444, &cmdlineData{task: task, arg: environDataArg}), "exe": fs.newExeSymlink(task, fs.NextIno()), "fd": fs.newFDDirInode(task), @@ -106,9 +107,9 @@ func (i *taskInode) Valid(ctx context.Context) bool { return i.task.ExitState() != kernel.TaskExitDead } -// Open implements kernfs.Inode. -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, &i.locks, &opts, kernfs.GenericDirectoryFDOptions{ +// Open implements kernfs.Inode.Open. +func (i *taskInode) Open(ctx context.Context, rp *vfs.ResolvingPath, d *kernfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), d, &i.OrderedChildren, &i.locks, &opts, kernfs.GenericDirectoryFDOptions{ SeekEnd: kernfs.SeekEndZero, }) if err != nil { @@ -117,18 +118,20 @@ func (i *taskInode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.D return fd.VFSFileDescription(), nil } -// SetStat implements Inode.SetStat not allowing inode attributes to be changed. +// SetStat implements kernfs.Inode.SetStat not allowing inode attributes to be changed. func (*taskInode) SetStat(context.Context, *vfs.Filesystem, *auth.Credentials, vfs.SetStatOptions) error { return syserror.EPERM } -// DecRef implements kernfs.Inode. +// DecRef implements kernfs.Inode.DecRef. func (i *taskInode) DecRef(context.Context) { i.taskInodeRefs.DecRef(i.Destroy) } // taskOwnedInode implements kernfs.Inode and overrides inode owner with task // effective user and group. +// +// +stateify savable type taskOwnedInode struct { kernfs.Inode @@ -168,7 +171,7 @@ func (fs *filesystem) newTaskOwnedDir(task *kernel.Task, ino uint64, perm linux. return d } -// Stat implements kernfs.Inode. +// Stat implements kernfs.Inode.Stat. func (i *taskOwnedInode) Stat(ctx context.Context, fs *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) { stat, err := i.Inode.Stat(ctx, fs, opts) if err != nil { @@ -186,7 +189,7 @@ func (i *taskOwnedInode) Stat(ctx context.Context, fs *vfs.Filesystem, opts vfs. return stat, nil } -// CheckPermissions implements kernfs.Inode. +// CheckPermissions implements kernfs.Inode.CheckPermissions. func (i *taskOwnedInode) CheckPermissions(_ context.Context, creds *auth.Credentials, ats vfs.AccessTypes) error { mode := i.Mode() uid, gid := i.getOwner(mode) diff --git a/pkg/sentry/fsimpl/proc/task_fds.go b/pkg/sentry/fsimpl/proc/task_fds.go index 94ec2ff69..0866cea2b 100644 --- a/pkg/sentry/fsimpl/proc/task_fds.go +++ b/pkg/sentry/fsimpl/proc/task_fds.go @@ -51,6 +51,7 @@ func taskFDExists(ctx context.Context, t *kernel.Task, fd int32) bool { return true } +// +stateify savable type fdDir struct { locks vfs.FileLocks @@ -62,7 +63,7 @@ type fdDir struct { produceSymlink bool } -// IterDirents implements kernfs.inodeDynamicLookup. +// IterDirents implements kernfs.inodeDynamicLookup.IterDirents. func (i *fdDir) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback, offset, relOffset int64) (int64, error) { var fds []int32 i.task.WithMuLocked(func(t *kernel.Task) { @@ -135,8 +136,8 @@ func (fs *filesystem) newFDDirInode(task *kernel.Task) *kernfs.Dentry { return dentry } -// Lookup implements kernfs.inodeDynamicLookup. -func (i *fdDirInode) Lookup(ctx context.Context, name string) (*vfs.Dentry, error) { +// Lookup implements kernfs.inodeDynamicLookup.Lookup. +func (i *fdDirInode) Lookup(ctx context.Context, name string) (*kernfs.Dentry, error) { fdInt, err := strconv.ParseInt(name, 10, 32) if err != nil { return nil, syserror.ENOENT @@ -145,13 +146,12 @@ func (i *fdDirInode) Lookup(ctx context.Context, name string) (*vfs.Dentry, erro if !taskFDExists(ctx, i.task, fd) { return nil, syserror.ENOENT } - taskDentry := i.fs.newFDSymlink(i.task, fd, i.fs.NextIno()) - return taskDentry.VFSDentry(), nil + return i.fs.newFDSymlink(i.task, fd, i.fs.NextIno()), 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, &i.locks, &opts, kernfs.GenericDirectoryFDOptions{ +// Open implements kernfs.Inode.Open. +func (i *fdDirInode) Open(ctx context.Context, rp *vfs.ResolvingPath, d *kernfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), d, &i.OrderedChildren, &i.locks, &opts, kernfs.GenericDirectoryFDOptions{ SeekEnd: kernfs.SeekEndZero, }) if err != nil { @@ -160,7 +160,7 @@ func (i *fdDirInode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs. return fd.VFSFileDescription(), nil } -// CheckPermissions implements kernfs.Inode. +// CheckPermissions implements kernfs.Inode.CheckPermissions. // // 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 @@ -182,7 +182,7 @@ func (i *fdDirInode) CheckPermissions(ctx context.Context, creds *auth.Credentia return err } -// DecRef implements kernfs.Inode. +// DecRef implements kernfs.Inode.DecRef. func (i *fdDirInode) DecRef(context.Context) { i.fdDirInodeRefs.DecRef(i.Destroy) } @@ -214,7 +214,7 @@ func (fs *filesystem) newFDSymlink(task *kernel.Task, fd int32, ino uint64) *ker return d } -func (s *fdSymlink) Readlink(ctx context.Context) (string, error) { +func (s *fdSymlink) Readlink(ctx context.Context, _ *vfs.Mount) (string, error) { file, _ := getTaskFD(s.task, s.fd) if file == nil { return "", syserror.ENOENT @@ -269,8 +269,8 @@ func (fs *filesystem) newFDInfoDirInode(task *kernel.Task) *kernfs.Dentry { return dentry } -// Lookup implements kernfs.inodeDynamicLookup. -func (i *fdInfoDirInode) Lookup(ctx context.Context, name string) (*vfs.Dentry, error) { +// Lookup implements kernfs.inodeDynamicLookup.Lookup. +func (i *fdInfoDirInode) Lookup(ctx context.Context, name string) (*kernfs.Dentry, error) { fdInt, err := strconv.ParseInt(name, 10, 32) if err != nil { return nil, syserror.ENOENT @@ -283,13 +283,12 @@ func (i *fdInfoDirInode) Lookup(ctx context.Context, name string) (*vfs.Dentry, task: i.task, fd: fd, } - dentry := i.fs.newTaskOwnedFile(i.task, i.fs.NextIno(), 0444, data) - return dentry.VFSDentry(), nil + return i.fs.newTaskOwnedFile(i.task, i.fs.NextIno(), 0444, data), 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, &i.locks, &opts, kernfs.GenericDirectoryFDOptions{ +// Open implements kernfs.Inode.Open. +func (i *fdInfoDirInode) Open(ctx context.Context, rp *vfs.ResolvingPath, d *kernfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), d, &i.OrderedChildren, &i.locks, &opts, kernfs.GenericDirectoryFDOptions{ SeekEnd: kernfs.SeekEndZero, }) if err != nil { @@ -298,7 +297,7 @@ func (i *fdInfoDirInode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd * return fd.VFSFileDescription(), nil } -// DecRef implements kernfs.Inode. +// DecRef implements kernfs.Inode.DecRef. func (i *fdInfoDirInode) DecRef(context.Context) { i.fdInfoDirInodeRefs.DecRef(i.Destroy) } diff --git a/pkg/sentry/fsimpl/proc/task_files.go b/pkg/sentry/fsimpl/proc/task_files.go index ce87b0d47..3fbf081a6 100644 --- a/pkg/sentry/fsimpl/proc/task_files.go +++ b/pkg/sentry/fsimpl/proc/task_files.go @@ -667,20 +667,24 @@ func (fs *filesystem) newExeSymlink(task *kernel.Task, ino uint64) *kernfs.Dentr 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() +// Readlink implements kernfs.Inode.Readlink. +func (s *exeSymlink) Readlink(ctx context.Context, _ *vfs.Mount) (string, error) { + exec, _, err := s.Getlink(ctx, nil) if err != nil { return "", err } defer exec.DecRef(ctx) - return exec.PathnameWithDeleted(ctx), nil + root := vfs.RootFromContext(ctx) + if !root.Ok() { + // It could have raced with process deletion. + return "", syserror.ESRCH + } + defer root.DecRef(ctx) + + vfsObj := exec.Mount().Filesystem().VirtualFilesystem() + name, _ := vfsObj.PathnameWithDeleted(ctx, root, exec) + return name, nil } // Getlink implements kernfs.Inode.Getlink. @@ -688,23 +692,12 @@ func (s *exeSymlink) Getlink(ctx context.Context, _ *vfs.Mount) (vfs.VirtualDent 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(ctx) - - 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 + return vfs.VirtualDentry{}, "", err } + var err error + var exec fsbridge.File s.task.WithMuLocked(func(t *kernel.Task) { mm := t.MemoryManager() if mm == nil { @@ -715,12 +708,78 @@ func (s *exeSymlink) executable() (file fsbridge.File, err error) { // 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 { + exec = mm.Executable() + if exec == nil { err = syserror.ESRCH } }) - return + if err != nil { + return vfs.VirtualDentry{}, "", err + } + defer exec.DecRef(ctx) + + vd := exec.(*fsbridge.VFSFile).FileDescription().VirtualDentry() + vd.IncRef() + return vd, "", nil +} + +// cwdSymlink is an symlink for the /proc/[pid]/cwd file. +// +// +stateify savable +type cwdSymlink struct { + implStatFS + kernfs.InodeAttrs + kernfs.InodeNoopRefCount + kernfs.InodeSymlink + + task *kernel.Task +} + +var _ kernfs.Inode = (*cwdSymlink)(nil) + +func (fs *filesystem) newCwdSymlink(task *kernel.Task, ino uint64) *kernfs.Dentry { + inode := &cwdSymlink{task: task} + inode.Init(task.Credentials(), linux.UNNAMED_MAJOR, fs.devMinor, ino, linux.ModeSymlink|0777) + + d := &kernfs.Dentry{} + d.Init(inode) + return d +} + +// Readlink implements kernfs.Inode.Readlink. +func (s *cwdSymlink) Readlink(ctx context.Context, _ *vfs.Mount) (string, error) { + cwd, _, err := s.Getlink(ctx, nil) + if err != nil { + return "", err + } + defer cwd.DecRef(ctx) + + root := vfs.RootFromContext(ctx) + if !root.Ok() { + // It could have raced with process deletion. + return "", syserror.ESRCH + } + defer root.DecRef(ctx) + + vfsObj := cwd.Mount().Filesystem().VirtualFilesystem() + name, _ := vfsObj.PathnameWithDeleted(ctx, root, cwd) + return name, nil +} + +// Getlink implements kernfs.Inode.Getlink. +func (s *cwdSymlink) Getlink(ctx context.Context, _ *vfs.Mount) (vfs.VirtualDentry, string, error) { + if !kernel.ContextCanTrace(ctx, s.task, false) { + return vfs.VirtualDentry{}, "", syserror.EACCES + } + if err := checkTaskState(s.task); err != nil { + return vfs.VirtualDentry{}, "", err + } + cwd := s.task.FSContext().WorkingDirectoryVFS2() + if !cwd.Ok() { + // It could have raced with process deletion. + return vfs.VirtualDentry{}, "", syserror.ESRCH + } + return cwd, "", nil } // mountInfoData is used to implement /proc/[pid]/mountinfo. @@ -785,6 +844,7 @@ func (i *mountsData) Generate(ctx context.Context, buf *bytes.Buffer) error { return nil } +// +stateify savable type namespaceSymlink struct { kernfs.StaticSymlink @@ -807,15 +867,15 @@ func (fs *filesystem) newNamespaceSymlink(task *kernel.Task, ino uint64, ns stri return d } -// Readlink implements Inode. -func (s *namespaceSymlink) Readlink(ctx context.Context) (string, error) { +// Readlink implements kernfs.Inode.Readlink. +func (s *namespaceSymlink) Readlink(ctx context.Context, mnt *vfs.Mount) (string, error) { if err := checkTaskState(s.task); err != nil { return "", err } - return s.StaticSymlink.Readlink(ctx) + return s.StaticSymlink.Readlink(ctx, mnt) } -// Getlink implements Inode.Getlink. +// Getlink implements kernfs.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 @@ -832,6 +892,8 @@ func (s *namespaceSymlink) Getlink(ctx context.Context, mnt *vfs.Mount) (vfs.Vir // namespaceInode is a synthetic inode created to represent a namespace in // /proc/[pid]/ns/*. +// +// +stateify savable type namespaceInode struct { implStatFS kernfs.InodeAttrs @@ -852,12 +914,12 @@ func (i *namespaceInode) Init(creds *auth.Credentials, devMajor, devMinor uint32 i.InodeAttrs.Init(creds, devMajor, devMinor, 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) { +// Open implements kernfs.Inode.Open. +func (i *namespaceInode) Open(ctx context.Context, rp *vfs.ResolvingPath, d *kernfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { fd := &namespaceFD{inode: i} i.IncRef() fd.LockFD.Init(&i.locks) - if err := fd.vfsfd.Init(fd, opts.Flags, rp.Mount(), vfsd, &vfs.FileDescriptionOptions{}); err != nil { + if err := fd.vfsfd.Init(fd, opts.Flags, rp.Mount(), d.VFSDentry(), &vfs.FileDescriptionOptions{}); err != nil { return nil, err } return &fd.vfsfd, nil @@ -865,6 +927,8 @@ func (i *namespaceInode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd * // namespace FD is a synthetic file that represents a namespace in // /proc/[pid]/ns/*. +// +// +stateify savable type namespaceFD struct { vfs.FileDescriptionDefaultImpl vfs.LockFD @@ -875,20 +939,20 @@ type namespaceFD struct { var _ vfs.FileDescriptionImpl = (*namespaceFD)(nil) -// Stat implements FileDescriptionImpl. +// Stat implements vfs.FileDescriptionImpl.Stat. func (fd *namespaceFD) Stat(ctx context.Context, opts vfs.StatOptions) (linux.Statx, error) { vfs := fd.vfsfd.VirtualDentry().Mount().Filesystem() return fd.inode.Stat(ctx, vfs, opts) } -// SetStat implements FileDescriptionImpl. +// SetStat implements vfs.FileDescriptionImpl.SetStat. 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. +// Release implements vfs.FileDescriptionImpl.Release. func (fd *namespaceFD) Release(ctx context.Context) { fd.inode.DecRef(ctx) } diff --git a/pkg/sentry/fsimpl/proc/task_net.go b/pkg/sentry/fsimpl/proc/task_net.go index 4e69782c7..e7f748655 100644 --- a/pkg/sentry/fsimpl/proc/task_net.go +++ b/pkg/sentry/fsimpl/proc/task_net.go @@ -616,6 +616,7 @@ type netSnmpData struct { var _ dynamicInode = (*netSnmpData)(nil) +// +stateify savable type snmpLine struct { prefix string header string @@ -660,7 +661,7 @@ func sprintSlice(s []uint64) string { return r[1 : len(r)-1] // Remove "[]" introduced by fmt of slice. } -// Generate implements vfs.DynamicBytesSource. +// Generate implements vfs.DynamicBytesSource.Generate. func (d *netSnmpData) Generate(ctx context.Context, buf *bytes.Buffer) error { types := []interface{}{ &inet.StatSNMPIP{}, @@ -709,7 +710,7 @@ type netRouteData struct { var _ dynamicInode = (*netRouteData)(nil) -// Generate implements vfs.DynamicBytesSource. +// Generate implements vfs.DynamicBytesSource.Generate. // See Linux's net/ipv4/fib_trie.c:fib_route_seq_show. func (d *netRouteData) Generate(ctx context.Context, buf *bytes.Buffer) error { fmt.Fprintf(buf, "%-127s\n", "Iface\tDestination\tGateway\tFlags\tRefCnt\tUse\tMetric\tMask\tMTU\tWindow\tIRTT") @@ -773,7 +774,7 @@ type netStatData struct { var _ dynamicInode = (*netStatData)(nil) -// Generate implements vfs.DynamicBytesSource. +// Generate implements vfs.DynamicBytesSource.Generate. // See Linux's net/ipv4/fib_trie.c:fib_route_seq_show. func (d *netStatData) Generate(ctx context.Context, buf *bytes.Buffer) error { buf.WriteString("TcpExt: SyncookiesSent SyncookiesRecv SyncookiesFailed " + diff --git a/pkg/sentry/fsimpl/proc/tasks.go b/pkg/sentry/fsimpl/proc/tasks.go index 3ea00ab87..d8f5dd509 100644 --- a/pkg/sentry/fsimpl/proc/tasks.go +++ b/pkg/sentry/fsimpl/proc/tasks.go @@ -52,8 +52,8 @@ type tasksInode struct { // '/proc/self' and '/proc/thread-self' have custom directory offsets in // Linux. So handle them outside of OrderedChildren. - selfSymlink *vfs.Dentry - threadSelfSymlink *vfs.Dentry + selfSymlink *kernfs.Dentry + threadSelfSymlink *kernfs.Dentry // cgroupControllers is a map of controller name to directory in the // cgroup hierarchy. These controllers are immutable and will be listed @@ -81,8 +81,8 @@ func (fs *filesystem) newTasksInode(k *kernel.Kernel, pidns *kernel.PIDNamespace inode := &tasksInode{ pidns: pidns, fs: fs, - selfSymlink: fs.newSelfSymlink(root, fs.NextIno(), pidns).VFSDentry(), - threadSelfSymlink: fs.newThreadSelfSymlink(root, fs.NextIno(), pidns).VFSDentry(), + selfSymlink: fs.newSelfSymlink(root, fs.NextIno(), pidns), + threadSelfSymlink: fs.newThreadSelfSymlink(root, fs.NextIno(), pidns), cgroupControllers: cgroupControllers, } inode.InodeAttrs.Init(root, linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), linux.ModeDirectory|0555) @@ -98,8 +98,8 @@ func (fs *filesystem) newTasksInode(k *kernel.Kernel, pidns *kernel.PIDNamespace return inode, dentry } -// Lookup implements kernfs.inodeDynamicLookup. -func (i *tasksInode) Lookup(ctx context.Context, name string) (*vfs.Dentry, error) { +// Lookup implements kernfs.inodeDynamicLookup.Lookup. +func (i *tasksInode) Lookup(ctx context.Context, name string) (*kernfs.Dentry, error) { // Try to lookup a corresponding task. tid, err := strconv.ParseUint(name, 10, 64) if err != nil { @@ -118,11 +118,10 @@ func (i *tasksInode) Lookup(ctx context.Context, name string) (*vfs.Dentry, erro return nil, syserror.ENOENT } - taskDentry := i.fs.newTaskInode(task, i.pidns, true, i.cgroupControllers) - return taskDentry.VFSDentry(), nil + return i.fs.newTaskInode(task, i.pidns, true, i.cgroupControllers), nil } -// IterDirents implements kernfs.inodeDynamicLookup. +// IterDirents implements kernfs.inodeDynamicLookup.IterDirents. func (i *tasksInode) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback, offset, _ int64) (int64, error) { // fs/proc/internal.h: #define FIRST_PROCESS_ENTRY 256 const FIRST_PROCESS_ENTRY = 256 @@ -200,9 +199,9 @@ func (i *tasksInode) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback return maxTaskID, nil } -// Open implements kernfs.Inode. -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, &i.locks, &opts, kernfs.GenericDirectoryFDOptions{ +// Open implements kernfs.Inode.Open. +func (i *tasksInode) Open(ctx context.Context, rp *vfs.ResolvingPath, d *kernfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), d, &i.OrderedChildren, &i.locks, &opts, kernfs.GenericDirectoryFDOptions{ SeekEnd: kernfs.SeekEndZero, }) if err != nil { @@ -229,7 +228,7 @@ func (i *tasksInode) Stat(ctx context.Context, vsfs *vfs.Filesystem, opts vfs.St return stat, nil } -// DecRef implements kernfs.Inode. +// DecRef implements kernfs.Inode.DecRef. func (i *tasksInode) DecRef(context.Context) { i.tasksInodeRefs.DecRef(i.Destroy) } @@ -237,6 +236,8 @@ func (i *tasksInode) DecRef(context.Context) { // 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. +// +// +stateify savable type staticFileSetStat struct { dynamicBytesFileSetAttr vfs.StaticData diff --git a/pkg/sentry/fsimpl/proc/tasks_files.go b/pkg/sentry/fsimpl/proc/tasks_files.go index 8c41729e4..f268c59b0 100644 --- a/pkg/sentry/fsimpl/proc/tasks_files.go +++ b/pkg/sentry/fsimpl/proc/tasks_files.go @@ -31,6 +31,7 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) +// +stateify savable type selfSymlink struct { implStatFS kernfs.InodeAttrs @@ -51,7 +52,7 @@ func (fs *filesystem) newSelfSymlink(creds *auth.Credentials, ino uint64, pidns return d } -func (s *selfSymlink) Readlink(ctx context.Context) (string, error) { +func (s *selfSymlink) Readlink(ctx context.Context, _ *vfs.Mount) (string, error) { t := kernel.TaskFromContext(ctx) if t == nil { // Who is reading this link? @@ -64,16 +65,17 @@ 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) +func (s *selfSymlink) Getlink(ctx context.Context, mnt *vfs.Mount) (vfs.VirtualDentry, string, error) { + target, err := s.Readlink(ctx, mnt) return vfs.VirtualDentry{}, target, err } -// SetStat implements Inode.SetStat not allowing inode attributes to be changed. +// SetStat implements kernfs.Inode.SetStat not allowing inode attributes to be changed. func (*selfSymlink) SetStat(context.Context, *vfs.Filesystem, *auth.Credentials, vfs.SetStatOptions) error { return syserror.EPERM } +// +stateify savable type threadSelfSymlink struct { implStatFS kernfs.InodeAttrs @@ -94,7 +96,7 @@ func (fs *filesystem) newThreadSelfSymlink(creds *auth.Credentials, ino uint64, return d } -func (s *threadSelfSymlink) Readlink(ctx context.Context) (string, error) { +func (s *threadSelfSymlink) Readlink(ctx context.Context, _ *vfs.Mount) (string, error) { t := kernel.TaskFromContext(ctx) if t == nil { // Who is reading this link? @@ -108,12 +110,12 @@ 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) +func (s *threadSelfSymlink) Getlink(ctx context.Context, mnt *vfs.Mount) (vfs.VirtualDentry, string, error) { + target, err := s.Readlink(ctx, mnt) return vfs.VirtualDentry{}, target, err } -// SetStat implements Inode.SetStat not allowing inode attributes to be changed. +// SetStat implements kernfs.Inode.SetStat not allowing inode attributes to be changed. func (*threadSelfSymlink) SetStat(context.Context, *vfs.Filesystem, *auth.Credentials, vfs.SetStatOptions) error { return syserror.EPERM } @@ -121,16 +123,20 @@ func (*threadSelfSymlink) SetStat(context.Context, *vfs.Filesystem, *auth.Creden // 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. +// +// +stateify savable type dynamicBytesFileSetAttr struct { kernfs.DynamicBytesFile } -// SetStat implements Inode.SetStat. +// SetStat implements kernfs.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. +// +// +stateify savable type cpuStats struct { // user is time spent in userspace tasks with non-positive niceness. user uint64 diff --git a/pkg/sentry/fsimpl/proc/tasks_sys.go b/pkg/sentry/fsimpl/proc/tasks_sys.go index 038a194c7..3312b0418 100644 --- a/pkg/sentry/fsimpl/proc/tasks_sys.go +++ b/pkg/sentry/fsimpl/proc/tasks_sys.go @@ -27,9 +27,11 @@ import ( "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/tcpip/network/ipv4" "gvisor.dev/gvisor/pkg/usermem" ) +// +stateify savable type tcpMemDir int const ( @@ -67,6 +69,7 @@ func (fs *filesystem) newSysNetDir(root *auth.Credentials, k *kernel.Kernel) *ke "tcp_rmem": fs.newDentry(root, fs.NextIno(), 0644, &tcpMemData{stack: stack, dir: tcpRMem}), "tcp_sack": fs.newDentry(root, fs.NextIno(), 0644, &tcpSackData{stack: stack}), "tcp_wmem": fs.newDentry(root, fs.NextIno(), 0644, &tcpMemData{stack: stack, dir: tcpWMem}), + "ip_forward": fs.newDentry(root, fs.NextIno(), 0444, &ipForwarding{stack: stack}), // The following files are simple stubs until they are implemented in // netstack, most of these files are configuration related. We use the @@ -174,7 +177,7 @@ type tcpSackData struct { var _ vfs.WritableDynamicBytesSource = (*tcpSackData)(nil) -// Generate implements vfs.DynamicBytesSource. +// Generate implements vfs.DynamicBytesSource.Generate. func (d *tcpSackData) Generate(ctx context.Context, buf *bytes.Buffer) error { if d.enabled == nil { sack, err := d.stack.TCPSACKEnabled() @@ -232,7 +235,7 @@ type tcpRecoveryData struct { var _ vfs.WritableDynamicBytesSource = (*tcpRecoveryData)(nil) -// Generate implements vfs.DynamicBytesSource. +// Generate implements vfs.DynamicBytesSource.Generate. func (d *tcpRecoveryData) Generate(ctx context.Context, buf *bytes.Buffer) error { recovery, err := d.stack.TCPRecovery() if err != nil { @@ -284,7 +287,7 @@ type tcpMemData struct { var _ vfs.WritableDynamicBytesSource = (*tcpMemData)(nil) -// Generate implements vfs.DynamicBytesSource. +// Generate implements vfs.DynamicBytesSource.Generate. func (d *tcpMemData) Generate(ctx context.Context, buf *bytes.Buffer) error { d.mu.Lock() defer d.mu.Unlock() @@ -354,3 +357,63 @@ func (d *tcpMemData) writeSizeLocked(size inet.TCPBufferSize) error { panic(fmt.Sprintf("unknown tcpMemFile type: %v", d.dir)) } } + +// ipForwarding implements vfs.WritableDynamicBytesSource for +// /proc/sys/net/ipv4/ip_forwarding. +// +// +stateify savable +type ipForwarding struct { + kernfs.DynamicBytesFile + + stack inet.Stack `state:"wait"` + enabled *bool +} + +var _ vfs.WritableDynamicBytesSource = (*ipForwarding)(nil) + +// Generate implements vfs.DynamicBytesSource.Generate. +func (ipf *ipForwarding) Generate(ctx context.Context, buf *bytes.Buffer) error { + if ipf.enabled == nil { + enabled := ipf.stack.Forwarding(ipv4.ProtocolNumber) + ipf.enabled = &enabled + } + + val := "0\n" + if *ipf.enabled { + // Technically, this is not quite compatible with Linux. Linux stores these + // as an integer, so if you write "2" into tcp_sack, you should get 2 back. + // Tough luck. + val = "1\n" + } + buf.WriteString(val) + + return nil +} + +// Write implements vfs.WritableDynamicBytesSource.Write. +func (ipf *ipForwarding) Write(ctx context.Context, src usermem.IOSequence, offset int64) (int64, error) { + if offset != 0 { + // No need to handle partial writes thus far. + return 0, syserror.EINVAL + } + if src.NumBytes() == 0 { + return 0, nil + } + + // Limit input size so as not to impact performance if input size is large. + src = src.TakeFirst(usermem.PageSize - 1) + + var v int32 + n, err := usermem.CopyInt32StringInVec(ctx, src.IO, src.Addrs, &v, src.Opts) + if err != nil { + return 0, err + } + if ipf.enabled == nil { + ipf.enabled = new(bool) + } + *ipf.enabled = v != 0 + if err := ipf.stack.SetForwarding(ipv4.ProtocolNumber, *ipf.enabled); err != nil { + return 0, err + } + return n, nil +} diff --git a/pkg/sentry/fsimpl/proc/tasks_sys_test.go b/pkg/sentry/fsimpl/proc/tasks_sys_test.go index be54897bb..6cee22823 100644 --- a/pkg/sentry/fsimpl/proc/tasks_sys_test.go +++ b/pkg/sentry/fsimpl/proc/tasks_sys_test.go @@ -20,8 +20,10 @@ import ( "testing" "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/contexttest" "gvisor.dev/gvisor/pkg/sentry/inet" + "gvisor.dev/gvisor/pkg/usermem" ) func newIPv6TestStack() *inet.TestStack { @@ -76,3 +78,72 @@ func TestIfinet6(t *testing.T) { t.Errorf("Got n.contents() = %v, want = %v", got, want) } } + +// TestIPForwarding tests the implementation of +// /proc/sys/net/ipv4/ip_forwarding +func TestConfigureIPForwarding(t *testing.T) { + ctx := context.Background() + s := inet.NewTestStack() + + var cases = []struct { + comment string + initial bool + str string + final bool + }{ + { + comment: `Forwarding is disabled; write 1 and enable forwarding`, + initial: false, + str: "1", + final: true, + }, + { + comment: `Forwarding is disabled; write 0 and disable forwarding`, + initial: false, + str: "0", + final: false, + }, + { + comment: `Forwarding is enabled; write 1 and enable forwarding`, + initial: true, + str: "1", + final: true, + }, + { + comment: `Forwarding is enabled; write 0 and disable forwarding`, + initial: true, + str: "0", + final: false, + }, + { + comment: `Forwarding is disabled; write 2404 and enable forwarding`, + initial: false, + str: "2404", + final: true, + }, + { + comment: `Forwarding is enabled; write 2404 and enable forwarding`, + initial: true, + str: "2404", + final: true, + }, + } + for _, c := range cases { + t.Run(c.comment, func(t *testing.T) { + s.IPForwarding = c.initial + + file := &ipForwarding{stack: s, enabled: &c.initial} + + // Write the values. + src := usermem.BytesIOSequence([]byte(c.str)) + if n, err := file.Write(ctx, src, 0); n != int64(len(c.str)) || err != nil { + t.Errorf("file.Write(ctx, nil, %q, 0) = (%d, %v); want (%d, nil)", c.str, n, err, len(c.str)) + } + + // Read the values from the stack and check them. + if got, want := s.IPForwarding, c.final; got != want { + t.Errorf("s.IPForwarding incorrect; got: %v, want: %v", got, want) + } + }) + } +} diff --git a/pkg/sentry/fsimpl/proc/tasks_test.go b/pkg/sentry/fsimpl/proc/tasks_test.go index f693f9060..6975af5a7 100644 --- a/pkg/sentry/fsimpl/proc/tasks_test.go +++ b/pkg/sentry/fsimpl/proc/tasks_test.go @@ -67,6 +67,7 @@ var ( taskStaticFiles = map[string]testutil.DirentType{ "auxv": linux.DT_REG, "cgroup": linux.DT_REG, + "cwd": linux.DT_LNK, "cmdline": linux.DT_REG, "comm": linux.DT_REG, "environ": linux.DT_REG, diff --git a/pkg/sentry/fsimpl/signalfd/BUILD b/pkg/sentry/fsimpl/signalfd/BUILD index 067c1657f..adb610213 100644 --- a/pkg/sentry/fsimpl/signalfd/BUILD +++ b/pkg/sentry/fsimpl/signalfd/BUILD @@ -8,7 +8,6 @@ go_library( visibility = ["//pkg/sentry:internal"], deps = [ "//pkg/abi/linux", - "//pkg/binary", "//pkg/context", "//pkg/sentry/kernel", "//pkg/sentry/vfs", diff --git a/pkg/sentry/fsimpl/signalfd/signalfd.go b/pkg/sentry/fsimpl/signalfd/signalfd.go index 6297e1df4..10f1452ef 100644 --- a/pkg/sentry/fsimpl/signalfd/signalfd.go +++ b/pkg/sentry/fsimpl/signalfd/signalfd.go @@ -16,7 +16,6 @@ package signalfd import ( "gvisor.dev/gvisor/pkg/abi/linux" - "gvisor.dev/gvisor/pkg/binary" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/vfs" @@ -26,7 +25,9 @@ import ( "gvisor.dev/gvisor/pkg/waiter" ) -// SignalFileDescription implements FileDescriptionImpl for signal fds. +// SignalFileDescription implements vfs.FileDescriptionImpl for signal fds. +// +// +stateify savable type SignalFileDescription struct { vfsfd vfs.FileDescription vfs.FileDescriptionDefaultImpl @@ -43,7 +44,7 @@ type SignalFileDescription struct { target *kernel.Task // mu protects mask. - mu sync.Mutex + mu sync.Mutex `state:"nosave"` // mask is the signal mask. Protected by mu. mask linux.SignalSet @@ -83,7 +84,7 @@ func (sfd *SignalFileDescription) SetMask(mask linux.SignalSet) { sfd.mask = mask } -// Read implements FileDescriptionImpl.Read. +// Read implements vfs.FileDescriptionImpl.Read. func (sfd *SignalFileDescription) Read(ctx context.Context, dst usermem.IOSequence, _ vfs.ReadOptions) (int64, error) { // Attempt to dequeue relevant signals. info, err := sfd.target.Sigtimedwait(sfd.Mask(), 0) @@ -93,8 +94,7 @@ func (sfd *SignalFileDescription) Read(ctx context.Context, dst usermem.IOSequen } // Copy out the signal info using the specified format. - var buf [128]byte - binary.Marshal(buf[:0], usermem.ByteOrder, &linux.SignalfdSiginfo{ + infoNative := linux.SignalfdSiginfo{ Signo: uint32(info.Signo), Errno: info.Errno, Code: info.Code, @@ -103,9 +103,13 @@ func (sfd *SignalFileDescription) Read(ctx context.Context, dst usermem.IOSequen Status: info.Status(), Overrun: uint32(info.Overrun()), Addr: info.Addr(), - }) - n, err := dst.CopyOut(ctx, buf[:]) - return int64(n), err + } + n, err := infoNative.WriteTo(dst.Writer(ctx)) + if err == usermem.ErrEndOfIOSequence { + // Partial copy-out ok. + err = nil + } + return n, err } // Readiness implements waiter.Waitable.Readiness. @@ -132,5 +136,5 @@ func (sfd *SignalFileDescription) EventUnregister(entry *waiter.Entry) { sfd.target.SignalUnregister(entry) } -// Release implements FileDescriptionImpl.Release() +// Release implements vfs.FileDescriptionImpl.Release. func (sfd *SignalFileDescription) Release(context.Context) {} diff --git a/pkg/sentry/fsimpl/sockfs/sockfs.go b/pkg/sentry/fsimpl/sockfs/sockfs.go index 94a998568..29e5371d6 100644 --- a/pkg/sentry/fsimpl/sockfs/sockfs.go +++ b/pkg/sentry/fsimpl/sockfs/sockfs.go @@ -28,14 +28,16 @@ import ( ) // filesystemType implements vfs.FilesystemType. +// +// +stateify savable type filesystemType struct{} -// GetFilesystem implements FilesystemType.GetFilesystem. +// GetFilesystem implements vfs.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. +// Name implements vfs.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 @@ -44,6 +46,7 @@ func (filesystemType) Name() string { return "sockfs" } +// +stateify savable type filesystem struct { kernfs.Filesystem @@ -80,6 +83,8 @@ func (fs *filesystem) PrependPath(ctx context.Context, vfsroot, vd vfs.VirtualDe } // inode implements kernfs.Inode. +// +// +stateify savable type inode struct { kernfs.InodeAttrs kernfs.InodeNoopRefCount @@ -88,7 +93,7 @@ type inode struct { } // Open implements kernfs.Inode.Open. -func (i *inode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { +func (i *inode) Open(ctx context.Context, rp *vfs.ResolvingPath, d *kernfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { return nil, syserror.ENXIO } diff --git a/pkg/sentry/fsimpl/sys/kcov.go b/pkg/sentry/fsimpl/sys/kcov.go index 73f3d3309..1a6749e53 100644 --- a/pkg/sentry/fsimpl/sys/kcov.go +++ b/pkg/sentry/fsimpl/sys/kcov.go @@ -36,6 +36,8 @@ func (fs *filesystem) newKcovFile(ctx context.Context, creds *auth.Credentials) } // kcovInode implements kernfs.Inode. +// +// +stateify savable type kcovInode struct { kernfs.InodeAttrs kernfs.InodeNoopRefCount @@ -44,7 +46,7 @@ type kcovInode struct { implStatFS } -func (i *kcovInode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { +func (i *kcovInode) Open(ctx context.Context, rp *vfs.ResolvingPath, d *kernfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { k := kernel.KernelFromContext(ctx) if k == nil { panic("KernelFromContext returned nil") @@ -54,7 +56,7 @@ func (i *kcovInode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.D kcov: k.NewKcov(), } - if err := fd.vfsfd.Init(fd, opts.Flags, rp.Mount(), vfsd, &vfs.FileDescriptionOptions{ + if err := fd.vfsfd.Init(fd, opts.Flags, rp.Mount(), d.VFSDentry(), &vfs.FileDescriptionOptions{ DenyPRead: true, DenyPWrite: true, }); err != nil { @@ -63,6 +65,7 @@ func (i *kcovInode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.D return &fd.vfsfd, nil } +// +stateify savable type kcovFD struct { vfs.FileDescriptionDefaultImpl vfs.NoLockFD @@ -101,7 +104,7 @@ func (fd *kcovFD) ConfigureMMap(ctx context.Context, opts *memmap.MMapOpts) erro func (fd *kcovFD) Release(ctx context.Context) { // kcov instances have reference counts in Linux, but this seems sufficient // for our purposes. - fd.kcov.Reset() + fd.kcov.Clear() } // SetStat implements vfs.FileDescriptionImpl.SetStat. diff --git a/pkg/sentry/fsimpl/sys/sys.go b/pkg/sentry/fsimpl/sys/sys.go index 39952d2d0..1568c581f 100644 --- a/pkg/sentry/fsimpl/sys/sys.go +++ b/pkg/sentry/fsimpl/sys/sys.go @@ -34,9 +34,13 @@ const Name = "sysfs" const defaultSysDirMode = linux.FileMode(0755) // FilesystemType implements vfs.FilesystemType. +// +// +stateify savable type FilesystemType struct{} // filesystem implements vfs.FilesystemImpl. +// +// +stateify savable type filesystem struct { kernfs.Filesystem @@ -117,6 +121,8 @@ func (fs *filesystem) Release(ctx context.Context) { } // dir implements kernfs.Inode. +// +// +stateify savable type dir struct { dirRefs kernfs.InodeAttrs @@ -148,8 +154,8 @@ func (*dir) SetStat(context.Context, *vfs.Filesystem, *auth.Credentials, vfs.Set } // Open implements kernfs.Inode.Open. -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, &d.locks, &opts, kernfs.GenericDirectoryFDOptions{ +func (d *dir) Open(ctx context.Context, rp *vfs.ResolvingPath, kd *kernfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), kd, &d.OrderedChildren, &d.locks, &opts, kernfs.GenericDirectoryFDOptions{ SeekEnd: kernfs.SeekEndStaticEntries, }) if err != nil { @@ -169,6 +175,8 @@ func (d *dir) StatFS(ctx context.Context, fs *vfs.Filesystem) (linux.Statfs, err } // cpuFile implements kernfs.Inode. +// +// +stateify savable type cpuFile struct { implStatFS kernfs.DynamicBytesFile @@ -190,6 +198,7 @@ func (fs *filesystem) newCPUFile(creds *auth.Credentials, maxCores uint, mode li return d } +// +stateify savable type implStatFS struct{} // StatFS implements kernfs.Inode.StatFS. diff --git a/pkg/sentry/fsimpl/timerfd/timerfd.go b/pkg/sentry/fsimpl/timerfd/timerfd.go index 86beaa0a8..8853c8ad2 100644 --- a/pkg/sentry/fsimpl/timerfd/timerfd.go +++ b/pkg/sentry/fsimpl/timerfd/timerfd.go @@ -26,8 +26,10 @@ import ( "gvisor.dev/gvisor/pkg/waiter" ) -// TimerFileDescription implements FileDescriptionImpl for timer fds. It also +// TimerFileDescription implements vfs.FileDescriptionImpl for timer fds. It also // implements ktime.TimerListener. +// +// +stateify savable type TimerFileDescription struct { vfsfd vfs.FileDescription vfs.FileDescriptionDefaultImpl @@ -62,7 +64,7 @@ func New(ctx context.Context, vfsObj *vfs.VirtualFilesystem, clock ktime.Clock, return &tfd.vfsfd, nil } -// Read implements FileDescriptionImpl.Read. +// Read implements vfs.FileDescriptionImpl.Read. func (tfd *TimerFileDescription) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { const sizeofUint64 = 8 if dst.NumBytes() < sizeofUint64 { @@ -128,7 +130,7 @@ func (tfd *TimerFileDescription) ResumeTimer() { tfd.timer.Resume() } -// Release implements FileDescriptionImpl.Release() +// Release implements vfs.FileDescriptionImpl.Release. func (tfd *TimerFileDescription) Release(context.Context) { tfd.timer.Destroy() } diff --git a/pkg/sentry/fsimpl/tmpfs/device_file.go b/pkg/sentry/fsimpl/tmpfs/device_file.go index ac54d420d..9129d35b7 100644 --- a/pkg/sentry/fsimpl/tmpfs/device_file.go +++ b/pkg/sentry/fsimpl/tmpfs/device_file.go @@ -22,6 +22,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/vfs" ) +// +stateify savable type deviceFile struct { inode inode kind vfs.DeviceKind diff --git a/pkg/sentry/fsimpl/tmpfs/directory.go b/pkg/sentry/fsimpl/tmpfs/directory.go index 070c75e68..e90669cf0 100644 --- a/pkg/sentry/fsimpl/tmpfs/directory.go +++ b/pkg/sentry/fsimpl/tmpfs/directory.go @@ -25,6 +25,7 @@ import ( "gvisor.dev/gvisor/pkg/syserror" ) +// +stateify savable type directory struct { // Since directories can't be hard-linked, each directory can only be // associated with a single dentry, which we can store in the directory @@ -44,7 +45,7 @@ type directory struct { // (with inode == nil) that represent the iteration position of // directoryFDs. childList is used to support directoryFD.IterDirents() // efficiently. childList is protected by iterMu. - iterMu sync.Mutex + iterMu sync.Mutex `state:"nosave"` childList dentryList } @@ -86,6 +87,7 @@ func (dir *directory) mayDelete(creds *auth.Credentials, child *dentry) error { return vfs.CheckDeleteSticky(creds, linux.FileMode(atomic.LoadUint32(&dir.inode.mode)), auth.KUID(atomic.LoadUint32(&child.inode.uid))) } +// +stateify savable type directoryFD struct { fileDescription vfs.DirectoryFileDescriptionDefaultImpl diff --git a/pkg/sentry/fsimpl/tmpfs/filesystem.go b/pkg/sentry/fsimpl/tmpfs/filesystem.go index 61d925a73..e39cd305b 100644 --- a/pkg/sentry/fsimpl/tmpfs/filesystem.go +++ b/pkg/sentry/fsimpl/tmpfs/filesystem.go @@ -673,11 +673,11 @@ func (fs *filesystem) SetStatAt(ctx context.Context, rp *vfs.ResolvingPath, opts fs.mu.RUnlock() return err } - if err := d.inode.setStat(ctx, rp.Credentials(), &opts); err != nil { - fs.mu.RUnlock() + err = d.inode.setStat(ctx, rp.Credentials(), &opts) + fs.mu.RUnlock() + if err != nil { return err } - fs.mu.RUnlock() if ev := vfs.InotifyEventFromStatMask(opts.Stat.Mask); ev != 0 { d.InotifyWithParent(ctx, ev, 0, vfs.InodeEvent) @@ -770,7 +770,7 @@ func (fs *filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error return nil } -// BoundEndpointAt implements FilesystemImpl.BoundEndpointAt. +// BoundEndpointAt implements vfs.FilesystemImpl.BoundEndpointAt. func (fs *filesystem) BoundEndpointAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.BoundEndpointOptions) (transport.BoundEndpoint, error) { fs.mu.RLock() defer fs.mu.RUnlock() @@ -822,11 +822,11 @@ func (fs *filesystem) SetXattrAt(ctx context.Context, rp *vfs.ResolvingPath, opt fs.mu.RUnlock() return err } - if err := d.inode.setXattr(rp.Credentials(), &opts); err != nil { - fs.mu.RUnlock() + err = d.inode.setXattr(rp.Credentials(), &opts) + fs.mu.RUnlock() + if err != nil { return err } - fs.mu.RUnlock() d.InotifyWithParent(ctx, linux.IN_ATTRIB, 0, vfs.InodeEvent) return nil @@ -840,11 +840,11 @@ func (fs *filesystem) RemoveXattrAt(ctx context.Context, rp *vfs.ResolvingPath, fs.mu.RUnlock() return err } - if err := d.inode.removeXattr(rp.Credentials(), name); err != nil { - fs.mu.RUnlock() + err = d.inode.removeXattr(rp.Credentials(), name) + fs.mu.RUnlock() + if err != nil { return err } - fs.mu.RUnlock() d.InotifyWithParent(ctx, linux.IN_ATTRIB, 0, vfs.InodeEvent) return nil @@ -865,8 +865,16 @@ func (fs *filesystem) PrependPath(ctx context.Context, vfsroot, vd vfs.VirtualDe } if d.parent == nil { if d.name != "" { - // This must be an anonymous memfd file. + // This file must have been created by + // newUnlinkedRegularFileDescription(). In Linux, + // mm/shmem.c:__shmem_file_setup() => + // fs/file_table.c:alloc_file_pseudo() sets the created + // dentry's dentry_operations to anon_ops, for which d_dname == + // simple_dname. fs/d_path.c:simple_dname() defines the + // dentry's pathname to be its name, prefixed with "/" and + // suffixed with " (deleted)". b.PrependComponent("/" + d.name) + b.AppendString(" (deleted)") return vfs.PrependPathSyntheticError{} } return vfs.PrependPathAtNonMountRootError{} diff --git a/pkg/sentry/fsimpl/tmpfs/named_pipe.go b/pkg/sentry/fsimpl/tmpfs/named_pipe.go index 5b0471ff4..d772db9e9 100644 --- a/pkg/sentry/fsimpl/tmpfs/named_pipe.go +++ b/pkg/sentry/fsimpl/tmpfs/named_pipe.go @@ -21,6 +21,7 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) +// +stateify savable type namedPipe struct { inode inode diff --git a/pkg/sentry/fsimpl/tmpfs/regular_file.go b/pkg/sentry/fsimpl/tmpfs/regular_file.go index 0710b65db..a199eb33d 100644 --- a/pkg/sentry/fsimpl/tmpfs/regular_file.go +++ b/pkg/sentry/fsimpl/tmpfs/regular_file.go @@ -36,12 +36,18 @@ import ( ) // regularFile is a regular (=S_IFREG) tmpfs file. +// +// +stateify savable type regularFile struct { inode inode // memFile is a platform.File used to allocate pages to this regularFile. memFile *pgalloc.MemoryFile + // memoryUsageKind is the memory accounting category under which pages backing + // this regularFile's contents are accounted. + memoryUsageKind usage.MemoryKind + // mapsMu protects mappings. mapsMu sync.Mutex `state:"nosave"` @@ -62,7 +68,7 @@ type regularFile struct { writableMappingPages uint64 // dataMu protects the fields below. - dataMu sync.RWMutex + dataMu sync.RWMutex `state:"nosave"` // data maps offsets into the file to offsets into memFile that store // the file's data. @@ -86,14 +92,75 @@ type regularFile struct { func (fs *filesystem) newRegularFile(kuid auth.KUID, kgid auth.KGID, mode linux.FileMode) *inode { file := ®ularFile{ - memFile: fs.memFile, - seals: linux.F_SEAL_SEAL, + memFile: fs.memFile, + memoryUsageKind: usage.Tmpfs, + seals: linux.F_SEAL_SEAL, } file.inode.init(file, fs, kuid, kgid, linux.S_IFREG|mode) file.inode.nlink = 1 // from parent directory return &file.inode } +// newUnlinkedRegularFileDescription creates a regular file on the tmpfs +// filesystem represented by mount and returns an FD representing that file. +// The new file is not reachable by path traversal from any other file. +// +// newUnlinkedRegularFileDescription is analogous to Linux's +// mm/shmem.c:__shmem_file_setup(). +// +// Preconditions: mount must be a tmpfs mount. +func newUnlinkedRegularFileDescription(ctx context.Context, creds *auth.Credentials, mount *vfs.Mount, name string) (*regularFileFD, error) { + fs, ok := mount.Filesystem().Impl().(*filesystem) + if !ok { + panic("tmpfs.newUnlinkedRegularFileDescription() called with non-tmpfs mount") + } + + inode := fs.newRegularFile(creds.EffectiveKUID, creds.EffectiveKGID, 0777) + d := fs.newDentry(inode) + defer d.DecRef(ctx) + d.name = name + + fd := ®ularFileFD{} + fd.Init(&inode.locks) + flags := uint32(linux.O_RDWR) + if err := fd.vfsfd.Init(fd, flags, mount, &d.vfsd, &vfs.FileDescriptionOptions{}); err != nil { + return nil, err + } + return fd, nil +} + +// NewZeroFile creates a new regular file and file description as for +// mmap(MAP_SHARED | MAP_ANONYMOUS). The file has the given size and is +// initially (implicitly) filled with zeroes. +// +// Preconditions: mount must be a tmpfs mount. +func NewZeroFile(ctx context.Context, creds *auth.Credentials, mount *vfs.Mount, size uint64) (*vfs.FileDescription, error) { + // Compare mm/shmem.c:shmem_zero_setup(). + fd, err := newUnlinkedRegularFileDescription(ctx, creds, mount, "dev/zero") + if err != nil { + return nil, err + } + rf := fd.inode().impl.(*regularFile) + rf.memoryUsageKind = usage.Anonymous + rf.size = size + return &fd.vfsfd, err +} + +// NewMemfd creates a new regular file and file description as for +// memfd_create. +// +// Preconditions: mount must be a tmpfs mount. +func NewMemfd(ctx context.Context, creds *auth.Credentials, mount *vfs.Mount, allowSeals bool, name string) (*vfs.FileDescription, error) { + fd, err := newUnlinkedRegularFileDescription(ctx, creds, mount, name) + if err != nil { + return nil, err + } + if allowSeals { + fd.inode().impl.(*regularFile).seals = 0 + } + return &fd.vfsfd, nil +} + // truncate grows or shrinks the file to the given size. It returns true if the // file size was updated. func (rf *regularFile) truncate(newSize uint64) (bool, error) { @@ -226,7 +293,7 @@ func (rf *regularFile) Translate(ctx context.Context, required, optional memmap. optional.End = pgend } - cerr := rf.data.Fill(ctx, required, optional, rf.memFile, usage.Tmpfs, func(_ context.Context, dsts safemem.BlockSeq, _ uint64) (uint64, error) { + cerr := rf.data.Fill(ctx, required, optional, rf.memFile, rf.memoryUsageKind, func(_ context.Context, dsts safemem.BlockSeq, _ uint64) (uint64, error) { // Newly-allocated pages are zeroed, so we don't need to do anything. return dsts.NumBytes(), nil }) @@ -260,13 +327,14 @@ func (*regularFile) InvalidateUnsavable(context.Context) error { return nil } +// +stateify savable type regularFileFD struct { fileDescription // off is the file offset. off is accessed using atomic memory operations. // offMu serializes operations that may mutate off. off int64 - offMu sync.Mutex + offMu sync.Mutex `state:"nosave"` } // Release implements vfs.FileDescriptionImpl.Release. @@ -575,7 +643,7 @@ func (rw *regularFileReadWriter) WriteFromBlocks(srcs safemem.BlockSeq) (uint64, case gap.Ok(): // Allocate memory for the write. gapMR := gap.Range().Intersect(pgMR) - fr, err := rw.file.memFile.Allocate(gapMR.Length(), usage.Tmpfs) + fr, err := rw.file.memFile.Allocate(gapMR.Length(), rw.file.memoryUsageKind) if err != nil { retErr = err goto exitLoop diff --git a/pkg/sentry/fsimpl/tmpfs/socket_file.go b/pkg/sentry/fsimpl/tmpfs/socket_file.go index 3ed650474..5699d5975 100644 --- a/pkg/sentry/fsimpl/tmpfs/socket_file.go +++ b/pkg/sentry/fsimpl/tmpfs/socket_file.go @@ -21,6 +21,8 @@ import ( ) // socketFile is a socket (=S_IFSOCK) tmpfs file. +// +// +stateify savable type socketFile struct { inode inode ep transport.BoundEndpoint diff --git a/pkg/sentry/fsimpl/tmpfs/symlink.go b/pkg/sentry/fsimpl/tmpfs/symlink.go index b0de5fabe..a102a2ee2 100644 --- a/pkg/sentry/fsimpl/tmpfs/symlink.go +++ b/pkg/sentry/fsimpl/tmpfs/symlink.go @@ -19,6 +19,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/kernel/auth" ) +// +stateify savable type symlink struct { inode inode target string // immutable diff --git a/pkg/sentry/fsimpl/tmpfs/tmpfs.go b/pkg/sentry/fsimpl/tmpfs/tmpfs.go index 4871e55d3..cefec8fde 100644 --- a/pkg/sentry/fsimpl/tmpfs/tmpfs.go +++ b/pkg/sentry/fsimpl/tmpfs/tmpfs.go @@ -51,9 +51,13 @@ import ( const Name = "tmpfs" // FilesystemType implements vfs.FilesystemType. +// +// +stateify savable type FilesystemType struct{} // filesystem implements vfs.FilesystemImpl. +// +// +stateify savable type filesystem struct { vfsfs vfs.Filesystem @@ -67,7 +71,7 @@ type filesystem struct { devMinor uint32 // mu serializes changes to the Dentry tree. - mu sync.RWMutex + mu sync.RWMutex `state:"nosave"` nextInoMinusOne uint64 // accessed using atomic memory operations } @@ -78,6 +82,8 @@ func (FilesystemType) Name() string { } // FilesystemOpts is used to pass configuration data to tmpfs. +// +// +stateify savable 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. @@ -221,6 +227,8 @@ var globalStatfs = linux.Statfs{ } // dentry implements vfs.DentryImpl. +// +// +stateify savable type dentry struct { vfsd vfs.Dentry @@ -300,6 +308,8 @@ func (d *dentry) Watches() *vfs.Watches { func (d *dentry) OnZeroWatches(context.Context) {} // inode represents a filesystem object. +// +// +stateify savable type inode struct { // fs is the owning filesystem. fs is immutable. fs *filesystem @@ -316,12 +326,12 @@ type inode struct { // Inode metadata. Writing multiple fields atomically requires holding // mu, othewise atomic operations can be used. - mu sync.Mutex - 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 ... - ino uint64 // immutable + mu sync.Mutex `state:"nosave"` + 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 ... + ino uint64 // immutable // Linux's tmpfs has no concept of btime. atime int64 // nanoseconds @@ -668,6 +678,8 @@ func (i *inode) checkXattrPermissions(creds *auth.Credentials, name string, ats // fileDescription is embedded by tmpfs implementations of // vfs.FileDescriptionImpl. +// +// +stateify savable type fileDescription struct { vfsfd vfs.FileDescription vfs.FileDescriptionDefaultImpl @@ -746,37 +758,6 @@ func (fd *fileDescription) RemoveXattr(ctx context.Context, name string) error { return nil } -// NewMemfd creates a new tmpfs regular file and file description that can back -// an anonymous fd created by memfd_create. -func NewMemfd(ctx context.Context, creds *auth.Credentials, mount *vfs.Mount, allowSeals bool, name string) (*vfs.FileDescription, error) { - fs, ok := mount.Filesystem().Impl().(*filesystem) - if !ok { - panic("NewMemfd() called with non-tmpfs mount") - } - - // Per Linux, mm/shmem.c:__shmem_file_setup(), memfd inodes are set up with - // S_IRWXUGO. - inode := fs.newRegularFile(creds.EffectiveKUID, creds.EffectiveKGID, 0777) - rf := inode.impl.(*regularFile) - if allowSeals { - rf.seals = 0 - } - - d := fs.newDentry(inode) - defer d.DecRef(ctx) - d.name = name - - // Per Linux, mm/shmem.c:__shmem_file_setup(), memfd files are set up with - // FMODE_READ | FMODE_WRITE. - var fd regularFileFD - fd.Init(&inode.locks) - flags := uint32(linux.O_RDWR) - if err := fd.vfsfd.Init(&fd, flags, mount, &d.vfsd, &vfs.FileDescriptionOptions{}); err != nil { - return nil, err - } - return &fd.vfsfd, nil -} - // LockPOSIX implements vfs.FileDescriptionImpl.LockPOSIX. func (fd *fileDescription) LockPOSIX(ctx context.Context, uid fslock.UniqueID, t fslock.LockType, start, length uint64, whence int16, block fslock.Blocker) error { return fd.Locks().LockPOSIX(ctx, &fd.vfsfd, uid, t, start, length, whence, block) diff --git a/pkg/sentry/fsimpl/verity/BUILD b/pkg/sentry/fsimpl/verity/BUILD index bc8e38431..0ca750281 100644 --- a/pkg/sentry/fsimpl/verity/BUILD +++ b/pkg/sentry/fsimpl/verity/BUILD @@ -1,4 +1,4 @@ -load("//tools:defs.bzl", "go_library") +load("//tools:defs.bzl", "go_library", "go_test") licenses(["notice"]) @@ -26,3 +26,22 @@ go_library( "//pkg/usermem", ], ) + +go_test( + name = "verity_test", + srcs = [ + "verity_test.go", + ], + library = ":verity", + deps = [ + "//pkg/abi/linux", + "//pkg/context", + "//pkg/fspath", + "//pkg/sentry/arch", + "//pkg/sentry/fsimpl/tmpfs", + "//pkg/sentry/kernel/auth", + "//pkg/sentry/kernel/contexttest", + "//pkg/sentry/vfs", + "//pkg/usermem", + ], +) diff --git a/pkg/sentry/fsimpl/verity/filesystem.go b/pkg/sentry/fsimpl/verity/filesystem.go index 2cf0a38c9..a560b0797 100644 --- a/pkg/sentry/fsimpl/verity/filesystem.go +++ b/pkg/sentry/fsimpl/verity/filesystem.go @@ -19,6 +19,7 @@ import ( "fmt" "io" "strconv" + "strings" "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" @@ -184,18 +185,14 @@ func (fs *filesystem) verifyChild(ctx context.Context, parent *dentry, child *de Start: child.lowerMerkleVD, }, &vfs.GetXattrOptions{ Name: merkleOffsetInParentXattr, - // Offset is a 32 bit integer. - Size: sizeOfInt32, + Size: sizeOfStringInt32, }) // The Merkle tree file for the child should have been created and // contains the expected xattrs. If the file or the xattr does not // exist, it indicates unexpected modifications to the file system. if err == syserror.ENOENT || err == syserror.ENODATA { - if noCrashOnVerificationFailure { - return nil, err - } - panic(fmt.Sprintf("Failed to get xattr %s for %s: %v", merkleOffsetInParentXattr, childPath, err)) + return nil, alertIntegrityViolation(err, fmt.Sprintf("Failed to get xattr %s for %s: %v", merkleOffsetInParentXattr, childPath, err)) } if err != nil { return nil, err @@ -204,10 +201,7 @@ func (fs *filesystem) verifyChild(ctx context.Context, parent *dentry, child *de // unexpected modifications to the file system. offset, err := strconv.Atoi(off) if err != nil { - if noCrashOnVerificationFailure { - return nil, syserror.EINVAL - } - panic(fmt.Sprintf("Failed to convert xattr %s for %s to int: %v", merkleOffsetInParentXattr, childPath, err)) + return nil, alertIntegrityViolation(err, fmt.Sprintf("Failed to convert xattr %s for %s to int: %v", merkleOffsetInParentXattr, childPath, err)) } // Open parent Merkle tree file to read and verify child's root hash. @@ -221,10 +215,7 @@ func (fs *filesystem) verifyChild(ctx context.Context, parent *dentry, child *de // The parent Merkle tree file should have been created. If it's // missing, it indicates an unexpected modification to the file system. if err == syserror.ENOENT { - if noCrashOnVerificationFailure { - return nil, err - } - panic(fmt.Sprintf("Failed to open parent Merkle file for %s: %v", childPath, err)) + return nil, alertIntegrityViolation(err, fmt.Sprintf("Failed to open parent Merkle file for %s: %v", childPath, err)) } if err != nil { return nil, err @@ -235,17 +226,14 @@ func (fs *filesystem) verifyChild(ctx context.Context, parent *dentry, child *de // the size of all its children's root hashes. dataSize, err := parentMerkleFD.GetXattr(ctx, &vfs.GetXattrOptions{ Name: merkleSizeXattr, - Size: sizeOfInt32, + Size: sizeOfStringInt32, }) // The Merkle tree file for the child should have been created and // contains the expected xattrs. If the file or the xattr does not // exist, it indicates unexpected modifications to the file system. if err == syserror.ENOENT || err == syserror.ENODATA { - if noCrashOnVerificationFailure { - return nil, err - } - panic(fmt.Sprintf("Failed to get xattr %s for %s: %v", merkleSizeXattr, childPath, err)) + return nil, alertIntegrityViolation(err, fmt.Sprintf("Failed to get xattr %s for %s: %v", merkleSizeXattr, childPath, err)) } if err != nil { return nil, err @@ -255,10 +243,7 @@ func (fs *filesystem) verifyChild(ctx context.Context, parent *dentry, child *de // unexpected modifications to the file system. parentSize, err := strconv.Atoi(dataSize) if err != nil { - if noCrashOnVerificationFailure { - return nil, syserror.EINVAL - } - panic(fmt.Sprintf("Failed to convert xattr %s for %s to int: %v", merkleSizeXattr, childPath, err)) + return nil, alertIntegrityViolation(syserror.EINVAL, fmt.Sprintf("Failed to convert xattr %s for %s to int: %v", merkleSizeXattr, childPath, err)) } fdReader := vfs.FileReadWriteSeeker{ @@ -270,11 +255,8 @@ func (fs *filesystem) verifyChild(ctx context.Context, parent *dentry, child *de // contain the root hash of the children in the parent Merkle tree when // Verify returns with success. var buf bytes.Buffer - if err := merkletree.Verify(&buf, &fdReader, &fdReader, int64(parentSize), int64(offset), int64(merkletree.DigestSize()), parent.rootHash, true /* dataAndTreeInSameFile */); err != nil && err != io.EOF { - if noCrashOnVerificationFailure { - return nil, syserror.EIO - } - panic(fmt.Sprintf("Verification for %s failed: %v", childPath, err)) + if _, err := merkletree.Verify(&buf, &fdReader, &fdReader, int64(parentSize), int64(offset), int64(merkletree.DigestSize()), parent.rootHash, true /* dataAndTreeInSameFile */); err != nil && err != io.EOF { + return nil, alertIntegrityViolation(syserror.EIO, fmt.Sprintf("Verification for %s failed: %v", childPath, err)) } // Cache child root hash when it's verified the first time. @@ -370,10 +352,7 @@ func (fs *filesystem) lookupAndVerifyLocked(ctx context.Context, parent *dentry, // corresponding Merkle tree is found. This indicates an // unexpected modification to the file system that // removed/renamed the child. - if noCrashOnVerificationFailure { - return nil, childErr - } - panic(fmt.Sprintf("Target file %s is expected but missing", parentPath+"/"+name)) + return nil, alertIntegrityViolation(childErr, fmt.Sprintf("Target file %s is expected but missing", parentPath+"/"+name)) } else if childErr == nil && childMerkleErr == syserror.ENOENT { // If in allowRuntimeEnable mode, and the Merkle tree file is // not created yet, we create an empty Merkle tree file, so that @@ -392,6 +371,7 @@ func (fs *filesystem) lookupAndVerifyLocked(ctx context.Context, parent *dentry, Path: fspath.Parse(childMerkleFilename), }, &vfs.OpenOptions{ Flags: linux.O_RDWR | linux.O_CREAT, + Mode: 0644, }) if err != nil { return nil, err @@ -409,10 +389,7 @@ func (fs *filesystem) lookupAndVerifyLocked(ctx context.Context, parent *dentry, // If runtime enable is not allowed. This indicates an // unexpected modification to the file system that // removed/renamed the Merkle tree file. - if noCrashOnVerificationFailure { - return nil, childMerkleErr - } - panic(fmt.Sprintf("Expected Merkle file for target %s but none found", parentPath+"/"+name)) + return nil, alertIntegrityViolation(childMerkleErr, fmt.Sprintf("Expected Merkle file for target %s but none found", parentPath+"/"+name)) } } else if childErr == syserror.ENOENT && childMerkleErr == syserror.ENOENT { // Both the child and the corresponding Merkle tree are missing. @@ -421,7 +398,7 @@ func (fs *filesystem) lookupAndVerifyLocked(ctx context.Context, parent *dentry, // TODO(b/167752508): Investigate possible ways to differentiate // cases that both files are deleted from cases that they never // exist in the file system. - panic(fmt.Sprintf("Failed to find file %s", parentPath+"/"+name)) + return nil, alertIntegrityViolation(childErr, fmt.Sprintf("Failed to find file %s", parentPath+"/"+name)) } mask := uint32(linux.STATX_TYPE | linux.STATX_MODE | linux.STATX_UID | linux.STATX_GID) @@ -580,8 +557,183 @@ func (fs *filesystem) MknodAt(ctx context.Context, rp *vfs.ResolvingPath, opts v // OpenAt implements vfs.FilesystemImpl.OpenAt. func (fs *filesystem) OpenAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.OpenOptions) (*vfs.FileDescription, error) { - //TODO(b/159261227): Implement OpenAt. - return nil, nil + // Verity fs is read-only. + if opts.Flags&(linux.O_WRONLY|linux.O_CREAT) != 0 { + return nil, syserror.EROFS + } + + var ds *[]*dentry + fs.renameMu.RLock() + defer fs.renameMuRUnlockAndCheckDrop(ctx, &ds) + + start := rp.Start().Impl().(*dentry) + if rp.Done() { + return start.openLocked(ctx, rp, &opts) + } + +afterTrailingSymlink: + parent, err := fs.walkParentDirLocked(ctx, rp, start, &ds) + if err != nil { + return nil, err + } + + // Check for search permission in the parent directory. + if err := parent.checkPermissions(rp.Credentials(), vfs.MayExec); err != nil { + return nil, err + } + + // Open existing child or follow symlink. + parent.dirMu.Lock() + child, err := fs.stepLocked(ctx, rp, parent, false /*mayFollowSymlinks*/, &ds) + parent.dirMu.Unlock() + if err != nil { + return nil, err + } + if child.isSymlink() && rp.ShouldFollowSymlink() { + target, err := child.readlink(ctx) + if err != nil { + return nil, err + } + if err := rp.HandleSymlink(target); err != nil { + return nil, err + } + start = parent + goto afterTrailingSymlink + } + return child.openLocked(ctx, rp, &opts) +} + +// Preconditions: fs.renameMu must be locked. +func (d *dentry) openLocked(ctx context.Context, rp *vfs.ResolvingPath, opts *vfs.OpenOptions) (*vfs.FileDescription, error) { + // Users should not open the Merkle tree files. Those are for verity fs + // use only. + if strings.Contains(d.name, merklePrefix) { + return nil, syserror.EPERM + } + ats := vfs.AccessTypesForOpenFlags(opts) + if err := d.checkPermissions(rp.Credentials(), ats); err != nil { + return nil, err + } + + // Verity fs is read-only. + if ats&vfs.MayWrite != 0 { + return nil, syserror.EROFS + } + + // Get the path to the target file. This is only used to provide path + // information in failure case. + path, err := d.fs.vfsfs.VirtualFilesystem().PathnameWithDeleted(ctx, d.fs.rootDentry.lowerVD, d.lowerVD) + if err != nil { + return nil, err + } + + // Open the file in the underlying file system. + lowerFD, err := rp.VirtualFilesystem().OpenAt(ctx, d.fs.creds, &vfs.PathOperation{ + Root: d.lowerVD, + Start: d.lowerVD, + }, opts) + + // The file should exist, as we succeeded in finding its dentry. If it's + // missing, it indicates an unexpected modification to the file system. + if err != nil { + if err == syserror.ENOENT { + return nil, alertIntegrityViolation(err, fmt.Sprintf("File %s expected but not found", path)) + } + return nil, err + } + + // lowerFD needs to be cleaned up if any error occurs. IncRef will be + // called if a verity FD is successfully created. + defer lowerFD.DecRef(ctx) + + // Open the Merkle tree file corresponding to the current file/directory + // to be used later for verifying Read/Walk. + merkleReader, err := rp.VirtualFilesystem().OpenAt(ctx, d.fs.creds, &vfs.PathOperation{ + Root: d.lowerMerkleVD, + Start: d.lowerMerkleVD, + }, &vfs.OpenOptions{ + Flags: linux.O_RDONLY, + }) + + // The Merkle tree file should exist, as we succeeded in finding its + // dentry. If it's missing, it indicates an unexpected modification to + // the file system. + if err != nil { + if err == syserror.ENOENT { + return nil, alertIntegrityViolation(err, fmt.Sprintf("Merkle file for %s expected but not found", path)) + } + return nil, err + } + + // merkleReader needs to be cleaned up if any error occurs. IncRef will + // be called if a verity FD is successfully created. + defer merkleReader.DecRef(ctx) + + lowerFlags := lowerFD.StatusFlags() + lowerFDOpts := lowerFD.Options() + var merkleWriter *vfs.FileDescription + var parentMerkleWriter *vfs.FileDescription + + // Only open the Merkle tree files for write if in allowRuntimeEnable + // mode. + if d.fs.allowRuntimeEnable { + merkleWriter, err = rp.VirtualFilesystem().OpenAt(ctx, d.fs.creds, &vfs.PathOperation{ + Root: d.lowerMerkleVD, + Start: d.lowerMerkleVD, + }, &vfs.OpenOptions{ + Flags: linux.O_WRONLY | linux.O_APPEND, + }) + if err != nil { + if err == syserror.ENOENT { + return nil, alertIntegrityViolation(err, fmt.Sprintf("Merkle file for %s expected but not found", path)) + } + return nil, err + } + // merkleWriter is cleaned up if any error occurs. IncRef will + // be called if a verity FD is created successfully. + defer merkleWriter.DecRef(ctx) + + if d.parent != nil { + parentMerkleWriter, err = rp.VirtualFilesystem().OpenAt(ctx, d.fs.creds, &vfs.PathOperation{ + Root: d.parent.lowerMerkleVD, + Start: d.parent.lowerMerkleVD, + }, &vfs.OpenOptions{ + Flags: linux.O_WRONLY | linux.O_APPEND, + }) + if err != nil { + if err == syserror.ENOENT { + parentPath, _ := d.fs.vfsfs.VirtualFilesystem().PathnameWithDeleted(ctx, d.fs.rootDentry.lowerVD, d.parent.lowerVD) + return nil, alertIntegrityViolation(err, fmt.Sprintf("Merkle file for %s expected but not found", parentPath)) + } + return nil, err + } + // parentMerkleWriter is cleaned up if any error occurs. IncRef + // will be called if a verity FD is created successfully. + defer parentMerkleWriter.DecRef(ctx) + } + } + + fd := &fileDescription{ + d: d, + lowerFD: lowerFD, + merkleReader: merkleReader, + merkleWriter: merkleWriter, + parentMerkleWriter: parentMerkleWriter, + isDir: d.isDir(), + } + + if err := fd.vfsfd.Init(fd, lowerFlags, rp.Mount(), &d.vfsd, &lowerFDOpts); err != nil { + return nil, err + } + lowerFD.IncRef() + merkleReader.IncRef() + if merkleWriter != nil { + merkleWriter.IncRef() + } + if parentMerkleWriter != nil { + parentMerkleWriter.IncRef() + } + return &fd.vfsfd, err } // ReadlinkAt implements vfs.FilesystemImpl.ReadlinkAt. @@ -657,7 +809,7 @@ func (fs *filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error return syserror.EROFS } -// BoundEndpointAt implements FilesystemImpl.BoundEndpointAt. +// BoundEndpointAt implements vfs.FilesystemImpl.BoundEndpointAt. func (fs *filesystem) BoundEndpointAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.BoundEndpointOptions) (transport.BoundEndpoint, error) { var ds *[]*dentry fs.renameMu.RLock() diff --git a/pkg/sentry/fsimpl/verity/verity.go b/pkg/sentry/fsimpl/verity/verity.go index 249cc1341..fc5eabbca 100644 --- a/pkg/sentry/fsimpl/verity/verity.go +++ b/pkg/sentry/fsimpl/verity/verity.go @@ -22,6 +22,7 @@ package verity import ( + "fmt" "strconv" "sync/atomic" @@ -29,7 +30,6 @@ import ( "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/fspath" "gvisor.dev/gvisor/pkg/marshal/primitive" - "gvisor.dev/gvisor/pkg/merkletree" "gvisor.dev/gvisor/pkg/sentry/arch" fslock "gvisor.dev/gvisor/pkg/sentry/fs/lock" @@ -57,8 +57,9 @@ const merkleOffsetInParentXattr = "user.merkle.offset" // whole file. For a directory, it's the size of all its children's root hashes. const merkleSizeXattr = "user.merkle.size" -// sizeOfInt32 is the size in bytes for a 32 bit integer in extended attributes. -const sizeOfInt32 = 4 +// sizeOfStringInt32 is the size for a 32 bit integer stored as string in +// extended attributes. The maximum value of a 32 bit integer is 10 digits. +const sizeOfStringInt32 = 10 // noCrashOnVerificationFailure indicates whether the sandbox should panic // whenever verification fails. If true, an error is returned instead of @@ -73,9 +74,13 @@ var noCrashOnVerificationFailure bool var verityMu sync.RWMutex // FilesystemType implements vfs.FilesystemType. +// +// +stateify savable type FilesystemType struct{} // filesystem implements vfs.FilesystemImpl. +// +// +stateify savable type filesystem struct { vfsfs vfs.Filesystem @@ -100,11 +105,13 @@ type filesystem struct { // renameMu synchronizes renaming with non-renaming operations in order // to ensure consistent lock ordering between dentry.dirMu in different // dentries. - renameMu sync.RWMutex + renameMu sync.RWMutex `state:"nosave"` } // InternalFilesystemOptions may be passed as // vfs.GetFilesystemOptions.InternalData to FilesystemType.GetFilesystem. +// +// +stateify savable type InternalFilesystemOptions struct { // RootMerkleFileName is the name of the verity root Merkle tree file. RootMerkleFileName string @@ -135,6 +142,16 @@ func (FilesystemType) Name() string { return Name } +// alertIntegrityViolation alerts a violation of integrity, which usually means +// unexpected modification to the file system is detected. In +// noCrashOnVerificationFailure mode, it returns an error, otherwise it panic. +func alertIntegrityViolation(err error, msg string) error { + if noCrashOnVerificationFailure { + return err + } + panic(msg) +} + // 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) { iopts, ok := opts.InternalData.(InternalFilesystemOptions) @@ -148,6 +165,7 @@ func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virt // verity, and should not be exposed or connected. mopts := &vfs.MountOptions{ GetFilesystemOptions: iopts.LowerGetFSOptions, + InternalMount: true, } mnt, err := vfsObj.MountDisconnected(ctx, creds, "", iopts.LowerName, mopts) if err != nil { @@ -204,15 +222,12 @@ func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virt return nil, nil, err } } else if err != nil { - // Failed to get dentry for the root Merkle file. This indicates - // an attack that removed/renamed the root Merkle file, or it's - // never generated. - if noCrashOnVerificationFailure { - fs.vfsfs.DecRef(ctx) - d.DecRef(ctx) - return nil, nil, err - } - panic("Failed to find root Merkle file") + // Failed to get dentry for the root Merkle file. This + // indicates an unexpected modification that removed/renamed + // the root Merkle file, or it's never generated. + fs.vfsfs.DecRef(ctx) + d.DecRef(ctx) + return nil, nil, alertIntegrityViolation(err, "Failed to find root Merkle file") } d.lowerMerkleVD = lowerMerkleVD @@ -250,6 +265,8 @@ func (fs *filesystem) Release(ctx context.Context) { } // dentry implements vfs.DentryImpl. +// +// +stateify savable type dentry struct { vfsd vfs.Dentry @@ -276,7 +293,7 @@ type dentry struct { // and dirents (if not nil) is a cache of dirents as returned by // directoryFDs representing this directory. children is protected by // dirMu. - dirMu sync.Mutex + dirMu sync.Mutex `state:"nosave"` children map[string]*dentry // lowerVD is the VirtualDentry in the underlying file system. @@ -420,6 +437,8 @@ func (d *dentry) readlink(ctx context.Context) (string, error) { // FileDescription is a wrapper of the underlying lowerFD, with support to build // Merkle trees through the Linux fs-verity API to verify contents read from // lowerFD. +// +// +stateify savable type fileDescription struct { vfsfd vfs.FileDescription vfs.FileDescriptionDefaultImpl @@ -539,7 +558,7 @@ func (fd *fileDescription) generateMerkle(ctx context.Context) ([]byte, uint64, // enableVerity enables verity features on fd by generating a Merkle tree file // and stores its root hash in its parent directory's Merkle tree. -func (fd *fileDescription) enableVerity(ctx context.Context, uio usermem.IO, args arch.SyscallArguments) (uintptr, error) { +func (fd *fileDescription) enableVerity(ctx context.Context, uio usermem.IO) (uintptr, error) { if !fd.d.fs.allowRuntimeEnable { return 0, syserror.EPERM } @@ -549,8 +568,12 @@ func (fd *fileDescription) enableVerity(ctx context.Context, uio usermem.IO, arg verityMu.Lock() defer verityMu.Unlock() - if fd.lowerFD == nil || fd.merkleReader == nil || fd.merkleWriter == nil || fd.parentMerkleWriter == nil { - panic("Unexpected verity fd: missing expected underlying fds") + // In allowRuntimeEnable mode, the underlying fd and read/write fd for + // the Merkle tree file should have all been initialized. For any file + // or directory other than the root, the parent Merkle tree file should + // have also been initialized. + if fd.lowerFD == nil || fd.merkleReader == nil || fd.merkleWriter == nil || (fd.parentMerkleWriter == nil && fd.d != fd.d.fs.rootDentry) { + return 0, alertIntegrityViolation(syserror.EIO, "Unexpected verity fd: missing expected underlying fds") } rootHash, dataSize, err := fd.generateMerkle(ctx) @@ -558,26 +581,28 @@ func (fd *fileDescription) enableVerity(ctx context.Context, uio usermem.IO, arg return 0, err } - stat, err := fd.parentMerkleWriter.Stat(ctx, vfs.StatOptions{}) - if err != nil { - return 0, err - } + if fd.parentMerkleWriter != nil { + stat, err := fd.parentMerkleWriter.Stat(ctx, vfs.StatOptions{}) + if err != nil { + return 0, err + } - // Write the root hash of fd to the parent directory's Merkle tree - // file, as it should be part of the parent Merkle tree data. - // parentMerkleWriter is open with O_APPEND, so it should write - // directly to the end of the file. - if _, err = fd.parentMerkleWriter.Write(ctx, usermem.BytesIOSequence(rootHash), vfs.WriteOptions{}); err != nil { - return 0, err - } + // Write the root hash of fd to the parent directory's Merkle + // tree file, as it should be part of the parent Merkle tree + // data. parentMerkleWriter is open with O_APPEND, so it + // should write directly to the end of the file. + if _, err = fd.parentMerkleWriter.Write(ctx, usermem.BytesIOSequence(rootHash), vfs.WriteOptions{}); err != nil { + return 0, err + } - // Record the offset of the root hash of fd in parent directory's - // Merkle tree file. - if err := fd.merkleWriter.SetXattr(ctx, &vfs.SetXattrOptions{ - Name: merkleOffsetInParentXattr, - Value: strconv.Itoa(int(stat.Size)), - }); err != nil { - return 0, err + // Record the offset of the root hash of fd in parent directory's + // Merkle tree file. + if err := fd.merkleWriter.SetXattr(ctx, &vfs.SetXattrOptions{ + Name: merkleOffsetInParentXattr, + Value: strconv.Itoa(int(stat.Size)), + }); err != nil { + return 0, err + } } // Record the size of the data being hashed for fd. @@ -591,7 +616,45 @@ func (fd *fileDescription) enableVerity(ctx context.Context, uio usermem.IO, arg return 0, nil } -func (fd *fileDescription) getFlags(ctx context.Context, uio usermem.IO, args arch.SyscallArguments) (uintptr, error) { +// measureVerity returns the root hash of fd, saved in args[2]. +func (fd *fileDescription) measureVerity(ctx context.Context, uio usermem.IO, verityDigest usermem.Addr) (uintptr, error) { + t := kernel.TaskFromContext(ctx) + var metadata linux.DigestMetadata + + // If allowRuntimeEnable is true, an empty fd.d.rootHash indicates that + // verity is not enabled for the file. If allowRuntimeEnable is false, + // this is an integrity violation because all files should have verity + // enabled, in which case fd.d.rootHash should be set. + if len(fd.d.rootHash) == 0 { + if fd.d.fs.allowRuntimeEnable { + return 0, syserror.ENODATA + } + return 0, alertIntegrityViolation(syserror.ENODATA, "Ioctl measureVerity: no root hash found") + } + + // The first part of VerityDigest is the metadata. + if _, err := metadata.CopyIn(t, verityDigest); err != nil { + return 0, err + } + if metadata.DigestSize < uint16(len(fd.d.rootHash)) { + return 0, syserror.EOVERFLOW + } + + // Populate the output digest size, since DigestSize is both input and + // output. + metadata.DigestSize = uint16(len(fd.d.rootHash)) + + // First copy the metadata. + if _, err := metadata.CopyOut(t, verityDigest); err != nil { + return 0, err + } + + // Now copy the root hash bytes to the memory after metadata. + _, err := t.CopyOutBytes(usermem.Addr(uintptr(verityDigest)+linux.SizeOfDigestMetadata), fd.d.rootHash) + return 0, err +} + +func (fd *fileDescription) verityFlags(ctx context.Context, uio usermem.IO, flags usermem.Addr) (uintptr, error) { f := int32(0) // All enabled files should store a root hash. This flag is not settable @@ -601,8 +664,7 @@ func (fd *fileDescription) getFlags(ctx context.Context, uio usermem.IO, args ar } t := kernel.TaskFromContext(ctx) - addr := args[2].Pointer() - _, err := primitive.CopyInt32Out(t, addr, f) + _, err := primitive.CopyInt32Out(t, flags, f) return 0, err } @@ -610,20 +672,72 @@ func (fd *fileDescription) getFlags(ctx context.Context, uio usermem.IO, args ar func (fd *fileDescription) Ioctl(ctx context.Context, uio usermem.IO, args arch.SyscallArguments) (uintptr, error) { switch cmd := args[1].Uint(); cmd { case linux.FS_IOC_ENABLE_VERITY: - return fd.enableVerity(ctx, uio, args) + return fd.enableVerity(ctx, uio) + case linux.FS_IOC_MEASURE_VERITY: + return fd.measureVerity(ctx, uio, args[2].Pointer()) case linux.FS_IOC_GETFLAGS: - return fd.getFlags(ctx, uio, args) + return fd.verityFlags(ctx, uio, args[2].Pointer()) default: - return fd.lowerFD.Ioctl(ctx, uio, args) + // TODO(b/169682228): Investigate which ioctl commands should + // be allowed. + return 0, syserror.ENOSYS + } +} + +// PRead implements vfs.FileDescriptionImpl.PRead. +func (fd *fileDescription) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { + // No need to verify if the file is not enabled yet in + // allowRuntimeEnable mode. + if fd.d.fs.allowRuntimeEnable && len(fd.d.rootHash) == 0 { + return fd.lowerFD.PRead(ctx, dst, offset, opts) + } + + // dataSize is the size of the whole file. + dataSize, err := fd.merkleReader.GetXattr(ctx, &vfs.GetXattrOptions{ + Name: merkleSizeXattr, + Size: sizeOfStringInt32, + }) + + // The Merkle tree file for the child should have been created and + // contains the expected xattrs. If the xattr does not exist, it + // indicates unexpected modifications to the file system. + if err == syserror.ENODATA { + return 0, alertIntegrityViolation(err, fmt.Sprintf("Failed to get xattr %s: %v", merkleSizeXattr, err)) + } + if err != nil { + return 0, err + } + + // The dataSize xattr should be an integer. If it's not, it indicates + // unexpected modifications to the file system. + size, err := strconv.Atoi(dataSize) + if err != nil { + return 0, alertIntegrityViolation(err, fmt.Sprintf("Failed to convert xattr %s to int: %v", merkleSizeXattr, err)) + } + + dataReader := vfs.FileReadWriteSeeker{ + FD: fd.lowerFD, + Ctx: ctx, + } + + merkleReader := vfs.FileReadWriteSeeker{ + FD: fd.merkleReader, + Ctx: ctx, + } + + n, err := merkletree.Verify(dst.Writer(ctx), &dataReader, &merkleReader, int64(size), offset, dst.NumBytes(), fd.d.rootHash, false /* dataAndTreeInSameFile */) + if err != nil { + return 0, alertIntegrityViolation(syserror.EINVAL, fmt.Sprintf("Verification failed: %v", err)) } + return n, err } // LockPOSIX implements vfs.FileDescriptionImpl.LockPOSIX. func (fd *fileDescription) LockPOSIX(ctx context.Context, uid fslock.UniqueID, t fslock.LockType, start, length uint64, whence int16, block fslock.Blocker) error { - return fd.Locks().LockPOSIX(ctx, &fd.vfsfd, uid, t, start, length, whence, block) + return fd.lowerFD.LockPOSIX(ctx, uid, t, start, length, whence, block) } // UnlockPOSIX implements vfs.FileDescriptionImpl.UnlockPOSIX. func (fd *fileDescription) UnlockPOSIX(ctx context.Context, uid fslock.UniqueID, start, length uint64, whence int16) error { - return fd.Locks().UnlockPOSIX(ctx, &fd.vfsfd, uid, start, length, whence) + return fd.lowerFD.UnlockPOSIX(ctx, uid, start, length, whence) } diff --git a/pkg/sentry/fsimpl/verity/verity_test.go b/pkg/sentry/fsimpl/verity/verity_test.go new file mode 100644 index 000000000..8bcc14131 --- /dev/null +++ b/pkg/sentry/fsimpl/verity/verity_test.go @@ -0,0 +1,429 @@ +// 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 verity + +import ( + "fmt" + "io" + "math/rand" + "testing" + "time" + + "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/tmpfs" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/kernel/contexttest" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/usermem" +) + +// rootMerkleFilename is the name of the root Merkle tree file. +const rootMerkleFilename = "root.verity" + +// maxDataSize is the maximum data size written to the file for test. +const maxDataSize = 100000 + +// newVerityRoot creates a new verity mount, and returns the root. The +// underlying file system is tmpfs. If the error is not nil, then cleanup +// should be called when the root is no longer needed. +func newVerityRoot(ctx context.Context) (*vfs.VirtualFilesystem, vfs.VirtualDentry, func(), error) { + rand.Seed(time.Now().UnixNano()) + vfsObj := &vfs.VirtualFilesystem{} + if err := vfsObj.Init(ctx); err != nil { + return nil, vfs.VirtualDentry{}, nil, fmt.Errorf("VFS init: %v", err) + } + + vfsObj.MustRegisterFilesystemType("verity", FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ + AllowUserMount: true, + }) + + vfsObj.MustRegisterFilesystemType("tmpfs", tmpfs.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ + AllowUserMount: true, + }) + + mntns, err := vfsObj.NewMountNamespace(ctx, auth.CredentialsFromContext(ctx), "", "verity", &vfs.MountOptions{ + GetFilesystemOptions: vfs.GetFilesystemOptions{ + InternalData: InternalFilesystemOptions{ + RootMerkleFileName: rootMerkleFilename, + LowerName: "tmpfs", + AllowRuntimeEnable: true, + NoCrashOnVerificationFailure: true, + }, + }, + }) + if err != nil { + return nil, vfs.VirtualDentry{}, nil, fmt.Errorf("NewMountNamespace: %v", err) + } + root := mntns.Root() + return vfsObj, root, func() { + root.DecRef(ctx) + mntns.DecRef(ctx) + }, nil +} + +// newFileFD creates a new file in the verity mount, and returns the FD. The FD +// points to a file that has random data generated. +func newFileFD(ctx context.Context, vfsObj *vfs.VirtualFilesystem, root vfs.VirtualDentry, filePath string, mode linux.FileMode) (*vfs.FileDescription, int, error) { + creds := auth.CredentialsFromContext(ctx) + lowerRoot := root.Dentry().Impl().(*dentry).lowerVD + + // Create the file in the underlying file system. + lowerFD, err := vfsObj.OpenAt(ctx, creds, &vfs.PathOperation{ + Root: lowerRoot, + Start: lowerRoot, + Path: fspath.Parse(filePath), + }, &vfs.OpenOptions{ + Flags: linux.O_RDWR | linux.O_CREAT | linux.O_EXCL, + Mode: linux.ModeRegular | mode, + }) + if err != nil { + return nil, 0, err + } + + // Generate random data to be written to the file. + dataSize := rand.Intn(maxDataSize) + 1 + data := make([]byte, dataSize) + rand.Read(data) + + // Write directly to the underlying FD, since verity FD is read-only. + n, err := lowerFD.Write(ctx, usermem.BytesIOSequence(data), vfs.WriteOptions{}) + if err != nil { + return nil, 0, err + } + + if n != int64(len(data)) { + return nil, 0, fmt.Errorf("lowerFD.Write got write length %d, want %d", n, len(data)) + } + + lowerFD.DecRef(ctx) + + // Now open the verity file descriptor. + fd, err := vfsObj.OpenAt(ctx, creds, &vfs.PathOperation{ + Root: root, + Start: root, + Path: fspath.Parse(filePath), + }, &vfs.OpenOptions{ + Flags: linux.O_RDONLY, + Mode: linux.ModeRegular | mode, + }) + return fd, dataSize, err +} + +// corruptRandomBit randomly flips a bit in the file represented by fd. +func corruptRandomBit(ctx context.Context, fd *vfs.FileDescription, size int) error { + // Flip a random bit in the underlying file. + randomPos := int64(rand.Intn(size)) + byteToModify := make([]byte, 1) + if _, err := fd.PRead(ctx, usermem.BytesIOSequence(byteToModify), randomPos, vfs.ReadOptions{}); err != nil { + return fmt.Errorf("lowerFD.PRead: %v", err) + } + byteToModify[0] ^= 1 + if _, err := fd.PWrite(ctx, usermem.BytesIOSequence(byteToModify), randomPos, vfs.WriteOptions{}); err != nil { + return fmt.Errorf("lowerFD.PWrite: %v", err) + } + return nil +} + +// TestOpen ensures that when a file is created, the corresponding Merkle tree +// file and the root Merkle tree file exist. +func TestOpen(t *testing.T) { + ctx := contexttest.Context(t) + vfsObj, root, cleanup, err := newVerityRoot(ctx) + if err != nil { + t.Fatalf("newVerityRoot: %v", err) + } + defer cleanup() + + filename := "verity-test-file" + if _, _, err := newFileFD(ctx, vfsObj, root, filename, 0644); err != nil { + t.Fatalf("newFileFD: %v", err) + } + + // Ensure that the corresponding Merkle tree file is created. + lowerRoot := root.Dentry().Impl().(*dentry).lowerVD + if _, err = vfsObj.OpenAt(ctx, auth.CredentialsFromContext(ctx), &vfs.PathOperation{ + Root: lowerRoot, + Start: lowerRoot, + Path: fspath.Parse(merklePrefix + filename), + }, &vfs.OpenOptions{ + Flags: linux.O_RDONLY, + }); err != nil { + t.Errorf("OpenAt Merkle tree file %s: %v", merklePrefix+filename, err) + } + + // Ensure the root merkle tree file is created. + if _, err = vfsObj.OpenAt(ctx, auth.CredentialsFromContext(ctx), &vfs.PathOperation{ + Root: lowerRoot, + Start: lowerRoot, + Path: fspath.Parse(merklePrefix + rootMerkleFilename), + }, &vfs.OpenOptions{ + Flags: linux.O_RDONLY, + }); err != nil { + t.Errorf("OpenAt root Merkle tree file %s: %v", merklePrefix+rootMerkleFilename, err) + } +} + +// TestUntouchedFileSucceeds ensures that read from an untouched verity file +// succeeds after enabling verity for it. +func TestReadUntouchedFileSucceeds(t *testing.T) { + ctx := contexttest.Context(t) + vfsObj, root, cleanup, err := newVerityRoot(ctx) + if err != nil { + t.Fatalf("newVerityRoot: %v", err) + } + defer cleanup() + + filename := "verity-test-file" + fd, size, err := newFileFD(ctx, vfsObj, root, filename, 0644) + if err != nil { + t.Fatalf("newFileFD: %v", err) + } + + // Enable verity on the file and confirm a normal read succeeds. + var args arch.SyscallArguments + args[1] = arch.SyscallArgument{Value: linux.FS_IOC_ENABLE_VERITY} + if _, err := fd.Ioctl(ctx, nil /* uio */, args); err != nil { + t.Fatalf("Ioctl: %v", err) + } + + buf := make([]byte, size) + n, err := fd.PRead(ctx, usermem.BytesIOSequence(buf), 0 /* offset */, vfs.ReadOptions{}) + if err != nil && err != io.EOF { + t.Fatalf("fd.PRead: %v", err) + } + + if n != int64(size) { + t.Errorf("fd.PRead got read length %d, want %d", n, size) + } +} + +// TestReopenUntouchedFileSucceeds ensures that reopen an untouched verity file +// succeeds after enabling verity for it. +func TestReopenUntouchedFileSucceeds(t *testing.T) { + ctx := contexttest.Context(t) + vfsObj, root, cleanup, err := newVerityRoot(ctx) + if err != nil { + t.Fatalf("newVerityRoot: %v", err) + } + defer cleanup() + + filename := "verity-test-file" + fd, _, err := newFileFD(ctx, vfsObj, root, filename, 0644) + if err != nil { + t.Fatalf("newFileFD: %v", err) + } + + // Enable verity on the file and confirms a normal read succeeds. + var args arch.SyscallArguments + args[1] = arch.SyscallArgument{Value: linux.FS_IOC_ENABLE_VERITY} + if _, err := fd.Ioctl(ctx, nil /* uio */, args); err != nil { + t.Fatalf("Ioctl: %v", err) + } + + // Ensure reopening the verity enabled file succeeds. + if _, err = vfsObj.OpenAt(ctx, auth.CredentialsFromContext(ctx), &vfs.PathOperation{ + Root: root, + Start: root, + Path: fspath.Parse(filename), + }, &vfs.OpenOptions{ + Flags: linux.O_RDONLY, + Mode: linux.ModeRegular, + }); err != nil { + t.Errorf("reopen enabled file failed: %v", err) + } +} + +// TestModifiedFileFails ensures that read from a modified verity file fails. +func TestModifiedFileFails(t *testing.T) { + ctx := contexttest.Context(t) + vfsObj, root, cleanup, err := newVerityRoot(ctx) + if err != nil { + t.Fatalf("newVerityRoot: %v", err) + } + defer cleanup() + + filename := "verity-test-file" + fd, size, err := newFileFD(ctx, vfsObj, root, filename, 0644) + if err != nil { + t.Fatalf("newFileFD: %v", err) + } + + // Enable verity on the file. + var args arch.SyscallArguments + args[1] = arch.SyscallArgument{Value: linux.FS_IOC_ENABLE_VERITY} + if _, err := fd.Ioctl(ctx, nil /* uio */, args); err != nil { + t.Fatalf("Ioctl: %v", err) + } + + // Open a new lowerFD that's read/writable. + lowerVD := fd.Impl().(*fileDescription).d.lowerVD + + lowerFD, err := vfsObj.OpenAt(ctx, auth.CredentialsFromContext(ctx), &vfs.PathOperation{ + Root: lowerVD, + Start: lowerVD, + }, &vfs.OpenOptions{ + Flags: linux.O_RDWR, + }) + if err != nil { + t.Fatalf("OpenAt: %v", err) + } + + if err := corruptRandomBit(ctx, lowerFD, size); err != nil { + t.Fatalf("corruptRandomBit: %v", err) + } + + // Confirm that read from the modified file fails. + buf := make([]byte, size) + if _, err := fd.PRead(ctx, usermem.BytesIOSequence(buf), 0 /* offset */, vfs.ReadOptions{}); err == nil { + t.Fatalf("fd.PRead succeeded with modified file") + } +} + +// TestModifiedMerkleFails ensures that read from a verity file fails if the +// corresponding Merkle tree file is modified. +func TestModifiedMerkleFails(t *testing.T) { + ctx := contexttest.Context(t) + vfsObj, root, cleanup, err := newVerityRoot(ctx) + if err != nil { + t.Fatalf("newVerityRoot: %v", err) + } + defer cleanup() + + filename := "verity-test-file" + fd, size, err := newFileFD(ctx, vfsObj, root, filename, 0644) + if err != nil { + t.Fatalf("newFileFD: %v", err) + } + + // Enable verity on the file. + var args arch.SyscallArguments + args[1] = arch.SyscallArgument{Value: linux.FS_IOC_ENABLE_VERITY} + if _, err := fd.Ioctl(ctx, nil /* uio */, args); err != nil { + t.Fatalf("Ioctl: %v", err) + } + + // Open a new lowerMerkleFD that's read/writable. + lowerMerkleVD := fd.Impl().(*fileDescription).d.lowerMerkleVD + + lowerMerkleFD, err := vfsObj.OpenAt(ctx, auth.CredentialsFromContext(ctx), &vfs.PathOperation{ + Root: lowerMerkleVD, + Start: lowerMerkleVD, + }, &vfs.OpenOptions{ + Flags: linux.O_RDWR, + }) + if err != nil { + t.Fatalf("OpenAt: %v", err) + } + + // Flip a random bit in the Merkle tree file. + stat, err := lowerMerkleFD.Stat(ctx, vfs.StatOptions{}) + if err != nil { + t.Fatalf("stat: %v", err) + } + merkleSize := int(stat.Size) + if err := corruptRandomBit(ctx, lowerMerkleFD, merkleSize); err != nil { + t.Fatalf("corruptRandomBit: %v", err) + } + + // Confirm that read from a file with modified Merkle tree fails. + buf := make([]byte, size) + if _, err := fd.PRead(ctx, usermem.BytesIOSequence(buf), 0 /* offset */, vfs.ReadOptions{}); err == nil { + fmt.Println(buf) + t.Fatalf("fd.PRead succeeded with modified Merkle file") + } +} + +// TestModifiedParentMerkleFails ensures that open a verity enabled file in a +// verity enabled directory fails if the hashes related to the target file in +// the parent Merkle tree file is modified. +func TestModifiedParentMerkleFails(t *testing.T) { + ctx := contexttest.Context(t) + vfsObj, root, cleanup, err := newVerityRoot(ctx) + if err != nil { + t.Fatalf("newVerityRoot: %v", err) + } + defer cleanup() + + filename := "verity-test-file" + fd, _, err := newFileFD(ctx, vfsObj, root, filename, 0644) + if err != nil { + t.Fatalf("newFileFD: %v", err) + } + + // Enable verity on the file. + var args arch.SyscallArguments + args[1] = arch.SyscallArgument{Value: linux.FS_IOC_ENABLE_VERITY} + if _, err := fd.Ioctl(ctx, nil /* uio */, args); err != nil { + t.Fatalf("Ioctl: %v", err) + } + + // Enable verity on the parent directory. + parentFD, err := vfsObj.OpenAt(ctx, auth.CredentialsFromContext(ctx), &vfs.PathOperation{ + Root: root, + Start: root, + }, &vfs.OpenOptions{ + Flags: linux.O_RDONLY, + }) + if err != nil { + t.Fatalf("OpenAt: %v", err) + } + + if _, err := parentFD.Ioctl(ctx, nil /* uio */, args); err != nil { + t.Fatalf("Ioctl: %v", err) + } + + // Open a new lowerMerkleFD that's read/writable. + parentLowerMerkleVD := fd.Impl().(*fileDescription).d.parent.lowerMerkleVD + + parentLowerMerkleFD, err := vfsObj.OpenAt(ctx, auth.CredentialsFromContext(ctx), &vfs.PathOperation{ + Root: parentLowerMerkleVD, + Start: parentLowerMerkleVD, + }, &vfs.OpenOptions{ + Flags: linux.O_RDWR, + }) + if err != nil { + t.Fatalf("OpenAt: %v", err) + } + + // Flip a random bit in the parent Merkle tree file. + // This parent directory contains only one child, so any random + // modification in the parent Merkle tree should cause verification + // failure when opening the child file. + stat, err := parentLowerMerkleFD.Stat(ctx, vfs.StatOptions{}) + if err != nil { + t.Fatalf("stat: %v", err) + } + parentMerkleSize := int(stat.Size) + if err := corruptRandomBit(ctx, parentLowerMerkleFD, parentMerkleSize); err != nil { + t.Fatalf("corruptRandomBit: %v", err) + } + + parentLowerMerkleFD.DecRef(ctx) + + // Ensure reopening the verity enabled file fails. + if _, err = vfsObj.OpenAt(ctx, auth.CredentialsFromContext(ctx), &vfs.PathOperation{ + Root: root, + Start: root, + Path: fspath.Parse(filename), + }, &vfs.OpenOptions{ + Flags: linux.O_RDONLY, + Mode: linux.ModeRegular, + }); err == nil { + t.Errorf("OpenAt file with modified parent Merkle succeeded") + } +} diff --git a/pkg/sentry/hostmm/BUILD b/pkg/sentry/hostmm/BUILD index 61c78569d..300b7ccce 100644 --- a/pkg/sentry/hostmm/BUILD +++ b/pkg/sentry/hostmm/BUILD @@ -7,11 +7,14 @@ go_library( srcs = [ "cgroup.go", "hostmm.go", + "membarrier.go", ], visibility = ["//pkg/sentry:internal"], deps = [ + "//pkg/abi/linux", "//pkg/fd", "//pkg/log", "//pkg/usermem", + "@org_golang_x_sys//unix:go_default_library", ], ) diff --git a/pkg/sentry/hostmm/membarrier.go b/pkg/sentry/hostmm/membarrier.go new file mode 100644 index 000000000..4468d75f1 --- /dev/null +++ b/pkg/sentry/hostmm/membarrier.go @@ -0,0 +1,90 @@ +// 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 hostmm + +import ( + "syscall" + + "golang.org/x/sys/unix" + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/log" +) + +var ( + haveMembarrierGlobal = false + haveMembarrierPrivateExpedited = false +) + +func init() { + supported, _, e := syscall.RawSyscall(unix.SYS_MEMBARRIER, linux.MEMBARRIER_CMD_QUERY, 0 /* flags */, 0 /* unused */) + if e != 0 { + if e != syscall.ENOSYS { + log.Warningf("membarrier(MEMBARRIER_CMD_QUERY) failed: %s", e.Error()) + } + return + } + // We don't use MEMBARRIER_CMD_GLOBAL_EXPEDITED because this sends IPIs to + // all CPUs running tasks that have previously invoked + // MEMBARRIER_CMD_REGISTER_GLOBAL_EXPEDITED, which presents a DOS risk. + // (MEMBARRIER_CMD_GLOBAL is synchronize_rcu(), i.e. it waits for an RCU + // grace period to elapse without bothering other CPUs. + // MEMBARRIER_CMD_PRIVATE_EXPEDITED sends IPIs only to CPUs running tasks + // sharing the caller's MM.) + if supported&linux.MEMBARRIER_CMD_GLOBAL != 0 { + haveMembarrierGlobal = true + } + if req := uintptr(linux.MEMBARRIER_CMD_PRIVATE_EXPEDITED | linux.MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED); supported&req == req { + if _, _, e := syscall.RawSyscall(unix.SYS_MEMBARRIER, linux.MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED, 0 /* flags */, 0 /* unused */); e != 0 { + log.Warningf("membarrier(MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED) failed: %s", e.Error()) + } else { + haveMembarrierPrivateExpedited = true + } + } +} + +// HaveGlobalMemoryBarrier returns true if GlobalMemoryBarrier is supported. +func HaveGlobalMemoryBarrier() bool { + return haveMembarrierGlobal +} + +// GlobalMemoryBarrier blocks until "all running threads [in the host OS] have +// passed through a state where all memory accesses to user-space addresses +// match program order between entry to and return from [GlobalMemoryBarrier]", +// as for membarrier(2). +// +// Preconditions: HaveGlobalMemoryBarrier() == true. +func GlobalMemoryBarrier() error { + if _, _, e := syscall.Syscall(unix.SYS_MEMBARRIER, linux.MEMBARRIER_CMD_GLOBAL, 0 /* flags */, 0 /* unused */); e != 0 { + return e + } + return nil +} + +// HaveProcessMemoryBarrier returns true if ProcessMemoryBarrier is supported. +func HaveProcessMemoryBarrier() bool { + return haveMembarrierPrivateExpedited +} + +// ProcessMemoryBarrier is equivalent to GlobalMemoryBarrier, but only +// synchronizes with threads sharing a virtual address space (from the host OS' +// perspective) with the calling thread. +// +// Preconditions: HaveProcessMemoryBarrier() == true. +func ProcessMemoryBarrier() error { + if _, _, e := syscall.RawSyscall(unix.SYS_MEMBARRIER, linux.MEMBARRIER_CMD_PRIVATE_EXPEDITED, 0 /* flags */, 0 /* unused */); e != 0 { + return e + } + return nil +} diff --git a/pkg/sentry/inet/BUILD b/pkg/sentry/inet/BUILD index 07bf39fed..5bba9de0b 100644 --- a/pkg/sentry/inet/BUILD +++ b/pkg/sentry/inet/BUILD @@ -15,6 +15,7 @@ go_library( ], deps = [ "//pkg/context", + "//pkg/tcpip", "//pkg/tcpip/stack", ], ) diff --git a/pkg/sentry/inet/inet.go b/pkg/sentry/inet/inet.go index c0b4831d1..fbe6d6aa6 100644 --- a/pkg/sentry/inet/inet.go +++ b/pkg/sentry/inet/inet.go @@ -15,7 +15,10 @@ // Package inet defines semantics for IP stacks. package inet -import "gvisor.dev/gvisor/pkg/tcpip/stack" +import ( + "gvisor.dev/gvisor/pkg/tcpip" + "gvisor.dev/gvisor/pkg/tcpip/stack" +) // Stack represents a TCP/IP stack. type Stack interface { @@ -80,6 +83,12 @@ type Stack interface { // RestoreCleanupEndpoints adds endpoints to cleanup tracking. This is useful // for restoring a stack after a save. RestoreCleanupEndpoints([]stack.TransportEndpoint) + + // Forwarding returns if packet forwarding between NICs is enabled. + Forwarding(protocol tcpip.NetworkProtocolNumber) bool + + // SetForwarding enables or disables packet forwarding between NICs. + SetForwarding(protocol tcpip.NetworkProtocolNumber, enable bool) error } // Interface contains information about a network interface. diff --git a/pkg/sentry/inet/test_stack.go b/pkg/sentry/inet/test_stack.go index 9771f01fc..1779cc6f3 100644 --- a/pkg/sentry/inet/test_stack.go +++ b/pkg/sentry/inet/test_stack.go @@ -14,7 +14,10 @@ package inet -import "gvisor.dev/gvisor/pkg/tcpip/stack" +import ( + "gvisor.dev/gvisor/pkg/tcpip" + "gvisor.dev/gvisor/pkg/tcpip/stack" +) // TestStack is a dummy implementation of Stack for tests. type TestStack struct { @@ -26,6 +29,7 @@ type TestStack struct { TCPSendBufSize TCPBufferSize TCPSACKFlag bool Recovery TCPLossRecovery + IPForwarding bool } // NewTestStack returns a TestStack with no network interfaces. The value of @@ -128,3 +132,14 @@ func (s *TestStack) CleanupEndpoints() []stack.TransportEndpoint { // RestoreCleanupEndpoints implements inet.Stack.RestoreCleanupEndpoints. func (s *TestStack) RestoreCleanupEndpoints([]stack.TransportEndpoint) {} + +// Forwarding implements inet.Stack.Forwarding. +func (s *TestStack) Forwarding(protocol tcpip.NetworkProtocolNumber) bool { + return s.IPForwarding +} + +// SetForwarding implements inet.Stack.SetForwarding. +func (s *TestStack) SetForwarding(protocol tcpip.NetworkProtocolNumber, enable bool) error { + s.IPForwarding = enable + return nil +} diff --git a/pkg/sentry/kernel/BUILD b/pkg/sentry/kernel/BUILD index a43c549f1..5de70aecb 100644 --- a/pkg/sentry/kernel/BUILD +++ b/pkg/sentry/kernel/BUILD @@ -69,8 +69,8 @@ go_template_instance( prefix = "socket", template = "//pkg/ilist:generic_list", types = { - "Element": "*SocketEntry", - "Linker": "*SocketEntry", + "Element": "*SocketRecordVFS1", + "Linker": "*SocketRecordVFS1", }, ) @@ -204,7 +204,6 @@ go_library( "//pkg/abi", "//pkg/abi/linux", "//pkg/amutex", - "//pkg/binary", "//pkg/bits", "//pkg/bpf", "//pkg/context", diff --git a/pkg/sentry/kernel/kcov.go b/pkg/sentry/kernel/kcov.go index aad63aa99..060c056df 100644 --- a/pkg/sentry/kernel/kcov.go +++ b/pkg/sentry/kernel/kcov.go @@ -89,6 +89,10 @@ func (kcov *Kcov) TaskWork(t *Task) { kcov.mu.Lock() defer kcov.mu.Unlock() + if kcov.mode != linux.KCOV_MODE_TRACE_PC { + return + } + rw := &kcovReadWriter{ mf: kcov.mfp.MemoryFile(), fr: kcov.mappable.FileRange(), @@ -142,7 +146,7 @@ func (kcov *Kcov) InitTrace(size uint64) error { } // EnableTrace performs the KCOV_ENABLE_TRACE ioctl. -func (kcov *Kcov) EnableTrace(ctx context.Context, traceMode uint8) error { +func (kcov *Kcov) EnableTrace(ctx context.Context, traceKind uint8) error { t := TaskFromContext(ctx) if t == nil { panic("kcovInode.EnableTrace() cannot be used outside of a task goroutine") @@ -156,9 +160,9 @@ func (kcov *Kcov) EnableTrace(ctx context.Context, traceMode uint8) error { return syserror.EINVAL } - switch traceMode { + switch traceKind { case linux.KCOV_TRACE_PC: - kcov.mode = traceMode + kcov.mode = linux.KCOV_MODE_TRACE_PC case linux.KCOV_TRACE_CMP: // We do not support KCOV_MODE_TRACE_CMP. return syserror.ENOTSUP @@ -171,6 +175,7 @@ func (kcov *Kcov) EnableTrace(ctx context.Context, traceMode uint8) error { } kcov.owningTask = t + t.SetKcov(kcov) t.RegisterWork(kcov) // Clear existing coverage data; the task expects to read only coverage data @@ -192,26 +197,35 @@ func (kcov *Kcov) DisableTrace(ctx context.Context) error { if t != kcov.owningTask { return syserror.EINVAL } - kcov.owningTask = nil kcov.mode = linux.KCOV_MODE_INIT - kcov.resetLocked() + kcov.owningTask = nil + kcov.mappable = nil return nil } -// Reset is called when the owning task exits. -func (kcov *Kcov) Reset() { +// Clear resets the mode and clears the owning task and memory mapping for kcov. +// It is called when the fd corresponding to kcov is closed. Note that the mode +// needs to be set so that the next call to kcov.TaskWork() will exit early. +func (kcov *Kcov) Clear() { kcov.mu.Lock() - kcov.resetLocked() + kcov.clearLocked() kcov.mu.Unlock() } -// The kcov instance is reset when the owning task exits or when tracing is -// disabled. -func (kcov *Kcov) resetLocked() { +func (kcov *Kcov) clearLocked() { + kcov.mode = linux.KCOV_MODE_INIT kcov.owningTask = nil - if kcov.mappable != nil { - kcov.mappable = nil - } + kcov.mappable = nil +} + +// OnTaskExit is called when the owning task exits. It is similar to +// kcov.Clear(), except the memory mapping is not cleared, so that the same +// mapping can be used in the future if kcov is enabled again by another task. +func (kcov *Kcov) OnTaskExit() { + kcov.mu.Lock() + kcov.mode = linux.KCOV_MODE_INIT + kcov.owningTask = nil + kcov.mu.Unlock() } // ConfigureMMap is called by the vfs.FileDescription for this kcov instance to diff --git a/pkg/sentry/kernel/kernel.go b/pkg/sentry/kernel/kernel.go index 22f9bb006..d6c21adb7 100644 --- a/pkg/sentry/kernel/kernel.go +++ b/pkg/sentry/kernel/kernel.go @@ -220,13 +220,18 @@ type Kernel struct { // danglingEndpoints is used to save / restore tcpip.DanglingEndpoints. danglingEndpoints struct{} `state:".([]tcpip.Endpoint)"` - // sockets is the list of all network sockets the system. Protected by - // extMu. + // sockets is the list of all network sockets in the system. + // Protected by extMu. + // TODO(gvisor.dev/issue/1624): Only used by VFS1. sockets socketList - // nextSocketEntry is the next entry number to use in sockets. Protected + // socketsVFS2 records all network sockets in the system. Protected by + // extMu. + socketsVFS2 map[*vfs.FileDescription]*SocketRecord + + // nextSocketRecord is the next entry number to use in sockets. Protected // by extMu. - nextSocketEntry uint64 + nextSocketRecord uint64 // deviceRegistry is used to save/restore device.SimpleDevices. deviceRegistry struct{} `state:".(*device.Registry)"` @@ -414,6 +419,8 @@ func (k *Kernel) Init(args InitKernelArgs) error { return fmt.Errorf("failed to create sockfs mount: %v", err) } k.socketMount = socketMount + + k.socketsVFS2 = make(map[*vfs.FileDescription]*SocketRecord) } return nil @@ -507,6 +514,10 @@ func (k *Kernel) SaveTo(w wire.Writer) error { // flushMountSourceRefs flushes the MountSources for all mounted filesystems // and open FDs. func (k *Kernel) flushMountSourceRefs(ctx context.Context) error { + if VFS2Enabled { + return nil // Not relevant. + } + // Flush all mount sources for currently mounted filesystems in each task. flushed := make(map[*fs.MountNamespace]struct{}) k.tasks.mu.RLock() @@ -533,11 +544,6 @@ func (k *Kernel) flushMountSourceRefs(ctx context.Context) error { // // Precondition: Must be called with the kernel paused. func (ts *TaskSet) forEachFDPaused(ctx context.Context, 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 { @@ -556,6 +562,10 @@ func (ts *TaskSet) forEachFDPaused(ctx context.Context, f func(*fs.File, *vfs.Fi func (ts *TaskSet) flushWritesToFiles(ctx context.Context) error { // TODO(gvisor.dev/issue/1663): Add save support for VFS2. + if VFS2Enabled { + return nil + } + return ts.forEachFDPaused(ctx, func(file *fs.File, _ *vfs.FileDescription) error { if flags := file.Flags(); !flags.Write { return nil @@ -1509,20 +1519,27 @@ func (k *Kernel) SupervisorContext() context.Context { } } -// SocketEntry represents a socket recorded in Kernel.sockets. It implements +// SocketRecord represents a socket recorded in Kernel.socketsVFS2. +// +// +stateify savable +type SocketRecord struct { + k *Kernel + Sock *refs.WeakRef // TODO(gvisor.dev/issue/1624): Only used by VFS1. + SockVFS2 *vfs.FileDescription // Only used by VFS2. + ID uint64 // Socket table entry number. +} + +// SocketRecordVFS1 represents a socket recorded in Kernel.sockets. It implements // refs.WeakRefUser for sockets stored in the socket table. // // +stateify savable -type SocketEntry struct { +type SocketRecordVFS1 struct { socketEntry - k *Kernel - Sock *refs.WeakRef - SockVFS2 *vfs.FileDescription - ID uint64 // Socket table entry number. + SocketRecord } // WeakRefGone implements refs.WeakRefUser.WeakRefGone. -func (s *SocketEntry) WeakRefGone(context.Context) { +func (s *SocketRecordVFS1) WeakRefGone(context.Context) { s.k.extMu.Lock() s.k.sockets.Remove(s) s.k.extMu.Unlock() @@ -1533,9 +1550,14 @@ func (s *SocketEntry) WeakRefGone(context.Context) { // Precondition: Caller must hold a reference to sock. func (k *Kernel) RecordSocket(sock *fs.File) { k.extMu.Lock() - id := k.nextSocketEntry - k.nextSocketEntry++ - s := &SocketEntry{k: k, ID: id} + id := k.nextSocketRecord + k.nextSocketRecord++ + s := &SocketRecordVFS1{ + SocketRecord: SocketRecord{ + k: k, + ID: id, + }, + } s.Sock = refs.NewWeakRef(sock, s) k.sockets.PushBack(s) k.extMu.Unlock() @@ -1547,29 +1569,45 @@ func (k *Kernel) RecordSocket(sock *fs.File) { // 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. +// vfs.FileDescription. func (k *Kernel) RecordSocketVFS2(sock *vfs.FileDescription) { k.extMu.Lock() - id := k.nextSocketEntry - k.nextSocketEntry++ - s := &SocketEntry{ + if _, ok := k.socketsVFS2[sock]; ok { + panic(fmt.Sprintf("Socket %p added twice", sock)) + } + id := k.nextSocketRecord + k.nextSocketRecord++ + s := &SocketRecord{ k: k, ID: id, SockVFS2: sock, } - k.sockets.PushBack(s) + k.socketsVFS2[sock] = s + k.extMu.Unlock() +} + +// DeleteSocketVFS2 removes a VFS2 socket from the system-wide socket table. +func (k *Kernel) DeleteSocketVFS2(sock *vfs.FileDescription) { + k.extMu.Lock() + delete(k.socketsVFS2, sock) k.extMu.Unlock() } // ListSockets returns a snapshot of all sockets. // -// Callers of ListSockets() in VFS2 should use SocketEntry.SockVFS2.TryIncRef() +// Callers of ListSockets() in VFS2 should use SocketRecord.SockVFS2.TryIncRef() // to get a reference on a socket in the table. -func (k *Kernel) ListSockets() []*SocketEntry { +func (k *Kernel) ListSockets() []*SocketRecord { k.extMu.Lock() - var socks []*SocketEntry - for s := k.sockets.Front(); s != nil; s = s.Next() { - socks = append(socks, s) + var socks []*SocketRecord + if VFS2Enabled { + for _, s := range k.socketsVFS2 { + socks = append(socks, s) + } + } else { + for s := k.sockets.Front(); s != nil; s = s.Next() { + socks = append(socks, &s.SocketRecord) + } } k.extMu.Unlock() return socks diff --git a/pkg/sentry/kernel/pipe/BUILD b/pkg/sentry/kernel/pipe/BUILD index 449643118..99134e634 100644 --- a/pkg/sentry/kernel/pipe/BUILD +++ b/pkg/sentry/kernel/pipe/BUILD @@ -21,6 +21,7 @@ go_library( "//pkg/amutex", "//pkg/buffer", "//pkg/context", + "//pkg/marshal/primitive", "//pkg/safemem", "//pkg/sentry/arch", "//pkg/sentry/device", diff --git a/pkg/sentry/kernel/pipe/pipe.go b/pkg/sentry/kernel/pipe/pipe.go index c410c96aa..67beb0ad6 100644 --- a/pkg/sentry/kernel/pipe/pipe.go +++ b/pkg/sentry/kernel/pipe/pipe.go @@ -17,6 +17,7 @@ package pipe import ( "fmt" + "io" "sync/atomic" "syscall" @@ -215,7 +216,7 @@ func (p *Pipe) readLocked(ctx context.Context, ops readOps) (int64, error) { if p.view.Size() == 0 { if !p.HasWriters() { // There are no writers, return EOF. - return 0, nil + return 0, io.EOF } return 0, syserror.ErrWouldBlock } diff --git a/pkg/sentry/kernel/pipe/pipe_util.go b/pkg/sentry/kernel/pipe/pipe_util.go index 6d58b682f..f665920cb 100644 --- a/pkg/sentry/kernel/pipe/pipe_util.go +++ b/pkg/sentry/kernel/pipe/pipe_util.go @@ -23,6 +23,7 @@ import ( "gvisor.dev/gvisor/pkg/amutex" "gvisor.dev/gvisor/pkg/buffer" "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/marshal/primitive" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/usermem" @@ -145,9 +146,14 @@ func (p *Pipe) Ioctl(ctx context.Context, io usermem.IO, args arch.SyscallArgume v = math.MaxInt32 // Silently truncate. } // Copy result to userspace. - _, err := usermem.CopyObjectOut(ctx, io, args[2].Pointer(), int32(v), usermem.IOOpts{ - AddressSpaceActive: true, - }) + iocc := primitive.IOCopyContext{ + IO: io, + Ctx: ctx, + Opts: usermem.IOOpts{ + AddressSpaceActive: true, + }, + } + _, err := primitive.CopyInt32Out(&iocc, args[2].Pointer(), int32(v)) return 0, err default: return 0, syscall.ENOTTY diff --git a/pkg/sentry/kernel/pipe/vfs.go b/pkg/sentry/kernel/pipe/vfs.go index f223d59e1..f61039f5b 100644 --- a/pkg/sentry/kernel/pipe/vfs.go +++ b/pkg/sentry/kernel/pipe/vfs.go @@ -67,6 +67,11 @@ func (vp *VFSPipe) ReaderWriterPair(mnt *vfs.Mount, vfsd *vfs.Dentry, statusFlag return vp.newFD(mnt, vfsd, linux.O_RDONLY|statusFlags, locks), vp.newFD(mnt, vfsd, linux.O_WRONLY|statusFlags, locks) } +// Allocate implements vfs.FileDescriptionImpl.Allocate. +func (*VFSPipe) Allocate(context.Context, uint64, uint64, uint64) error { + return syserror.ESPIPE +} + // Open opens the pipe represented by vp. func (vp *VFSPipe) Open(ctx context.Context, mnt *vfs.Mount, vfsd *vfs.Dentry, statusFlags uint32, locks *vfs.FileLocks) (*vfs.FileDescription, error) { vp.mu.Lock() diff --git a/pkg/sentry/kernel/seccomp.go b/pkg/sentry/kernel/seccomp.go index c38c5a40c..387edfa91 100644 --- a/pkg/sentry/kernel/seccomp.go +++ b/pkg/sentry/kernel/seccomp.go @@ -18,7 +18,6 @@ import ( "syscall" "gvisor.dev/gvisor/pkg/abi/linux" - "gvisor.dev/gvisor/pkg/binary" "gvisor.dev/gvisor/pkg/bpf" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/syserror" @@ -27,25 +26,18 @@ import ( const maxSyscallFilterInstructions = 1 << 15 -// seccompData is equivalent to struct seccomp_data, which contains the data -// passed to seccomp-bpf filters. -type seccompData struct { - // nr is the system call number. - nr int32 - - // arch is an AUDIT_ARCH_* value indicating the system call convention. - arch uint32 - - // instructionPointer is the value of the instruction pointer at the time - // of the system call. - instructionPointer uint64 - - // args contains the first 6 system call arguments. - args [6]uint64 -} - -func (d *seccompData) asBPFInput() bpf.Input { - return bpf.InputBytes{binary.Marshal(nil, usermem.ByteOrder, d), usermem.ByteOrder} +// dataAsBPFInput returns a serialized BPF program, only valid on the current task +// goroutine. +// +// Note: this is called for every syscall, which is a very hot path. +func dataAsBPFInput(t *Task, d *linux.SeccompData) bpf.Input { + buf := t.CopyScratchBuffer(d.SizeBytes()) + d.MarshalUnsafe(buf) + return bpf.InputBytes{ + Data: buf, + // Go-marshal always uses the native byte order. + Order: usermem.ByteOrder, + } } func seccompSiginfo(t *Task, errno, sysno int32, ip usermem.Addr) *arch.SignalInfo { @@ -112,20 +104,20 @@ func (t *Task) checkSeccompSyscall(sysno int32, args arch.SyscallArguments, ip u } func (t *Task) evaluateSyscallFilters(sysno int32, args arch.SyscallArguments, ip usermem.Addr) uint32 { - data := seccompData{ - nr: sysno, - arch: t.tc.st.AuditNumber, - instructionPointer: uint64(ip), + data := linux.SeccompData{ + Nr: sysno, + Arch: t.tc.st.AuditNumber, + InstructionPointer: uint64(ip), } // data.args is []uint64 and args is []arch.SyscallArgument (uintptr), so // we can't do any slicing tricks or even use copy/append here. for i, arg := range args { - if i >= len(data.args) { + if i >= len(data.Args) { break } - data.args[i] = arg.Uint64() + data.Args[i] = arg.Uint64() } - input := data.asBPFInput() + input := dataAsBPFInput(t, &data) ret := uint32(linux.SECCOMP_RET_ALLOW) f := t.syscallFilters.Load() diff --git a/pkg/sentry/kernel/signalfd/BUILD b/pkg/sentry/kernel/signalfd/BUILD index 3eb78e91b..76d472292 100644 --- a/pkg/sentry/kernel/signalfd/BUILD +++ b/pkg/sentry/kernel/signalfd/BUILD @@ -8,7 +8,6 @@ go_library( visibility = ["//pkg/sentry:internal"], deps = [ "//pkg/abi/linux", - "//pkg/binary", "//pkg/context", "//pkg/sentry/fs", "//pkg/sentry/fs/anon", diff --git a/pkg/sentry/kernel/signalfd/signalfd.go b/pkg/sentry/kernel/signalfd/signalfd.go index b07e1c1bd..78f718cfe 100644 --- a/pkg/sentry/kernel/signalfd/signalfd.go +++ b/pkg/sentry/kernel/signalfd/signalfd.go @@ -17,7 +17,6 @@ package signalfd import ( "gvisor.dev/gvisor/pkg/abi/linux" - "gvisor.dev/gvisor/pkg/binary" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fs/anon" @@ -103,8 +102,7 @@ func (s *SignalOperations) Read(ctx context.Context, _ *fs.File, dst usermem.IOS } // Copy out the signal info using the specified format. - var buf [128]byte - binary.Marshal(buf[:0], usermem.ByteOrder, &linux.SignalfdSiginfo{ + infoNative := linux.SignalfdSiginfo{ Signo: uint32(info.Signo), Errno: info.Errno, Code: info.Code, @@ -113,9 +111,13 @@ func (s *SignalOperations) Read(ctx context.Context, _ *fs.File, dst usermem.IOS Status: info.Status(), Overrun: uint32(info.Overrun()), Addr: info.Addr(), - }) - n, err := dst.CopyOut(ctx, buf[:]) - return int64(n), err + } + n, err := infoNative.WriteTo(dst.Writer(ctx)) + if err == usermem.ErrEndOfIOSequence { + // Partial copy-out ok. + err = nil + } + return n, err } // Readiness implements waiter.Waitable.Readiness. diff --git a/pkg/sentry/kernel/task.go b/pkg/sentry/kernel/task.go index a436610c9..f796e0fa3 100644 --- a/pkg/sentry/kernel/task.go +++ b/pkg/sentry/kernel/task.go @@ -917,7 +917,7 @@ func (t *Task) SetKcov(k *Kcov) { // ResetKcov clears the kcov instance associated with t. func (t *Task) ResetKcov() { if t.kcov != nil { - t.kcov.Reset() + t.kcov.OnTaskExit() t.kcov = nil } } diff --git a/pkg/sentry/kernel/task_context.go b/pkg/sentry/kernel/task_context.go index 9fa528384..d1136461a 100644 --- a/pkg/sentry/kernel/task_context.go +++ b/pkg/sentry/kernel/task_context.go @@ -126,7 +126,11 @@ func (t *Task) SyscallTable() *SyscallTable { // Preconditions: The caller must be running on the task goroutine, or t.mu // must be locked. func (t *Task) Stack() *arch.Stack { - return &arch.Stack{t.Arch(), t.MemoryManager(), usermem.Addr(t.Arch().Stack())} + return &arch.Stack{ + Arch: t.Arch(), + IO: t.MemoryManager(), + Bottom: usermem.Addr(t.Arch().Stack()), + } } // LoadTaskImage loads a specified file into a new TaskContext. diff --git a/pkg/sentry/kernel/task_signals.go b/pkg/sentry/kernel/task_signals.go index feaa38596..ebdb83061 100644 --- a/pkg/sentry/kernel/task_signals.go +++ b/pkg/sentry/kernel/task_signals.go @@ -259,7 +259,11 @@ func (t *Task) deliverSignalToHandler(info *arch.SignalInfo, act arch.SignalAct) // Set up the signal handler. If we have a saved signal mask, the signal // handler should run with the current mask, but sigreturn should restore // the saved one. - st := &arch.Stack{t.Arch(), mm, sp} + st := &arch.Stack{ + Arch: t.Arch(), + IO: mm, + Bottom: sp, + } mask := t.signalMask if t.haveSavedSignalMask { mask = t.savedSignalMask diff --git a/pkg/sentry/kernel/task_usermem.go b/pkg/sentry/kernel/task_usermem.go index 14d765af1..ce134bf54 100644 --- a/pkg/sentry/kernel/task_usermem.go +++ b/pkg/sentry/kernel/task_usermem.go @@ -18,6 +18,7 @@ import ( "math" "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/marshal" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" ) @@ -280,29 +281,29 @@ func (t *Task) IovecsIOSequence(addr usermem.Addr, iovcnt int, opts usermem.IOOp }, nil } -// CopyContext wraps a task to allow copying memory to and from the -// task memory with user specified usermem.IOOpts. -type CopyContext struct { +// copyContext implements marshal.CopyContext. It wraps a task to allow copying +// memory to and from the task memory with custom usermem.IOOpts. +type copyContext struct { *Task opts usermem.IOOpts } // AsCopyContext wraps the task and returns it as CopyContext. -func (t *Task) AsCopyContext(opts usermem.IOOpts) *CopyContext { - return &CopyContext{t, opts} +func (t *Task) AsCopyContext(opts usermem.IOOpts) marshal.CopyContext { + return ©Context{t, opts} } // CopyInString copies a string in from the task's memory. -func (t *CopyContext) CopyInString(addr usermem.Addr, maxLen int) (string, error) { +func (t *copyContext) CopyInString(addr usermem.Addr, maxLen int) (string, error) { return usermem.CopyStringIn(t, t.MemoryManager(), addr, maxLen, t.opts) } // CopyInBytes copies task memory into dst from an IO context. -func (t *CopyContext) CopyInBytes(addr usermem.Addr, dst []byte) (int, error) { +func (t *copyContext) CopyInBytes(addr usermem.Addr, dst []byte) (int, error) { return t.MemoryManager().CopyIn(t, addr, dst, t.opts) } // CopyOutBytes copies src into task memoryfrom an IO context. -func (t *CopyContext) CopyOutBytes(addr usermem.Addr, src []byte) (int, error) { +func (t *copyContext) CopyOutBytes(addr usermem.Addr, src []byte) (int, error) { return t.MemoryManager().CopyOut(t, addr, src, t.opts) } diff --git a/pkg/sentry/kernel/vdso.go b/pkg/sentry/kernel/vdso.go index e44a139b3..9bc452e67 100644 --- a/pkg/sentry/kernel/vdso.go +++ b/pkg/sentry/kernel/vdso.go @@ -17,7 +17,6 @@ package kernel import ( "fmt" - "gvisor.dev/gvisor/pkg/binary" "gvisor.dev/gvisor/pkg/safemem" "gvisor.dev/gvisor/pkg/sentry/memmap" "gvisor.dev/gvisor/pkg/sentry/pgalloc" @@ -28,6 +27,8 @@ import ( // // They are exposed to the VDSO via a parameter page managed by VDSOParamPage, // which also includes a sequence counter. +// +// +marshal type vdsoParams struct { monotonicReady uint64 monotonicBaseCycles int64 @@ -68,6 +69,13 @@ type VDSOParamPage struct { // checked in state_test_util tests, causing this field to change across // save / restore. seq uint64 + + // copyScratchBuffer is a temporary buffer used to marshal the params before + // copying it to the real parameter page. The parameter page is typically + // updated at a moderate frequency of ~O(seconds) throughout the lifetime of + // the sentry, so reusing this buffer is a good tradeoff between memory + // usage and the cost of allocation. + copyScratchBuffer []byte } // NewVDSOParamPage returns a VDSOParamPage. @@ -79,7 +87,11 @@ type VDSOParamPage struct { // * VDSOParamPage must be the only writer to fr. // * mfp.MemoryFile().MapInternal(fr) must return a single safemem.Block. func NewVDSOParamPage(mfp pgalloc.MemoryFileProvider, fr memmap.FileRange) *VDSOParamPage { - return &VDSOParamPage{mfp: mfp, fr: fr} + return &VDSOParamPage{ + mfp: mfp, + fr: fr, + copyScratchBuffer: make([]byte, (*vdsoParams)(nil).SizeBytes()), + } } // access returns a mapping of the param page. @@ -133,7 +145,8 @@ func (v *VDSOParamPage) Write(f func() vdsoParams) error { // Get the new params. p := f() - buf := binary.Marshal(nil, usermem.ByteOrder, p) + buf := v.copyScratchBuffer[:p.SizeBytes()] + p.MarshalUnsafe(buf) // Skip the sequence counter. if _, err := safemem.Copy(paramPage.DropFirst(8), safemem.BlockFromSafeSlice(buf)); err != nil { diff --git a/pkg/sentry/loader/loader.go b/pkg/sentry/loader/loader.go index 15c88aa7c..c69b62db9 100644 --- a/pkg/sentry/loader/loader.go +++ b/pkg/sentry/loader/loader.go @@ -122,7 +122,7 @@ func allocStack(ctx context.Context, m *mm.MemoryManager, a arch.Context) (*arch if err != nil { return nil, err } - return &arch.Stack{a, m, ar.End}, nil + return &arch.Stack{Arch: a, IO: m, Bottom: ar.End}, nil } const ( @@ -247,20 +247,20 @@ func Load(ctx context.Context, args LoadArgs, extraAuxv []arch.AuxEntry, vdso *V } // Push the original filename to the stack, for AT_EXECFN. - execfn, err := stack.Push(args.Filename) - if err != nil { + if _, err := stack.PushNullTerminatedByteSlice([]byte(args.Filename)); err != nil { return 0, nil, "", syserr.NewDynamic(fmt.Sprintf("Failed to push exec filename: %v", err), syserr.FromError(err).ToLinux()) } + execfn := stack.Bottom // Push 16 random bytes on the stack which AT_RANDOM will point to. var b [16]byte if _, err := rand.Read(b[:]); err != nil { return 0, nil, "", syserr.NewDynamic(fmt.Sprintf("Failed to read random bytes: %v", err), syserr.FromError(err).ToLinux()) } - random, err := stack.Push(b) - if err != nil { + if _, err = stack.PushNullTerminatedByteSlice(b[:]); err != nil { return 0, nil, "", syserr.NewDynamic(fmt.Sprintf("Failed to push random bytes: %v", err), syserr.FromError(err).ToLinux()) } + random := stack.Bottom c := auth.CredentialsFromContext(ctx) diff --git a/pkg/sentry/mm/mm.go b/pkg/sentry/mm/mm.go index 3e85964e4..43567b92c 100644 --- a/pkg/sentry/mm/mm.go +++ b/pkg/sentry/mm/mm.go @@ -235,6 +235,14 @@ type MemoryManager struct { // vdsoSigReturnAddr is the address of 'vdso_sigreturn'. vdsoSigReturnAddr uint64 + + // membarrierPrivateEnabled is non-zero if EnableMembarrierPrivate has + // previously been called. Since, as of this writing, + // MEMBARRIER_CMD_PRIVATE_EXPEDITED is implemented as a global memory + // barrier, membarrierPrivateEnabled has no other effect. + // + // membarrierPrivateEnabled is accessed using atomic memory operations. + membarrierPrivateEnabled uint32 } // vma represents a virtual memory area. @@ -242,7 +250,7 @@ type MemoryManager struct { // +stateify savable type vma struct { // mappable is the virtual memory object mapped by this vma. If mappable is - // nil, the vma represents a private anonymous mapping. + // nil, the vma represents an anonymous mapping. mappable memmap.Mappable // off is the offset into mappable at which this vma begins. If mappable is diff --git a/pkg/sentry/mm/mm_test.go b/pkg/sentry/mm/mm_test.go index fdc308542..acac3d357 100644 --- a/pkg/sentry/mm/mm_test.go +++ b/pkg/sentry/mm/mm_test.go @@ -51,7 +51,8 @@ func TestUsageASUpdates(t *testing.T) { defer mm.DecUsers(ctx) addr, err := mm.MMap(ctx, memmap.MMapOpts{ - Length: 2 * usermem.PageSize, + Length: 2 * usermem.PageSize, + Private: true, }) if err != nil { t.Fatalf("MMap got err %v want nil", err) diff --git a/pkg/sentry/mm/special_mappable.go b/pkg/sentry/mm/special_mappable.go index f4c93baeb..2dbe5b751 100644 --- a/pkg/sentry/mm/special_mappable.go +++ b/pkg/sentry/mm/special_mappable.go @@ -136,9 +136,12 @@ func (m *SpecialMappable) Length() uint64 { // NewSharedAnonMappable returns a SpecialMappable that implements the // semantics of mmap(MAP_SHARED|MAP_ANONYMOUS) and mappings of /dev/zero. // -// TODO(jamieliu): The use of SpecialMappable is a lazy code reuse hack. Linux -// uses an ephemeral file created by mm/shmem.c:shmem_zero_setup(); we should -// do the same to get non-zero device and inode IDs. +// TODO(gvisor.dev/issue/1624): Linux uses an ephemeral file created by +// mm/shmem.c:shmem_zero_setup(), and VFS2 does something analogous. VFS1 uses +// a SpecialMappable instead, incorrectly getting device and inode IDs of zero +// and causing memory for shared anonymous mappings to be allocated up-front +// instead of on first touch; this is to avoid exacerbating the fs.MountSource +// leak (b/143656263). Delete this function along with VFS1. func NewSharedAnonMappable(length uint64, mfp pgalloc.MemoryFileProvider) (*SpecialMappable, error) { if length == 0 { return nil, syserror.EINVAL diff --git a/pkg/sentry/mm/syscalls.go b/pkg/sentry/mm/syscalls.go index 4c9a575e7..0a66b1cdd 100644 --- a/pkg/sentry/mm/syscalls.go +++ b/pkg/sentry/mm/syscalls.go @@ -17,6 +17,7 @@ package mm import ( "fmt" mrand "math/rand" + "sync/atomic" "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" @@ -24,7 +25,6 @@ import ( "gvisor.dev/gvisor/pkg/sentry/kernel/futex" "gvisor.dev/gvisor/pkg/sentry/limits" "gvisor.dev/gvisor/pkg/sentry/memmap" - "gvisor.dev/gvisor/pkg/sentry/pgalloc" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" ) @@ -93,18 +93,6 @@ func (mm *MemoryManager) MMap(ctx context.Context, opts memmap.MMapOpts) (userme } } else { opts.Offset = 0 - if !opts.Private { - if opts.MappingIdentity != nil { - return 0, syserror.EINVAL - } - m, err := NewSharedAnonMappable(opts.Length, pgalloc.MemoryFileProviderFromContext(ctx)) - if err != nil { - return 0, err - } - defer m.DecRef(ctx) - opts.MappingIdentity = m - opts.Mappable = m - } } if opts.Addr.RoundDown() != opts.Addr { @@ -1287,3 +1275,15 @@ func (mm *MemoryManager) VirtualDataSize() uint64 { defer mm.mappingMu.RUnlock() return mm.dataAS } + +// EnableMembarrierPrivate causes future calls to IsMembarrierPrivateEnabled to +// return true. +func (mm *MemoryManager) EnableMembarrierPrivate() { + atomic.StoreUint32(&mm.membarrierPrivateEnabled, 1) +} + +// IsMembarrierPrivateEnabled returns true if mm.EnableMembarrierPrivate() has +// previously been called. +func (mm *MemoryManager) IsMembarrierPrivateEnabled() bool { + return atomic.LoadUint32(&mm.membarrierPrivateEnabled) != 0 +} diff --git a/pkg/sentry/platform/BUILD b/pkg/sentry/platform/BUILD index 209b28053..db7d55ef2 100644 --- a/pkg/sentry/platform/BUILD +++ b/pkg/sentry/platform/BUILD @@ -15,6 +15,7 @@ go_library( "//pkg/context", "//pkg/seccomp", "//pkg/sentry/arch", + "//pkg/sentry/hostmm", "//pkg/sentry/memmap", "//pkg/usermem", ], diff --git a/pkg/sentry/platform/kvm/BUILD b/pkg/sentry/platform/kvm/BUILD index 3970dd81d..9fe23c417 100644 --- a/pkg/sentry/platform/kvm/BUILD +++ b/pkg/sentry/platform/kvm/BUILD @@ -9,12 +9,12 @@ go_library( "bluepill.go", "bluepill_allocator.go", "bluepill_amd64.go", - "bluepill_amd64.s", "bluepill_amd64_unsafe.go", "bluepill_arm64.go", "bluepill_arm64.s", "bluepill_arm64_unsafe.go", "bluepill_fault.go", + "bluepill_impl_amd64.s", "bluepill_unsafe.go", "context.go", "filters_amd64.go", @@ -56,6 +56,7 @@ go_library( "//pkg/sentry/time", "//pkg/sync", "//pkg/usermem", + "@org_golang_x_sys//unix:go_default_library", ], ) @@ -81,3 +82,11 @@ go_test( "//pkg/usermem", ], ) + +genrule( + name = "bluepill_impl_amd64", + srcs = ["bluepill_amd64.s"], + outs = ["bluepill_impl_amd64.s"], + cmd = "(echo -e '// build +amd64\\n' && $(location //pkg/sentry/platform/ring0/gen_offsets) && cat $(SRCS)) > $@", + tools = ["//pkg/sentry/platform/ring0/gen_offsets"], +) diff --git a/pkg/sentry/platform/kvm/bluepill_amd64.s b/pkg/sentry/platform/kvm/bluepill_amd64.s index 2bc34a435..025ea93b5 100644 --- a/pkg/sentry/platform/kvm/bluepill_amd64.s +++ b/pkg/sentry/platform/kvm/bluepill_amd64.s @@ -19,11 +19,6 @@ // This is guaranteed to be zero. #define VCPU_CPU 0x0 -// CPU_SELF is the self reference in ring0's percpu. -// -// This is guaranteed to be zero. -#define CPU_SELF 0x0 - // Context offsets. // // Only limited use of the context is done in the assembly stub below, most is @@ -44,7 +39,7 @@ begin: LEAQ VCPU_CPU(AX), BX BYTE CLI; check_vcpu: - MOVQ CPU_SELF(GS), CX + MOVQ ENTRY_CPU_SELF(GS), CX CMPQ BX, CX JE right_vCPU wrong_vcpu: diff --git a/pkg/sentry/platform/kvm/bluepill_amd64_unsafe.go b/pkg/sentry/platform/kvm/bluepill_amd64_unsafe.go index 03a98512e..0a54dd30d 100644 --- a/pkg/sentry/platform/kvm/bluepill_amd64_unsafe.go +++ b/pkg/sentry/platform/kvm/bluepill_amd64_unsafe.go @@ -83,5 +83,34 @@ func bluepillStopGuest(c *vCPU) { // //go:nosplit func bluepillReadyStopGuest(c *vCPU) bool { - return c.runData.readyForInterruptInjection != 0 + if c.runData.readyForInterruptInjection == 0 { + return false + } + + if c.runData.ifFlag == 0 { + // This is impossible if readyForInterruptInjection is 1. + throw("interrupts are disabled") + } + + // Disable interrupts if we are in the kernel space. + // + // When the Sentry switches into the kernel mode, it disables + // interrupts. But when goruntime switches on a goroutine which has + // been saved in the host mode, it restores flags and this enables + // interrupts. See the comment of UserFlagsSet for more details. + uregs := userRegs{} + err := c.getUserRegisters(&uregs) + if err != 0 { + throw("failed to get user registers") + } + + if ring0.IsKernelFlags(uregs.RFLAGS) { + uregs.RFLAGS &^= ring0.KernelFlagsClear + err = c.setUserRegisters(&uregs) + if err != 0 { + throw("failed to set user registers") + } + return false + } + return true } diff --git a/pkg/sentry/platform/kvm/filters_amd64.go b/pkg/sentry/platform/kvm/filters_amd64.go index 7d949f1dd..d3d216aa5 100644 --- a/pkg/sentry/platform/kvm/filters_amd64.go +++ b/pkg/sentry/platform/kvm/filters_amd64.go @@ -17,14 +17,23 @@ package kvm import ( "syscall" + "golang.org/x/sys/unix" + + "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/seccomp" ) // SyscallFilters returns syscalls made exclusively by the KVM platform. func (*KVM) SyscallFilters() seccomp.SyscallRules { return seccomp.SyscallRules{ - syscall.SYS_ARCH_PRCTL: {}, - syscall.SYS_IOCTL: {}, + syscall.SYS_ARCH_PRCTL: {}, + syscall.SYS_IOCTL: {}, + unix.SYS_MEMBARRIER: []seccomp.Rule{ + { + seccomp.EqualTo(linux.MEMBARRIER_CMD_PRIVATE_EXPEDITED), + seccomp.EqualTo(0), + }, + }, syscall.SYS_MMAP: {}, syscall.SYS_RT_SIGSUSPEND: {}, syscall.SYS_RT_SIGTIMEDWAIT: {}, diff --git a/pkg/sentry/platform/kvm/filters_arm64.go b/pkg/sentry/platform/kvm/filters_arm64.go index 9245d07c2..21abc2a3d 100644 --- a/pkg/sentry/platform/kvm/filters_arm64.go +++ b/pkg/sentry/platform/kvm/filters_arm64.go @@ -17,13 +17,22 @@ package kvm import ( "syscall" + "golang.org/x/sys/unix" + + "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/seccomp" ) // SyscallFilters returns syscalls made exclusively by the KVM platform. func (*KVM) SyscallFilters() seccomp.SyscallRules { return seccomp.SyscallRules{ - syscall.SYS_IOCTL: {}, + syscall.SYS_IOCTL: {}, + unix.SYS_MEMBARRIER: []seccomp.Rule{ + { + seccomp.EqualTo(linux.MEMBARRIER_CMD_PRIVATE_EXPEDITED), + seccomp.EqualTo(0), + }, + }, syscall.SYS_MMAP: {}, syscall.SYS_RT_SIGSUSPEND: {}, syscall.SYS_RT_SIGTIMEDWAIT: {}, diff --git a/pkg/sentry/platform/kvm/kvm.go b/pkg/sentry/platform/kvm/kvm.go index ae813e24e..dd45ad10b 100644 --- a/pkg/sentry/platform/kvm/kvm.go +++ b/pkg/sentry/platform/kvm/kvm.go @@ -63,6 +63,9 @@ type runData struct { type KVM struct { platform.NoCPUPreemptionDetection + // KVM never changes mm_structs. + platform.UseHostProcessMemoryBarrier + // machine is the backing VM. machine *machine } @@ -156,15 +159,7 @@ func (*KVM) MaxUserAddress() usermem.Addr { func (k *KVM) NewAddressSpace(_ interface{}) (platform.AddressSpace, <-chan struct{}, error) { // Allocate page tables and install system mappings. pageTables := pagetables.New(newAllocator()) - applyPhysicalRegions(func(pr physicalRegion) bool { - // Map the kernel in the upper half. - pageTables.Map( - usermem.Addr(ring0.KernelStartAddress|pr.virtual), - pr.length, - pagetables.MapOpts{AccessType: usermem.AnyAccess}, - pr.physical) - return true // Keep iterating. - }) + k.machine.mapUpperHalf(pageTables) // Return the new address space. return &addressSpace{ diff --git a/pkg/sentry/platform/kvm/kvm_const.go b/pkg/sentry/platform/kvm/kvm_const.go index 5c4b18899..5f627a016 100644 --- a/pkg/sentry/platform/kvm/kvm_const.go +++ b/pkg/sentry/platform/kvm/kvm_const.go @@ -32,6 +32,7 @@ const ( _KVM_SET_REGS = 0x4090ae82 _KVM_SET_SREGS = 0x4138ae84 _KVM_GET_REGS = 0x8090ae81 + _KVM_GET_SREGS = 0x8138ae83 _KVM_GET_SUPPORTED_CPUID = 0xc008ae05 _KVM_SET_CPUID2 = 0x4008ae90 _KVM_SET_SIGNAL_MASK = 0x4004ae8b diff --git a/pkg/sentry/platform/kvm/machine.go b/pkg/sentry/platform/kvm/machine.go index 372a4cbd7..75da253c5 100644 --- a/pkg/sentry/platform/kvm/machine.go +++ b/pkg/sentry/platform/kvm/machine.go @@ -155,7 +155,7 @@ func (m *machine) newVCPU() *vCPU { fd: int(fd), machine: m, } - c.CPU.Init(&m.kernel, c) + c.CPU.Init(&m.kernel, c.id, c) m.vCPUsByID[c.id] = c // Ensure the signal mask is correct. @@ -183,9 +183,6 @@ func newMachine(vm int) (*machine, error) { // Create the machine. m := &machine{fd: vm} m.available.L = &m.mu - m.kernel.Init(ring0.KernelOpts{ - PageTables: pagetables.New(newAllocator()), - }) // Pull the maximum vCPUs. maxVCPUs, _, errno := syscall.RawSyscall(syscall.SYS_IOCTL, uintptr(m.fd), _KVM_CHECK_EXTENSION, _KVM_CAP_MAX_VCPUS) @@ -197,6 +194,9 @@ func newMachine(vm int) (*machine, error) { log.Debugf("The maximum number of vCPUs is %d.", m.maxVCPUs) m.vCPUsByTID = make(map[uint64]*vCPU) m.vCPUsByID = make([]*vCPU, m.maxVCPUs) + m.kernel.Init(ring0.KernelOpts{ + PageTables: pagetables.New(newAllocator()), + }, m.maxVCPUs) // Pull the maximum slots. maxSlots, _, errno := syscall.RawSyscall(syscall.SYS_IOCTL, uintptr(m.fd), _KVM_CHECK_EXTENSION, _KVM_CAP_MAX_MEMSLOTS) @@ -219,15 +219,9 @@ func newMachine(vm int) (*machine, error) { pagetables.MapOpts{AccessType: usermem.AnyAccess}, pr.physical) - // And keep everything in the upper half. - m.kernel.PageTables.Map( - usermem.Addr(ring0.KernelStartAddress|pr.virtual), - pr.length, - pagetables.MapOpts{AccessType: usermem.AnyAccess}, - pr.physical) - return true // Keep iterating. }) + m.mapUpperHalf(m.kernel.PageTables) var physicalRegionsReadOnly []physicalRegion var physicalRegionsAvailable []physicalRegion @@ -365,6 +359,11 @@ func (m *machine) Destroy() { // Get gets an available vCPU. // // This will return with the OS thread locked. +// +// It is guaranteed that if any OS thread TID is in guest, m.vCPUs[TID] points +// to the vCPU in which the OS thread TID is running. So if Get() returns with +// the corrent context in guest, the vCPU of it must be the same as what +// Get() returns. func (m *machine) Get() *vCPU { m.mu.RLock() runtime.LockOSThread() diff --git a/pkg/sentry/platform/kvm/machine_amd64.go b/pkg/sentry/platform/kvm/machine_amd64.go index acc823ba6..54e721bb1 100644 --- a/pkg/sentry/platform/kvm/machine_amd64.go +++ b/pkg/sentry/platform/kvm/machine_amd64.go @@ -144,6 +144,7 @@ func (c *vCPU) initArchState() error { // Set the entrypoint for the kernel. kernelUserRegs.RIP = uint64(reflect.ValueOf(ring0.Start).Pointer()) kernelUserRegs.RAX = uint64(reflect.ValueOf(&c.CPU).Pointer()) + kernelUserRegs.RSP = c.StackTop() kernelUserRegs.RFLAGS = ring0.KernelFlagsSet // Set the system registers. @@ -152,8 +153,8 @@ func (c *vCPU) initArchState() error { } // Set the user registers. - if err := c.setUserRegisters(&kernelUserRegs); err != nil { - return err + if errno := c.setUserRegisters(&kernelUserRegs); errno != 0 { + return fmt.Errorf("error setting user registers: %v", errno) } // Allocate some floating point state save area for the local vCPU. @@ -345,3 +346,43 @@ func rdonlyRegionsForSetMem() (phyRegions []physicalRegion) { func availableRegionsForSetMem() (phyRegions []physicalRegion) { return physicalRegions } + +var execRegions []region + +func init() { + applyVirtualRegions(func(vr virtualRegion) { + if excludeVirtualRegion(vr) || vr.filename == "[vsyscall]" { + return + } + + if vr.accessType.Execute { + execRegions = append(execRegions, vr.region) + } + }) +} + +func (m *machine) mapUpperHalf(pageTable *pagetables.PageTables) { + for _, r := range execRegions { + physical, length, ok := translateToPhysical(r.virtual) + if !ok || length < r.length { + panic("impossilbe translation") + } + pageTable.Map( + usermem.Addr(ring0.KernelStartAddress|r.virtual), + r.length, + pagetables.MapOpts{AccessType: usermem.Execute}, + physical) + } + for start, end := range m.kernel.EntryRegions() { + regionLen := end - start + physical, length, ok := translateToPhysical(start) + if !ok || length < regionLen { + panic("impossible translation") + } + pageTable.Map( + usermem.Addr(ring0.KernelStartAddress|start), + regionLen, + pagetables.MapOpts{AccessType: usermem.ReadWrite}, + physical) + } +} diff --git a/pkg/sentry/platform/kvm/machine_amd64_unsafe.go b/pkg/sentry/platform/kvm/machine_amd64_unsafe.go index 290f035dd..330f29065 100644 --- a/pkg/sentry/platform/kvm/machine_amd64_unsafe.go +++ b/pkg/sentry/platform/kvm/machine_amd64_unsafe.go @@ -137,15 +137,17 @@ func (c *vCPU) setSignalMask() error { } // setUserRegisters sets user registers in the vCPU. -func (c *vCPU) setUserRegisters(uregs *userRegs) error { +// +//go:nosplit +func (c *vCPU) setUserRegisters(uregs *userRegs) syscall.Errno { if _, _, errno := syscall.RawSyscall( syscall.SYS_IOCTL, uintptr(c.fd), _KVM_SET_REGS, uintptr(unsafe.Pointer(uregs))); errno != 0 { - return fmt.Errorf("error setting user registers: %v", errno) + return errno } - return nil + return 0 } // getUserRegisters reloads user registers in the vCPU. @@ -175,3 +177,17 @@ func (c *vCPU) setSystemRegisters(sregs *systemRegs) error { } return nil } + +// getSystemRegisters sets system registers. +// +//go:nosplit +func (c *vCPU) getSystemRegisters(sregs *systemRegs) syscall.Errno { + if _, _, errno := syscall.RawSyscall( + syscall.SYS_IOCTL, + uintptr(c.fd), + _KVM_GET_SREGS, + uintptr(unsafe.Pointer(sregs))); errno != 0 { + return errno + } + return 0 +} diff --git a/pkg/sentry/platform/kvm/machine_arm64.go b/pkg/sentry/platform/kvm/machine_arm64.go index 9db171af9..2df762991 100644 --- a/pkg/sentry/platform/kvm/machine_arm64.go +++ b/pkg/sentry/platform/kvm/machine_arm64.go @@ -19,6 +19,7 @@ package kvm import ( "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/platform" + "gvisor.dev/gvisor/pkg/sentry/platform/ring0" "gvisor.dev/gvisor/pkg/sentry/platform/ring0/pagetables" "gvisor.dev/gvisor/pkg/usermem" ) @@ -48,6 +49,18 @@ const ( poolPCIDs = 8 ) +func (m *machine) mapUpperHalf(pageTable *pagetables.PageTables) { + applyPhysicalRegions(func(pr physicalRegion) bool { + pageTable.Map( + usermem.Addr(ring0.KernelStartAddress|pr.virtual), + pr.length, + pagetables.MapOpts{AccessType: usermem.AnyAccess}, + pr.physical) + + return true // Keep iterating. + }) +} + // Get all read-only physicalRegions. func rdonlyRegionsForSetMem() (phyRegions []physicalRegion) { var rdonlyRegions []region diff --git a/pkg/sentry/platform/kvm/machine_arm64_unsafe.go b/pkg/sentry/platform/kvm/machine_arm64_unsafe.go index 905712076..537419657 100644 --- a/pkg/sentry/platform/kvm/machine_arm64_unsafe.go +++ b/pkg/sentry/platform/kvm/machine_arm64_unsafe.go @@ -79,7 +79,7 @@ func (c *vCPU) initArchState() error { } // tcr_el1 - data = _TCR_TXSZ_VA48 | _TCR_CACHE_FLAGS | _TCR_SHARED | _TCR_TG_FLAGS | _TCR_ASID16 | _TCR_IPS_40BITS + data = _TCR_TXSZ_VA48 | _TCR_CACHE_FLAGS | _TCR_SHARED | _TCR_TG_FLAGS | _TCR_ASID16 | _TCR_IPS_40BITS | _TCR_A1 reg.id = _KVM_ARM64_REGS_TCR_EL1 if err := c.setOneRegister(®); err != nil { return err @@ -103,7 +103,7 @@ func (c *vCPU) initArchState() error { c.SetTtbr0Kvm(uintptr(data)) // ttbr1_el1 - data = c.machine.kernel.PageTables.TTBR1_EL1(false, 0) + data = c.machine.kernel.PageTables.TTBR1_EL1(false, 1) reg.id = _KVM_ARM64_REGS_TTBR1_EL1 if err := c.setOneRegister(®); err != nil { diff --git a/pkg/sentry/platform/platform.go b/pkg/sentry/platform/platform.go index 530e779b0..dcfe839a7 100644 --- a/pkg/sentry/platform/platform.go +++ b/pkg/sentry/platform/platform.go @@ -25,6 +25,7 @@ import ( "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/seccomp" "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/hostmm" "gvisor.dev/gvisor/pkg/sentry/memmap" "gvisor.dev/gvisor/pkg/usermem" ) @@ -52,6 +53,10 @@ type Platform interface { // can reliably return ErrContextCPUPreempted. DetectsCPUPreemption() bool + // HaveGlobalMemoryBarrier returns true if the GlobalMemoryBarrier method + // is supported. + HaveGlobalMemoryBarrier() bool + // MapUnit returns the alignment used for optional mappings into this // platform's AddressSpaces. Higher values indicate lower per-page costs // for AddressSpace.MapFile. As a special case, a MapUnit of 0 indicates @@ -97,6 +102,15 @@ type Platform interface { // called. PreemptAllCPUs() error + // GlobalMemoryBarrier blocks until all threads running application code + // (via Context.Switch) and all task goroutines "have passed through a + // state where all memory accesses to user-space addresses match program + // order between entry to and return from [GlobalMemoryBarrier]", as for + // membarrier(2). + // + // Preconditions: HaveGlobalMemoryBarrier() == true. + GlobalMemoryBarrier() error + // SyscallFilters returns syscalls made exclusively by this platform. SyscallFilters() seccomp.SyscallRules } @@ -115,6 +129,43 @@ func (NoCPUPreemptionDetection) PreemptAllCPUs() error { panic("This platform does not support CPU preemption detection") } +// UseHostGlobalMemoryBarrier implements Platform.HaveGlobalMemoryBarrier and +// Platform.GlobalMemoryBarrier by invoking equivalent functionality on the +// host. +type UseHostGlobalMemoryBarrier struct{} + +// HaveGlobalMemoryBarrier implements Platform.HaveGlobalMemoryBarrier. +func (UseHostGlobalMemoryBarrier) HaveGlobalMemoryBarrier() bool { + return hostmm.HaveGlobalMemoryBarrier() +} + +// GlobalMemoryBarrier implements Platform.GlobalMemoryBarrier. +func (UseHostGlobalMemoryBarrier) GlobalMemoryBarrier() error { + return hostmm.GlobalMemoryBarrier() +} + +// UseHostProcessMemoryBarrier implements Platform.HaveGlobalMemoryBarrier and +// Platform.GlobalMemoryBarrier by invoking a process-local memory barrier. +// This is faster than UseHostGlobalMemoryBarrier, but is only appropriate for +// platforms for which application code executes while using the sentry's +// mm_struct. +type UseHostProcessMemoryBarrier struct{} + +// HaveGlobalMemoryBarrier implements Platform.HaveGlobalMemoryBarrier. +func (UseHostProcessMemoryBarrier) HaveGlobalMemoryBarrier() bool { + // Fall back to a global memory barrier if a process-local one isn't + // available. + return hostmm.HaveProcessMemoryBarrier() || hostmm.HaveGlobalMemoryBarrier() +} + +// GlobalMemoryBarrier implements Platform.GlobalMemoryBarrier. +func (UseHostProcessMemoryBarrier) GlobalMemoryBarrier() error { + if hostmm.HaveProcessMemoryBarrier() { + return hostmm.ProcessMemoryBarrier() + } + return hostmm.GlobalMemoryBarrier() +} + // MemoryManager represents an abstraction above the platform address space // which manages memory mappings and their contents. type MemoryManager interface { diff --git a/pkg/sentry/platform/ptrace/ptrace.go b/pkg/sentry/platform/ptrace/ptrace.go index b52d0fbd8..f56aa3b79 100644 --- a/pkg/sentry/platform/ptrace/ptrace.go +++ b/pkg/sentry/platform/ptrace/ptrace.go @@ -192,6 +192,7 @@ func (c *context) PullFullState(as platform.AddressSpace, ac arch.Context) {} type PTrace struct { platform.MMapMinAddr platform.NoCPUPreemptionDetection + platform.UseHostGlobalMemoryBarrier } // New returns a new ptrace-based implementation of the platform interface. diff --git a/pkg/sentry/platform/ring0/defs_amd64.go b/pkg/sentry/platform/ring0/defs_amd64.go index 9c6c2cf5c..f617519fa 100644 --- a/pkg/sentry/platform/ring0/defs_amd64.go +++ b/pkg/sentry/platform/ring0/defs_amd64.go @@ -76,15 +76,42 @@ type KernelOpts struct { type KernelArchState struct { KernelOpts + // cpuEntries is array of kernelEntry for all cpus + cpuEntries []kernelEntry + // globalIDT is our set of interrupt gates. - globalIDT idt64 + globalIDT *idt64 } -// CPUArchState contains CPU-specific arch state. -type CPUArchState struct { +// kernelEntry contains minimal CPU-specific arch state +// that can be mapped at the upper of the address space. +// Malicious APP might steal info from it via CPU bugs. +type kernelEntry struct { // stack is the stack used for interrupts on this CPU. stack [256]byte + // scratch space for temporary usage. + scratch0 uint64 + scratch1 uint64 + + // stackTop is the top of the stack. + stackTop uint64 + + // cpuSelf is back reference to CPU. + cpuSelf *CPU + + // kernelCR3 is the cr3 used for sentry kernel. + kernelCR3 uintptr + + // gdt is the CPU's descriptor table. + gdt descriptorTable + + // tss is the CPU's task state. + tss TaskState64 +} + +// CPUArchState contains CPU-specific arch state. +type CPUArchState struct { // errorCode is the error code from the last exception. errorCode uintptr @@ -97,11 +124,7 @@ type CPUArchState struct { // exception. errorType uintptr - // gdt is the CPU's descriptor table. - gdt descriptorTable - - // tss is the CPU's task state. - tss TaskState64 + *kernelEntry } // ErrorCode returns the last error code. diff --git a/pkg/sentry/platform/ring0/defs_arm64.go b/pkg/sentry/platform/ring0/defs_arm64.go index 0e2ab716c..508236e46 100644 --- a/pkg/sentry/platform/ring0/defs_arm64.go +++ b/pkg/sentry/platform/ring0/defs_arm64.go @@ -77,6 +77,9 @@ type CPUArchState struct { // lazyVFP is the value of cpacr_el1. lazyVFP uintptr + + // appASID is the asid value of guest application. + appASID uintptr } // ErrorCode returns the last error code. diff --git a/pkg/sentry/platform/ring0/entry_amd64.go b/pkg/sentry/platform/ring0/entry_amd64.go index 7fa43c2f5..d87b1fd00 100644 --- a/pkg/sentry/platform/ring0/entry_amd64.go +++ b/pkg/sentry/platform/ring0/entry_amd64.go @@ -36,12 +36,15 @@ func sysenter() // This must be called prior to sysret/iret. func swapgs() +// jumpToKernel jumps to the kernel version of the current RIP. +func jumpToKernel() + // sysret returns to userspace from a system call. // // The return code is the vector that interrupted execution. // // See stubs.go for a note regarding the frame size of this function. -func sysret(*CPU, *arch.Registers) Vector +func sysret(cpu *CPU, regs *arch.Registers, userCR3 uintptr) Vector // "iret is the cadillac of CPL switching." // @@ -50,7 +53,7 @@ func sysret(*CPU, *arch.Registers) Vector // iret is nearly identical to sysret, except an iret is used to fully restore // all user state. This must be called in cases where all registers need to be // restored. -func iret(*CPU, *arch.Registers) Vector +func iret(cpu *CPU, regs *arch.Registers, userCR3 uintptr) Vector // exception is the generic exception entry. // diff --git a/pkg/sentry/platform/ring0/entry_amd64.s b/pkg/sentry/platform/ring0/entry_amd64.s index 02df38331..f59747df3 100644 --- a/pkg/sentry/platform/ring0/entry_amd64.s +++ b/pkg/sentry/platform/ring0/entry_amd64.s @@ -63,6 +63,15 @@ MOVQ offset+PTRACE_RSI(reg), SI; \ MOVQ offset+PTRACE_RDI(reg), DI; +// WRITE_CR3() writes the given CR3 value. +// +// The code corresponds to: +// +// mov %rax, %cr3 +// +#define WRITE_CR3() \ + BYTE $0x0f; BYTE $0x22; BYTE $0xd8; + // SWAP_GS swaps the kernel GS (CPU). #define SWAP_GS() \ BYTE $0x0F; BYTE $0x01; BYTE $0xf8; @@ -75,15 +84,9 @@ #define SYSRET64() \ BYTE $0x48; BYTE $0x0f; BYTE $0x07; -// LOAD_KERNEL_ADDRESS loads a kernel address. -#define LOAD_KERNEL_ADDRESS(from, to) \ - MOVQ from, to; \ - ORQ ·KernelStartAddress(SB), to; - // LOAD_KERNEL_STACK loads the kernel stack. -#define LOAD_KERNEL_STACK(from) \ - LOAD_KERNEL_ADDRESS(CPU_SELF(from), SP); \ - LEAQ CPU_STACK_TOP(SP), SP; +#define LOAD_KERNEL_STACK(entry) \ + MOVQ ENTRY_STACK_TOP(entry), SP; // See kernel.go. TEXT ·Halt(SB),NOSPLIT,$0 @@ -95,58 +98,93 @@ TEXT ·swapgs(SB),NOSPLIT,$0 SWAP_GS() RET +// jumpToKernel changes execution to the kernel address space. +// +// This works by changing the return value to the kernel version. +TEXT ·jumpToKernel(SB),NOSPLIT,$0 + MOVQ 0(SP), AX + ORQ ·KernelStartAddress(SB), AX // Future return value. + MOVQ AX, 0(SP) + RET + // See entry_amd64.go. TEXT ·sysret(SB),NOSPLIT,$0-24 - // Save original state. - LOAD_KERNEL_ADDRESS(cpu+0(FP), BX) - LOAD_KERNEL_ADDRESS(regs+8(FP), AX) + CALL ·jumpToKernel(SB) + // Save original state and stack. sysenter() or exception() + // from APP(gr3) will switch to this stack, set the return + // value (vector: 32(SP)) and then do RET, which will also + // automatically return to the lower half. + MOVQ cpu+0(FP), BX + MOVQ regs+8(FP), AX + MOVQ userCR3+16(FP), CX MOVQ SP, CPU_REGISTERS+PTRACE_RSP(BX) MOVQ BP, CPU_REGISTERS+PTRACE_RBP(BX) MOVQ AX, CPU_REGISTERS+PTRACE_RAX(BX) + // save SP AX userCR3 on the kernel stack. + MOVQ CPU_ENTRY(BX), BX + LOAD_KERNEL_STACK(BX) + PUSHQ PTRACE_RSP(AX) + PUSHQ PTRACE_RAX(AX) + PUSHQ CX + // Restore user register state. REGISTERS_LOAD(AX, 0) MOVQ PTRACE_RIP(AX), CX // Needed for SYSRET. MOVQ PTRACE_FLAGS(AX), R11 // Needed for SYSRET. - MOVQ PTRACE_RSP(AX), SP // Restore the stack directly. - MOVQ PTRACE_RAX(AX), AX // Restore AX (scratch). + + // restore userCR3, AX, SP. + POPQ AX // Get userCR3. + WRITE_CR3() // Switch to userCR3. + POPQ AX // Restore AX. + POPQ SP // Restore SP. SYSRET64() // See entry_amd64.go. TEXT ·iret(SB),NOSPLIT,$0-24 - // Save original state. - LOAD_KERNEL_ADDRESS(cpu+0(FP), BX) - LOAD_KERNEL_ADDRESS(regs+8(FP), AX) + CALL ·jumpToKernel(SB) + // Save original state and stack. sysenter() or exception() + // from APP(gr3) will switch to this stack, set the return + // value (vector: 32(SP)) and then do RET, which will also + // automatically return to the lower half. + MOVQ cpu+0(FP), BX + MOVQ regs+8(FP), AX + MOVQ userCR3+16(FP), CX MOVQ SP, CPU_REGISTERS+PTRACE_RSP(BX) MOVQ BP, CPU_REGISTERS+PTRACE_RBP(BX) MOVQ AX, CPU_REGISTERS+PTRACE_RAX(BX) // Build an IRET frame & restore state. + MOVQ CPU_ENTRY(BX), BX LOAD_KERNEL_STACK(BX) - MOVQ PTRACE_SS(AX), BX; PUSHQ BX - MOVQ PTRACE_RSP(AX), CX; PUSHQ CX - MOVQ PTRACE_FLAGS(AX), DX; PUSHQ DX - MOVQ PTRACE_CS(AX), DI; PUSHQ DI - MOVQ PTRACE_RIP(AX), SI; PUSHQ SI - REGISTERS_LOAD(AX, 0) // Restore most registers. - MOVQ PTRACE_RAX(AX), AX // Restore AX (scratch). + PUSHQ PTRACE_SS(AX) + PUSHQ PTRACE_RSP(AX) + PUSHQ PTRACE_FLAGS(AX) + PUSHQ PTRACE_CS(AX) + PUSHQ PTRACE_RIP(AX) + PUSHQ PTRACE_RAX(AX) // Save AX on kernel stack. + PUSHQ CX // Save userCR3 on kernel stack. + REGISTERS_LOAD(AX, 0) // Restore most registers. + POPQ AX // Get userCR3. + WRITE_CR3() // Switch to userCR3. + POPQ AX // Restore AX. IRET() // See entry_amd64.go. TEXT ·resume(SB),NOSPLIT,$0 // See iret, above. - MOVQ CPU_REGISTERS+PTRACE_SS(GS), BX; PUSHQ BX - MOVQ CPU_REGISTERS+PTRACE_RSP(GS), CX; PUSHQ CX - MOVQ CPU_REGISTERS+PTRACE_FLAGS(GS), DX; PUSHQ DX - MOVQ CPU_REGISTERS+PTRACE_CS(GS), DI; PUSHQ DI - MOVQ CPU_REGISTERS+PTRACE_RIP(GS), SI; PUSHQ SI - REGISTERS_LOAD(GS, CPU_REGISTERS) - MOVQ CPU_REGISTERS+PTRACE_RAX(GS), AX + MOVQ ENTRY_CPU_SELF(GS), AX // Load vCPU. + PUSHQ CPU_REGISTERS+PTRACE_SS(AX) + PUSHQ CPU_REGISTERS+PTRACE_RSP(AX) + PUSHQ CPU_REGISTERS+PTRACE_FLAGS(AX) + PUSHQ CPU_REGISTERS+PTRACE_CS(AX) + PUSHQ CPU_REGISTERS+PTRACE_RIP(AX) + REGISTERS_LOAD(AX, CPU_REGISTERS) + MOVQ CPU_REGISTERS+PTRACE_RAX(AX), AX IRET() // See entry_amd64.go. TEXT ·Start(SB),NOSPLIT,$0 - LOAD_KERNEL_STACK(AX) // Set the stack. PUSHQ $0x0 // Previous frame pointer. MOVQ SP, BP // Set frame pointer. PUSHQ AX // First argument (CPU). @@ -155,53 +193,60 @@ TEXT ·Start(SB),NOSPLIT,$0 // See entry_amd64.go. TEXT ·sysenter(SB),NOSPLIT,$0 - // Interrupts are always disabled while we're executing in kernel mode - // and always enabled while executing in user mode. Therefore, we can - // reliably look at the flags in R11 to determine where this syscall - // was from. - TESTL $_RFLAGS_IF, R11 + // _RFLAGS_IOPL0 is always set in the user mode and it is never set in + // the kernel mode. See the comment of UserFlagsSet for more details. + TESTL $_RFLAGS_IOPL0, R11 JZ kernel - user: SWAP_GS() - XCHGQ CPU_REGISTERS+PTRACE_RSP(GS), SP // Swap stacks. - XCHGQ CPU_REGISTERS+PTRACE_RAX(GS), AX // Swap for AX (regs). + MOVQ AX, ENTRY_SCRATCH0(GS) // Save user AX on scratch. + MOVQ ENTRY_KERNEL_CR3(GS), AX // Get kernel cr3 on AX. + WRITE_CR3() // Switch to kernel cr3. + + MOVQ ENTRY_CPU_SELF(GS), AX // Load vCPU. + MOVQ CPU_REGISTERS+PTRACE_RAX(AX), AX // Get user regs. REGISTERS_SAVE(AX, 0) // Save all except IP, FLAGS, SP, AX. - MOVQ CPU_REGISTERS+PTRACE_RAX(GS), BX // Load saved AX value. - MOVQ BX, PTRACE_RAX(AX) // Save everything else. - MOVQ BX, PTRACE_ORIGRAX(AX) MOVQ CX, PTRACE_RIP(AX) MOVQ R11, PTRACE_FLAGS(AX) - MOVQ CPU_REGISTERS+PTRACE_RSP(GS), BX; MOVQ BX, PTRACE_RSP(AX) - MOVQ $0, CPU_ERROR_CODE(GS) // Clear error code. - MOVQ $1, CPU_ERROR_TYPE(GS) // Set error type to user. + MOVQ SP, PTRACE_RSP(AX) + MOVQ ENTRY_SCRATCH0(GS), CX // Load saved user AX value. + MOVQ CX, PTRACE_RAX(AX) // Save everything else. + MOVQ CX, PTRACE_ORIGRAX(AX) + + MOVQ ENTRY_CPU_SELF(GS), AX // Load vCPU. + MOVQ CPU_REGISTERS+PTRACE_RSP(AX), SP // Get stacks. + MOVQ $0, CPU_ERROR_CODE(AX) // Clear error code. + MOVQ $1, CPU_ERROR_TYPE(AX) // Set error type to user. // Return to the kernel, where the frame is: // - // vector (sp+24) + // vector (sp+32) + // userCR3 (sp+24) // regs (sp+16) // cpu (sp+8) // vcpu.Switch (sp+0) // - MOVQ CPU_REGISTERS+PTRACE_RBP(GS), BP // Original base pointer. - MOVQ $Syscall, 24(SP) // Output vector. + MOVQ CPU_REGISTERS+PTRACE_RBP(AX), BP // Original base pointer. + MOVQ $Syscall, 32(SP) // Output vector. RET kernel: // We can't restore the original stack, but we can access the registers // in the CPU state directly. No need for temporary juggling. - MOVQ AX, CPU_REGISTERS+PTRACE_ORIGRAX(GS) - MOVQ AX, CPU_REGISTERS+PTRACE_RAX(GS) - REGISTERS_SAVE(GS, CPU_REGISTERS) - MOVQ CX, CPU_REGISTERS+PTRACE_RIP(GS) - MOVQ R11, CPU_REGISTERS+PTRACE_FLAGS(GS) - MOVQ SP, CPU_REGISTERS+PTRACE_RSP(GS) - MOVQ $0, CPU_ERROR_CODE(GS) // Clear error code. - MOVQ $0, CPU_ERROR_TYPE(GS) // Set error type to kernel. + MOVQ AX, ENTRY_SCRATCH0(GS) + MOVQ ENTRY_CPU_SELF(GS), AX // Load vCPU. + REGISTERS_SAVE(AX, CPU_REGISTERS) + MOVQ CX, CPU_REGISTERS+PTRACE_RIP(AX) + MOVQ R11, CPU_REGISTERS+PTRACE_FLAGS(AX) + MOVQ SP, CPU_REGISTERS+PTRACE_RSP(AX) + MOVQ ENTRY_SCRATCH0(GS), BX + MOVQ BX, CPU_REGISTERS+PTRACE_ORIGRAX(AX) + MOVQ BX, CPU_REGISTERS+PTRACE_RAX(AX) + MOVQ $0, CPU_ERROR_CODE(AX) // Clear error code. + MOVQ $0, CPU_ERROR_TYPE(AX) // Set error type to kernel. // Call the syscall trampoline. LOAD_KERNEL_STACK(GS) - MOVQ CPU_SELF(GS), AX // Load vCPU. PUSHQ AX // First argument (vCPU). CALL ·kernelSyscall(SB) // Call the trampoline. POPQ AX // Pop vCPU. @@ -230,16 +275,21 @@ TEXT ·exception(SB),NOSPLIT,$0 // ERROR_CODE (sp+8) // VECTOR (sp+0) // - TESTL $_RFLAGS_IF, 32(SP) + TESTL $_RFLAGS_IOPL0, 32(SP) JZ kernel user: SWAP_GS() ADDQ $-8, SP // Adjust for flags. MOVQ $_KERNEL_FLAGS, 0(SP); BYTE $0x9d; // Reset flags (POPFQ). - XCHGQ CPU_REGISTERS+PTRACE_RAX(GS), AX // Swap for user regs. + PUSHQ AX // Save user AX on stack. + MOVQ ENTRY_KERNEL_CR3(GS), AX // Get kernel cr3 on AX. + WRITE_CR3() // Switch to kernel cr3. + + MOVQ ENTRY_CPU_SELF(GS), AX // Load vCPU. + MOVQ CPU_REGISTERS+PTRACE_RAX(AX), AX // Get user regs. REGISTERS_SAVE(AX, 0) // Save all except IP, FLAGS, SP, AX. - MOVQ CPU_REGISTERS+PTRACE_RAX(GS), BX // Restore original AX. + POPQ BX // Restore original AX. MOVQ BX, PTRACE_RAX(AX) // Save it. MOVQ BX, PTRACE_ORIGRAX(AX) MOVQ 16(SP), BX; MOVQ BX, PTRACE_RIP(AX) @@ -249,34 +299,36 @@ user: MOVQ 48(SP), SI; MOVQ SI, PTRACE_SS(AX) // Copy out and return. + MOVQ ENTRY_CPU_SELF(GS), AX // Load vCPU. MOVQ 0(SP), BX // Load vector. MOVQ 8(SP), CX // Load error code. - MOVQ CPU_REGISTERS+PTRACE_RSP(GS), SP // Original stack (kernel version). - MOVQ CPU_REGISTERS+PTRACE_RBP(GS), BP // Original base pointer. - MOVQ CX, CPU_ERROR_CODE(GS) // Set error code. - MOVQ $1, CPU_ERROR_TYPE(GS) // Set error type to user. - MOVQ BX, 24(SP) // Output vector. + MOVQ CPU_REGISTERS+PTRACE_RSP(AX), SP // Original stack (kernel version). + MOVQ CPU_REGISTERS+PTRACE_RBP(AX), BP // Original base pointer. + MOVQ CX, CPU_ERROR_CODE(AX) // Set error code. + MOVQ $1, CPU_ERROR_TYPE(AX) // Set error type to user. + MOVQ BX, 32(SP) // Output vector. RET kernel: // As per above, we can save directly. - MOVQ AX, CPU_REGISTERS+PTRACE_RAX(GS) - MOVQ AX, CPU_REGISTERS+PTRACE_ORIGRAX(GS) - REGISTERS_SAVE(GS, CPU_REGISTERS) - MOVQ 16(SP), AX; MOVQ AX, CPU_REGISTERS+PTRACE_RIP(GS) - MOVQ 32(SP), BX; MOVQ BX, CPU_REGISTERS+PTRACE_FLAGS(GS) - MOVQ 40(SP), CX; MOVQ CX, CPU_REGISTERS+PTRACE_RSP(GS) + PUSHQ AX + MOVQ ENTRY_CPU_SELF(GS), AX // Load vCPU. + REGISTERS_SAVE(AX, CPU_REGISTERS) + POPQ BX + MOVQ BX, CPU_REGISTERS+PTRACE_RAX(AX) + MOVQ BX, CPU_REGISTERS+PTRACE_ORIGRAX(AX) + MOVQ 16(SP), BX; MOVQ BX, CPU_REGISTERS+PTRACE_RIP(AX) + MOVQ 32(SP), BX; MOVQ BX, CPU_REGISTERS+PTRACE_FLAGS(AX) + MOVQ 40(SP), BX; MOVQ BX, CPU_REGISTERS+PTRACE_RSP(AX) // Set the error code and adjust the stack. - MOVQ 8(SP), AX // Load the error code. - MOVQ AX, CPU_ERROR_CODE(GS) // Copy out to the CPU. - MOVQ $0, CPU_ERROR_TYPE(GS) // Set error type to kernel. + MOVQ 8(SP), BX // Load the error code. + MOVQ BX, CPU_ERROR_CODE(AX) // Copy out to the CPU. + MOVQ $0, CPU_ERROR_TYPE(AX) // Set error type to kernel. MOVQ 0(SP), BX // BX contains the vector. - ADDQ $48, SP // Drop the exception frame. // Call the exception trampoline. LOAD_KERNEL_STACK(GS) - MOVQ CPU_SELF(GS), AX // Load vCPU. PUSHQ BX // Second argument (vector). PUSHQ AX // First argument (vCPU). CALL ·kernelException(SB) // Call the trampoline. diff --git a/pkg/sentry/platform/ring0/entry_arm64.s b/pkg/sentry/platform/ring0/entry_arm64.s index 1e477cc49..f3d934996 100644 --- a/pkg/sentry/platform/ring0/entry_arm64.s +++ b/pkg/sentry/platform/ring0/entry_arm64.s @@ -302,17 +302,23 @@ // SWITCH_TO_APP_PAGETABLE sets a new pagetable for a container application. #define SWITCH_TO_APP_PAGETABLE(from) \ - MOVD CPU_TTBR0_APP(from), RSV_REG; \ - WORD $0xd5182012; \ // MSR R18, TTBR0_EL1 + MRS TTBR1_EL1, R0; \ + MOVD CPU_APP_ASID(from), R1; \ + BFI $48, R1, $16, R0; \ + MSR R0, TTBR1_EL1; \ // set the ASID in TTBR1_EL1 (since TCR.A1 is set) ISB $15; \ - DSB $15; + MOVD CPU_TTBR0_APP(from), RSV_REG; \ + MSR RSV_REG, TTBR0_EL1; // SWITCH_TO_KVM_PAGETABLE sets the kvm pagetable. #define SWITCH_TO_KVM_PAGETABLE(from) \ - MOVD CPU_TTBR0_KVM(from), RSV_REG; \ - WORD $0xd5182012; \ // MSR R18, TTBR0_EL1 + MRS TTBR1_EL1, R0; \ + MOVD $1, R1; \ + BFI $48, R1, $16, R0; \ + MSR R0, TTBR1_EL1; \ ISB $15; \ - DSB $15; + MOVD CPU_TTBR0_KVM(from), RSV_REG; \ + MSR RSV_REG, TTBR0_EL1; #define VFP_ENABLE \ MOVD $FPEN_ENABLE, R0; \ @@ -328,23 +334,20 @@ #define KERNEL_ENTRY_FROM_EL0 \ SUB $16, RSP, RSP; \ // step1, save r18, r9 into kernel temporary stack. STP (RSV_REG, RSV_REG_APP), 16*0(RSP); \ - WORD $0xd538d092; \ //MRS TPIDR_EL1, R18, step2, switch user pagetable. - SWITCH_TO_KVM_PAGETABLE(RSV_REG); \ - WORD $0xd538d092; \ //MRS TPIDR_EL1, R18 - MOVD CPU_APP_ADDR(RSV_REG), RSV_REG_APP; \ // step3, load app context pointer. - REGISTERS_SAVE(RSV_REG_APP, 0); \ // step4, save app context. + WORD $0xd538d092; \ // MRS TPIDR_EL1, R18 + MOVD CPU_APP_ADDR(RSV_REG), RSV_REG_APP; \ // step2, load app context pointer. + REGISTERS_SAVE(RSV_REG_APP, 0); \ // step3, save app context. MOVD RSV_REG_APP, R20; \ LDP 16*0(RSP), (RSV_REG, RSV_REG_APP); \ ADD $16, RSP, RSP; \ MOVD RSV_REG, PTRACE_R18(R20); \ MOVD RSV_REG_APP, PTRACE_R9(R20); \ - MOVD R20, RSV_REG_APP; \ WORD $0xd5384003; \ // MRS SPSR_EL1, R3 - MOVD R3, PTRACE_PSTATE(RSV_REG_APP); \ + MOVD R3, PTRACE_PSTATE(R20); \ MRS ELR_EL1, R3; \ - MOVD R3, PTRACE_PC(RSV_REG_APP); \ + MOVD R3, PTRACE_PC(R20); \ WORD $0xd5384103; \ // MRS SP_EL0, R3 - MOVD R3, PTRACE_SP(RSV_REG_APP); + MOVD R3, PTRACE_SP(R20); // KERNEL_ENTRY_FROM_EL1 is the entry code of the vcpu from el1 to el1. #define KERNEL_ENTRY_FROM_EL1 \ @@ -359,6 +362,13 @@ MOVD R4, CPU_REGISTERS+PTRACE_SP(RSV_REG); \ LOAD_KERNEL_STACK(RSV_REG); // Load the temporary stack. +// storeAppASID writes the application's asid value. +TEXT ·storeAppASID(SB),NOSPLIT,$0-8 + MOVD asid+0(FP), R1 + MRS TPIDR_EL1, RSV_REG + MOVD R1, CPU_APP_ASID(RSV_REG) + RET + // Halt halts execution. TEXT ·Halt(SB),NOSPLIT,$0 // Clear bluepill. @@ -416,7 +426,7 @@ TEXT ·Current(SB),NOSPLIT,$0-8 MOVD R8, ret+0(FP) RET -#define STACK_FRAME_SIZE 16 +#define STACK_FRAME_SIZE 32 // kernelExitToEl0 is the entrypoint for application in guest_el0. // Prepare the vcpu environment for container application. @@ -451,6 +461,14 @@ TEXT ·kernelExitToEl0(SB),NOSPLIT,$0 MOVD PTRACE_PSTATE(RSV_REG_APP), R1 WORD $0xd5184001 //MSR R1, SPSR_EL1 + // need use kernel space address to excute below code, since + // after SWITCH_TO_APP_PAGETABLE the ASID is changed to app's + // ASID. + WORD $0x10000061 // ADR R1, do_exit_to_el0 + ORR $0xffff000000000000, R1, R1 + JMP (R1) + +do_exit_to_el0: // RSV_REG & RSV_REG_APP will be loaded at the end. REGISTERS_LOAD(RSV_REG_APP, 0) @@ -460,15 +478,16 @@ TEXT ·kernelExitToEl0(SB),NOSPLIT,$0 SUB $STACK_FRAME_SIZE, RSP, RSP STP (RSV_REG, RSV_REG_APP), 16*0(RSP) + STP (R0, R1), 16*1(RSP) WORD $0xd538d092 //MRS TPIDR_EL1, R18 SWITCH_TO_APP_PAGETABLE(RSV_REG) + LDP 16*1(RSP), (R0, R1) LDP 16*0(RSP), (RSV_REG, RSV_REG_APP) ADD $STACK_FRAME_SIZE, RSP, RSP - ISB $15 ERET() // kernelExitToEl1 is the entrypoint for sentry in guest_el1. @@ -484,6 +503,9 @@ TEXT ·kernelExitToEl1(SB),NOSPLIT,$0 MOVD CPU_REGISTERS+PTRACE_SP(RSV_REG), R1 MOVD R1, RSP + SWITCH_TO_KVM_PAGETABLE(RSV_REG) + MRS TPIDR_EL1, RSV_REG + REGISTERS_LOAD(RSV_REG, CPU_REGISTERS) MOVD CPU_REGISTERS+PTRACE_R9(RSV_REG), RSV_REG_APP diff --git a/pkg/sentry/platform/ring0/gen_offsets/BUILD b/pkg/sentry/platform/ring0/gen_offsets/BUILD index 549f3d228..9742308d8 100644 --- a/pkg/sentry/platform/ring0/gen_offsets/BUILD +++ b/pkg/sentry/platform/ring0/gen_offsets/BUILD @@ -24,7 +24,10 @@ go_binary( "defs_impl_arm64.go", "main.go", ], - visibility = ["//pkg/sentry/platform/ring0:__pkg__"], + visibility = [ + "//pkg/sentry/platform/kvm:__pkg__", + "//pkg/sentry/platform/ring0:__pkg__", + ], deps = [ "//pkg/cpuid", "//pkg/sentry/arch", diff --git a/pkg/sentry/platform/ring0/kernel.go b/pkg/sentry/platform/ring0/kernel.go index 021693791..264be23d3 100644 --- a/pkg/sentry/platform/ring0/kernel.go +++ b/pkg/sentry/platform/ring0/kernel.go @@ -19,8 +19,8 @@ package ring0 // N.B. that constraints on KernelOpts must be satisfied. // //go:nosplit -func (k *Kernel) Init(opts KernelOpts) { - k.init(opts) +func (k *Kernel) Init(opts KernelOpts, maxCPUs int) { + k.init(opts, maxCPUs) } // Halt halts execution. @@ -49,6 +49,11 @@ func (defaultHooks) KernelException(Vector) { // kernelSyscall is a trampoline. // +// When in amd64, it is called with %rip on the upper half, so it can +// NOT access to any global data which is not mapped on upper and must +// call to function pointers or interfaces to switch to the lower half +// so that callee can access to global data. +// // +checkescape:hard,stack // //go:nosplit @@ -58,6 +63,11 @@ func kernelSyscall(c *CPU) { // kernelException is a trampoline. // +// When in amd64, it is called with %rip on the upper half, so it can +// NOT access to any global data which is not mapped on upper and must +// call to function pointers or interfaces to switch to the lower half +// so that callee can access to global data. +// // +checkescape:hard,stack // //go:nosplit @@ -68,10 +78,10 @@ func kernelException(c *CPU, vector Vector) { // Init initializes a new CPU. // // Init allows embedding in other objects. -func (c *CPU) Init(k *Kernel, hooks Hooks) { - c.self = c // Set self reference. - c.kernel = k // Set kernel reference. - c.init() // Perform architectural init. +func (c *CPU) Init(k *Kernel, cpuID int, hooks Hooks) { + c.self = c // Set self reference. + c.kernel = k // Set kernel reference. + c.init(cpuID) // Perform architectural init. // Require hooks. if hooks != nil { diff --git a/pkg/sentry/platform/ring0/kernel_amd64.go b/pkg/sentry/platform/ring0/kernel_amd64.go index d37981dbf..3a9dff4cc 100644 --- a/pkg/sentry/platform/ring0/kernel_amd64.go +++ b/pkg/sentry/platform/ring0/kernel_amd64.go @@ -18,13 +18,42 @@ package ring0 import ( "encoding/binary" + "reflect" + + "gvisor.dev/gvisor/pkg/usermem" ) // init initializes architecture-specific state. -func (k *Kernel) init(opts KernelOpts) { +func (k *Kernel) init(opts KernelOpts, maxCPUs int) { // Save the root page tables. k.PageTables = opts.PageTables + entrySize := reflect.TypeOf(kernelEntry{}).Size() + var ( + entries []kernelEntry + padding = 1 + ) + for { + entries = make([]kernelEntry, maxCPUs+padding-1) + totalSize := entrySize * uintptr(maxCPUs+padding-1) + addr := reflect.ValueOf(&entries[0]).Pointer() + if addr&(usermem.PageSize-1) == 0 && totalSize >= usermem.PageSize { + // The runtime forces power-of-2 alignment for allocations, and we are therefore + // safe once the first address is aligned and the chunk is at least a full page. + break + } + padding = padding << 1 + } + k.cpuEntries = entries + + k.globalIDT = &idt64{} + if reflect.TypeOf(idt64{}).Size() != usermem.PageSize { + panic("Size of globalIDT should be PageSize") + } + if reflect.ValueOf(k.globalIDT).Pointer()&(usermem.PageSize-1) != 0 { + panic("Allocated globalIDT should be page aligned") + } + // Setup the IDT, which is uniform. for v, handler := range handlers { // Allow Breakpoint and Overflow to be called from all @@ -39,8 +68,26 @@ func (k *Kernel) init(opts KernelOpts) { } } +func (k *Kernel) EntryRegions() map[uintptr]uintptr { + regions := make(map[uintptr]uintptr) + + addr := reflect.ValueOf(&k.cpuEntries[0]).Pointer() + size := reflect.TypeOf(kernelEntry{}).Size() * uintptr(len(k.cpuEntries)) + end, _ := usermem.Addr(addr + size).RoundUp() + regions[uintptr(usermem.Addr(addr).RoundDown())] = uintptr(end) + + addr = reflect.ValueOf(k.globalIDT).Pointer() + size = reflect.TypeOf(idt64{}).Size() + end, _ = usermem.Addr(addr + size).RoundUp() + regions[uintptr(usermem.Addr(addr).RoundDown())] = uintptr(end) + + return regions +} + // init initializes architecture-specific state. -func (c *CPU) init() { +func (c *CPU) init(cpuID int) { + c.kernelEntry = &c.kernel.cpuEntries[cpuID] + c.cpuSelf = c // Null segment. c.gdt[0].setNull() @@ -65,6 +112,7 @@ func (c *CPU) init() { // Set the kernel stack pointer in the TSS (virtual address). stackAddr := c.StackTop() + c.stackTop = stackAddr c.tss.rsp0Lo = uint32(stackAddr) c.tss.rsp0Hi = uint32(stackAddr >> 32) c.tss.ist1Lo = uint32(stackAddr) @@ -183,7 +231,7 @@ func IsCanonical(addr uint64) bool { //go:nosplit func (c *CPU) SwitchToUser(switchOpts SwitchOpts) (vector Vector) { userCR3 := switchOpts.PageTables.CR3(!switchOpts.Flush, switchOpts.UserPCID) - kernelCR3 := c.kernel.PageTables.CR3(true, switchOpts.KernelPCID) + c.kernelCR3 = uintptr(c.kernel.PageTables.CR3(true, switchOpts.KernelPCID)) // Sanitize registers. regs := switchOpts.Registers @@ -197,15 +245,11 @@ func (c *CPU) SwitchToUser(switchOpts SwitchOpts) (vector Vector) { WriteFS(uintptr(regs.Fs_base)) // escapes: no. Set application FS. WriteGS(uintptr(regs.Gs_base)) // escapes: no. Set application GS. LoadFloatingPoint(switchOpts.FloatingPointState) // escapes: no. Copy in floating point. - jumpToKernel() // Switch to upper half. - writeCR3(uintptr(userCR3)) // Change to user address space. if switchOpts.FullRestore { - vector = iret(c, regs) + vector = iret(c, regs, uintptr(userCR3)) } else { - vector = sysret(c, regs) + vector = sysret(c, regs, uintptr(userCR3)) } - writeCR3(uintptr(kernelCR3)) // Return to kernel address space. - jumpToUser() // Return to lower half. SaveFloatingPoint(switchOpts.FloatingPointState) // escapes: no. Copy out floating point. WriteFS(uintptr(c.registers.Fs_base)) // escapes: no. Restore kernel FS. return @@ -219,7 +263,7 @@ func (c *CPU) SwitchToUser(switchOpts SwitchOpts) (vector Vector) { //go:nosplit func start(c *CPU) { // Save per-cpu & FS segment. - WriteGS(kernelAddr(c)) + WriteGS(kernelAddr(c.kernelEntry)) WriteFS(uintptr(c.registers.Fs_base)) // Initialize floating point. diff --git a/pkg/sentry/platform/ring0/kernel_arm64.go b/pkg/sentry/platform/ring0/kernel_arm64.go index d0afa1aaa..0ca98a7c7 100644 --- a/pkg/sentry/platform/ring0/kernel_arm64.go +++ b/pkg/sentry/platform/ring0/kernel_arm64.go @@ -25,13 +25,13 @@ func HaltAndResume() func HaltEl1SvcAndResume() // init initializes architecture-specific state. -func (k *Kernel) init(opts KernelOpts) { +func (k *Kernel) init(opts KernelOpts, maxCPUs int) { // Save the root page tables. k.PageTables = opts.PageTables } // init initializes architecture-specific state. -func (c *CPU) init() { +func (c *CPU) init(cpuID int) { // Set the kernel stack pointer(virtual address). c.registers.Sp = uint64(c.StackTop()) @@ -53,6 +53,11 @@ func IsCanonical(addr uint64) bool { //go:nosplit func (c *CPU) SwitchToUser(switchOpts SwitchOpts) (vector Vector) { + storeAppASID(uintptr(switchOpts.UserASID)) + if switchOpts.Flush { + FlushTlbAll() + } + regs := switchOpts.Registers regs.Pstate &= ^uint64(PsrFlagsClear) diff --git a/pkg/sentry/platform/ring0/lib_amd64.go b/pkg/sentry/platform/ring0/lib_amd64.go index ca968a036..0ec5c3bc5 100644 --- a/pkg/sentry/platform/ring0/lib_amd64.go +++ b/pkg/sentry/platform/ring0/lib_amd64.go @@ -61,21 +61,9 @@ func wrgsbase(addr uintptr) // wrgsmsr writes to the GS_BASE MSR. func wrgsmsr(addr uintptr) -// writeCR3 writes the CR3 value. -func writeCR3(phys uintptr) - -// readCR3 reads the current CR3 value. -func readCR3() uintptr - // readCR2 reads the current CR2 value. func readCR2() uintptr -// jumpToKernel jumps to the kernel version of the current RIP. -func jumpToKernel() - -// jumpToUser jumps to the user version of the current RIP. -func jumpToUser() - // fninit initializes the floating point unit. func fninit() diff --git a/pkg/sentry/platform/ring0/lib_amd64.s b/pkg/sentry/platform/ring0/lib_amd64.s index 75d742750..2fe83568a 100644 --- a/pkg/sentry/platform/ring0/lib_amd64.s +++ b/pkg/sentry/platform/ring0/lib_amd64.s @@ -127,53 +127,6 @@ TEXT ·wrgsmsr(SB),NOSPLIT,$0-8 BYTE $0x0f; BYTE $0x30; // WRMSR RET -// jumpToUser changes execution to the user address. -// -// This works by changing the return value to the user version. -TEXT ·jumpToUser(SB),NOSPLIT,$0 - MOVQ 0(SP), AX - MOVQ ·KernelStartAddress(SB), BX - NOTQ BX - ANDQ BX, SP // Switch the stack. - ANDQ BX, BP // Switch the frame pointer. - ANDQ BX, AX // Future return value. - MOVQ AX, 0(SP) - RET - -// jumpToKernel changes execution to the kernel address space. -// -// This works by changing the return value to the kernel version. -TEXT ·jumpToKernel(SB),NOSPLIT,$0 - MOVQ 0(SP), AX - MOVQ ·KernelStartAddress(SB), BX - ORQ BX, SP // Switch the stack. - ORQ BX, BP // Switch the frame pointer. - ORQ BX, AX // Future return value. - MOVQ AX, 0(SP) - RET - -// writeCR3 writes the given CR3 value. -// -// The code corresponds to: -// -// mov %rax, %cr3 -// -TEXT ·writeCR3(SB),NOSPLIT,$0-8 - MOVQ cr3+0(FP), AX - BYTE $0x0f; BYTE $0x22; BYTE $0xd8; - RET - -// readCR3 reads the current CR3 value. -// -// The code corresponds to: -// -// mov %cr3, %rax -// -TEXT ·readCR3(SB),NOSPLIT,$0-8 - BYTE $0x0f; BYTE $0x20; BYTE $0xd8; - MOVQ AX, ret+0(FP) - RET - // readCR2 reads the current CR2 value. // // The code corresponds to: diff --git a/pkg/sentry/platform/ring0/lib_arm64.go b/pkg/sentry/platform/ring0/lib_arm64.go index 00e52c8af..2f1abcb0f 100644 --- a/pkg/sentry/platform/ring0/lib_arm64.go +++ b/pkg/sentry/platform/ring0/lib_arm64.go @@ -16,6 +16,15 @@ package ring0 +// storeAppASID writes the application's asid value. +func storeAppASID(asid uintptr) + +// LocalFlushTlbAll same as FlushTlbAll, but only applies to the calling CPU. +func LocalFlushTlbAll() + +// FlushTlbAll flush all tlb. +func FlushTlbAll() + // CPACREL1 returns the value of the CPACR_EL1 register. func CPACREL1() (value uintptr) diff --git a/pkg/sentry/platform/ring0/lib_arm64.s b/pkg/sentry/platform/ring0/lib_arm64.s index 86bfbe46f..8aabf7d0e 100644 --- a/pkg/sentry/platform/ring0/lib_arm64.s +++ b/pkg/sentry/platform/ring0/lib_arm64.s @@ -15,6 +15,20 @@ #include "funcdata.h" #include "textflag.h" +TEXT ·LocalFlushTlbAll(SB),NOSPLIT,$0 + DSB $6 // dsb(nshst) + WORD $0xd508871f // __tlbi(vmalle1) + DSB $7 // dsb(nsh) + ISB $15 + RET + +TEXT ·FlushTlbAll(SB),NOSPLIT,$0 + DSB $10 // dsb(ishst) + WORD $0xd508831f // __tlbi(vmalle1is) + DSB $11 // dsb(ish) + ISB $15 + RET + TEXT ·GetTLS(SB),NOSPLIT,$0-8 MRS TPIDR_EL0, R1 MOVD R1, ret+0(FP) diff --git a/pkg/sentry/platform/ring0/offsets_amd64.go b/pkg/sentry/platform/ring0/offsets_amd64.go index b8ab120a0..290d94bd6 100644 --- a/pkg/sentry/platform/ring0/offsets_amd64.go +++ b/pkg/sentry/platform/ring0/offsets_amd64.go @@ -30,14 +30,22 @@ func Emit(w io.Writer) { c := &CPU{} fmt.Fprintf(w, "\n// CPU offsets.\n") - fmt.Fprintf(w, "#define CPU_SELF 0x%02x\n", reflect.ValueOf(&c.self).Pointer()-reflect.ValueOf(c).Pointer()) fmt.Fprintf(w, "#define CPU_REGISTERS 0x%02x\n", reflect.ValueOf(&c.registers).Pointer()-reflect.ValueOf(c).Pointer()) - fmt.Fprintf(w, "#define CPU_STACK_TOP 0x%02x\n", reflect.ValueOf(&c.stack[0]).Pointer()-reflect.ValueOf(c).Pointer()+uintptr(len(c.stack))) fmt.Fprintf(w, "#define CPU_ERROR_CODE 0x%02x\n", reflect.ValueOf(&c.errorCode).Pointer()-reflect.ValueOf(c).Pointer()) fmt.Fprintf(w, "#define CPU_ERROR_TYPE 0x%02x\n", reflect.ValueOf(&c.errorType).Pointer()-reflect.ValueOf(c).Pointer()) + fmt.Fprintf(w, "#define CPU_ENTRY 0x%02x\n", reflect.ValueOf(&c.kernelEntry).Pointer()-reflect.ValueOf(c).Pointer()) + + e := &kernelEntry{} + fmt.Fprintf(w, "\n// CPU entry offsets.\n") + fmt.Fprintf(w, "#define ENTRY_SCRATCH0 0x%02x\n", reflect.ValueOf(&e.scratch0).Pointer()-reflect.ValueOf(e).Pointer()) + fmt.Fprintf(w, "#define ENTRY_SCRATCH1 0x%02x\n", reflect.ValueOf(&e.scratch1).Pointer()-reflect.ValueOf(e).Pointer()) + fmt.Fprintf(w, "#define ENTRY_STACK_TOP 0x%02x\n", reflect.ValueOf(&e.stackTop).Pointer()-reflect.ValueOf(e).Pointer()) + fmt.Fprintf(w, "#define ENTRY_CPU_SELF 0x%02x\n", reflect.ValueOf(&e.cpuSelf).Pointer()-reflect.ValueOf(e).Pointer()) + fmt.Fprintf(w, "#define ENTRY_KERNEL_CR3 0x%02x\n", reflect.ValueOf(&e.kernelCR3).Pointer()-reflect.ValueOf(e).Pointer()) fmt.Fprintf(w, "\n// Bits.\n") fmt.Fprintf(w, "#define _RFLAGS_IF 0x%02x\n", _RFLAGS_IF) + fmt.Fprintf(w, "#define _RFLAGS_IOPL0 0x%02x\n", _RFLAGS_IOPL0) fmt.Fprintf(w, "#define _KERNEL_FLAGS 0x%02x\n", KernelFlagsSet) fmt.Fprintf(w, "\n// Vectors.\n") diff --git a/pkg/sentry/platform/ring0/offsets_arm64.go b/pkg/sentry/platform/ring0/offsets_arm64.go index f3de962f0..1d86b4bcf 100644 --- a/pkg/sentry/platform/ring0/offsets_arm64.go +++ b/pkg/sentry/platform/ring0/offsets_arm64.go @@ -41,6 +41,7 @@ func Emit(w io.Writer) { fmt.Fprintf(w, "#define CPU_VECTOR_CODE 0x%02x\n", reflect.ValueOf(&c.vecCode).Pointer()-reflect.ValueOf(c).Pointer()) fmt.Fprintf(w, "#define CPU_APP_ADDR 0x%02x\n", reflect.ValueOf(&c.appAddr).Pointer()-reflect.ValueOf(c).Pointer()) fmt.Fprintf(w, "#define CPU_LAZY_VFP 0x%02x\n", reflect.ValueOf(&c.lazyVFP).Pointer()-reflect.ValueOf(c).Pointer()) + fmt.Fprintf(w, "#define CPU_APP_ASID 0x%02x\n", reflect.ValueOf(&c.appASID).Pointer()-reflect.ValueOf(c).Pointer()) fmt.Fprintf(w, "\n// Bits.\n") fmt.Fprintf(w, "#define _KERNEL_FLAGS 0x%02x\n", KernelFlagsSet) diff --git a/pkg/sentry/platform/ring0/x86.go b/pkg/sentry/platform/ring0/x86.go index 9da0ea685..34fbc1c35 100644 --- a/pkg/sentry/platform/ring0/x86.go +++ b/pkg/sentry/platform/ring0/x86.go @@ -39,7 +39,9 @@ const ( _RFLAGS_AC = 1 << 18 _RFLAGS_NT = 1 << 14 - _RFLAGS_IOPL = 3 << 12 + _RFLAGS_IOPL0 = 1 << 12 + _RFLAGS_IOPL1 = 1 << 13 + _RFLAGS_IOPL = _RFLAGS_IOPL0 | _RFLAGS_IOPL1 _RFLAGS_DF = 1 << 10 _RFLAGS_IF = 1 << 9 _RFLAGS_STEP = 1 << 8 @@ -67,15 +69,45 @@ const ( KernelFlagsSet = _RFLAGS_RESERVED // UserFlagsSet are always set in userspace. - UserFlagsSet = _RFLAGS_RESERVED | _RFLAGS_IF + // + // _RFLAGS_IOPL is a set of two bits and it shows the I/O privilege + // level. The Current Privilege Level (CPL) of the task must be less + // than or equal to the IOPL in order for the task or program to access + // I/O ports. + // + // Here, _RFLAGS_IOPL0 is used only to determine whether the task is + // running in the kernel or userspace mode. In the user mode, the CPL is + // always 3 and it doesn't matter what IOPL is set if it is bellow CPL. + // + // We need to have one bit which will be always different in user and + // kernel modes. And we have to remember that even though we have + // KernelFlagsClear, we still can see some of these flags in the kernel + // mode. This can happen when the goruntime switches on a goroutine + // which has been saved in the host mode. On restore, the popf + // instruction is used to restore flags and this means that all flags + // what the goroutine has in the host mode will be restored in the + // kernel mode. + // + // _RFLAGS_IOPL0 is never set in host and kernel modes and we always set + // it in the user mode. So if this flag is set, the task is running in + // the user mode and if it isn't set, the task is running in the kernel + // mode. + UserFlagsSet = _RFLAGS_RESERVED | _RFLAGS_IF | _RFLAGS_IOPL0 // KernelFlagsClear should always be clear in the kernel. KernelFlagsClear = _RFLAGS_STEP | _RFLAGS_IF | _RFLAGS_IOPL | _RFLAGS_AC | _RFLAGS_NT // UserFlagsClear are always cleared in userspace. - UserFlagsClear = _RFLAGS_NT | _RFLAGS_IOPL + UserFlagsClear = _RFLAGS_NT | _RFLAGS_IOPL1 ) +// IsKernelFlags returns true if rflags coresponds to the kernel mode. +// +// go:nosplit +func IsKernelFlags(rflags uint64) bool { + return rflags&_RFLAGS_IOPL0 == 0 +} + // Vector is an exception vector. type Vector uintptr @@ -104,7 +136,7 @@ const ( VirtualizationException SecurityException = 0x1e SyscallInt80 = 0x80 - _NR_INTERRUPTS = SyscallInt80 + 1 + _NR_INTERRUPTS = 0x100 ) // System call vectors. diff --git a/pkg/sentry/socket/hostinet/BUILD b/pkg/sentry/socket/hostinet/BUILD index 632e33452..b6ebe29d6 100644 --- a/pkg/sentry/socket/hostinet/BUILD +++ b/pkg/sentry/socket/hostinet/BUILD @@ -39,6 +39,9 @@ go_library( "//pkg/sentry/vfs", "//pkg/syserr", "//pkg/syserror", + "//pkg/tcpip", + "//pkg/tcpip/network/ipv4", + "//pkg/tcpip/network/ipv6", "//pkg/tcpip/stack", "//pkg/usermem", "//pkg/waiter", diff --git a/pkg/sentry/socket/hostinet/socket_vfs2.go b/pkg/sentry/socket/hostinet/socket_vfs2.go index 8a1d52ebf..163af329b 100644 --- a/pkg/sentry/socket/hostinet/socket_vfs2.go +++ b/pkg/sentry/socket/hostinet/socket_vfs2.go @@ -52,6 +52,7 @@ var _ = socket.SocketVFS2(&socketVFS2{}) func newVFS2Socket(t *kernel.Task, family int, stype linux.SockType, protocol int, fd int, flags uint32) (*vfs.FileDescription, *syserr.Error) { mnt := t.Kernel().SocketMount() d := sockfs.NewDentry(t.Credentials(), mnt) + defer d.DecRef(t) s := &socketVFS2{ socketOpsCommon: socketOpsCommon{ @@ -77,6 +78,13 @@ func newVFS2Socket(t *kernel.Task, family int, stype linux.SockType, protocol in return vfsfd, nil } +// Release implements vfs.FileDescriptionImpl.Release. +func (s *socketVFS2) Release(ctx context.Context) { + t := kernel.TaskFromContext(ctx) + t.Kernel().DeleteSocketVFS2(&s.vfsfd) + s.socketOpsCommon.Release(ctx) +} + // Readiness implements waiter.Waitable.Readiness. func (s *socketVFS2) Readiness(mask waiter.EventMask) waiter.EventMask { return s.socketOpsCommon.Readiness(mask) @@ -97,11 +105,6 @@ func (s *socketVFS2) Ioctl(ctx context.Context, uio usermem.IO, args arch.Syscal return ioctl(ctx, s.fd, uio, args) } -// Allocate implements vfs.FileDescriptionImpl.Allocate. -func (s *socketVFS2) Allocate(ctx context.Context, mode, offset, length uint64) error { - return syserror.ENODEV -} - // PRead implements vfs.FileDescriptionImpl.PRead. func (s *socketVFS2) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { return 0, syserror.ESPIPE diff --git a/pkg/sentry/socket/hostinet/stack.go b/pkg/sentry/socket/hostinet/stack.go index fda3dcb35..faa61160e 100644 --- a/pkg/sentry/socket/hostinet/stack.go +++ b/pkg/sentry/socket/hostinet/stack.go @@ -30,6 +30,9 @@ import ( "gvisor.dev/gvisor/pkg/sentry/inet" "gvisor.dev/gvisor/pkg/syserr" "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/tcpip" + "gvisor.dev/gvisor/pkg/tcpip/network/ipv4" + "gvisor.dev/gvisor/pkg/tcpip/network/ipv6" "gvisor.dev/gvisor/pkg/tcpip/stack" "gvisor.dev/gvisor/pkg/usermem" ) @@ -59,6 +62,8 @@ type Stack struct { tcpSACKEnabled bool netDevFile *os.File netSNMPFile *os.File + ipv4Forwarding bool + ipv6Forwarding bool } // NewStack returns an empty Stack containing no configuration. @@ -118,6 +123,13 @@ func (s *Stack) Configure() error { s.netSNMPFile = f } + s.ipv6Forwarding = false + if ipForwarding, err := ioutil.ReadFile("/proc/sys/net/ipv6/conf/all/forwarding"); err == nil { + s.ipv6Forwarding = strings.TrimSpace(string(ipForwarding)) != "0" + } else { + log.Warningf("Failed to read if ipv6 forwarding is enabled, setting to false") + } + return nil } @@ -468,3 +480,21 @@ func (s *Stack) CleanupEndpoints() []stack.TransportEndpoint { return nil } // RestoreCleanupEndpoints implements inet.Stack.RestoreCleanupEndpoints. func (s *Stack) RestoreCleanupEndpoints([]stack.TransportEndpoint) {} + +// Forwarding implements inet.Stack.Forwarding. +func (s *Stack) Forwarding(protocol tcpip.NetworkProtocolNumber) bool { + switch protocol { + case ipv4.ProtocolNumber: + return s.ipv4Forwarding + case ipv6.ProtocolNumber: + return s.ipv6Forwarding + default: + log.Warningf("Forwarding(%v) failed: unsupported protocol", protocol) + return false + } +} + +// SetForwarding implements inet.Stack.SetForwarding. +func (s *Stack) SetForwarding(protocol tcpip.NetworkProtocolNumber, enable bool) error { + return syserror.EACCES +} diff --git a/pkg/sentry/socket/netfilter/extensions.go b/pkg/sentry/socket/netfilter/extensions.go index 0336a32d8..549787955 100644 --- a/pkg/sentry/socket/netfilter/extensions.go +++ b/pkg/sentry/socket/netfilter/extensions.go @@ -19,6 +19,8 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/binary" + "gvisor.dev/gvisor/pkg/syserr" + "gvisor.dev/gvisor/pkg/tcpip" "gvisor.dev/gvisor/pkg/tcpip/stack" "gvisor.dev/gvisor/pkg/usermem" ) @@ -37,7 +39,7 @@ type matchMaker interface { // name is the matcher name as stored in the xt_entry_match struct. name() string - // marshal converts from an stack.Matcher to an ABI struct. + // marshal converts from a stack.Matcher to an ABI struct. marshal(matcher stack.Matcher) []byte // unmarshal converts from the ABI matcher struct to an @@ -93,3 +95,71 @@ func unmarshalMatcher(match linux.XTEntryMatch, filter stack.IPHeaderFilter, buf } return matchMaker.unmarshal(buf, filter) } + +// targetMaker knows how to (un)marshal a target. Once registered, +// marshalTarget and unmarshalTarget can be used. +type targetMaker interface { + // id uniquely identifies the target. + id() stack.TargetID + + // marshal converts from a stack.Target to an ABI struct. + marshal(target stack.Target) []byte + + // unmarshal converts from the ABI matcher struct to a stack.Target. + unmarshal(buf []byte, filter stack.IPHeaderFilter) (stack.Target, *syserr.Error) +} + +// targetMakers maps the TargetID of supported targets to the targetMaker that +// marshals and unmarshals it. It is immutable after package initialization. +var targetMakers = map[stack.TargetID]targetMaker{} + +func targetRevision(name string, netProto tcpip.NetworkProtocolNumber, rev uint8) (uint8, bool) { + tid := stack.TargetID{ + Name: name, + NetworkProtocol: netProto, + Revision: rev, + } + if _, ok := targetMakers[tid]; !ok { + return 0, false + } + + // Return the highest supported revision unless rev is higher. + for _, other := range targetMakers { + otherID := other.id() + if name == otherID.Name && netProto == otherID.NetworkProtocol && otherID.Revision > rev { + rev = uint8(otherID.Revision) + } + } + return rev, true +} + +// registerTargetMaker should be called by target extensions to register them +// with the netfilter package. +func registerTargetMaker(tm targetMaker) { + if _, ok := targetMakers[tm.id()]; ok { + panic(fmt.Sprintf("multiple targets registered with name %q.", tm.id())) + } + targetMakers[tm.id()] = tm +} + +func marshalTarget(target stack.Target) []byte { + targetMaker, ok := targetMakers[target.ID()] + if !ok { + panic(fmt.Sprintf("unknown target of type %T with id %+v.", target, target.ID())) + } + return targetMaker.marshal(target) +} + +func unmarshalTarget(target linux.XTEntryTarget, filter stack.IPHeaderFilter, buf []byte) (stack.Target, *syserr.Error) { + tid := stack.TargetID{ + Name: target.Name.String(), + NetworkProtocol: filter.NetworkProtocol(), + Revision: target.Revision, + } + targetMaker, ok := targetMakers[tid] + if !ok { + nflog("unsupported target with name %q", target.Name.String()) + return nil, syserr.ErrInvalidArgument + } + return targetMaker.unmarshal(buf, filter) +} diff --git a/pkg/sentry/socket/netfilter/ipv4.go b/pkg/sentry/socket/netfilter/ipv4.go index e4c55a100..b560fae0d 100644 --- a/pkg/sentry/socket/netfilter/ipv4.go +++ b/pkg/sentry/socket/netfilter/ipv4.go @@ -181,18 +181,23 @@ func modifyEntries4(stk *stack.Stack, optVal []byte, replace *linux.IPTReplace, nflog("entry doesn't have enough room for its target (only %d bytes remain)", len(optVal)) return nil, syserr.ErrInvalidArgument } - target, err := parseTarget(filter, optVal[:targetSize]) - if err != nil { - nflog("failed to parse target: %v", err) - return nil, syserr.ErrInvalidArgument - } - optVal = optVal[targetSize:] - table.Rules = append(table.Rules, stack.Rule{ + rule := stack.Rule{ Filter: filter, - Target: target, Matchers: matchers, - }) + } + + { + target, err := parseTarget(filter, optVal[:targetSize], false /* ipv6 */) + if err != nil { + nflog("failed to parse target: %v", err) + return nil, err + } + rule.Target = target + } + optVal = optVal[targetSize:] + + table.Rules = append(table.Rules, rule) offsets[offset] = int(entryIdx) offset += uint32(entry.NextOffset) diff --git a/pkg/sentry/socket/netfilter/ipv6.go b/pkg/sentry/socket/netfilter/ipv6.go index 3b2c1becd..4253f7bf4 100644 --- a/pkg/sentry/socket/netfilter/ipv6.go +++ b/pkg/sentry/socket/netfilter/ipv6.go @@ -184,18 +184,23 @@ func modifyEntries6(stk *stack.Stack, optVal []byte, replace *linux.IPTReplace, nflog("entry doesn't have enough room for its target (only %d bytes remain)", len(optVal)) return nil, syserr.ErrInvalidArgument } - target, err := parseTarget(filter, optVal[:targetSize]) - if err != nil { - nflog("failed to parse target: %v", err) - return nil, syserr.ErrInvalidArgument - } - optVal = optVal[targetSize:] - table.Rules = append(table.Rules, stack.Rule{ + rule := stack.Rule{ Filter: filter, - Target: target, Matchers: matchers, - }) + } + + { + target, err := parseTarget(filter, optVal[:targetSize], true /* ipv6 */) + if err != nil { + nflog("failed to parse target: %v", err) + return nil, err + } + rule.Target = target + } + optVal = optVal[targetSize:] + + table.Rules = append(table.Rules, rule) offsets[offset] = int(entryIdx) offset += uint32(entry.NextOffset) diff --git a/pkg/sentry/socket/netfilter/netfilter.go b/pkg/sentry/socket/netfilter/netfilter.go index 3e1735079..904a12e38 100644 --- a/pkg/sentry/socket/netfilter/netfilter.go +++ b/pkg/sentry/socket/netfilter/netfilter.go @@ -25,6 +25,7 @@ import ( "gvisor.dev/gvisor/pkg/log" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/syserr" + "gvisor.dev/gvisor/pkg/tcpip" "gvisor.dev/gvisor/pkg/tcpip/stack" "gvisor.dev/gvisor/pkg/usermem" ) @@ -195,7 +196,7 @@ func SetEntries(stk *stack.Stack, optVal []byte, ipv6 bool) *syserr.Error { // Check the user chains. for ruleIdx, rule := range table.Rules { - if _, ok := rule.Target.(stack.UserChainTarget); !ok { + if _, ok := rule.Target.(*stack.UserChainTarget); !ok { continue } @@ -216,7 +217,7 @@ func SetEntries(stk *stack.Stack, optVal []byte, ipv6 bool) *syserr.Error { // Set each jump to point to the appropriate rule. Right now they hold byte // offsets. for ruleIdx, rule := range table.Rules { - jump, ok := rule.Target.(JumpTarget) + jump, ok := rule.Target.(*JumpTarget) if !ok { continue } @@ -307,7 +308,7 @@ func validUnderflow(rule stack.Rule, ipv6 bool) bool { return false } switch rule.Target.(type) { - case stack.AcceptTarget, stack.DropTarget: + case *stack.AcceptTarget, *stack.DropTarget: return true default: return false @@ -318,7 +319,7 @@ func isUnconditionalAccept(rule stack.Rule, ipv6 bool) bool { if !validUnderflow(rule, ipv6) { return false } - _, ok := rule.Target.(stack.AcceptTarget) + _, ok := rule.Target.(*stack.AcceptTarget) return ok } @@ -337,3 +338,20 @@ func hookFromLinux(hook int) stack.Hook { } panic(fmt.Sprintf("Unknown hook %d does not correspond to a builtin chain", hook)) } + +// TargetRevision returns a linux.XTGetRevision for a given target. It sets +// Revision to the highest supported value, unless the provided revision number +// is larger. +func TargetRevision(t *kernel.Task, revPtr usermem.Addr, netProto tcpip.NetworkProtocolNumber) (linux.XTGetRevision, *syserr.Error) { + // Read in the target name and version. + var rev linux.XTGetRevision + if _, err := rev.CopyIn(t, revPtr); err != nil { + return linux.XTGetRevision{}, syserr.FromError(err) + } + maxSupported, ok := targetRevision(rev.Name.String(), netProto, rev.Revision) + if !ok { + return linux.XTGetRevision{}, syserr.ErrProtocolNotSupported + } + rev.Revision = maxSupported + return rev, nil +} diff --git a/pkg/sentry/socket/netfilter/targets.go b/pkg/sentry/socket/netfilter/targets.go index 87e41abd8..0e14447fe 100644 --- a/pkg/sentry/socket/netfilter/targets.go +++ b/pkg/sentry/socket/netfilter/targets.go @@ -15,255 +15,357 @@ package netfilter import ( - "errors" "fmt" "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/binary" + "gvisor.dev/gvisor/pkg/syserr" "gvisor.dev/gvisor/pkg/tcpip" "gvisor.dev/gvisor/pkg/tcpip/header" "gvisor.dev/gvisor/pkg/tcpip/stack" "gvisor.dev/gvisor/pkg/usermem" ) -// errorTargetName is used to mark targets as error targets. Error targets -// shouldn't be reached - an error has occurred if we fall through to one. -const errorTargetName = "ERROR" +func init() { + // Standard targets include ACCEPT, DROP, RETURN, and JUMP. + registerTargetMaker(&standardTargetMaker{ + NetworkProtocol: header.IPv4ProtocolNumber, + }) + registerTargetMaker(&standardTargetMaker{ + NetworkProtocol: header.IPv6ProtocolNumber, + }) + + // Both user chains and actual errors are represented in iptables by + // error targets. + registerTargetMaker(&errorTargetMaker{ + NetworkProtocol: header.IPv4ProtocolNumber, + }) + registerTargetMaker(&errorTargetMaker{ + NetworkProtocol: header.IPv6ProtocolNumber, + }) + + registerTargetMaker(&redirectTargetMaker{ + NetworkProtocol: header.IPv4ProtocolNumber, + }) + registerTargetMaker(&nfNATTargetMaker{ + NetworkProtocol: header.IPv6ProtocolNumber, + }) +} -// 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" +type standardTargetMaker struct { + NetworkProtocol tcpip.NetworkProtocolNumber +} -func marshalTarget(target stack.Target) []byte { +func (sm *standardTargetMaker) id() stack.TargetID { + // Standard targets have the empty string as a name and no revisions. + return stack.TargetID{ + NetworkProtocol: sm.NetworkProtocol, + } +} +func (*standardTargetMaker) marshal(target stack.Target) []byte { + // Translate verdicts the same way as the iptables tool. + var verdict int32 switch tg := target.(type) { - case stack.AcceptTarget: - return marshalStandardTarget(stack.RuleAccept) - case stack.DropTarget: - return marshalStandardTarget(stack.RuleDrop) - case stack.ErrorTarget: - return marshalErrorTarget(errorTargetName) - case stack.UserChainTarget: - return marshalErrorTarget(tg.Name) - case stack.ReturnTarget: - return marshalStandardTarget(stack.RuleReturn) - case stack.RedirectTarget: - return marshalRedirectTarget(tg) - case JumpTarget: - return marshalJumpTarget(tg) + case *stack.AcceptTarget: + verdict = -linux.NF_ACCEPT - 1 + case *stack.DropTarget: + verdict = -linux.NF_DROP - 1 + case *stack.ReturnTarget: + verdict = linux.NF_RETURN + case *JumpTarget: + verdict = int32(tg.Offset) default: panic(fmt.Errorf("unknown target of type %T", target)) } -} - -func marshalStandardTarget(verdict stack.RuleVerdict) []byte { - nflog("convert to binary: marshalling standard target") // The target's name will be the empty string. - target := linux.XTStandardTarget{ + xt := linux.XTStandardTarget{ Target: linux.XTEntryTarget{ TargetSize: linux.SizeOfXTStandardTarget, }, - Verdict: translateFromStandardVerdict(verdict), + Verdict: verdict, } ret := make([]byte, 0, linux.SizeOfXTStandardTarget) - return binary.Marshal(ret, usermem.ByteOrder, target) + return binary.Marshal(ret, usermem.ByteOrder, xt) +} + +func (*standardTargetMaker) unmarshal(buf []byte, filter stack.IPHeaderFilter) (stack.Target, *syserr.Error) { + if len(buf) != linux.SizeOfXTStandardTarget { + nflog("buf has wrong size for standard target %d", len(buf)) + return nil, syserr.ErrInvalidArgument + } + var standardTarget linux.XTStandardTarget + buf = buf[:linux.SizeOfXTStandardTarget] + binary.Unmarshal(buf, usermem.ByteOrder, &standardTarget) + + if standardTarget.Verdict < 0 { + // A Verdict < 0 indicates a non-jump verdict. + return translateToStandardTarget(standardTarget.Verdict, filter.NetworkProtocol()) + } + // A verdict >= 0 indicates a jump. + return &JumpTarget{ + Offset: uint32(standardTarget.Verdict), + NetworkProtocol: filter.NetworkProtocol(), + }, nil +} + +type errorTargetMaker struct { + NetworkProtocol tcpip.NetworkProtocolNumber +} + +func (em *errorTargetMaker) id() stack.TargetID { + // Error targets have no revision. + return stack.TargetID{ + Name: stack.ErrorTargetName, + NetworkProtocol: em.NetworkProtocol, + } } -func marshalErrorTarget(errorName string) []byte { +func (*errorTargetMaker) marshal(target stack.Target) []byte { + var errorName string + switch tg := target.(type) { + case *stack.ErrorTarget: + errorName = stack.ErrorTargetName + case *stack.UserChainTarget: + errorName = tg.Name + default: + panic(fmt.Sprintf("errorMakerTarget cannot marshal unknown type %T", target)) + } + // This is an error target named error - target := linux.XTErrorTarget{ + xt := linux.XTErrorTarget{ Target: linux.XTEntryTarget{ TargetSize: linux.SizeOfXTErrorTarget, }, } - copy(target.Name[:], errorName) - copy(target.Target.Name[:], errorTargetName) + copy(xt.Name[:], errorName) + copy(xt.Target.Name[:], stack.ErrorTargetName) ret := make([]byte, 0, linux.SizeOfXTErrorTarget) - return binary.Marshal(ret, usermem.ByteOrder, target) + return binary.Marshal(ret, usermem.ByteOrder, xt) } -func marshalRedirectTarget(rt stack.RedirectTarget) []byte { +func (*errorTargetMaker) unmarshal(buf []byte, filter stack.IPHeaderFilter) (stack.Target, *syserr.Error) { + if len(buf) != linux.SizeOfXTErrorTarget { + nflog("buf has insufficient size for error target %d", len(buf)) + return nil, syserr.ErrInvalidArgument + } + var errorTarget linux.XTErrorTarget + buf = buf[:linux.SizeOfXTErrorTarget] + binary.Unmarshal(buf, usermem.ByteOrder, &errorTarget) + + // Error targets are used in 2 cases: + // * An actual error case. These rules have an error + // named stack.ErrorTargetName. The last entry of the table + // is usually an error case to catch any packets that + // somehow fall through every rule. + // * To mark the start of a user defined chain. These + // rules have an error with the name of the chain. + switch name := errorTarget.Name.String(); name { + case stack.ErrorTargetName: + return &stack.ErrorTarget{NetworkProtocol: filter.NetworkProtocol()}, nil + default: + // User defined chain. + return &stack.UserChainTarget{ + Name: name, + NetworkProtocol: filter.NetworkProtocol(), + }, nil + } +} + +type redirectTargetMaker struct { + NetworkProtocol tcpip.NetworkProtocolNumber +} + +func (rm *redirectTargetMaker) id() stack.TargetID { + return stack.TargetID{ + Name: stack.RedirectTargetName, + NetworkProtocol: rm.NetworkProtocol, + } +} + +func (*redirectTargetMaker) marshal(target stack.Target) []byte { + rt := target.(*stack.RedirectTarget) // This is a redirect target named redirect - target := linux.XTRedirectTarget{ + xt := linux.XTRedirectTarget{ Target: linux.XTEntryTarget{ TargetSize: linux.SizeOfXTRedirectTarget, }, } - copy(target.Target.Name[:], redirectTargetName) + copy(xt.Target.Name[:], stack.RedirectTargetName) ret := make([]byte, 0, linux.SizeOfXTRedirectTarget) - target.NfRange.RangeSize = 1 - if rt.RangeProtoSpecified { - target.NfRange.RangeIPV4.Flags |= linux.NF_NAT_RANGE_PROTO_SPECIFIED + xt.NfRange.RangeSize = 1 + xt.NfRange.RangeIPV4.Flags |= linux.NF_NAT_RANGE_PROTO_SPECIFIED + xt.NfRange.RangeIPV4.MinPort = htons(rt.Port) + xt.NfRange.RangeIPV4.MaxPort = xt.NfRange.RangeIPV4.MinPort + return binary.Marshal(ret, usermem.ByteOrder, xt) +} + +func (*redirectTargetMaker) unmarshal(buf []byte, filter stack.IPHeaderFilter) (stack.Target, *syserr.Error) { + if len(buf) < linux.SizeOfXTRedirectTarget { + nflog("redirectTargetMaker: buf has insufficient size for redirect target %d", len(buf)) + return nil, syserr.ErrInvalidArgument + } + + if p := filter.Protocol; p != header.TCPProtocolNumber && p != header.UDPProtocolNumber { + nflog("redirectTargetMaker: bad proto %d", p) + return nil, syserr.ErrInvalidArgument + } + + var redirectTarget linux.XTRedirectTarget + buf = buf[:linux.SizeOfXTRedirectTarget] + binary.Unmarshal(buf, usermem.ByteOrder, &redirectTarget) + + // Copy linux.XTRedirectTarget to stack.RedirectTarget. + target := stack.RedirectTarget{NetworkProtocol: filter.NetworkProtocol()} + + // RangeSize should be 1. + nfRange := redirectTarget.NfRange + if nfRange.RangeSize != 1 { + nflog("redirectTargetMaker: bad rangesize %d", nfRange.RangeSize) + return nil, syserr.ErrInvalidArgument + } + + // 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 { + nflog("redirectTargetMaker: invalid range flags %d", nfRange.RangeIPV4.Flags) + return nil, syserr.ErrInvalidArgument + } + + // TODO(gvisor.dev/issue/170): Port range is not supported yet. + if nfRange.RangeIPV4.MinPort != nfRange.RangeIPV4.MaxPort { + nflog("redirectTargetMaker: MinPort != MaxPort (%d, %d)", nfRange.RangeIPV4.MinPort, nfRange.RangeIPV4.MaxPort) + return nil, syserr.ErrInvalidArgument } - // Convert port from little endian to big endian. - port := make([]byte, 2) - binary.LittleEndian.PutUint16(port, rt.MinPort) - target.NfRange.RangeIPV4.MinPort = binary.BigEndian.Uint16(port) - binary.LittleEndian.PutUint16(port, rt.MaxPort) - target.NfRange.RangeIPV4.MaxPort = binary.BigEndian.Uint16(port) - return binary.Marshal(ret, usermem.ByteOrder, target) + if nfRange.RangeIPV4.MinIP != nfRange.RangeIPV4.MaxIP { + nflog("redirectTargetMaker: MinIP != MaxIP (%d, %d)", nfRange.RangeIPV4.MinPort, nfRange.RangeIPV4.MaxPort) + return nil, syserr.ErrInvalidArgument + } + + target.Addr = tcpip.Address(nfRange.RangeIPV4.MinIP[:]) + target.Port = ntohs(nfRange.RangeIPV4.MinPort) + + return &target, nil } -func marshalJumpTarget(jt JumpTarget) []byte { - nflog("convert to binary: marshalling jump target") +type nfNATTarget struct { + Target linux.XTEntryTarget + Range linux.NFNATRange +} - // The target's name will be the empty string. - target := linux.XTStandardTarget{ +const nfNATMarhsalledSize = linux.SizeOfXTEntryTarget + linux.SizeOfNFNATRange + +type nfNATTargetMaker struct { + NetworkProtocol tcpip.NetworkProtocolNumber +} + +func (rm *nfNATTargetMaker) id() stack.TargetID { + return stack.TargetID{ + Name: stack.RedirectTargetName, + NetworkProtocol: rm.NetworkProtocol, + } +} + +func (*nfNATTargetMaker) marshal(target stack.Target) []byte { + rt := target.(*stack.RedirectTarget) + nt := nfNATTarget{ Target: linux.XTEntryTarget{ - TargetSize: linux.SizeOfXTStandardTarget, + TargetSize: nfNATMarhsalledSize, + }, + Range: linux.NFNATRange{ + Flags: linux.NF_NAT_RANGE_PROTO_SPECIFIED, }, - // Verdict is overloaded by the ABI. When positive, it holds - // the jump offset from the start of the table. - Verdict: int32(jt.Offset), } + copy(nt.Target.Name[:], stack.RedirectTargetName) + copy(nt.Range.MinAddr[:], rt.Addr) + copy(nt.Range.MaxAddr[:], rt.Addr) - ret := make([]byte, 0, linux.SizeOfXTStandardTarget) - return binary.Marshal(ret, usermem.ByteOrder, target) + nt.Range.MinProto = htons(rt.Port) + nt.Range.MaxProto = nt.Range.MinProto + + ret := make([]byte, 0, nfNATMarhsalledSize) + return binary.Marshal(ret, usermem.ByteOrder, nt) } -// translateFromStandardVerdict translates verdicts the same way as the iptables -// tool. -func translateFromStandardVerdict(verdict stack.RuleVerdict) int32 { - switch verdict { - case stack.RuleAccept: - return -linux.NF_ACCEPT - 1 - case stack.RuleDrop: - return -linux.NF_DROP - 1 - case stack.RuleReturn: - return linux.NF_RETURN - default: - // TODO(gvisor.dev/issue/170): Support Jump. - panic(fmt.Sprintf("unknown standard verdict: %d", verdict)) +func (*nfNATTargetMaker) unmarshal(buf []byte, filter stack.IPHeaderFilter) (stack.Target, *syserr.Error) { + if size := nfNATMarhsalledSize; len(buf) < size { + nflog("nfNATTargetMaker: buf has insufficient size (%d) for nfNAT target (%d)", len(buf), size) + return nil, syserr.ErrInvalidArgument } + + if p := filter.Protocol; p != header.TCPProtocolNumber && p != header.UDPProtocolNumber { + nflog("nfNATTargetMaker: bad proto %d", p) + return nil, syserr.ErrInvalidArgument + } + + var natRange linux.NFNATRange + buf = buf[linux.SizeOfXTEntryTarget:nfNATMarhsalledSize] + binary.Unmarshal(buf, usermem.ByteOrder, &natRange) + + // We don't support port or address ranges. + if natRange.MinAddr != natRange.MaxAddr { + nflog("nfNATTargetMaker: MinAddr and MaxAddr are different") + return nil, syserr.ErrInvalidArgument + } + if natRange.MinProto != natRange.MaxProto { + nflog("nfNATTargetMaker: MinProto and MaxProto are different") + return nil, syserr.ErrInvalidArgument + } + + // TODO(gvisor.dev/issue/3549): Check for other flags. + // For now, redirect target only supports destination change. + if natRange.Flags != linux.NF_NAT_RANGE_PROTO_SPECIFIED { + nflog("nfNATTargetMaker: invalid range flags %d", natRange.Flags) + return nil, syserr.ErrInvalidArgument + } + + target := stack.RedirectTarget{ + NetworkProtocol: filter.NetworkProtocol(), + Addr: tcpip.Address(natRange.MinAddr[:]), + Port: ntohs(natRange.MinProto), + } + + return &target, nil } // translateToStandardTarget translates from the value in a // linux.XTStandardTarget to an stack.Verdict. -func translateToStandardTarget(val int32) (stack.Target, error) { +func translateToStandardTarget(val int32, netProto tcpip.NetworkProtocolNumber) (stack.Target, *syserr.Error) { // TODO(gvisor.dev/issue/170): Support other verdicts. switch val { case -linux.NF_ACCEPT - 1: - return stack.AcceptTarget{}, nil + return &stack.AcceptTarget{NetworkProtocol: netProto}, nil case -linux.NF_DROP - 1: - return stack.DropTarget{}, nil + return &stack.DropTarget{NetworkProtocol: netProto}, nil case -linux.NF_QUEUE - 1: - return nil, errors.New("unsupported iptables verdict QUEUE") + nflog("unsupported iptables verdict QUEUE") + return nil, syserr.ErrInvalidArgument case linux.NF_RETURN: - return stack.ReturnTarget{}, nil + return &stack.ReturnTarget{NetworkProtocol: netProto}, nil default: - return nil, fmt.Errorf("unknown iptables verdict %d", val) + nflog("unknown iptables verdict %d", val) + return nil, syserr.ErrInvalidArgument } } // parseTarget parses a target from optVal. optVal should contain only the // target. -func parseTarget(filter stack.IPHeaderFilter, optVal []byte) (stack.Target, error) { +func parseTarget(filter stack.IPHeaderFilter, optVal []byte, ipv6 bool) (stack.Target, *syserr.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)) + nflog("optVal has insufficient size for entry target %d", len(optVal)) + return nil, syserr.ErrInvalidArgument } var target linux.XTEntryTarget buf := optVal[:linux.SizeOfXTEntryTarget] binary.Unmarshal(buf, usermem.ByteOrder, &target) - switch target.Name.String() { - case "": - // Standard target. - if len(optVal) != linux.SizeOfXTStandardTarget { - return nil, fmt.Errorf("optVal has wrong size for standard target %d", len(optVal)) - } - var standardTarget linux.XTStandardTarget - buf = optVal[:linux.SizeOfXTStandardTarget] - binary.Unmarshal(buf, usermem.ByteOrder, &standardTarget) - - if standardTarget.Verdict < 0 { - // A Verdict < 0 indicates a non-jump verdict. - return translateToStandardTarget(standardTarget.Verdict) - } - // A verdict >= 0 indicates a jump. - return JumpTarget{Offset: uint32(standardTarget.Verdict)}, nil - - case errorTargetName: - // Error target. - if len(optVal) != linux.SizeOfXTErrorTarget { - return nil, fmt.Errorf("optVal has insufficient size for error target %d", len(optVal)) - } - var errorTarget linux.XTErrorTarget - buf = optVal[:linux.SizeOfXTErrorTarget] - binary.Unmarshal(buf, usermem.ByteOrder, &errorTarget) - - // Error targets are used in 2 cases: - // * An actual error case. These rules have an error - // named errorTargetName. The last entry of the table - // is usually an error case to catch any packets that - // somehow fall through every rule. - // * To mark the start of a user defined chain. These - // rules have an error with the name of the chain. - switch name := errorTarget.Name.String(); name { - case errorTargetName: - nflog("set entries: error target") - return stack.ErrorTarget{}, nil - default: - // User defined chain. - nflog("set entries: user-defined target %q", name) - return stack.UserChainTarget{Name: name}, nil - } - - case redirectTargetName: - // Redirect target. - if len(optVal) < linux.SizeOfXTRedirectTarget { - return nil, fmt.Errorf("netfilter.SetEntries: optVal has insufficient size for redirect target %d", len(optVal)) - } - - if p := filter.Protocol; p != header.TCPProtocolNumber && p != header.UDPProtocolNumber { - return nil, fmt.Errorf("netfilter.SetEntries: bad proto %d", p) - } - - 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: bad rangesize %d", nfRange.RangeSize) - } - - // 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 range flags %d", nfRange.RangeIPV4.Flags) - } - 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: minport != maxport (%d, %d)", nfRange.RangeIPV4.MinPort, nfRange.RangeIPV4.MaxPort) - } - - // 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()) + return unmarshalTarget(target, filter, optVal) } // JumpTarget implements stack.Target. @@ -274,9 +376,31 @@ type JumpTarget struct { // RuleNum is the rule to jump to. RuleNum int + + // NetworkProtocol is the network protocol the target is used with. + NetworkProtocol tcpip.NetworkProtocolNumber +} + +// ID implements Target.ID. +func (jt *JumpTarget) ID() stack.TargetID { + return stack.TargetID{ + NetworkProtocol: jt.NetworkProtocol, + } } // Action implements stack.Target.Action. -func (jt JumpTarget) Action(*stack.PacketBuffer, *stack.ConnTrack, stack.Hook, *stack.GSO, *stack.Route, tcpip.Address) (stack.RuleVerdict, int) { +func (jt *JumpTarget) Action(*stack.PacketBuffer, *stack.ConnTrack, stack.Hook, *stack.GSO, *stack.Route, tcpip.Address) (stack.RuleVerdict, int) { return stack.RuleJump, jt.RuleNum } + +func ntohs(port uint16) uint16 { + buf := make([]byte, 2) + binary.BigEndian.PutUint16(buf, port) + return usermem.ByteOrder.Uint16(buf) +} + +func htons(port uint16) uint16 { + buf := make([]byte, 2) + usermem.ByteOrder.PutUint16(buf, port) + return binary.BigEndian.Uint16(buf) +} diff --git a/pkg/sentry/socket/netfilter/tcp_matcher.go b/pkg/sentry/socket/netfilter/tcp_matcher.go index 0bfd6c1f4..844acfede 100644 --- a/pkg/sentry/socket/netfilter/tcp_matcher.go +++ b/pkg/sentry/socket/netfilter/tcp_matcher.go @@ -97,17 +97,33 @@ func (*TCPMatcher) Name() string { // Match implements Matcher.Match. func (tm *TCPMatcher) Match(hook stack.Hook, pkt *stack.PacketBuffer, interfaceName string) (bool, bool) { - netHeader := header.IPv4(pkt.NetworkHeader().View()) + // TODO(gvisor.dev/issue/170): Proto checks should ultimately be moved + // into the stack.Check codepath as matchers are added. + switch pkt.NetworkProtocolNumber { + case header.IPv4ProtocolNumber: + netHeader := header.IPv4(pkt.NetworkHeader().View()) + if netHeader.TransportProtocol() != header.TCPProtocolNumber { + return false, false + } - if netHeader.TransportProtocol() != header.TCPProtocolNumber { - return false, false - } + // We don't match fragments. + if frag := netHeader.FragmentOffset(); frag != 0 { + if frag == 1 { + return false, true + } + return false, false + } - // We dont't match fragments. - if frag := netHeader.FragmentOffset(); frag != 0 { - if frag == 1 { - return false, true + case header.IPv6ProtocolNumber: + // As in Linux, we do not perform an IPv6 fragment check. See + // xt_action_param.fragoff in + // include/linux/netfilter/x_tables.h. + if header.IPv6(pkt.NetworkHeader().View()).TransportProtocol() != header.TCPProtocolNumber { + return false, false } + + default: + // We don't know the network protocol. return false, false } diff --git a/pkg/sentry/socket/netfilter/udp_matcher.go b/pkg/sentry/socket/netfilter/udp_matcher.go index 7ed05461d..63201201c 100644 --- a/pkg/sentry/socket/netfilter/udp_matcher.go +++ b/pkg/sentry/socket/netfilter/udp_matcher.go @@ -94,19 +94,33 @@ func (*UDPMatcher) Name() string { // Match implements Matcher.Match. func (um *UDPMatcher) Match(hook stack.Hook, pkt *stack.PacketBuffer, interfaceName string) (bool, bool) { - netHeader := header.IPv4(pkt.NetworkHeader().View()) - // TODO(gvisor.dev/issue/170): Proto checks should ultimately be moved // into the stack.Check codepath as matchers are added. - if netHeader.TransportProtocol() != header.UDPProtocolNumber { - return false, false - } + switch pkt.NetworkProtocolNumber { + case header.IPv4ProtocolNumber: + netHeader := header.IPv4(pkt.NetworkHeader().View()) + if netHeader.TransportProtocol() != header.UDPProtocolNumber { + return false, false + } - // We dont't match fragments. - if frag := netHeader.FragmentOffset(); frag != 0 { - if frag == 1 { - return false, true + // We don't match fragments. + if frag := netHeader.FragmentOffset(); frag != 0 { + if frag == 1 { + return false, true + } + return false, false } + + case header.IPv6ProtocolNumber: + // As in Linux, we do not perform an IPv6 fragment check. See + // xt_action_param.fragoff in + // include/linux/netfilter/x_tables.h. + if header.IPv6(pkt.NetworkHeader().View()).TransportProtocol() != header.UDPProtocolNumber { + return false, false + } + + default: + // We don't know the network protocol. return false, false } diff --git a/pkg/sentry/socket/netlink/provider_vfs2.go b/pkg/sentry/socket/netlink/provider_vfs2.go index bb205be0d..e8930f031 100644 --- a/pkg/sentry/socket/netlink/provider_vfs2.go +++ b/pkg/sentry/socket/netlink/provider_vfs2.go @@ -52,6 +52,7 @@ func (*socketProviderVFS2) Socket(t *kernel.Task, stype linux.SockType, protocol vfsfd := &s.vfsfd mnt := t.Kernel().SocketMount() d := sockfs.NewDentry(t.Credentials(), mnt) + defer d.DecRef(t) if err := vfsfd.Init(s, linux.O_RDWR, mnt, d, &vfs.FileDescriptionOptions{ DenyPRead: true, DenyPWrite: true, diff --git a/pkg/sentry/socket/netlink/socket_vfs2.go b/pkg/sentry/socket/netlink/socket_vfs2.go index a38d25da9..c83b23242 100644 --- a/pkg/sentry/socket/netlink/socket_vfs2.go +++ b/pkg/sentry/socket/netlink/socket_vfs2.go @@ -82,6 +82,13 @@ func NewVFS2(t *kernel.Task, skType linux.SockType, protocol Protocol) (*SocketV return fd, nil } +// Release implements vfs.FileDescriptionImpl.Release. +func (s *SocketVFS2) Release(ctx context.Context) { + t := kernel.TaskFromContext(ctx) + t.Kernel().DeleteSocketVFS2(&s.vfsfd) + s.socketOpsCommon.Release(ctx) +} + // Readiness implements waiter.Waitable.Readiness. func (s *SocketVFS2) Readiness(mask waiter.EventMask) waiter.EventMask { return s.socketOpsCommon.Readiness(mask) diff --git a/pkg/sentry/socket/netstack/netstack.go b/pkg/sentry/socket/netstack/netstack.go index 2e568bc3d..87e30d742 100644 --- a/pkg/sentry/socket/netstack/netstack.go +++ b/pkg/sentry/socket/netstack/netstack.go @@ -158,6 +158,9 @@ var Metrics = tcpip.Stats{ OutgoingPacketErrors: mustCreateMetric("/netstack/ip/outgoing_packet_errors", "Total number of IP packets which failed to write to a link-layer endpoint."), MalformedPacketsReceived: mustCreateMetric("/netstack/ip/malformed_packets_received", "Total number of IP packets which failed IP header validation checks."), MalformedFragmentsReceived: mustCreateMetric("/netstack/ip/malformed_fragments_received", "Total number of IP fragments which failed IP fragment validation checks."), + IPTablesPreroutingDropped: mustCreateMetric("/netstack/ip/iptables/prerouting_dropped", "Total number of IP packets dropped in the Prerouting chain."), + IPTablesInputDropped: mustCreateMetric("/netstack/ip/iptables/input_dropped", "Total number of IP packets dropped in the Input chain."), + IPTablesOutputDropped: mustCreateMetric("/netstack/ip/iptables/output_dropped", "Total number of IP packets dropped in the Output chain."), }, TCP: tcpip.TCPStats{ ActiveConnectionOpenings: mustCreateMetric("/netstack/tcp/active_connection_openings", "Number of connections opened successfully via Connect."), @@ -195,7 +198,6 @@ var Metrics = tcpip.Stats{ PacketsSent: mustCreateMetric("/netstack/udp/packets_sent", "Number of UDP datagrams sent."), PacketSendErrors: mustCreateMetric("/netstack/udp/packet_send_errors", "Number of UDP datagrams failed to be sent."), ChecksumErrors: mustCreateMetric("/netstack/udp/checksum_errors", "Number of UDP datagrams dropped due to bad checksums."), - InvalidSourceAddress: mustCreateMetric("/netstack/udp/invalid_source", "Number of UDP datagrams dropped due to invalid source address."), }, } @@ -482,8 +484,35 @@ func (s *socketOpsCommon) fetchReadView() *syserr.Error { } // Release implements fs.FileOperations.Release. -func (s *socketOpsCommon) Release(context.Context) { +func (s *socketOpsCommon) Release(ctx context.Context) { + e, ch := waiter.NewChannelEntry(nil) + s.EventRegister(&e, waiter.EventHUp|waiter.EventErr) + defer s.EventUnregister(&e) + s.Endpoint.Close() + + // SO_LINGER option is valid only for TCP. For other socket types + // return after endpoint close. + if family, skType, _ := s.Type(); skType != linux.SOCK_STREAM || (family != linux.AF_INET && family != linux.AF_INET6) { + return + } + + var v tcpip.LingerOption + if err := s.Endpoint.GetSockOpt(&v); err != nil { + return + } + + // The case for zero timeout is handled in tcp endpoint close function. + // Close is blocked until either: + // 1. The endpoint state is not in any of the states: FIN-WAIT1, + // CLOSING and LAST_ACK. + // 2. Timeout is reached. + if v.Enabled && v.Timeout != 0 { + t := kernel.TaskFromContext(ctx) + start := t.Kernel().MonotonicClock().Now() + deadline := start.Add(v.Timeout) + t.BlockWithDeadline(ch, true, deadline) + } } // Read implements fs.FileOperations.Read. @@ -1155,7 +1184,16 @@ func getSockOptSocket(t *kernel.Task, s socket.SocketOps, ep commonEndpoint, fam return nil, syserr.ErrInvalidArgument } - linger := linux.Linger{} + var v tcpip.LingerOption + var linger linux.Linger + if err := ep.GetSockOpt(&v); err != nil { + return nil, syserr.TranslateNetstackError(err) + } + + if v.Enabled { + linger.OnOff = 1 + } + linger.Linger = int32(v.Timeout.Seconds()) return &linger, nil case linux.SO_SNDTIMEO: @@ -1474,8 +1512,17 @@ func getSockOptIPv6(t *kernel.Task, s socket.SocketOps, ep commonEndpoint, name return &vP, nil case linux.IP6T_ORIGINAL_DST: - // TODO(gvisor.dev/issue/170): ip6tables. - return nil, syserr.ErrInvalidArgument + if outLen < int(binary.Size(linux.SockAddrInet6{})) { + return nil, syserr.ErrInvalidArgument + } + + var v tcpip.OriginalDestinationOption + if err := ep.GetSockOpt(&v); err != nil { + return nil, syserr.TranslateNetstackError(err) + } + + a, _ := ConvertAddress(linux.AF_INET6, tcpip.FullAddress(v)) + return a.(*linux.SockAddrInet6), nil case linux.IP6T_SO_GET_INFO: if outLen < linux.SizeOfIPTGetinfo { @@ -1517,6 +1564,26 @@ func getSockOptIPv6(t *kernel.Task, s socket.SocketOps, ep commonEndpoint, name } return &entries, nil + case linux.IP6T_SO_GET_REVISION_TARGET: + if outLen < linux.SizeOfXTGetRevision { + return nil, syserr.ErrInvalidArgument + } + + // Only valid for raw IPv6 sockets. + if family, skType, _ := s.Type(); family != linux.AF_INET6 || skType != linux.SOCK_RAW { + return nil, syserr.ErrProtocolNotAvailable + } + + stack := inet.StackFromContext(t) + if stack == nil { + return nil, syserr.ErrNoDevice + } + ret, err := netfilter.TargetRevision(t, outPtr, header.IPv6ProtocolNumber) + if err != nil { + return nil, err + } + return &ret, nil + default: emitUnimplementedEventIPv6(t, name) } @@ -1680,6 +1747,26 @@ func getSockOptIP(t *kernel.Task, s socket.SocketOps, ep commonEndpoint, name in } return &entries, nil + case linux.IPT_SO_GET_REVISION_TARGET: + if outLen < linux.SizeOfXTGetRevision { + return nil, syserr.ErrInvalidArgument + } + + // Only valid for raw IPv4 sockets. + if family, skType, _ := s.Type(); family != linux.AF_INET || skType != linux.SOCK_RAW { + return nil, syserr.ErrProtocolNotAvailable + } + + stack := inet.StackFromContext(t) + if stack == nil { + return nil, syserr.ErrNoDevice + } + ret, err := netfilter.TargetRevision(t, outPtr, header.IPv4ProtocolNumber) + if err != nil { + return nil, err + } + return &ret, nil + default: emitUnimplementedEventIP(t, name) } @@ -1732,10 +1819,16 @@ func SetSockOpt(t *kernel.Task, s socket.SocketOps, ep commonEndpoint, level int case linux.SOL_IP: return setSockOptIP(t, s, ep, name, optVal) + case linux.SOL_PACKET: + // gVisor doesn't support any SOL_PACKET options just return not + // supported. Returning nil here will result in tcpdump thinking AF_PACKET + // features are supported and proceed to use them and break. + t.Kernel().EmitUnimplementedEvent(t) + return syserr.ErrProtocolNotAvailable + case linux.SOL_UDP, linux.SOL_ICMPV6, - linux.SOL_RAW, - linux.SOL_PACKET: + linux.SOL_RAW: t.Kernel().EmitUnimplementedEvent(t) } @@ -1884,7 +1977,10 @@ func setSockOptSocket(t *kernel.Task, s socket.SocketOps, ep commonEndpoint, nam socket.SetSockOptEmitUnimplementedEvent(t, name) } - return nil + return syserr.TranslateNetstackError( + ep.SetSockOpt(&tcpip.LingerOption{ + Enabled: v.OnOff != 0, + Timeout: time.Second * time.Duration(v.Linger)})) case linux.SO_DETACH_FILTER: // optval is ignored. diff --git a/pkg/sentry/socket/netstack/netstack_vfs2.go b/pkg/sentry/socket/netstack/netstack_vfs2.go index c0212ad76..4c6791fff 100644 --- a/pkg/sentry/socket/netstack/netstack_vfs2.go +++ b/pkg/sentry/socket/netstack/netstack_vfs2.go @@ -79,6 +79,13 @@ func NewVFS2(t *kernel.Task, family int, skType linux.SockType, protocol int, qu return vfsfd, nil } +// Release implements vfs.FileDescriptionImpl.Release. +func (s *SocketVFS2) Release(ctx context.Context) { + t := kernel.TaskFromContext(ctx) + t.Kernel().DeleteSocketVFS2(&s.vfsfd) + s.socketOpsCommon.Release(ctx) +} + // Readiness implements waiter.Waitable.Readiness. func (s *SocketVFS2) Readiness(mask waiter.EventMask) waiter.EventMask { return s.socketOpsCommon.Readiness(mask) diff --git a/pkg/sentry/socket/netstack/stack.go b/pkg/sentry/socket/netstack/stack.go index 36144e1eb..1028d2a6e 100644 --- a/pkg/sentry/socket/netstack/stack.go +++ b/pkg/sentry/socket/netstack/stack.go @@ -412,3 +412,24 @@ func (s *Stack) CleanupEndpoints() []stack.TransportEndpoint { func (s *Stack) RestoreCleanupEndpoints(es []stack.TransportEndpoint) { s.Stack.RestoreCleanupEndpoints(es) } + +// Forwarding implements inet.Stack.Forwarding. +func (s *Stack) Forwarding(protocol tcpip.NetworkProtocolNumber) bool { + switch protocol { + case ipv4.ProtocolNumber, ipv6.ProtocolNumber: + return s.Stack.Forwarding(protocol) + default: + panic(fmt.Sprintf("Forwarding(%v) failed: unsupported protocol", protocol)) + } +} + +// SetForwarding implements inet.Stack.SetForwarding. +func (s *Stack) SetForwarding(protocol tcpip.NetworkProtocolNumber, enable bool) error { + switch protocol { + case ipv4.ProtocolNumber, ipv6.ProtocolNumber: + s.Stack.SetForwarding(protocol, enable) + default: + panic(fmt.Sprintf("SetForwarding(%v) failed: unsupported protocol", protocol)) + } + return nil +} diff --git a/pkg/sentry/socket/unix/BUILD b/pkg/sentry/socket/unix/BUILD index a89583dad..cc7408698 100644 --- a/pkg/sentry/socket/unix/BUILD +++ b/pkg/sentry/socket/unix/BUILD @@ -7,10 +7,21 @@ go_template_instance( name = "socket_refs", out = "socket_refs.go", package = "unix", - prefix = "socketOpsCommon", + prefix = "socketOperations", template = "//pkg/refs_vfs2:refs_template", types = { - "T": "socketOpsCommon", + "T": "SocketOperations", + }, +) + +go_template_instance( + name = "socket_vfs2_refs", + out = "socket_vfs2_refs.go", + package = "unix", + prefix = "socketVFS2", + template = "//pkg/refs_vfs2:refs_template", + types = { + "T": "SocketVFS2", }, ) @@ -20,6 +31,7 @@ go_library( "device.go", "io.go", "socket_refs.go", + "socket_vfs2_refs.go", "unix.go", "unix_vfs2.go", ], diff --git a/pkg/sentry/socket/unix/transport/unix.go b/pkg/sentry/socket/unix/transport/unix.go index cbbdd000f..d6fc03520 100644 --- a/pkg/sentry/socket/unix/transport/unix.go +++ b/pkg/sentry/socket/unix/transport/unix.go @@ -746,6 +746,9 @@ type baseEndpoint struct { // path is not empty if the endpoint has been bound, // or may be used if the endpoint is connected. path string + + // linger is used for SO_LINGER socket option. + linger tcpip.LingerOption } // EventRegister implements waiter.Waitable.EventRegister. @@ -841,8 +844,14 @@ func (e *baseEndpoint) SendMsg(ctx context.Context, data [][]byte, c ControlMess return n, err } -// SetSockOpt sets a socket option. Currently not supported. -func (e *baseEndpoint) SetSockOpt(tcpip.SettableSocketOption) *tcpip.Error { +// SetSockOpt sets a socket option. +func (e *baseEndpoint) SetSockOpt(opt tcpip.SettableSocketOption) *tcpip.Error { + switch v := opt.(type) { + case *tcpip.LingerOption: + e.Lock() + e.linger = *v + e.Unlock() + } return nil } @@ -945,8 +954,17 @@ func (e *baseEndpoint) GetSockOptInt(opt tcpip.SockOptInt) (int, *tcpip.Error) { // GetSockOpt implements tcpip.Endpoint.GetSockOpt. func (e *baseEndpoint) GetSockOpt(opt tcpip.GettableSocketOption) *tcpip.Error { - log.Warningf("Unsupported socket option: %T", opt) - return tcpip.ErrUnknownProtocolOption + switch o := opt.(type) { + case *tcpip.LingerOption: + e.Lock() + *o = e.linger + e.Unlock() + return nil + + default: + log.Warningf("Unsupported socket option: %T", opt) + return tcpip.ErrUnknownProtocolOption + } } // LastError implements Endpoint.LastError. diff --git a/pkg/sentry/socket/unix/unix.go b/pkg/sentry/socket/unix/unix.go index 917055cea..f80011ce4 100644 --- a/pkg/sentry/socket/unix/unix.go +++ b/pkg/sentry/socket/unix/unix.go @@ -55,6 +55,7 @@ type SocketOperations struct { fsutil.FileNoopFlush `state:"nosave"` fsutil.FileUseInodeUnstableAttr `state:"nosave"` + socketOperationsRefs socketOpsCommon } @@ -84,11 +85,27 @@ func NewWithDirent(ctx context.Context, d *fs.Dirent, ep transport.Endpoint, sty return fs.NewFile(ctx, d, flags, &s) } +// DecRef implements RefCounter.DecRef. +func (s *SocketOperations) DecRef(ctx context.Context) { + s.socketOperationsRefs.DecRef(func() { + s.ep.Close(ctx) + if s.abstractNamespace != nil { + s.abstractNamespace.Remove(s.abstractName, s) + } + }) +} + +// Release implemements fs.FileOperations.Release. +func (s *SocketOperations) Release(ctx context.Context) { + // Release only decrements a reference on s because s may be referenced in + // the abstract socket namespace. + s.DecRef(ctx) +} + // socketOpsCommon contains the socket operations common to VFS1 and VFS2. // // +stateify savable type socketOpsCommon struct { - socketOpsCommonRefs socket.SendReceiveTimeout ep transport.Endpoint @@ -101,23 +118,6 @@ type socketOpsCommon struct { abstractNamespace *kernel.AbstractSocketNamespace } -// DecRef implements RefCounter.DecRef. -func (s *socketOpsCommon) DecRef(ctx context.Context) { - s.socketOpsCommonRefs.DecRef(func() { - s.ep.Close(ctx) - if s.abstractNamespace != nil { - s.abstractNamespace.Remove(s.abstractName, s) - } - }) -} - -// Release implemements fs.FileOperations.Release. -func (s *socketOpsCommon) Release(ctx context.Context) { - // Release only decrements a reference on s because s may be referenced in - // the abstract socket namespace. - s.DecRef(ctx) -} - func (s *socketOpsCommon) isPacket() bool { switch s.stype { case linux.SOCK_DGRAM, linux.SOCK_SEQPACKET: diff --git a/pkg/sentry/socket/unix/unix_vfs2.go b/pkg/sentry/socket/unix/unix_vfs2.go index 3688f22d2..3345124cc 100644 --- a/pkg/sentry/socket/unix/unix_vfs2.go +++ b/pkg/sentry/socket/unix/unix_vfs2.go @@ -37,12 +37,15 @@ import ( // SocketVFS2 implements socket.SocketVFS2 (and by extension, // vfs.FileDescriptionImpl) for Unix sockets. +// +// +stateify savable type SocketVFS2 struct { vfsfd vfs.FileDescription vfs.FileDescriptionDefaultImpl vfs.DentryMetadataFileDescriptionImpl vfs.LockFD + socketVFS2Refs socketOpsCommon } @@ -53,6 +56,7 @@ var _ = socket.SocketVFS2(&SocketVFS2{}) func NewSockfsFile(t *kernel.Task, ep transport.Endpoint, stype linux.SockType) (*vfs.FileDescription, *syserr.Error) { mnt := t.Kernel().SocketMount() d := sockfs.NewDentry(t.Credentials(), mnt) + defer d.DecRef(t) fd, err := NewFileDescription(ep, stype, linux.O_RDWR, mnt, d, &vfs.FileLocks{}) if err != nil { @@ -88,6 +92,25 @@ func NewFileDescription(ep transport.Endpoint, stype linux.SockType, flags uint3 return vfsfd, nil } +// DecRef implements RefCounter.DecRef. +func (s *SocketVFS2) DecRef(ctx context.Context) { + s.socketVFS2Refs.DecRef(func() { + t := kernel.TaskFromContext(ctx) + t.Kernel().DeleteSocketVFS2(&s.vfsfd) + s.ep.Close(ctx) + if s.abstractNamespace != nil { + s.abstractNamespace.Remove(s.abstractName, s) + } + }) +} + +// Release implements vfs.FileDescriptionImpl.Release. +func (s *SocketVFS2) Release(ctx context.Context) { + // Release only decrements a reference on s because s may be referenced in + // the abstract socket namespace. + s.DecRef(ctx) +} + // GetSockOpt implements the linux syscall getsockopt(2) for sockets backed by // a transport.Endpoint. func (s *SocketVFS2) GetSockOpt(t *kernel.Task, level, name int, outPtr usermem.Addr, outLen int) (marshal.Marshallable, *syserr.Error) { diff --git a/pkg/sentry/state/state.go b/pkg/sentry/state/state.go index a06c9b8ab..245d2c5cf 100644 --- a/pkg/sentry/state/state.go +++ b/pkg/sentry/state/state.go @@ -61,8 +61,10 @@ func (opts SaveOpts) Save(k *kernel.Kernel, w *watchdog.Watchdog) error { log.Infof("Sandbox save started, pausing all tasks.") k.Pause() k.ReceiveTaskStates() - defer k.Unpause() - defer log.Infof("Tasks resumed after save.") + defer func() { + k.Unpause() + log.Infof("Tasks resumed after save.") + }() w.Stop() defer w.Start() diff --git a/pkg/sentry/strace/strace.go b/pkg/sentry/strace/strace.go index 52281ccc2..396744597 100644 --- a/pkg/sentry/strace/strace.go +++ b/pkg/sentry/strace/strace.go @@ -17,7 +17,6 @@ package strace import ( - "encoding/binary" "fmt" "strconv" "strings" @@ -294,7 +293,7 @@ func itimerval(t *kernel.Task, addr usermem.Addr) string { } interval := timeval(t, addr) - value := timeval(t, addr+usermem.Addr(binary.Size(linux.Timeval{}))) + value := timeval(t, addr+usermem.Addr((*linux.Timeval)(nil).SizeBytes())) return fmt.Sprintf("%#x {interval=%s, value=%s}", addr, interval, value) } @@ -304,7 +303,7 @@ func itimerspec(t *kernel.Task, addr usermem.Addr) string { } interval := timespec(t, addr) - value := timespec(t, addr+usermem.Addr(binary.Size(linux.Timespec{}))) + value := timespec(t, addr+usermem.Addr((*linux.Timespec)(nil).SizeBytes())) return fmt.Sprintf("%#x {interval=%s, value=%s}", addr, interval, value) } diff --git a/pkg/sentry/syscalls/linux/BUILD b/pkg/sentry/syscalls/linux/BUILD index 75752b2e6..a2e441448 100644 --- a/pkg/sentry/syscalls/linux/BUILD +++ b/pkg/sentry/syscalls/linux/BUILD @@ -21,6 +21,7 @@ go_library( "sys_identity.go", "sys_inotify.go", "sys_lseek.go", + "sys_membarrier.go", "sys_mempolicy.go", "sys_mmap.go", "sys_mount.go", diff --git a/pkg/sentry/syscalls/linux/linux64.go b/pkg/sentry/syscalls/linux/linux64.go index 5f26697d2..9c9def7cd 100644 --- a/pkg/sentry/syscalls/linux/linux64.go +++ b/pkg/sentry/syscalls/linux/linux64.go @@ -376,7 +376,7 @@ var AMD64 = &kernel.SyscallTable{ 321: syscalls.CapError("bpf", linux.CAP_SYS_ADMIN, "", nil), 322: syscalls.Supported("execveat", Execveat), 323: syscalls.ErrorWithEvent("userfaultfd", syserror.ENOSYS, "", []string{"gvisor.dev/issue/266"}), // TODO(b/118906345) - 324: syscalls.ErrorWithEvent("membarrier", syserror.ENOSYS, "", []string{"gvisor.dev/issue/267"}), // TODO(gvisor.dev/issue/267) + 324: syscalls.PartiallySupported("membarrier", Membarrier, "Not supported on all platforms.", nil), 325: syscalls.PartiallySupported("mlock2", Mlock2, "Stub implementation. The sandbox lacks appropriate permissions.", nil), // Syscalls implemented after 325 are "backports" from versions @@ -527,8 +527,8 @@ var ARM64 = &kernel.SyscallTable{ 96: syscalls.Supported("set_tid_address", SetTidAddress), 97: syscalls.PartiallySupported("unshare", Unshare, "Mount, cgroup namespaces not supported. Network namespaces supported but must be empty.", nil), 98: syscalls.PartiallySupported("futex", Futex, "Robust futexes not supported.", nil), - 99: syscalls.Error("set_robust_list", syserror.ENOSYS, "Obsolete.", nil), - 100: syscalls.Error("get_robust_list", syserror.ENOSYS, "Obsolete.", nil), + 99: syscalls.Supported("set_robust_list", SetRobustList), + 100: syscalls.Supported("get_robust_list", GetRobustList), 101: syscalls.Supported("nanosleep", Nanosleep), 102: syscalls.Supported("getitimer", Getitimer), 103: syscalls.Supported("setitimer", Setitimer), @@ -695,7 +695,7 @@ var ARM64 = &kernel.SyscallTable{ 280: syscalls.CapError("bpf", linux.CAP_SYS_ADMIN, "", nil), 281: syscalls.Supported("execveat", Execveat), 282: syscalls.ErrorWithEvent("userfaultfd", syserror.ENOSYS, "", []string{"gvisor.dev/issue/266"}), // TODO(b/118906345) - 283: syscalls.ErrorWithEvent("membarrier", syserror.ENOSYS, "", []string{"gvisor.dev/issue/267"}), // TODO(gvisor.dev/issue/267) + 283: syscalls.PartiallySupported("membarrier", Membarrier, "Not supported on all platforms.", nil), 284: syscalls.PartiallySupported("mlock2", Mlock2, "Stub implementation. The sandbox lacks appropriate permissions.", nil), // Syscalls after 284 are "backports" from versions of Linux after 4.4. diff --git a/pkg/sentry/syscalls/linux/sys_membarrier.go b/pkg/sentry/syscalls/linux/sys_membarrier.go new file mode 100644 index 000000000..288cd8512 --- /dev/null +++ b/pkg/sentry/syscalls/linux/sys_membarrier.go @@ -0,0 +1,70 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package linux + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/syserror" +) + +// Membarrier implements syscall membarrier(2). +func Membarrier(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + cmd := args[0].Int() + flags := args[1].Int() + + p := t.Kernel().Platform + if !p.HaveGlobalMemoryBarrier() { + // Event for applications that want membarrier on a configuration that + // doesn't support them. + t.Kernel().EmitUnimplementedEvent(t) + return 0, nil, syserror.ENOSYS + } + + if flags != 0 { + return 0, nil, syserror.EINVAL + } + + switch cmd { + case linux.MEMBARRIER_CMD_QUERY: + const supportedCommands = linux.MEMBARRIER_CMD_GLOBAL | + linux.MEMBARRIER_CMD_GLOBAL_EXPEDITED | + linux.MEMBARRIER_CMD_REGISTER_GLOBAL_EXPEDITED | + linux.MEMBARRIER_CMD_PRIVATE_EXPEDITED | + linux.MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED + return supportedCommands, nil, nil + case linux.MEMBARRIER_CMD_PRIVATE_EXPEDITED: + if !t.MemoryManager().IsMembarrierPrivateEnabled() { + return 0, nil, syserror.EPERM + } + fallthrough + case linux.MEMBARRIER_CMD_GLOBAL, linux.MEMBARRIER_CMD_GLOBAL_EXPEDITED: + return 0, nil, p.GlobalMemoryBarrier() + case linux.MEMBARRIER_CMD_REGISTER_GLOBAL_EXPEDITED: + // no-op + return 0, nil, nil + case linux.MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED: + t.MemoryManager().EnableMembarrierPrivate() + return 0, nil, nil + case linux.MEMBARRIER_CMD_PRIVATE_EXPEDITED_SYNC_CORE, linux.MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED_SYNC_CORE: + // We're aware of these, but they aren't implemented since no platform + // supports them yet. + t.Kernel().EmitUnimplementedEvent(t) + fallthrough + default: + return 0, nil, syserror.EINVAL + } +} diff --git a/pkg/sentry/syscalls/linux/sys_mmap.go b/pkg/sentry/syscalls/linux/sys_mmap.go index 8ab062bca..cd8dfdfa4 100644 --- a/pkg/sentry/syscalls/linux/sys_mmap.go +++ b/pkg/sentry/syscalls/linux/sys_mmap.go @@ -100,6 +100,15 @@ func Mmap(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallC if err := file.ConfigureMMap(t, &opts); err != nil { return 0, nil, err } + } else if shared { + // Back shared anonymous mappings with a special mappable. + opts.Offset = 0 + m, err := mm.NewSharedAnonMappable(opts.Length, t.Kernel()) + if err != nil { + return 0, nil, err + } + opts.MappingIdentity = m // transfers ownership of m to opts + opts.Mappable = m } rv, err := t.MemoryManager().MMap(t, opts) diff --git a/pkg/sentry/syscalls/linux/sys_sched.go b/pkg/sentry/syscalls/linux/sys_sched.go index cd6f4dd94..bfcf44b6f 100644 --- a/pkg/sentry/syscalls/linux/sys_sched.go +++ b/pkg/sentry/syscalls/linux/sys_sched.go @@ -30,7 +30,7 @@ const ( // // +marshal type SchedParam struct { - schedPriority int64 + schedPriority int32 } // SchedGetparam implements linux syscall sched_getparam(2). diff --git a/pkg/sentry/syscalls/linux/vfs2/mmap.go b/pkg/sentry/syscalls/linux/vfs2/mmap.go index dc05c2994..9d9dbf775 100644 --- a/pkg/sentry/syscalls/linux/vfs2/mmap.go +++ b/pkg/sentry/syscalls/linux/vfs2/mmap.go @@ -17,6 +17,7 @@ package vfs2 import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/tmpfs" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/memmap" "gvisor.dev/gvisor/pkg/syserror" @@ -85,6 +86,17 @@ func Mmap(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallC if err := file.ConfigureMMap(t, &opts); err != nil { return 0, nil, err } + } else if shared { + // Back shared anonymous mappings with an anonymous tmpfs file. + opts.Offset = 0 + file, err := tmpfs.NewZeroFile(t, t.Credentials(), t.Kernel().ShmMount(), opts.Length) + if err != nil { + return 0, nil, err + } + defer file.DecRef(t) + if err := file.ConfigureMMap(t, &opts); err != nil { + return 0, nil, err + } } rv, err := t.MemoryManager().MMap(t, opts) diff --git a/pkg/sentry/syscalls/linux/vfs2/splice.go b/pkg/sentry/syscalls/linux/vfs2/splice.go index f55d74cd2..bf5c1171f 100644 --- a/pkg/sentry/syscalls/linux/vfs2/splice.go +++ b/pkg/sentry/syscalls/linux/vfs2/splice.go @@ -23,6 +23,7 @@ import ( "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/sentry/vfs" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" @@ -146,11 +147,6 @@ func Splice(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal panic("at least one end of splice must be a pipe") } - if n == 0 && err == io.EOF { - // We reached the end of the file. Eat the error and exit the loop. - err = nil - break - } if n != 0 || err != syserror.ErrWouldBlock || nonBlock { break } @@ -171,15 +167,16 @@ func Splice(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal } } - if n == 0 { - return 0, nil, err + if n != 0 { + // On Linux, inotify behavior is not very consistent with splice(2). We try + // our best to emulate Linux for very basic calls to splice, where for some + // reason, events are generated for output files, but not input files. + outFile.Dentry().InotifyWithParent(t, linux.IN_MODIFY, 0, vfs.PathEvent) } - // On Linux, inotify behavior is not very consistent with splice(2). We try - // our best to emulate Linux for very basic calls to splice, where for some - // reason, events are generated for output files, but not input files. - outFile.Dentry().InotifyWithParent(t, linux.IN_MODIFY, 0, vfs.PathEvent) - return uintptr(n), nil, nil + // We can only pass a single file to handleIOError, so pick inFile arbitrarily. + // This is used only for debugging purposes. + return uintptr(n), nil, slinux.HandleIOErrorVFS2(t, n != 0, err, syserror.ERESTARTSYS, "splice", outFile) } // Tee implements Linux syscall tee(2). @@ -251,11 +248,20 @@ func Tee(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallCo break } } - if n == 0 { - return 0, nil, err + + if n != 0 { + outFile.Dentry().InotifyWithParent(t, linux.IN_MODIFY, 0, vfs.PathEvent) + + // If a partial write is completed, the error is dropped. Log it here. + if err != nil && err != io.EOF && err != syserror.ErrWouldBlock { + log.Debugf("tee completed a partial write with error: %v", err) + err = nil + } } - outFile.Dentry().InotifyWithParent(t, linux.IN_MODIFY, 0, vfs.PathEvent) - return uintptr(n), nil, nil + + // We can only pass a single file to handleIOError, so pick inFile arbitrarily. + // This is used only for debugging purposes. + return uintptr(n), nil, slinux.HandleIOErrorVFS2(t, n != 0, err, syserror.ERESTARTSYS, "tee", inFile) } // Sendfile implements linux system call sendfile(2). @@ -348,11 +354,6 @@ func Sendfile(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysc for n < count { var spliceN int64 spliceN, err = outPipeFD.SpliceFromNonPipe(t, inFile, offset, count) - if spliceN == 0 && err == io.EOF { - // We reached the end of the file. Eat the error and exit the loop. - err = nil - break - } if offset != -1 { offset += spliceN } @@ -375,13 +376,6 @@ func Sendfile(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysc } else { readN, err = inFile.Read(t, usermem.BytesIOSequence(buf), vfs.ReadOptions{}) } - if readN == 0 && err != nil { - if err == io.EOF { - // We reached the end of the file. Eat the error before exiting the loop. - err = nil - } - break - } n += readN // Write all of the bytes that we read. This may need @@ -432,13 +426,20 @@ func Sendfile(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysc } } - if n == 0 { - return 0, nil, err + if n != 0 { + inFile.Dentry().InotifyWithParent(t, linux.IN_ACCESS, 0, vfs.PathEvent) + outFile.Dentry().InotifyWithParent(t, linux.IN_MODIFY, 0, vfs.PathEvent) + + if err != nil && err != io.EOF && err != syserror.ErrWouldBlock { + // If a partial write is completed, the error is dropped. Log it here. + log.Debugf("sendfile completed a partial write with error: %v", err) + err = nil + } } - inFile.Dentry().InotifyWithParent(t, linux.IN_ACCESS, 0, vfs.PathEvent) - outFile.Dentry().InotifyWithParent(t, linux.IN_MODIFY, 0, vfs.PathEvent) - return uintptr(n), nil, nil + // We can only pass a single file to handleIOError, so pick inFile arbitrarily. + // This is used only for debugging purposes. + return uintptr(n), nil, slinux.HandleIOErrorVFS2(t, n != 0, err, syserror.ERESTARTSYS, "sendfile", inFile) } // dualWaiter is used to wait on one or both vfs.FileDescriptions. It is not diff --git a/pkg/sentry/syscalls/linux/vfs2/vfs2.go b/pkg/sentry/syscalls/linux/vfs2/vfs2.go index 0df3bd449..c50fd97eb 100644 --- a/pkg/sentry/syscalls/linux/vfs2/vfs2.go +++ b/pkg/sentry/syscalls/linux/vfs2/vfs2.go @@ -163,6 +163,7 @@ func Override() { // Override ARM64. s = linux.ARM64 + s.Table[2] = syscalls.PartiallySupported("io_submit", IoSubmit, "Generally supported with exceptions. User ring optimizations are not implemented.", []string{"gvisor.dev/issue/204"}) s.Table[5] = syscalls.Supported("setxattr", SetXattr) s.Table[6] = syscalls.Supported("lsetxattr", Lsetxattr) s.Table[7] = syscalls.Supported("fsetxattr", Fsetxattr) @@ -200,6 +201,7 @@ func Override() { s.Table[44] = syscalls.Supported("fstatfs", Fstatfs) s.Table[45] = syscalls.Supported("truncate", Truncate) s.Table[46] = syscalls.Supported("ftruncate", Ftruncate) + s.Table[47] = syscalls.PartiallySupported("fallocate", Fallocate, "Not all options are supported.", nil) s.Table[48] = syscalls.Supported("faccessat", Faccessat) s.Table[49] = syscalls.Supported("chdir", Chdir) s.Table[50] = syscalls.Supported("fchdir", Fchdir) @@ -221,12 +223,14 @@ func Override() { s.Table[68] = syscalls.Supported("pwrite64", Pwrite64) s.Table[69] = syscalls.Supported("preadv", Preadv) s.Table[70] = syscalls.Supported("pwritev", Pwritev) + s.Table[71] = syscalls.Supported("sendfile", Sendfile) s.Table[72] = syscalls.Supported("pselect", Pselect) s.Table[73] = syscalls.Supported("ppoll", Ppoll) s.Table[74] = syscalls.Supported("signalfd4", Signalfd4) s.Table[76] = syscalls.Supported("splice", Splice) s.Table[77] = syscalls.Supported("tee", Tee) s.Table[78] = syscalls.Supported("readlinkat", Readlinkat) + s.Table[79] = syscalls.Supported("newfstatat", Newfstatat) s.Table[80] = syscalls.Supported("fstat", Fstat) s.Table[81] = syscalls.Supported("sync", Sync) s.Table[82] = syscalls.Supported("fsync", Fsync) @@ -251,8 +255,10 @@ func Override() { s.Table[210] = syscalls.Supported("shutdown", Shutdown) s.Table[211] = syscalls.Supported("sendmsg", SendMsg) s.Table[212] = syscalls.Supported("recvmsg", RecvMsg) + s.Table[213] = syscalls.Supported("readahead", Readahead) s.Table[221] = syscalls.Supported("execve", Execve) s.Table[222] = syscalls.Supported("mmap", Mmap) + s.Table[223] = syscalls.PartiallySupported("fadvise64", Fadvise64, "Not all options are supported.", nil) s.Table[242] = syscalls.Supported("accept4", Accept4) s.Table[243] = syscalls.Supported("recvmmsg", RecvMMsg) s.Table[267] = syscalls.Supported("syncfs", Syncfs) diff --git a/pkg/sentry/vfs/BUILD b/pkg/sentry/vfs/BUILD index 8093ca55c..c855608db 100644 --- a/pkg/sentry/vfs/BUILD +++ b/pkg/sentry/vfs/BUILD @@ -92,7 +92,6 @@ go_library( visibility = ["//pkg/sentry:internal"], deps = [ "//pkg/abi/linux", - "//pkg/binary", "//pkg/context", "//pkg/fd", "//pkg/fdnotifier", diff --git a/pkg/sentry/vfs/anonfs.go b/pkg/sentry/vfs/anonfs.go index 9c4db3047..bdfd3ca8f 100644 --- a/pkg/sentry/vfs/anonfs.go +++ b/pkg/sentry/vfs/anonfs.go @@ -52,6 +52,8 @@ const ( ) // anonFilesystemType implements FilesystemType. +// +// +stateify savable type anonFilesystemType struct{} // GetFilesystem implements FilesystemType.GetFilesystem. @@ -69,12 +71,15 @@ func (anonFilesystemType) Name() string { // // Since all Dentries in anonFilesystem are non-directories, all FilesystemImpl // methods that would require an anonDentry to be a directory return ENOTDIR. +// +// +stateify savable type anonFilesystem struct { vfsfs Filesystem devMinor uint32 } +// +stateify savable type anonDentry struct { vfsd Dentry diff --git a/pkg/sentry/vfs/dentry.go b/pkg/sentry/vfs/dentry.go index a69a5b2f1..320ab7ce1 100644 --- a/pkg/sentry/vfs/dentry.go +++ b/pkg/sentry/vfs/dentry.go @@ -89,6 +89,8 @@ func (d *Dentry) Impl() DentryImpl { // DentryImpl contains implementation details for a Dentry. Implementations of // DentryImpl should contain their associated Dentry by value as their first // field. +// +// +stateify savable type DentryImpl interface { // IncRef increments the Dentry's reference count. A Dentry with a non-zero // reference count must remain coherent with the state of the filesystem. diff --git a/pkg/sentry/vfs/device.go b/pkg/sentry/vfs/device.go index 1e9dffc8f..dde2ad79b 100644 --- a/pkg/sentry/vfs/device.go +++ b/pkg/sentry/vfs/device.go @@ -22,6 +22,8 @@ import ( ) // DeviceKind indicates whether a device is a block or character device. +// +// +stateify savable type DeviceKind uint32 const ( @@ -44,6 +46,7 @@ func (kind DeviceKind) String() string { } } +// +stateify savable type devTuple struct { kind DeviceKind major uint32 diff --git a/pkg/sentry/vfs/epoll.go b/pkg/sentry/vfs/epoll.go index 754e76aec..8f36c3e3b 100644 --- a/pkg/sentry/vfs/epoll.go +++ b/pkg/sentry/vfs/epoll.go @@ -27,6 +27,8 @@ import ( var epollCycleMu sync.Mutex // EpollInstance represents an epoll instance, as described by epoll(7). +// +// +stateify savable type EpollInstance struct { vfsfd FileDescription FileDescriptionDefaultImpl @@ -38,11 +40,11 @@ type EpollInstance struct { // interest is the set of file descriptors that are registered with the // EpollInstance for monitoring. interest is protected by interestMu. - interestMu sync.Mutex + interestMu sync.Mutex `state:"nosave"` interest map[epollInterestKey]*epollInterest // mu protects fields in registered epollInterests. - mu sync.Mutex + mu sync.Mutex `state:"nosave"` // ready is the set of file descriptors that may be "ready" for I/O. Note // that this must be an ordered list, not a map: "If more than maxevents @@ -55,6 +57,7 @@ type EpollInstance struct { ready epollInterestList } +// +stateify savable type epollInterestKey struct { // file is the registered FileDescription. No reference is held on file; // instead, when the last reference is dropped, FileDescription.DecRef() @@ -67,6 +70,8 @@ type epollInterestKey struct { } // epollInterest represents an EpollInstance's interest in a file descriptor. +// +// +stateify savable type epollInterest struct { // epoll is the owning EpollInstance. epoll is immutable. epoll *EpollInstance diff --git a/pkg/sentry/vfs/file_description.go b/pkg/sentry/vfs/file_description.go index 2b29a3c3f..1eba0270f 100644 --- a/pkg/sentry/vfs/file_description.go +++ b/pkg/sentry/vfs/file_description.go @@ -37,11 +37,13 @@ import ( // FileDescription methods require that a reference is held. // // FileDescription is analogous to Linux's struct file. +// +// +stateify savable type FileDescription struct { FileDescriptionRefs // flagsMu protects statusFlags and asyncHandler below. - flagsMu sync.Mutex + flagsMu sync.Mutex `state:"nosave"` // statusFlags contains status flags, "initialized by open(2) and possibly // modified by fcntl()" - fcntl(2). statusFlags can be read using atomic @@ -56,7 +58,7 @@ type FileDescription struct { // epolls is the set of epollInterests registered for this FileDescription. // epolls is protected by epollMu. - epollMu sync.Mutex + epollMu sync.Mutex `state:"nosave"` epolls map[*epollInterest]struct{} // vd is the filesystem location at which this FileDescription was opened. @@ -88,6 +90,8 @@ type FileDescription struct { } // FileDescriptionOptions contains options to FileDescription.Init(). +// +// +stateify savable type FileDescriptionOptions struct { // If AllowDirectIO is true, allow O_DIRECT to be set on the file. AllowDirectIO bool @@ -326,6 +330,9 @@ type FileDescriptionImpl interface { // Allocate grows the file to offset + length bytes. // Only mode == 0 is supported currently. // + // Allocate should return EISDIR on directories, ESPIPE on pipes, and ENODEV on + // other files where it is not supported. + // // Preconditions: The FileDescription was opened for writing. Allocate(ctx context.Context, mode, offset, length uint64) error @@ -448,6 +455,8 @@ type FileDescriptionImpl interface { } // Dirent holds the information contained in struct linux_dirent64. +// +// +stateify savable type Dirent struct { // Name is the filename. Name string diff --git a/pkg/sentry/vfs/file_description_impl_util.go b/pkg/sentry/vfs/file_description_impl_util.go index 68b80a951..48ca9de44 100644 --- a/pkg/sentry/vfs/file_description_impl_util.go +++ b/pkg/sentry/vfs/file_description_impl_util.go @@ -42,6 +42,8 @@ import ( // FileDescriptionDefaultImpl may be embedded by implementations of // FileDescriptionImpl to obtain implementations of many FileDescriptionImpl // methods with default behavior analogous to Linux's. +// +// +stateify savable type FileDescriptionDefaultImpl struct{} // OnClose implements FileDescriptionImpl.OnClose analogously to @@ -57,7 +59,11 @@ func (FileDescriptionDefaultImpl) StatFS(ctx context.Context) (linux.Statfs, err } // Allocate implements FileDescriptionImpl.Allocate analogously to -// fallocate called on regular file, directory or FIFO in Linux. +// fallocate called on an invalid type of file in Linux. +// +// Note that directories can rely on this implementation even though they +// should technically return EISDIR. Allocate should never be called for a +// directory, because it requires a writable fd. func (FileDescriptionDefaultImpl) Allocate(ctx context.Context, mode, offset, length uint64) error { return syserror.ENODEV } @@ -162,6 +168,8 @@ func (FileDescriptionDefaultImpl) RemoveXattr(ctx context.Context, name string) // DirectoryFileDescriptionDefaultImpl may be embedded by implementations of // FileDescriptionImpl that always represent directories to obtain // implementations of non-directory I/O methods that return EISDIR. +// +// +stateify savable type DirectoryFileDescriptionDefaultImpl struct{} // Allocate implements DirectoryFileDescriptionDefaultImpl.Allocate. @@ -192,6 +200,8 @@ func (DirectoryFileDescriptionDefaultImpl) Write(ctx context.Context, src userme // DentryMetadataFileDescriptionImpl may be embedded by implementations of // FileDescriptionImpl for which FileDescriptionOptions.UseDentryMetadata is // true to obtain implementations of Stat and SetStat that panic. +// +// +stateify savable type DentryMetadataFileDescriptionImpl struct{} // Stat implements FileDescriptionImpl.Stat. @@ -206,12 +216,16 @@ func (DentryMetadataFileDescriptionImpl) SetStat(ctx context.Context, opts SetSt // DynamicBytesSource represents a data source for a // DynamicBytesFileDescriptionImpl. +// +// +stateify savable type DynamicBytesSource interface { // Generate writes the file's contents to buf. Generate(ctx context.Context, buf *bytes.Buffer) error } // StaticData implements DynamicBytesSource over a static string. +// +// +stateify savable type StaticData struct { Data string } @@ -238,14 +252,24 @@ type WritableDynamicBytesSource interface { // // DynamicBytesFileDescriptionImpl.SetDataSource() must be called before first // use. +// +// +stateify savable type DynamicBytesFileDescriptionImpl struct { data DynamicBytesSource // immutable - mu sync.Mutex // protects the following fields - buf bytes.Buffer + mu sync.Mutex `state:"nosave"` // protects the following fields + buf bytes.Buffer `state:".([]byte)"` off int64 lastRead int64 // offset at which the last Read, PRead, or Seek ended } +func (fd *DynamicBytesFileDescriptionImpl) saveBuf() []byte { + return fd.buf.Bytes() +} + +func (fd *DynamicBytesFileDescriptionImpl) loadBuf(p []byte) { + fd.buf.Write(p) +} + // SetDataSource must be called exactly once on fd before first use. func (fd *DynamicBytesFileDescriptionImpl) SetDataSource(data DynamicBytesSource) { fd.data = data @@ -378,6 +402,8 @@ func GenericConfigureMMap(fd *FileDescription, m memmap.Mappable, opts *memmap.M // LockFD may be used by most implementations of FileDescriptionImpl.Lock* // functions. Caller must call Init(). +// +// +stateify savable type LockFD struct { locks *FileLocks } @@ -405,6 +431,8 @@ func (fd *LockFD) UnlockBSD(ctx context.Context, uid fslock.UniqueID) error { // NoLockFD implements Lock*/Unlock* portion of FileDescriptionImpl interface // returning ENOLCK. +// +// +stateify savable type NoLockFD struct{} // LockBSD implements vfs.FileDescriptionImpl.LockBSD. diff --git a/pkg/sentry/vfs/filesystem.go b/pkg/sentry/vfs/filesystem.go index 7dae4e7e8..c93d94634 100644 --- a/pkg/sentry/vfs/filesystem.go +++ b/pkg/sentry/vfs/filesystem.go @@ -506,6 +506,8 @@ type FilesystemImpl interface { // PrependPathAtVFSRootError is returned by implementations of // FilesystemImpl.PrependPath() when they encounter the contextual VFS root. +// +// +stateify savable type PrependPathAtVFSRootError struct{} // Error implements error.Error. @@ -516,6 +518,8 @@ func (PrependPathAtVFSRootError) Error() string { // PrependPathAtNonMountRootError is returned by implementations of // FilesystemImpl.PrependPath() when they encounter an independent ancestor // Dentry that is not the Mount root. +// +// +stateify savable type PrependPathAtNonMountRootError struct{} // Error implements error.Error. @@ -526,6 +530,8 @@ func (PrependPathAtNonMountRootError) Error() string { // PrependPathSyntheticError is returned by implementations of // FilesystemImpl.PrependPath() for which prepended names do not represent real // paths. +// +// +stateify savable type PrependPathSyntheticError struct{} // Error implements error.Error. diff --git a/pkg/sentry/vfs/filesystem_type.go b/pkg/sentry/vfs/filesystem_type.go index f2298f7f6..bc19db1d5 100644 --- a/pkg/sentry/vfs/filesystem_type.go +++ b/pkg/sentry/vfs/filesystem_type.go @@ -55,10 +55,13 @@ type registeredFilesystemType struct { // RegisterFilesystemTypeOptions contains options to // VirtualFilesystem.RegisterFilesystem(). +// +// +stateify savable type RegisterFilesystemTypeOptions struct { - // If AllowUserMount is true, allow calls to VirtualFilesystem.MountAt() - // for which MountOptions.InternalMount == false to use this filesystem - // type. + // AllowUserMount determines whether users are allowed to mount a file system + // of this type, i.e. through mount(2). If AllowUserMount is true, allow calls + // to VirtualFilesystem.MountAt() for which MountOptions.InternalMount == false + // to use this filesystem type. AllowUserMount bool // If AllowUserList is true, make this filesystem type visible in diff --git a/pkg/sentry/vfs/genericfstree/genericfstree.go b/pkg/sentry/vfs/genericfstree/genericfstree.go index 8882fa84a..2d27d9d35 100644 --- a/pkg/sentry/vfs/genericfstree/genericfstree.go +++ b/pkg/sentry/vfs/genericfstree/genericfstree.go @@ -27,6 +27,8 @@ import ( ) // Dentry is a required type parameter that is a struct with the given fields. +// +// +stateify savable type Dentry struct { // vfsd is the embedded vfs.Dentry corresponding to this vfs.DentryImpl. vfsd vfs.Dentry diff --git a/pkg/sentry/vfs/inotify.go b/pkg/sentry/vfs/inotify.go index aff220a61..3f0b8f45b 100644 --- a/pkg/sentry/vfs/inotify.go +++ b/pkg/sentry/vfs/inotify.go @@ -37,6 +37,8 @@ const inotifyEventBaseSize = 16 // // The way events are labelled appears somewhat arbitrary, but they must match // Linux so that IN_EXCL_UNLINK behaves as it does in Linux. +// +// +stateify savable type EventType uint8 // PathEvent and InodeEvent correspond to FSNOTIFY_EVENT_PATH and diff --git a/pkg/sentry/vfs/lock.go b/pkg/sentry/vfs/lock.go index 42666eebf..55783d4eb 100644 --- a/pkg/sentry/vfs/lock.go +++ b/pkg/sentry/vfs/lock.go @@ -33,6 +33,8 @@ import ( // Note that in Linux these two types of locks are _not_ cooperative, because // race and deadlock conditions make merging them prohibitive. We do the same // and keep them oblivious to each other. +// +// +stateify savable type FileLocks struct { // bsd is a set of BSD-style advisory file wide locks, see flock(2). bsd fslock.Locks diff --git a/pkg/sentry/vfs/mount.go b/pkg/sentry/vfs/mount.go index 9da09d4c1..dfc3ae6c0 100644 --- a/pkg/sentry/vfs/mount.go +++ b/pkg/sentry/vfs/mount.go @@ -65,7 +65,7 @@ type Mount struct { // // Invariant: key.parent != nil iff key.point != nil. key.point belongs to // key.parent.fs. - key mountKey + key mountKey `state:".(VirtualDentry)"` // ns is the namespace in which this Mount was mounted. ns is protected by // VirtualFilesystem.mountMu. @@ -345,6 +345,7 @@ func (vfs *VirtualFilesystem) UmountAt(ctx context.Context, creds *auth.Credenti return nil } +// +stateify savable type umountRecursiveOptions struct { // If eager is true, ensure that future calls to Mount.tryIncMountedRef() // on umounted mounts fail. @@ -414,7 +415,7 @@ func (vfs *VirtualFilesystem) connectLocked(mnt *Mount, vd VirtualDentry, mntns } } mnt.IncRef() // dropped by callers of umountRecursiveLocked - mnt.storeKey(vd) + mnt.setKey(vd) if vd.mount.children == nil { vd.mount.children = make(map[*Mount]struct{}) } @@ -439,13 +440,13 @@ func (vfs *VirtualFilesystem) connectLocked(mnt *Mount, vd VirtualDentry, mntns // * vfs.mounts.seq must be in a writer critical section. // * mnt.parent() != nil. func (vfs *VirtualFilesystem) disconnectLocked(mnt *Mount) VirtualDentry { - vd := mnt.loadKey() + vd := mnt.getKey() if checkInvariants { if vd.mount != nil { panic("VFS.disconnectLocked called on disconnected mount") } } - mnt.storeKey(VirtualDentry{}) + mnt.loadKey(VirtualDentry{}) delete(vd.mount.children, mnt) atomic.AddUint32(&vd.dentry.mounts, math.MaxUint32) // -1 mnt.ns.mountpoints[vd.dentry]-- diff --git a/pkg/sentry/vfs/mount_test.go b/pkg/sentry/vfs/mount_test.go index 3335e4057..cb8c56bd3 100644 --- a/pkg/sentry/vfs/mount_test.go +++ b/pkg/sentry/vfs/mount_test.go @@ -38,7 +38,7 @@ func TestMountTableInsertLookup(t *testing.T) { mt.Init() mount := &Mount{} - mount.storeKey(VirtualDentry{&Mount{}, &Dentry{}}) + mount.setKey(VirtualDentry{&Mount{}, &Dentry{}}) mt.Insert(mount) if m := mt.Lookup(mount.parent(), mount.point()); m != mount { @@ -79,7 +79,7 @@ const enableComparativeBenchmarks = false func newBenchMount() *Mount { mount := &Mount{} - mount.storeKey(VirtualDentry{&Mount{}, &Dentry{}}) + mount.loadKey(VirtualDentry{&Mount{}, &Dentry{}}) return mount } @@ -94,7 +94,7 @@ func BenchmarkMountTableParallelLookup(b *testing.B) { for i := 0; i < numMounts; i++ { mount := newBenchMount() mt.Insert(mount) - keys = append(keys, mount.loadKey()) + keys = append(keys, mount.saveKey()) } var ready sync.WaitGroup @@ -146,7 +146,7 @@ func BenchmarkMountMapParallelLookup(b *testing.B) { keys := make([]VirtualDentry, 0, numMounts) for i := 0; i < numMounts; i++ { mount := newBenchMount() - key := mount.loadKey() + key := mount.saveKey() ms[key] = mount keys = append(keys, key) } @@ -201,7 +201,7 @@ func BenchmarkMountSyncMapParallelLookup(b *testing.B) { keys := make([]VirtualDentry, 0, numMounts) for i := 0; i < numMounts; i++ { mount := newBenchMount() - key := mount.loadKey() + key := mount.getKey() ms.Store(key, mount) keys = append(keys, key) } @@ -283,7 +283,7 @@ func BenchmarkMountMapNegativeLookup(b *testing.B) { ms := make(map[VirtualDentry]*Mount) for i := 0; i < numMounts; i++ { mount := newBenchMount() - ms[mount.loadKey()] = mount + ms[mount.getKey()] = mount } negkeys := make([]VirtualDentry, 0, numMounts) for i := 0; i < numMounts; i++ { @@ -318,7 +318,7 @@ func BenchmarkMountSyncMapNegativeLookup(b *testing.B) { var ms sync.Map for i := 0; i < numMounts; i++ { mount := newBenchMount() - ms.Store(mount.loadKey(), mount) + ms.Store(mount.saveKey(), mount) } negkeys := make([]VirtualDentry, 0, numMounts) for i := 0; i < numMounts; i++ { @@ -372,7 +372,7 @@ func BenchmarkMountMapInsert(b *testing.B) { b.ResetTimer() for i := range mounts { mount := mounts[i] - ms[mount.loadKey()] = mount + ms[mount.saveKey()] = mount } } @@ -392,7 +392,7 @@ func BenchmarkMountSyncMapInsert(b *testing.B) { b.ResetTimer() for i := range mounts { mount := mounts[i] - ms.Store(mount.loadKey(), mount) + ms.Store(mount.saveKey(), mount) } } @@ -425,13 +425,13 @@ func BenchmarkMountMapRemove(b *testing.B) { ms := make(map[VirtualDentry]*Mount) for i := range mounts { mount := mounts[i] - ms[mount.loadKey()] = mount + ms[mount.saveKey()] = mount } b.ResetTimer() for i := range mounts { mount := mounts[i] - delete(ms, mount.loadKey()) + delete(ms, mount.saveKey()) } } @@ -447,12 +447,12 @@ func BenchmarkMountSyncMapRemove(b *testing.B) { var ms sync.Map for i := range mounts { mount := mounts[i] - ms.Store(mount.loadKey(), mount) + ms.Store(mount.saveKey(), mount) } b.ResetTimer() for i := range mounts { mount := mounts[i] - ms.Delete(mount.loadKey()) + ms.Delete(mount.saveKey()) } } diff --git a/pkg/sentry/vfs/mount_unsafe.go b/pkg/sentry/vfs/mount_unsafe.go index da2a2e9c4..b7d122d22 100644 --- a/pkg/sentry/vfs/mount_unsafe.go +++ b/pkg/sentry/vfs/mount_unsafe.go @@ -34,6 +34,8 @@ import ( // structurally identical to VirtualDentry, but stores its fields as // unsafe.Pointer since mutators synchronize with VFS path traversal using // seqcounts. +// +// This is explicitly not savable. type mountKey struct { parent unsafe.Pointer // *Mount point unsafe.Pointer // *Dentry @@ -47,19 +49,23 @@ func (mnt *Mount) point() *Dentry { return (*Dentry)(atomic.LoadPointer(&mnt.key.point)) } -func (mnt *Mount) loadKey() VirtualDentry { +func (mnt *Mount) getKey() VirtualDentry { return VirtualDentry{ mount: mnt.parent(), dentry: mnt.point(), } } +func (mnt *Mount) saveKey() VirtualDentry { return mnt.getKey() } + // Invariant: mnt.key.parent == nil. vd.Ok(). -func (mnt *Mount) storeKey(vd VirtualDentry) { +func (mnt *Mount) setKey(vd VirtualDentry) { atomic.StorePointer(&mnt.key.parent, unsafe.Pointer(vd.mount)) atomic.StorePointer(&mnt.key.point, unsafe.Pointer(vd.dentry)) } +func (mnt *Mount) loadKey(vd VirtualDentry) { mnt.setKey(vd) } + // mountTable maps (mount parent, mount point) pairs to mounts. It supports // efficient concurrent lookup, even in the presence of concurrent mutators // (provided mutation is sufficiently uncommon). @@ -92,6 +98,7 @@ type mountTable struct { // length and cap in separate uint32s) for ~free. size uint64 + // FIXME(gvisor.dev/issue/1663): Slots need to be saved. slots unsafe.Pointer `state:"nosave"` // []mountSlot; never nil after Init } diff --git a/pkg/sentry/vfs/options.go b/pkg/sentry/vfs/options.go index b33d36cb1..bc79e5ecc 100644 --- a/pkg/sentry/vfs/options.go +++ b/pkg/sentry/vfs/options.go @@ -21,6 +21,8 @@ import ( // GetDentryOptions contains options to VirtualFilesystem.GetDentryAt() and // FilesystemImpl.GetDentryAt(). +// +// +stateify savable type GetDentryOptions struct { // If CheckSearchable is true, FilesystemImpl.GetDentryAt() must check that // the returned Dentry is a directory for which creds has search @@ -30,6 +32,8 @@ type GetDentryOptions struct { // MkdirOptions contains options to VirtualFilesystem.MkdirAt() and // FilesystemImpl.MkdirAt(). +// +// +stateify savable type MkdirOptions struct { // Mode is the file mode bits for the created directory. Mode linux.FileMode @@ -56,6 +60,8 @@ type MkdirOptions struct { // MknodOptions contains options to VirtualFilesystem.MknodAt() and // FilesystemImpl.MknodAt(). +// +// +stateify savable type MknodOptions struct { // Mode is the file type and mode bits for the created file. Mode linux.FileMode @@ -72,6 +78,8 @@ type MknodOptions struct { // 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. +// +// +stateify savable type MountFlags struct { // NoExec is equivalent to MS_NOEXEC. NoExec bool @@ -93,6 +101,8 @@ type MountFlags struct { } // MountOptions contains options to VirtualFilesystem.MountAt(). +// +// +stateify savable type MountOptions struct { // Flags contains flags as specified for mount(2), e.g. MS_NOEXEC. Flags MountFlags @@ -103,13 +113,17 @@ type MountOptions struct { // GetFilesystemOptions contains options to FilesystemType.GetFilesystem(). GetFilesystemOptions GetFilesystemOptions - // If InternalMount is true, allow the use of filesystem types for which - // RegisterFilesystemTypeOptions.AllowUserMount == false. + // InternalMount indicates whether the mount operation is coming from the + // application, i.e. through mount(2). If InternalMount is true, allow the use + // of filesystem types for which RegisterFilesystemTypeOptions.AllowUserMount + // == false. InternalMount bool } // OpenOptions contains options to VirtualFilesystem.OpenAt() and // FilesystemImpl.OpenAt(). +// +// +stateify savable type OpenOptions struct { // Flags contains access mode and flags as specified for open(2). // @@ -135,6 +149,8 @@ type OpenOptions struct { // ReadOptions contains options to FileDescription.PRead(), // FileDescriptionImpl.PRead(), FileDescription.Read(), and // FileDescriptionImpl.Read(). +// +// +stateify savable type ReadOptions struct { // Flags contains flags as specified for preadv2(2). Flags uint32 @@ -142,6 +158,8 @@ type ReadOptions struct { // RenameOptions contains options to VirtualFilesystem.RenameAt() and // FilesystemImpl.RenameAt(). +// +// +stateify savable type RenameOptions struct { // Flags contains flags as specified for renameat2(2). Flags uint32 @@ -153,6 +171,8 @@ type RenameOptions struct { // SetStatOptions contains options to VirtualFilesystem.SetStatAt(), // FilesystemImpl.SetStatAt(), FileDescription.SetStat(), and // FileDescriptionImpl.SetStat(). +// +// +stateify savable type SetStatOptions struct { // Stat is the metadata that should be set. Only fields indicated by // Stat.Mask should be set. @@ -174,6 +194,8 @@ type SetStatOptions struct { // BoundEndpointOptions contains options to VirtualFilesystem.BoundEndpointAt() // and FilesystemImpl.BoundEndpointAt(). +// +// +stateify savable type BoundEndpointOptions struct { // Addr is the path of the file whose socket endpoint is being retrieved. // It is generally irrelevant: most endpoints are stored at a dentry that @@ -193,6 +215,8 @@ type BoundEndpointOptions struct { // GetXattrOptions contains options to VirtualFilesystem.GetXattrAt(), // FilesystemImpl.GetXattrAt(), FileDescription.GetXattr(), and // FileDescriptionImpl.GetXattr(). +// +// +stateify savable type GetXattrOptions struct { // Name is the name of the extended attribute to retrieve. Name string @@ -207,6 +231,8 @@ type GetXattrOptions struct { // SetXattrOptions contains options to VirtualFilesystem.SetXattrAt(), // FilesystemImpl.SetXattrAt(), FileDescription.SetXattr(), and // FileDescriptionImpl.SetXattr(). +// +// +stateify savable type SetXattrOptions struct { // Name is the name of the extended attribute being mutated. Name string @@ -221,6 +247,8 @@ type SetXattrOptions struct { // StatOptions contains options to VirtualFilesystem.StatAt(), // FilesystemImpl.StatAt(), FileDescription.Stat(), and // FileDescriptionImpl.Stat(). +// +// +stateify savable type StatOptions struct { // Mask is the set of fields in the returned Statx that the FilesystemImpl // or FileDescriptionImpl should provide. Bits are as in linux.Statx.Mask. @@ -238,6 +266,8 @@ type StatOptions struct { } // UmountOptions contains options to VirtualFilesystem.UmountAt(). +// +// +stateify savable type UmountOptions struct { // Flags contains flags as specified for umount2(2). Flags uint32 @@ -246,6 +276,8 @@ type UmountOptions struct { // WriteOptions contains options to FileDescription.PWrite(), // FileDescriptionImpl.PWrite(), FileDescription.Write(), and // FileDescriptionImpl.Write(). +// +// +stateify savable type WriteOptions struct { // Flags contains flags as specified for pwritev2(2). Flags uint32 diff --git a/pkg/sentry/vfs/permissions.go b/pkg/sentry/vfs/permissions.go index 00eeb8842..d48520d58 100644 --- a/pkg/sentry/vfs/permissions.go +++ b/pkg/sentry/vfs/permissions.go @@ -26,6 +26,8 @@ import ( ) // AccessTypes is a bitmask of Unix file permissions. +// +// +stateify savable type AccessTypes uint16 // Bits in AccessTypes. diff --git a/pkg/sentry/vfs/resolving_path.go b/pkg/sentry/vfs/resolving_path.go index 3304372d9..e4fd55012 100644 --- a/pkg/sentry/vfs/resolving_path.go +++ b/pkg/sentry/vfs/resolving_path.go @@ -35,6 +35,8 @@ import ( // FilesystemImpl methods. // // ResolvingPath is loosely analogous to Linux's struct nameidata. +// +// +stateify savable type ResolvingPath struct { vfs *VirtualFilesystem root VirtualDentry // refs borrowed from PathOperation @@ -88,6 +90,7 @@ func init() { // so error "constants" are really mutable vars, necessitating somewhat // expensive interface object comparisons. +// +stateify savable type resolveMountRootOrJumpError struct{} // Error implements error.Error. @@ -95,6 +98,7 @@ func (resolveMountRootOrJumpError) Error() string { return "resolving mount root or jump" } +// +stateify savable type resolveMountPointError struct{} // Error implements error.Error. @@ -102,6 +106,7 @@ func (resolveMountPointError) Error() string { return "resolving mount point" } +// +stateify savable type resolveAbsSymlinkError struct{} // Error implements error.Error. diff --git a/pkg/sentry/vfs/vfs.go b/pkg/sentry/vfs/vfs.go index 1ebf355ef..5bd756ea5 100644 --- a/pkg/sentry/vfs/vfs.go +++ b/pkg/sentry/vfs/vfs.go @@ -163,6 +163,8 @@ func (vfs *VirtualFilesystem) Init(ctx context.Context) error { // PathOperation is passed to VFS methods by pointer to reduce memory copying: // it's somewhat large and should never escape. (Options structs are passed by // pointer to VFS and FileDescription methods for the same reason.) +// +// +stateify savable type PathOperation struct { // Root is the VFS root. References on Root are borrowed from the provider // of the PathOperation. |