diff options
Diffstat (limited to 'pkg/sentry')
443 files changed, 27081 insertions, 7101 deletions
diff --git a/pkg/sentry/arch/BUILD b/pkg/sentry/arch/BUILD index e27f21e5e..901e0f320 100644 --- a/pkg/sentry/arch/BUILD +++ b/pkg/sentry/arch/BUILD @@ -11,7 +11,6 @@ go_library( "arch_amd64.go", "arch_amd64.s", "arch_arm64.go", - "arch_state_aarch64.go", "arch_state_x86.go", "arch_x86.go", "arch_x86_impl.go", @@ -26,11 +25,11 @@ go_library( "syscalls_amd64.go", "syscalls_arm64.go", ], + marshal = True, visibility = ["//:sandbox"], deps = [ ":registers_go_proto", "//pkg/abi/linux", - "//pkg/binary", "//pkg/context", "//pkg/cpuid", "//pkg/log", @@ -38,6 +37,7 @@ go_library( "//pkg/sync", "//pkg/syserror", "//pkg/usermem", + "//tools/go_marshal/marshal", ], ) diff --git a/pkg/sentry/arch/arch.go b/pkg/sentry/arch/arch.go index 1d11cc472..a903d031c 100644 --- a/pkg/sentry/arch/arch.go +++ b/pkg/sentry/arch/arch.go @@ -88,6 +88,9 @@ type Context interface { // SyscallNo returns the syscall number. SyscallNo() uintptr + // SyscallSaveOrig save orignal register value. + SyscallSaveOrig() + // SyscallArgs returns the syscall arguments in an array. SyscallArgs() SyscallArguments diff --git a/pkg/sentry/arch/arch_aarch64.go b/pkg/sentry/arch/arch_aarch64.go index 3b6987665..343f81f59 100644 --- a/pkg/sentry/arch/arch_aarch64.go +++ b/pkg/sentry/arch/arch_aarch64.go @@ -17,44 +17,55 @@ package arch import ( + "encoding/binary" "fmt" "io" - "syscall" - "gvisor.dev/gvisor/pkg/binary" + "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/cpuid" "gvisor.dev/gvisor/pkg/log" rpb "gvisor.dev/gvisor/pkg/sentry/arch/registers_go_proto" "gvisor.dev/gvisor/pkg/syserror" - "gvisor.dev/gvisor/pkg/usermem" ) +// Registers represents the CPU registers for this architecture. +type Registers = linux.PtraceRegs + const ( // SyscallWidth is the width of insturctions. SyscallWidth = 4 + + // fpsimdMagic is the magic number which is used in fpsimd_context. + fpsimdMagic = 0x46508001 + + // fpsimdContextSize is the size of fpsimd_context. + fpsimdContextSize = 0x210 ) +// ARMTrapFlag is the mask for the trap flag. +const ARMTrapFlag = uint64(1) << 21 + // aarch64FPState is aarch64 floating point state. type aarch64FPState []byte -// initAarch64FPState (defined in asm files) sets up initial state. -func initAarch64FPState(data *FloatingPointData) { - // TODO(gvisor.dev/issue/1238): floating-point is not supported. +// initAarch64FPState sets up initial state. +func initAarch64FPState(data aarch64FPState) { + binary.LittleEndian.PutUint32(data, fpsimdMagic) + binary.LittleEndian.PutUint32(data[4:], fpsimdContextSize) } func newAarch64FPStateSlice() []byte { - return alignedBytes(4096, 32)[:4096] + return alignedBytes(4096, 16)[:fpsimdContextSize] } // newAarch64FPState returns an initialized floating point state. // // The returned state is large enough to store all floating point state // supported by host, even if the app won't use much of it due to a restricted -// FeatureSet. Since they may still be able to see state not advertised by -// CPUID we must ensure it does not contain any sentry state. +// FeatureSet. func newAarch64FPState() aarch64FPState { f := aarch64FPState(newAarch64FPStateSlice()) - initAarch64FPState(f.FloatingPointData()) + initAarch64FPState(f) return f } @@ -81,13 +92,19 @@ func NewFloatingPointData() *FloatingPointData { // file ensures it's only built on aarch64). type State struct { // The system registers. - Regs syscall.PtraceRegs `state:".(syscallPtraceRegs)"` + Regs Registers // Our floating point state. aarch64FPState `state:"wait"` + // TLS pointer + TPValue uint64 + // FeatureSet is a pointer to the currently active feature set. FeatureSet *cpuid.FeatureSet + + // OrigR0 stores the value of register R0. + OrigR0 uint64 } // Proto returns a protobuf representation of the system registers in State. @@ -133,10 +150,12 @@ func (s State) Proto() *rpb.Registers { // Fork creates and returns an identical copy of the state. func (s *State) Fork() State { - // TODO(gvisor.dev/issue/1238): floating-point is not supported. return State{ - Regs: s.Regs, - FeatureSet: s.FeatureSet, + Regs: s.Regs, + aarch64FPState: s.aarch64FPState.fork(), + TPValue: s.TPValue, + FeatureSet: s.FeatureSet, + OrigR0: s.OrigR0, } } @@ -209,25 +228,27 @@ func (s *State) RegisterMap() (map[string]uintptr, error) { // PtraceGetRegs implements Context.PtraceGetRegs. func (s *State) PtraceGetRegs(dst io.Writer) (int, error) { - return dst.Write(binary.Marshal(nil, usermem.ByteOrder, s.ptraceGetRegs())) + regs := s.ptraceGetRegs() + n, err := regs.WriteTo(dst) + return int(n), err } -func (s *State) ptraceGetRegs() syscall.PtraceRegs { +func (s *State) ptraceGetRegs() Registers { return s.Regs } -var ptraceRegsSize = int(binary.Size(syscall.PtraceRegs{})) +var registersSize = (*Registers)(nil).SizeBytes() // PtraceSetRegs implements Context.PtraceSetRegs. func (s *State) PtraceSetRegs(src io.Reader) (int, error) { - var regs syscall.PtraceRegs - buf := make([]byte, ptraceRegsSize) + var regs Registers + buf := make([]byte, registersSize) if _, err := io.ReadFull(src, buf); err != nil { return 0, err } - binary.Unmarshal(buf, usermem.ByteOrder, ®s) + regs.UnmarshalUnsafe(buf) s.Regs = regs - return ptraceRegsSize, nil + return registersSize, nil } // PtraceGetFPRegs implements Context.PtraceGetFPRegs. @@ -246,13 +267,14 @@ func (s *State) PtraceSetFPRegs(src io.Reader) (int, error) { const ( _NT_PRSTATUS = 1 _NT_PRFPREG = 2 + _NT_ARM_TLS = 0x401 ) // PtraceGetRegSet implements Context.PtraceGetRegSet. func (s *State) PtraceGetRegSet(regset uintptr, dst io.Writer, maxlen int) (int, error) { switch regset { case _NT_PRSTATUS: - if maxlen < ptraceRegsSize { + if maxlen < registersSize { return 0, syserror.EFAULT } return s.PtraceGetRegs(dst) @@ -265,7 +287,7 @@ func (s *State) PtraceGetRegSet(regset uintptr, dst io.Writer, maxlen int) (int, func (s *State) PtraceSetRegSet(regset uintptr, src io.Reader, maxlen int) (int, error) { switch regset { case _NT_PRSTATUS: - if maxlen < ptraceRegsSize { + if maxlen < registersSize { return 0, syserror.EFAULT } return s.PtraceSetRegs(src) @@ -285,8 +307,10 @@ func New(arch Arch, fs *cpuid.FeatureSet) Context { case ARM64: return &context64{ State{ - FeatureSet: fs, + aarch64FPState: newAarch64FPState(), + FeatureSet: fs, }, + []aarch64FPState(nil), } } panic(fmt.Sprintf("unknown architecture %v", arch)) diff --git a/pkg/sentry/arch/arch_amd64.go b/pkg/sentry/arch/arch_amd64.go index 85d6acc0f..3b3a0a272 100644 --- a/pkg/sentry/arch/arch_amd64.go +++ b/pkg/sentry/arch/arch_amd64.go @@ -22,7 +22,6 @@ import ( "math/rand" "syscall" - "gvisor.dev/gvisor/pkg/binary" "gvisor.dev/gvisor/pkg/cpuid" "gvisor.dev/gvisor/pkg/sentry/limits" "gvisor.dev/gvisor/pkg/usermem" @@ -301,8 +300,10 @@ func (c *context64) PtracePeekUser(addr uintptr) (interface{}, error) { // PTRACE_PEEKUSER and PTRACE_POKEUSER are only effective on regs and // u_debugreg, returning 0 or silently no-oping for other fields // respectively. - if addr < uintptr(ptraceRegsSize) { - buf := binary.Marshal(nil, usermem.ByteOrder, c.ptraceGetRegs()) + if addr < uintptr(registersSize) { + regs := c.ptraceGetRegs() + buf := make([]byte, regs.SizeBytes()) + regs.MarshalUnsafe(buf) return c.Native(uintptr(usermem.ByteOrder.Uint64(buf[addr:]))), nil } // Note: x86 debug registers are missing. @@ -314,8 +315,10 @@ func (c *context64) PtracePokeUser(addr, data uintptr) error { if addr&7 != 0 || addr >= userStructSize { return syscall.EIO } - if addr < uintptr(ptraceRegsSize) { - buf := binary.Marshal(nil, usermem.ByteOrder, c.ptraceGetRegs()) + if addr < uintptr(registersSize) { + regs := c.ptraceGetRegs() + buf := make([]byte, regs.SizeBytes()) + regs.MarshalUnsafe(buf) usermem.ByteOrder.PutUint64(buf[addr:], uint64(data)) _, err := c.PtraceSetRegs(bytes.NewBuffer(buf)) return err diff --git a/pkg/sentry/arch/arch_arm64.go b/pkg/sentry/arch/arch_arm64.go index ac98897b5..ada7ac7b8 100644 --- a/pkg/sentry/arch/arch_arm64.go +++ b/pkg/sentry/arch/arch_arm64.go @@ -12,6 +12,8 @@ // See the License for the specific language governing permissions and // limitations under the License. +// +build arm64 + package arch import ( @@ -53,6 +55,11 @@ const ( preferredPIELoadAddr usermem.Addr = maxAddr64 / 6 * 5 ) +var ( + // CPUIDInstruction doesn't exist on ARM64. + CPUIDInstruction = []byte{} +) + // These constants are selected as heuristics to help make the Platform's // potentially limited address space conform as closely to Linux as possible. const ( @@ -68,6 +75,7 @@ const ( // context64 represents an ARM64 context. type context64 struct { State + sigFPState []aarch64FPState // fpstate to be restored on sigreturn. } // Arch implements Context.Arch. @@ -75,10 +83,19 @@ func (c *context64) Arch() Arch { return ARM64 } +func (c *context64) copySigFPState() []aarch64FPState { + var sigfps []aarch64FPState + for _, s := range c.sigFPState { + sigfps = append(sigfps, s.fork()) + } + return sigfps +} + // Fork returns an exact copy of this context. func (c *context64) Fork() Context { return &context64{ - State: c.State.Fork(), + State: c.State.Fork(), + sigFPState: c.copySigFPState(), } } @@ -125,16 +142,17 @@ func (c *context64) SetStack(value uintptr) { // TLS returns the current TLS pointer. func (c *context64) TLS() uintptr { - // TODO(gvisor.dev/issue/1238): TLS is not supported. - // MRS_TPIDR_EL0 - return 0 + return uintptr(c.TPValue) } // SetTLS sets the current TLS pointer. Returns false if value is invalid. func (c *context64) SetTLS(value uintptr) bool { - // TODO(gvisor.dev/issue/1238): TLS is not supported. - // MSR_TPIDR_EL0 - return false + if value >= uintptr(maxAddr64) { + return false + } + + c.TPValue = uint64(value) + return true } // SetOldRSeqInterruptedIP implements Context.SetOldRSeqInterruptedIP. diff --git a/pkg/sentry/arch/arch_state_x86.go b/pkg/sentry/arch/arch_state_x86.go index e35c9214a..19ce99d25 100644 --- a/pkg/sentry/arch/arch_state_x86.go +++ b/pkg/sentry/arch/arch_state_x86.go @@ -12,13 +12,12 @@ // See the License for the specific language governing permissions and // limitations under the License. -// +build amd64 i386 +// +build amd64 386 package arch import ( "fmt" - "syscall" "gvisor.dev/gvisor/pkg/cpuid" "gvisor.dev/gvisor/pkg/usermem" @@ -90,44 +89,3 @@ func (s *State) afterLoadFPState() { // Copy to the new, aligned location. copy(s.x86FPState, old) } - -// +stateify savable -type syscallPtraceRegs struct { - R15 uint64 - R14 uint64 - R13 uint64 - R12 uint64 - Rbp uint64 - Rbx uint64 - R11 uint64 - R10 uint64 - R9 uint64 - R8 uint64 - Rax uint64 - Rcx uint64 - Rdx uint64 - Rsi uint64 - Rdi uint64 - Orig_rax uint64 - Rip uint64 - Cs uint64 - Eflags uint64 - Rsp uint64 - Ss uint64 - Fs_base uint64 - Gs_base uint64 - Ds uint64 - Es uint64 - Fs uint64 - Gs uint64 -} - -// saveRegs is invoked by stateify. -func (s *State) saveRegs() syscallPtraceRegs { - return syscallPtraceRegs(s.Regs) -} - -// loadRegs is invoked by stateify. -func (s *State) loadRegs(r syscallPtraceRegs) { - s.Regs = syscall.PtraceRegs(r) -} diff --git a/pkg/sentry/arch/arch_x86.go b/pkg/sentry/arch/arch_x86.go index 88b40a9d1..dc458b37f 100644 --- a/pkg/sentry/arch/arch_x86.go +++ b/pkg/sentry/arch/arch_x86.go @@ -12,7 +12,7 @@ // See the License for the specific language governing permissions and // limitations under the License. -// +build amd64 i386 +// +build amd64 386 package arch @@ -21,7 +21,7 @@ import ( "io" "syscall" - "gvisor.dev/gvisor/pkg/binary" + "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/cpuid" "gvisor.dev/gvisor/pkg/log" rpb "gvisor.dev/gvisor/pkg/sentry/arch/registers_go_proto" @@ -30,6 +30,9 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) +// Registers represents the CPU registers for this architecture. +type Registers = linux.PtraceRegs + // System-related constants for x86. const ( // SyscallWidth is the width of syscall, sysenter, and int 80 insturctions. @@ -267,10 +270,12 @@ func (s *State) RegisterMap() (map[string]uintptr, error) { // PtraceGetRegs implements Context.PtraceGetRegs. func (s *State) PtraceGetRegs(dst io.Writer) (int, error) { - return dst.Write(binary.Marshal(nil, usermem.ByteOrder, s.ptraceGetRegs())) + regs := s.ptraceGetRegs() + n, err := regs.WriteTo(dst) + return int(n), err } -func (s *State) ptraceGetRegs() syscall.PtraceRegs { +func (s *State) ptraceGetRegs() Registers { regs := s.Regs // These may not be initialized. if regs.Cs == 0 || regs.Ss == 0 || regs.Eflags == 0 { @@ -306,16 +311,16 @@ func (s *State) ptraceGetRegs() syscall.PtraceRegs { return regs } -var ptraceRegsSize = int(binary.Size(syscall.PtraceRegs{})) +var registersSize = (*Registers)(nil).SizeBytes() // PtraceSetRegs implements Context.PtraceSetRegs. func (s *State) PtraceSetRegs(src io.Reader) (int, error) { - var regs syscall.PtraceRegs - buf := make([]byte, ptraceRegsSize) + var regs Registers + buf := make([]byte, registersSize) if _, err := io.ReadFull(src, buf); err != nil { return 0, err } - binary.Unmarshal(buf, usermem.ByteOrder, ®s) + regs.UnmarshalUnsafe(buf) // Truncate segment registers to 16 bits. regs.Cs = uint64(uint16(regs.Cs)) regs.Ds = uint64(uint16(regs.Ds)) @@ -369,7 +374,7 @@ func (s *State) PtraceSetRegs(src io.Reader) (int, error) { } regs.Eflags = (s.Regs.Eflags &^ eflagsPtraceMutable) | (regs.Eflags & eflagsPtraceMutable) s.Regs = regs - return ptraceRegsSize, nil + return registersSize, nil } // isUserSegmentSelector returns true if the given segment selector specifies a @@ -538,7 +543,7 @@ const ( func (s *State) PtraceGetRegSet(regset uintptr, dst io.Writer, maxlen int) (int, error) { switch regset { case _NT_PRSTATUS: - if maxlen < ptraceRegsSize { + if maxlen < registersSize { return 0, syserror.EFAULT } return s.PtraceGetRegs(dst) @@ -558,7 +563,7 @@ func (s *State) PtraceGetRegSet(regset uintptr, dst io.Writer, maxlen int) (int, func (s *State) PtraceSetRegSet(regset uintptr, src io.Reader, maxlen int) (int, error) { switch regset { case _NT_PRSTATUS: - if maxlen < ptraceRegsSize { + if maxlen < registersSize { return 0, syserror.EFAULT } return s.PtraceSetRegs(src) diff --git a/pkg/sentry/arch/arch_x86_impl.go b/pkg/sentry/arch/arch_x86_impl.go index 04ac283c6..0c73fcbfb 100644 --- a/pkg/sentry/arch/arch_x86_impl.go +++ b/pkg/sentry/arch/arch_x86_impl.go @@ -12,13 +12,11 @@ // See the License for the specific language governing permissions and // limitations under the License. -// +build amd64 i386 +// +build amd64 386 package arch import ( - "syscall" - "gvisor.dev/gvisor/pkg/cpuid" ) @@ -28,7 +26,7 @@ import ( // +stateify savable type State struct { // The system registers. - Regs syscall.PtraceRegs `state:".(syscallPtraceRegs)"` + Regs Registers // Our floating point state. x86FPState `state:"wait"` diff --git a/pkg/sentry/arch/signal.go b/pkg/sentry/arch/signal.go index 8b03d0187..c9fb55d00 100644 --- a/pkg/sentry/arch/signal.go +++ b/pkg/sentry/arch/signal.go @@ -22,6 +22,7 @@ import ( // SignalAct represents the action that should be taken when a signal is // delivered, and is equivalent to struct sigaction. // +// +marshal // +stateify savable type SignalAct struct { Handler uint64 @@ -43,6 +44,7 @@ func (s *SignalAct) DeserializeTo(other *SignalAct) { // SignalStack represents information about a user stack, and is equivalent to // stack_t. // +// +marshal // +stateify savable type SignalStack struct { Addr uint64 @@ -64,6 +66,7 @@ func (s *SignalStack) DeserializeTo(other *SignalStack) { // SignalInfo represents information about a signal being delivered, and is // equivalent to struct siginfo in linux kernel(linux/include/uapi/asm-generic/siginfo.h). // +// +marshal // +stateify savable type SignalInfo struct { Signo int32 // Signal number diff --git a/pkg/sentry/arch/signal_act.go b/pkg/sentry/arch/signal_act.go index f9ca2e74e..32173aa20 100644 --- a/pkg/sentry/arch/signal_act.go +++ b/pkg/sentry/arch/signal_act.go @@ -14,6 +14,8 @@ package arch +import "gvisor.dev/gvisor/tools/go_marshal/marshal" + // Special values for SignalAct.Handler. const ( // SignalActDefault is SIG_DFL and specifies that the default behavior for @@ -71,6 +73,8 @@ func (s SignalAct) HasRestorer() bool { // NativeSignalAct is a type that is equivalent to struct sigaction in the // guest architecture. type NativeSignalAct interface { + marshal.Marshallable + // SerializeFrom copies the data in the host SignalAct s into this object. SerializeFrom(s *SignalAct) diff --git a/pkg/sentry/arch/signal_arm64.go b/pkg/sentry/arch/signal_arm64.go index 4f4cc46a8..642c79dda 100644 --- a/pkg/sentry/arch/signal_arm64.go +++ b/pkg/sentry/arch/signal_arm64.go @@ -19,6 +19,7 @@ import ( "syscall" "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/log" "gvisor.dev/gvisor/pkg/usermem" ) @@ -30,14 +31,29 @@ type SignalContext64 struct { Sp uint64 Pc uint64 Pstate uint64 - _pad [8]byte // __attribute__((__aligned__(16))) - Reserved [4096]uint8 + _pad [8]byte // __attribute__((__aligned__(16))) + Fpsimd64 FpsimdContext // size = 528 + Reserved [3568]uint8 +} + +type aarch64Ctx struct { + Magic uint32 + Size uint32 +} + +// FpsimdContext is equivalent to struct fpsimd_context on arm64 +// (arch/arm64/include/uapi/asm/sigcontext.h). +type FpsimdContext struct { + Head aarch64Ctx + Fpsr uint32 + Fpcr uint32 + Vregs [64]uint64 // actually [32]uint128 } // UContext64 is equivalent to ucontext on arm64(arch/arm64/include/uapi/asm/ucontext.h). type UContext64 struct { Flags uint64 - Link *UContext64 + Link uint64 Stack SignalStack Sigset linux.SignalSet // glibc uses a 1024-bit sigset_t @@ -83,9 +99,12 @@ func (c *context64) SignalSetup(st *Stack, act *SignalAct, info *SignalInfo, alt if ucSize < 0 { panic("can't get size of UContext64") } - // st.Arch.Width() is for the restorer address. sizeof(siginfo) == 128. - frameSize := int(st.Arch.Width()) + ucSize + 128 - frameBottom := (sp-usermem.Addr(frameSize)) & ^usermem.Addr(15) - 8 + + // frameSize = ucSize + sizeof(siginfo). + // sizeof(siginfo) == 128. + // R30 stores the restorer address. + frameSize := ucSize + 128 + frameBottom := (sp - usermem.Addr(frameSize)) & ^usermem.Addr(15) sp = frameBottom + usermem.Addr(frameSize) st.Bottom = sp @@ -115,12 +134,48 @@ func (c *context64) SignalSetup(st *Stack, act *SignalAct, info *SignalInfo, alt c.Regs.Regs[0] = uint64(info.Signo) c.Regs.Regs[1] = uint64(infoAddr) c.Regs.Regs[2] = uint64(ucAddr) + c.Regs.Regs[30] = uint64(act.Restorer) + // Save the thread's floating point state. + c.sigFPState = append(c.sigFPState, c.aarch64FPState) + // Signal handler gets a clean floating point state. + c.aarch64FPState = newAarch64FPState() return nil } // SignalRestore implements Context.SignalRestore. -// Only used on intel. func (c *context64) SignalRestore(st *Stack, rt bool) (linux.SignalSet, SignalStack, error) { - return 0, SignalStack{}, nil + // Copy out the stack frame. + var uc UContext64 + if _, err := st.Pop(&uc); err != nil { + return 0, SignalStack{}, err + } + var info SignalInfo + if _, err := st.Pop(&info); err != nil { + return 0, SignalStack{}, err + } + + // Restore registers. + c.Regs.Regs = uc.MContext.Regs + c.Regs.Pc = uc.MContext.Pc + c.Regs.Sp = uc.MContext.Sp + c.Regs.Pstate = uc.MContext.Pstate + + // Restore floating point state. + l := len(c.sigFPState) + if l > 0 { + c.aarch64FPState = c.sigFPState[l-1] + // NOTE(cl/133042258): State save requires that any slice + // elements from '[len:cap]' to be zero value. + c.sigFPState[l-1] = nil + c.sigFPState = c.sigFPState[0 : l-1] + } else { + // This might happen if sigreturn(2) calls are unbalanced with + // respect to signal handler entries. This is not expected so + // don't bother to do anything fancy with the floating point + // state. + log.Warningf("sigreturn unable to restore application fpstate") + } + + return uc.Sigset, uc.Stack, nil } diff --git a/pkg/sentry/arch/signal_stack.go b/pkg/sentry/arch/signal_stack.go index 1a6056171..0fa738a1d 100644 --- a/pkg/sentry/arch/signal_stack.go +++ b/pkg/sentry/arch/signal_stack.go @@ -12,12 +12,13 @@ // See the License for the specific language governing permissions and // limitations under the License. -// +build i386 amd64 arm64 +// +build 386 amd64 arm64 package arch import ( "gvisor.dev/gvisor/pkg/usermem" + "gvisor.dev/gvisor/tools/go_marshal/marshal" ) const ( @@ -55,6 +56,8 @@ func (s *SignalStack) Contains(sp usermem.Addr) bool { // NativeSignalStack is a type that is equivalent to stack_t in the guest // architecture. type NativeSignalStack interface { + marshal.Marshallable + // SerializeFrom copies the data in the host SignalStack s into this // object. SerializeFrom(s *SignalStack) diff --git a/pkg/sentry/arch/stack.go b/pkg/sentry/arch/stack.go index 09bceabc9..1108fa0bd 100644 --- a/pkg/sentry/arch/stack.go +++ b/pkg/sentry/arch/stack.go @@ -97,7 +97,6 @@ func (s *Stack) Push(vals ...interface{}) (usermem.Addr, error) { if c < 0 { return 0, fmt.Errorf("bad binary.Size for %T", v) } - // TODO(b/38173783): Use a real context.Context. n, err := usermem.CopyObjectOut(context.Background(), s.IO, s.Bottom-usermem.Addr(c), norm, usermem.IOOpts{}) if err != nil || c != n { return 0, err @@ -121,11 +120,9 @@ func (s *Stack) Pop(vals ...interface{}) (usermem.Addr, error) { var err error if isVaddr { value := s.Arch.Native(uintptr(0)) - // TODO(b/38173783): Use a real context.Context. n, err = usermem.CopyObjectIn(context.Background(), s.IO, s.Bottom, value, usermem.IOOpts{}) *vaddr = usermem.Addr(s.Arch.Value(value)) } else { - // TODO(b/38173783): Use a real context.Context. n, err = usermem.CopyObjectIn(context.Background(), s.IO, s.Bottom, v, usermem.IOOpts{}) } if err != nil { diff --git a/pkg/sentry/arch/syscalls_amd64.go b/pkg/sentry/arch/syscalls_amd64.go index 8b4f23007..3859f41ee 100644 --- a/pkg/sentry/arch/syscalls_amd64.go +++ b/pkg/sentry/arch/syscalls_amd64.go @@ -18,6 +18,13 @@ package arch const restartSyscallNr = uintptr(219) +// SyscallSaveOrig save the value of the register which is clobbered in +// syscall handler(doSyscall()). +// +// Noop on x86. +func (c *context64) SyscallSaveOrig() { +} + // SyscallNo returns the syscall number according to the 64-bit convention. func (c *context64) SyscallNo() uintptr { return uintptr(c.Regs.Orig_rax) diff --git a/pkg/sentry/arch/syscalls_arm64.go b/pkg/sentry/arch/syscalls_arm64.go index 00d5ef461..95dfd1e90 100644 --- a/pkg/sentry/arch/syscalls_arm64.go +++ b/pkg/sentry/arch/syscalls_arm64.go @@ -18,6 +18,17 @@ package arch const restartSyscallNr = uintptr(128) +// SyscallSaveOrig save the value of the register R0 which is clobbered in +// syscall handler(doSyscall()). +// +// In linux, at the entry of the syscall handler(el0_svc_common()), value of R0 +// is saved to the pt_regs.orig_x0 in kernel code. But currently, the orig_x0 +// was not accessible to the userspace application, so we have to do the same +// operation in the sentry code to save the R0 value into the App context. +func (c *context64) SyscallSaveOrig() { + c.OrigR0 = c.Regs.Regs[0] +} + // SyscallNo returns the syscall number according to the 64-bit convention. func (c *context64) SyscallNo() uintptr { return uintptr(c.Regs.Regs[8]) @@ -40,7 +51,7 @@ func (c *context64) SyscallNo() uintptr { // R30: the link register. func (c *context64) SyscallArgs() SyscallArguments { return SyscallArguments{ - SyscallArgument{Value: uintptr(c.Regs.Regs[0])}, + SyscallArgument{Value: uintptr(c.OrigR0)}, SyscallArgument{Value: uintptr(c.Regs.Regs[1])}, SyscallArgument{Value: uintptr(c.Regs.Regs[2])}, SyscallArgument{Value: uintptr(c.Regs.Regs[3])}, @@ -50,13 +61,21 @@ func (c *context64) SyscallArgs() SyscallArguments { } // RestartSyscall implements Context.RestartSyscall. +// Prepare for system call restart, OrigR0 will be restored to R0. +// Please see the linux code as reference: +// arch/arm64/kernel/signal.c:do_signal() func (c *context64) RestartSyscall() { c.Regs.Pc -= SyscallWidth - c.Regs.Regs[8] = uint64(restartSyscallNr) + // R0 will be backed up into OrigR0 when entering doSyscall(). + // Please see the linux code as reference: + // arch/arm64/kernel/syscall.c:el0_svc_common(). + // Here we restore it back. + c.Regs.Regs[0] = uint64(c.OrigR0) } // RestartSyscallWithRestartBlock implements Context.RestartSyscallWithRestartBlock. func (c *context64) RestartSyscallWithRestartBlock() { c.Regs.Pc -= SyscallWidth + c.Regs.Regs[0] = uint64(c.OrigR0) c.Regs.Regs[8] = uint64(restartSyscallNr) } diff --git a/pkg/sentry/contexttest/contexttest.go b/pkg/sentry/contexttest/contexttest.go index 031fc64ec..8e5658c7a 100644 --- a/pkg/sentry/contexttest/contexttest.go +++ b/pkg/sentry/contexttest/contexttest.go @@ -97,7 +97,7 @@ type hostClock struct { } // Now implements ktime.Clock.Now. -func (hostClock) Now() ktime.Time { +func (*hostClock) Now() ktime.Time { return ktime.FromNanoseconds(time.Now().UnixNano()) } @@ -127,7 +127,7 @@ func (t *TestContext) Value(key interface{}) interface{} { case uniqueid.CtxInotifyCookie: return atomic.AddUint32(&lastInotifyCookie, 1) case ktime.CtxRealtimeClock: - return hostClock{} + return &hostClock{} default: if val, ok := t.otherValues[key]; ok { return val diff --git a/pkg/sentry/control/BUILD b/pkg/sentry/control/BUILD index e69496477..0c9a62f0d 100644 --- a/pkg/sentry/control/BUILD +++ b/pkg/sentry/control/BUILD @@ -16,10 +16,16 @@ go_library( ], deps = [ "//pkg/abi/linux", + "//pkg/context", "//pkg/fd", + "//pkg/fspath", "//pkg/log", + "//pkg/sentry/fdimport", "//pkg/sentry/fs", "//pkg/sentry/fs/host", + "//pkg/sentry/fs/user", + "//pkg/sentry/fsbridge", + "//pkg/sentry/fsimpl/host", "//pkg/sentry/kernel", "//pkg/sentry/kernel/auth", "//pkg/sentry/kernel/time", @@ -27,10 +33,13 @@ go_library( "//pkg/sentry/state", "//pkg/sentry/strace", "//pkg/sentry/usage", + "//pkg/sentry/vfs", "//pkg/sentry/watchdog", "//pkg/sync", + "//pkg/syserror", "//pkg/tcpip/link/sniffer", "//pkg/urpc", + "@org_golang_x_sys//unix:go_default_library", ], ) diff --git a/pkg/sentry/control/pprof.go b/pkg/sentry/control/pprof.go index 151808911..663e51989 100644 --- a/pkg/sentry/control/pprof.go +++ b/pkg/sentry/control/pprof.go @@ -117,9 +117,9 @@ func (p *Profile) HeapProfile(o *ProfileOpts, _ *struct{}) error { return nil } -// Goroutine is an RPC stub which dumps out the stack trace for all running -// goroutines. -func (p *Profile) Goroutine(o *ProfileOpts, _ *struct{}) error { +// GoroutineProfile is an RPC stub which dumps out the stack trace for all +// running goroutines. +func (p *Profile) GoroutineProfile(o *ProfileOpts, _ *struct{}) error { if len(o.FilePayload.Files) < 1 { return errNoOutput } @@ -131,6 +131,34 @@ func (p *Profile) Goroutine(o *ProfileOpts, _ *struct{}) error { return nil } +// BlockProfile is an RPC stub which dumps out the stack trace that led to +// blocking on synchronization primitives. +func (p *Profile) BlockProfile(o *ProfileOpts, _ *struct{}) error { + if len(o.FilePayload.Files) < 1 { + return errNoOutput + } + output := o.FilePayload.Files[0] + defer output.Close() + if err := pprof.Lookup("block").WriteTo(output, 0); err != nil { + return err + } + return nil +} + +// MutexProfile is an RPC stub which dumps out the stack trace of holders of +// contended mutexes. +func (p *Profile) MutexProfile(o *ProfileOpts, _ *struct{}) error { + if len(o.FilePayload.Files) < 1 { + return errNoOutput + } + output := o.FilePayload.Files[0] + defer output.Close() + if err := pprof.Lookup("mutex").WriteTo(output, 0); err != nil { + return err + } + return nil +} + // StartTrace is an RPC stub which starts collection of an execution trace. func (p *Profile) StartTrace(o *ProfileOpts, _ *struct{}) error { if len(o.FilePayload.Files) < 1 { diff --git a/pkg/sentry/control/proc.go b/pkg/sentry/control/proc.go index ced51c66c..8767430b7 100644 --- a/pkg/sentry/control/proc.go +++ b/pkg/sentry/control/proc.go @@ -23,14 +23,23 @@ import ( "text/tabwriter" "time" + "golang.org/x/sys/unix" "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/sentry/fdimport" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fs/host" + "gvisor.dev/gvisor/pkg/sentry/fs/user" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" + hostvfs2 "gvisor.dev/gvisor/pkg/sentry/fsimpl/host" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" "gvisor.dev/gvisor/pkg/sentry/limits" "gvisor.dev/gvisor/pkg/sentry/usage" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/urpc" ) @@ -60,6 +69,12 @@ type ExecArgs struct { // process's MountNamespace. MountNamespace *fs.MountNamespace + // MountNamespaceVFS2 is the mount namespace to execute the new process in. + // A reference on MountNamespace must be held for the lifetime of the + // ExecArgs. If MountNamespace is nil, it will default to the init + // process's MountNamespace. + MountNamespaceVFS2 *vfs.MountNamespace + // WorkingDirectory defines the working directory for the new process. WorkingDirectory string `json:"wd"` @@ -71,15 +86,13 @@ type ExecArgs struct { // the root group if not set explicitly. KGID auth.KGID - // ExtraKGIDs is the list of additional groups to which the user - // belongs. + // ExtraKGIDs is the list of additional groups to which the user belongs. ExtraKGIDs []auth.KGID // Capabilities is the list of capabilities to give to the process. Capabilities *auth.TaskCapabilities - // StdioIsPty indicates that FDs 0, 1, and 2 are connected to a host - // pty FD. + // StdioIsPty indicates that FDs 0, 1, and 2 are connected to a host pty FD. StdioIsPty bool // FilePayload determines the files to give to the new process. @@ -104,7 +117,7 @@ func (args ExecArgs) String() string { // Exec runs a new task. func (proc *Proc) Exec(args *ExecArgs, waitStatus *uint32) error { - newTG, _, _, err := proc.execAsync(args) + newTG, _, _, _, err := proc.execAsync(args) if err != nil { return err } @@ -117,26 +130,18 @@ func (proc *Proc) Exec(args *ExecArgs, waitStatus *uint32) error { // ExecAsync runs a new task, but doesn't wait for it to finish. It is defined // as a function rather than a method to avoid exposing execAsync as an RPC. -func ExecAsync(proc *Proc, args *ExecArgs) (*kernel.ThreadGroup, kernel.ThreadID, *host.TTYFileOperations, error) { +func ExecAsync(proc *Proc, args *ExecArgs) (*kernel.ThreadGroup, kernel.ThreadID, *host.TTYFileOperations, *hostvfs2.TTYFileDescription, error) { return proc.execAsync(args) } // execAsync runs a new task, but doesn't wait for it to finish. It returns the // newly created thread group and its PID. If the stdio FDs are TTYs, then a // TTYFileOperations that wraps the TTY is also returned. -func (proc *Proc) execAsync(args *ExecArgs) (*kernel.ThreadGroup, kernel.ThreadID, *host.TTYFileOperations, error) { +func (proc *Proc) execAsync(args *ExecArgs) (*kernel.ThreadGroup, kernel.ThreadID, *host.TTYFileOperations, *hostvfs2.TTYFileDescription, error) { // Import file descriptors. fdTable := proc.Kernel.NewFDTable() defer fdTable.DecRef() - // No matter what happens, we should close all files in the FilePayload - // before returning. Any files that are imported will be duped. - defer func() { - for _, f := range args.FilePayload.Files { - f.Close() - } - }() - creds := auth.NewUserCredentials( args.KUID, args.KGID, @@ -150,6 +155,7 @@ func (proc *Proc) execAsync(args *ExecArgs) (*kernel.ThreadGroup, kernel.ThreadI Envv: args.Envv, WorkingDirectory: args.WorkingDirectory, MountNamespace: args.MountNamespace, + MountNamespaceVFS2: args.MountNamespaceVFS2, Credentials: creds, FDTable: fdTable, Umask: 0022, @@ -166,81 +172,86 @@ func (proc *Proc) execAsync(args *ExecArgs) (*kernel.ThreadGroup, kernel.ThreadI // be donated to the new process in CreateProcess. initArgs.MountNamespace.IncRef() } + if initArgs.MountNamespaceVFS2 != nil { + // initArgs must hold a reference on MountNamespaceVFS2, which will + // be donated to the new process in CreateProcess. + initArgs.MountNamespaceVFS2.IncRef() + } ctx := initArgs.NewContext(proc.Kernel) if initArgs.Filename == "" { - // Get the full path to the filename from the PATH env variable. - paths := fs.GetPath(initArgs.Envv) - mns := initArgs.MountNamespace - if mns == nil { - mns = proc.Kernel.GlobalInit().Leader().MountNamespace() - } - f, err := mns.ResolveExecutablePath(ctx, initArgs.WorkingDirectory, initArgs.Argv[0], paths) - if err != nil { - return nil, 0, nil, fmt.Errorf("error finding executable %q in PATH %v: %v", initArgs.Argv[0], paths, err) - } - initArgs.Filename = f - } - - mounter := fs.FileOwnerFromContext(ctx) - - var ttyFile *fs.File - for appFD, hostFile := range args.FilePayload.Files { - var appFile *fs.File - - if args.StdioIsPty && appFD < 3 { - // Import the file as a host TTY file. - if ttyFile == nil { - var err error - appFile, err = host.ImportFile(ctx, int(hostFile.Fd()), mounter, true /* isTTY */) - if err != nil { - return nil, 0, nil, err - } - defer appFile.DecRef() - - // Remember this in the TTY file, as we will - // use it for the other stdio FDs. - ttyFile = appFile - } else { - // Re-use the existing TTY file, as all three - // stdio FDs must point to the same fs.File in - // order to share TTY state, specifically the - // foreground process group id. - appFile = ttyFile + if kernel.VFS2Enabled { + // Get the full path to the filename from the PATH env variable. + if initArgs.MountNamespaceVFS2 == nil { + // Set initArgs so that 'ctx' returns the namespace. + // + // MountNamespaceVFS2 adds a reference to the namespace, which is + // transferred to the new process. + initArgs.MountNamespaceVFS2 = proc.Kernel.GlobalInit().Leader().MountNamespaceVFS2() + } + file, err := getExecutableFD(ctx, creds, proc.Kernel.VFS(), initArgs.MountNamespaceVFS2, initArgs.Envv, initArgs.WorkingDirectory, initArgs.Argv[0]) + if err != nil { + return nil, 0, nil, nil, fmt.Errorf("error finding executable %q in environment %v: %v", initArgs.Argv[0], initArgs.Envv, err) } + initArgs.File = fsbridge.NewVFSFile(file) } else { - // Import the file as a regular host file. - var err error - appFile, err = host.ImportFile(ctx, int(hostFile.Fd()), mounter, false /* isTTY */) + if initArgs.MountNamespace == nil { + // Set initArgs so that 'ctx' returns the namespace. + initArgs.MountNamespace = proc.Kernel.GlobalInit().Leader().MountNamespace() + + // initArgs must hold a reference on MountNamespace, which will + // be donated to the new process in CreateProcess. + initArgs.MountNamespace.IncRef() + } + f, err := user.ResolveExecutablePath(ctx, creds, initArgs.MountNamespace, initArgs.Envv, initArgs.WorkingDirectory, initArgs.Argv[0]) if err != nil { - return nil, 0, nil, err + return nil, 0, nil, nil, fmt.Errorf("error finding executable %q in PATH %v: %v", initArgs.Argv[0], initArgs.Envv, err) } - defer appFile.DecRef() + initArgs.Filename = f } + } - // Add the file to the FD map. - if err := fdTable.NewFDAt(ctx, int32(appFD), appFile, kernel.FDFlags{}); err != nil { - return nil, 0, nil, err + fds := make([]int, len(args.FilePayload.Files)) + for i, file := range args.FilePayload.Files { + if kernel.VFS2Enabled { + // Need to dup to remove ownership from os.File. + dup, err := unix.Dup(int(file.Fd())) + if err != nil { + return nil, 0, nil, nil, fmt.Errorf("duplicating payload files: %w", err) + } + fds[i] = dup + } else { + // VFS1 dups the file on import. + fds[i] = int(file.Fd()) + } + } + ttyFile, ttyFileVFS2, err := fdimport.Import(ctx, fdTable, args.StdioIsPty, fds) + if err != nil { + if kernel.VFS2Enabled { + for _, fd := range fds { + unix.Close(fd) + } } + return nil, 0, nil, nil, err } tg, tid, err := proc.Kernel.CreateProcess(initArgs) if err != nil { - return nil, 0, nil, err + return nil, 0, nil, nil, err } - var ttyFileOps *host.TTYFileOperations - if ttyFile != nil { - // Set the foreground process group on the TTY before starting - // the process. - ttyFileOps = ttyFile.FileOperations.(*host.TTYFileOperations) - ttyFileOps.InitForegroundProcessGroup(tg.ProcessGroup()) + // Set the foreground process group on the TTY before starting the process. + switch { + case ttyFile != nil: + ttyFile.InitForegroundProcessGroup(tg.ProcessGroup()) + case ttyFileVFS2 != nil: + ttyFileVFS2.InitForegroundProcessGroup(tg.ProcessGroup()) } // Start the newly created process. proc.Kernel.StartProcess(tg) - return tg, tid, ttyFileOps, nil + return tg, tid, ttyFile, ttyFileVFS2, nil } // PsArgs is the set of arguments to ps. @@ -411,3 +422,31 @@ func ttyName(tty *kernel.TTY) string { } return fmt.Sprintf("pts/%d", tty.Index) } + +// getExecutableFD resolves the given executable name and returns a +// vfs.FileDescription for the executable file. +func getExecutableFD(ctx context.Context, creds *auth.Credentials, vfsObj *vfs.VirtualFilesystem, mns *vfs.MountNamespace, envv []string, wd, name string) (*vfs.FileDescription, error) { + path, err := user.ResolveExecutablePathVFS2(ctx, creds, mns, envv, wd, name) + if err != nil { + return nil, err + } + + root := vfs.RootFromContext(ctx) + defer root.DecRef() + + pop := vfs.PathOperation{ + Root: root, + Start: root, // binPath is absolute, Start can be anything. + Path: fspath.Parse(path), + FollowFinalSymlink: true, + } + opts := &vfs.OpenOptions{ + Flags: linux.O_RDONLY, + FileExec: true, + } + f, err := vfsObj.OpenAt(ctx, creds, &pop, opts) + if err == syserror.ENOENT || err == syserror.EACCES { + return nil, nil + } + return f, err +} diff --git a/pkg/sentry/fdimport/BUILD b/pkg/sentry/fdimport/BUILD new file mode 100644 index 000000000..5e41ceb4e --- /dev/null +++ b/pkg/sentry/fdimport/BUILD @@ -0,0 +1,19 @@ +load("//tools:defs.bzl", "go_library") + +package(licenses = ["notice"]) + +go_library( + name = "fdimport", + srcs = [ + "fdimport.go", + ], + visibility = ["//pkg/sentry:internal"], + deps = [ + "//pkg/context", + "//pkg/sentry/fs", + "//pkg/sentry/fs/host", + "//pkg/sentry/fsimpl/host", + "//pkg/sentry/kernel", + "//pkg/sentry/vfs", + ], +) diff --git a/pkg/sentry/fdimport/fdimport.go b/pkg/sentry/fdimport/fdimport.go new file mode 100644 index 000000000..a4199f9e9 --- /dev/null +++ b/pkg/sentry/fdimport/fdimport.go @@ -0,0 +1,129 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package fdimport + +import ( + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/fs/host" + hostvfs2 "gvisor.dev/gvisor/pkg/sentry/fsimpl/host" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/vfs" +) + +// Import imports a slice of FDs into the given FDTable. If console is true, +// sets up TTY for the first 3 FDs in the slice representing stdin, stdout, +// stderr. Upon success, Import takes ownership of all FDs. +func Import(ctx context.Context, fdTable *kernel.FDTable, console bool, fds []int) (*host.TTYFileOperations, *hostvfs2.TTYFileDescription, error) { + if kernel.VFS2Enabled { + ttyFile, err := importVFS2(ctx, fdTable, console, fds) + return nil, ttyFile, err + } + ttyFile, err := importFS(ctx, fdTable, console, fds) + return ttyFile, nil, err +} + +func importFS(ctx context.Context, fdTable *kernel.FDTable, console bool, fds []int) (*host.TTYFileOperations, error) { + var ttyFile *fs.File + for appFD, hostFD := range fds { + var appFile *fs.File + + if console && appFD < 3 { + // Import the file as a host TTY file. + if ttyFile == nil { + var err error + appFile, err = host.ImportFile(ctx, hostFD, true /* isTTY */) + if err != nil { + return nil, err + } + defer appFile.DecRef() + + // Remember this in the TTY file, as we will + // use it for the other stdio FDs. + ttyFile = appFile + } else { + // Re-use the existing TTY file, as all three + // stdio FDs must point to the same fs.File in + // order to share TTY state, specifically the + // foreground process group id. + appFile = ttyFile + } + } else { + // Import the file as a regular host file. + var err error + appFile, err = host.ImportFile(ctx, hostFD, false /* isTTY */) + if err != nil { + return nil, err + } + defer appFile.DecRef() + } + + // Add the file to the FD map. + if err := fdTable.NewFDAt(ctx, int32(appFD), appFile, kernel.FDFlags{}); err != nil { + return nil, err + } + } + + if ttyFile == nil { + return nil, nil + } + return ttyFile.FileOperations.(*host.TTYFileOperations), nil +} + +func importVFS2(ctx context.Context, fdTable *kernel.FDTable, console bool, stdioFDs []int) (*hostvfs2.TTYFileDescription, error) { + k := kernel.KernelFromContext(ctx) + + var ttyFile *vfs.FileDescription + for appFD, hostFD := range stdioFDs { + var appFile *vfs.FileDescription + + if console && appFD < 3 { + // Import the file as a host TTY file. + if ttyFile == nil { + var err error + appFile, err = hostvfs2.ImportFD(ctx, k.HostMount(), hostFD, true /* isTTY */) + if err != nil { + return nil, err + } + defer appFile.DecRef() + + // Remember this in the TTY file, as we will use it for the other stdio + // FDs. + ttyFile = appFile + } else { + // Re-use the existing TTY file, as all three stdio FDs must point to + // the same fs.File in order to share TTY state, specifically the + // foreground process group id. + appFile = ttyFile + } + } else { + var err error + appFile, err = hostvfs2.ImportFD(ctx, k.HostMount(), hostFD, false /* isTTY */) + if err != nil { + return nil, err + } + defer appFile.DecRef() + } + + if err := fdTable.NewFDAtVFS2(ctx, int32(appFD), appFile, kernel.FDFlags{}); err != nil { + return nil, err + } + } + + if ttyFile == nil { + return nil, nil + } + return ttyFile.Impl().(*hostvfs2.TTYFileDescription), nil +} diff --git a/pkg/sentry/fs/copy_up.go b/pkg/sentry/fs/copy_up.go index b060a12ff..ab1424c95 100644 --- a/pkg/sentry/fs/copy_up.go +++ b/pkg/sentry/fs/copy_up.go @@ -222,8 +222,8 @@ func copyUpLocked(ctx context.Context, parent *Dirent, next *Dirent) error { } childUpper, err := parentUpper.Lookup(ctx, next.name) if err != nil { - log.Warningf("copy up failed to lookup directory: %v", err) - cleanupUpper(ctx, parentUpper, next.name) + werr := fmt.Errorf("copy up failed to lookup directory: %v", err) + cleanupUpper(ctx, parentUpper, next.name, werr) return syserror.EIO } defer childUpper.DecRef() @@ -242,8 +242,8 @@ func copyUpLocked(ctx context.Context, parent *Dirent, next *Dirent) error { } childUpper, err := parentUpper.Lookup(ctx, next.name) if err != nil { - log.Warningf("copy up failed to lookup symlink: %v", err) - cleanupUpper(ctx, parentUpper, next.name) + werr := fmt.Errorf("copy up failed to lookup symlink: %v", err) + cleanupUpper(ctx, parentUpper, next.name, werr) return syserror.EIO } defer childUpper.DecRef() @@ -256,23 +256,23 @@ func copyUpLocked(ctx context.Context, parent *Dirent, next *Dirent) error { // Bring file attributes up to date. This does not include size, which will be // brought up to date with copyContentsLocked. if err := copyAttributesLocked(ctx, childUpperInode, next.Inode.overlay.lower); err != nil { - log.Warningf("copy up failed to copy up attributes: %v", err) - cleanupUpper(ctx, parentUpper, next.name) + werr := fmt.Errorf("copy up failed to copy up attributes: %v", err) + cleanupUpper(ctx, parentUpper, next.name, werr) return syserror.EIO } // Copy the entire file. if err := copyContentsLocked(ctx, childUpperInode, next.Inode.overlay.lower, attrs.Size); err != nil { - log.Warningf("copy up failed to copy up contents: %v", err) - cleanupUpper(ctx, parentUpper, next.name) + werr := fmt.Errorf("copy up failed to copy up contents: %v", err) + cleanupUpper(ctx, parentUpper, next.name, werr) return syserror.EIO } lowerMappable := next.Inode.overlay.lower.Mappable() upperMappable := childUpperInode.Mappable() if lowerMappable != nil && upperMappable == nil { - log.Warningf("copy up failed: cannot ensure memory mapping coherence") - cleanupUpper(ctx, parentUpper, next.name) + werr := fmt.Errorf("copy up failed: cannot ensure memory mapping coherence") + cleanupUpper(ctx, parentUpper, next.name, werr) return syserror.EIO } @@ -324,12 +324,14 @@ func copyUpLocked(ctx context.Context, parent *Dirent, next *Dirent) error { return nil } -// cleanupUpper removes name from parent, and panics if it is unsuccessful. -func cleanupUpper(ctx context.Context, parent *Inode, name string) { +// cleanupUpper is called when copy-up fails. It logs the copy-up error and +// attempts to remove name from parent. If that fails, then it panics. +func cleanupUpper(ctx context.Context, parent *Inode, name string, copyUpErr error) { + log.Warningf(copyUpErr.Error()) if err := parent.InodeOperations.Remove(ctx, parent, name); err != nil { // Unfortunately we don't have much choice. We shouldn't // willingly give the caller access to a nonsense filesystem. - panic(fmt.Sprintf("overlay filesystem is in an inconsistent state: failed to remove %q from upper filesystem: %v", name, err)) + panic(fmt.Sprintf("overlay filesystem is in an inconsistent state: copyUp got error: %v; then cleanup failed to remove %q from upper filesystem: %v.", copyUpErr, name, err)) } } diff --git a/pkg/sentry/fs/dev/BUILD b/pkg/sentry/fs/dev/BUILD index 4c4b7d5cc..9379a4d7b 100644 --- a/pkg/sentry/fs/dev/BUILD +++ b/pkg/sentry/fs/dev/BUILD @@ -9,6 +9,7 @@ go_library( "device.go", "fs.go", "full.go", + "net_tun.go", "null.go", "random.go", "tty.go", @@ -19,15 +20,20 @@ go_library( "//pkg/context", "//pkg/rand", "//pkg/safemem", + "//pkg/sentry/arch", "//pkg/sentry/device", "//pkg/sentry/fs", "//pkg/sentry/fs/fsutil", "//pkg/sentry/fs/ramfs", "//pkg/sentry/fs/tmpfs", + "//pkg/sentry/inet", + "//pkg/sentry/kernel", "//pkg/sentry/memmap", "//pkg/sentry/mm", "//pkg/sentry/pgalloc", + "//pkg/sentry/socket/netstack", "//pkg/syserror", + "//pkg/tcpip/link/tun", "//pkg/usermem", "//pkg/waiter", ], diff --git a/pkg/sentry/fs/dev/dev.go b/pkg/sentry/fs/dev/dev.go index 35bd23991..acbd401a0 100644 --- a/pkg/sentry/fs/dev/dev.go +++ b/pkg/sentry/fs/dev/dev.go @@ -22,6 +22,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fs/ramfs" "gvisor.dev/gvisor/pkg/sentry/fs/tmpfs" + "gvisor.dev/gvisor/pkg/sentry/inet" "gvisor.dev/gvisor/pkg/usermem" ) @@ -66,8 +67,8 @@ func newMemDevice(ctx context.Context, iops fs.InodeOperations, msrc *fs.MountSo }) } -func newDirectory(ctx context.Context, msrc *fs.MountSource) *fs.Inode { - iops := ramfs.NewDir(ctx, nil, fs.RootOwner, fs.FilePermsFromMode(0555)) +func newDirectory(ctx context.Context, contents map[string]*fs.Inode, msrc *fs.MountSource) *fs.Inode { + iops := ramfs.NewDir(ctx, contents, fs.RootOwner, fs.FilePermsFromMode(0555)) return fs.NewInode(ctx, iops, msrc, fs.StableAttr{ DeviceID: devDevice.DeviceID(), InodeID: devDevice.NextIno(), @@ -111,7 +112,7 @@ func New(ctx context.Context, msrc *fs.MountSource) *fs.Inode { // A devpts is typically mounted at /dev/pts to provide // pseudoterminal support. Place an empty directory there for // the devpts to be mounted over. - "pts": newDirectory(ctx, msrc), + "pts": newDirectory(ctx, nil, msrc), // Similarly, applications expect a ptmx device at /dev/ptmx // connected to the terminals provided by /dev/pts/. Rather // than creating a device directly (which requires a hairy @@ -126,6 +127,12 @@ func New(ctx context.Context, msrc *fs.MountSource) *fs.Inode { "tty": newCharacterDevice(ctx, newTTYDevice(ctx, fs.RootOwner, 0666), msrc, ttyDevMajor, ttyDevMinor), } + if isNetTunSupported(inet.StackFromContext(ctx)) { + contents["net"] = newDirectory(ctx, map[string]*fs.Inode{ + "tun": newCharacterDevice(ctx, newNetTunDevice(ctx, fs.RootOwner, 0666), msrc, netTunDevMajor, netTunDevMinor), + }, msrc) + } + iops := ramfs.NewDir(ctx, contents, fs.RootOwner, fs.FilePermsFromMode(0555)) return fs.NewInode(ctx, iops, msrc, fs.StableAttr{ DeviceID: devDevice.DeviceID(), diff --git a/pkg/sentry/fs/dev/net_tun.go b/pkg/sentry/fs/dev/net_tun.go new file mode 100644 index 000000000..dc7ad075a --- /dev/null +++ b/pkg/sentry/fs/dev/net_tun.go @@ -0,0 +1,177 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package dev + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/fs/fsutil" + "gvisor.dev/gvisor/pkg/sentry/inet" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/socket/netstack" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/tcpip/link/tun" + "gvisor.dev/gvisor/pkg/usermem" + "gvisor.dev/gvisor/pkg/waiter" +) + +const ( + netTunDevMajor = 10 + netTunDevMinor = 200 +) + +// +stateify savable +type netTunInodeOperations struct { + fsutil.InodeGenericChecker `state:"nosave"` + fsutil.InodeNoExtendedAttributes `state:"nosave"` + fsutil.InodeNoopAllocate `state:"nosave"` + fsutil.InodeNoopRelease `state:"nosave"` + fsutil.InodeNoopTruncate `state:"nosave"` + fsutil.InodeNoopWriteOut `state:"nosave"` + fsutil.InodeNotDirectory `state:"nosave"` + fsutil.InodeNotMappable `state:"nosave"` + fsutil.InodeNotSocket `state:"nosave"` + fsutil.InodeNotSymlink `state:"nosave"` + fsutil.InodeVirtual `state:"nosave"` + + fsutil.InodeSimpleAttributes +} + +var _ fs.InodeOperations = (*netTunInodeOperations)(nil) + +func newNetTunDevice(ctx context.Context, owner fs.FileOwner, mode linux.FileMode) *netTunInodeOperations { + return &netTunInodeOperations{ + InodeSimpleAttributes: fsutil.NewInodeSimpleAttributes(ctx, owner, fs.FilePermsFromMode(mode), linux.TMPFS_MAGIC), + } +} + +// GetFile implements fs.InodeOperations.GetFile. +func (iops *netTunInodeOperations) GetFile(ctx context.Context, d *fs.Dirent, flags fs.FileFlags) (*fs.File, error) { + return fs.NewFile(ctx, d, flags, &netTunFileOperations{}), nil +} + +// +stateify savable +type netTunFileOperations struct { + fsutil.FileNoSeek `state:"nosave"` + fsutil.FileNoMMap `state:"nosave"` + fsutil.FileNoSplice `state:"nosave"` + fsutil.FileNoopFlush `state:"nosave"` + fsutil.FileNoopFsync `state:"nosave"` + fsutil.FileNotDirReaddir `state:"nosave"` + fsutil.FileUseInodeUnstableAttr `state:"nosave"` + + device tun.Device +} + +var _ fs.FileOperations = (*netTunFileOperations)(nil) + +// Release implements fs.FileOperations.Release. +func (fops *netTunFileOperations) Release() { + fops.device.Release() +} + +// Ioctl implements fs.FileOperations.Ioctl. +func (fops *netTunFileOperations) Ioctl(ctx context.Context, file *fs.File, io usermem.IO, args arch.SyscallArguments) (uintptr, error) { + request := args[1].Uint() + data := args[2].Pointer() + + switch request { + case linux.TUNSETIFF: + t := kernel.TaskFromContext(ctx) + if t == nil { + panic("Ioctl should be called from a task context") + } + if !t.HasCapability(linux.CAP_NET_ADMIN) { + return 0, syserror.EPERM + } + stack, ok := t.NetworkContext().(*netstack.Stack) + if !ok { + return 0, syserror.EINVAL + } + + var req linux.IFReq + if _, err := usermem.CopyObjectIn(ctx, io, data, &req, usermem.IOOpts{ + AddressSpaceActive: true, + }); err != nil { + return 0, err + } + flags := usermem.ByteOrder.Uint16(req.Data[:]) + return 0, fops.device.SetIff(stack.Stack, req.Name(), flags) + + case linux.TUNGETIFF: + var req linux.IFReq + + copy(req.IFName[:], fops.device.Name()) + + // Linux adds IFF_NOFILTER (the same value as IFF_NO_PI unfortunately) when + // there is no sk_filter. See __tun_chr_ioctl() in net/drivers/tun.c. + flags := fops.device.Flags() | linux.IFF_NOFILTER + usermem.ByteOrder.PutUint16(req.Data[:], flags) + + _, err := usermem.CopyObjectOut(ctx, io, data, &req, usermem.IOOpts{ + AddressSpaceActive: true, + }) + return 0, err + + default: + return 0, syserror.ENOTTY + } +} + +// Write implements fs.FileOperations.Write. +func (fops *netTunFileOperations) Write(ctx context.Context, file *fs.File, src usermem.IOSequence, offset int64) (int64, error) { + data := make([]byte, src.NumBytes()) + if _, err := src.CopyIn(ctx, data); err != nil { + return 0, err + } + return fops.device.Write(data) +} + +// Read implements fs.FileOperations.Read. +func (fops *netTunFileOperations) Read(ctx context.Context, file *fs.File, dst usermem.IOSequence, offset int64) (int64, error) { + data, err := fops.device.Read() + if err != nil { + return 0, err + } + n, err := dst.CopyOut(ctx, data) + if n > 0 && n < len(data) { + // Not an error for partial copying. Packet truncated. + err = nil + } + return int64(n), err +} + +// Readiness implements watier.Waitable.Readiness. +func (fops *netTunFileOperations) Readiness(mask waiter.EventMask) waiter.EventMask { + return fops.device.Readiness(mask) +} + +// EventRegister implements watier.Waitable.EventRegister. +func (fops *netTunFileOperations) EventRegister(e *waiter.Entry, mask waiter.EventMask) { + fops.device.EventRegister(e, mask) +} + +// EventUnregister implements watier.Waitable.EventUnregister. +func (fops *netTunFileOperations) EventUnregister(e *waiter.Entry) { + fops.device.EventUnregister(e) +} + +// isNetTunSupported returns whether /dev/net/tun device is supported for s. +func isNetTunSupported(s inet.Stack) bool { + _, ok := s.(*netstack.Stack) + return ok +} diff --git a/pkg/sentry/fs/dirent.go b/pkg/sentry/fs/dirent.go index acab0411a..65be12175 100644 --- a/pkg/sentry/fs/dirent.go +++ b/pkg/sentry/fs/dirent.go @@ -17,7 +17,6 @@ package fs import ( "fmt" "path" - "sort" "sync/atomic" "syscall" @@ -121,9 +120,6 @@ type Dirent struct { // deleted may be set atomically when removed. deleted int32 - // frozen indicates this entry can't walk to unknown nodes. - frozen bool - // mounted is true if Dirent is a mount point, similar to include/linux/dcache.h:DCACHE_MOUNTED. mounted bool @@ -253,8 +249,7 @@ func (d *Dirent) IsNegative() bool { return d.Inode == nil } -// hashChild will hash child into the children list of its new parent d, carrying over -// any "frozen" state from d. +// hashChild will hash child into the children list of its new parent d. // // Returns (*WeakRef, true) if hashing child caused a Dirent to be unhashed. The caller must // validate the returned unhashed weak reference. Common cases: @@ -282,9 +277,6 @@ func (d *Dirent) hashChild(child *Dirent) (*refs.WeakRef, bool) { d.IncRef() } - // Carry over parent's frozen state. - child.frozen = d.frozen - return d.hashChildParentSet(child) } @@ -320,9 +312,9 @@ func (d *Dirent) SyncAll(ctx context.Context) { // There is nothing to sync for a read-only filesystem. if !d.Inode.MountSource.Flags.ReadOnly { - // FIXME(b/34856369): This should be a mount traversal, not a - // Dirent traversal, because some Inodes that need to be synced - // may no longer be reachable by name (after sys_unlink). + // NOTE(b/34856369): This should be a mount traversal, not a Dirent + // traversal, because some Inodes that need to be synced may no longer + // be reachable by name (after sys_unlink). // // Write out metadata, dirty page cached pages, and sync disk/remote // caches. @@ -400,38 +392,6 @@ func (d *Dirent) MountRoot() *Dirent { return mountRoot } -// Freeze prevents this dirent from walking to more nodes. Freeze is applied -// recursively to all children. -// -// If this particular Dirent represents a Virtual node, then Walks and Creates -// may proceed as before. -// -// Freeze can only be called before the application starts running, otherwise -// the root it might be out of sync with the application root if modified by -// sys_chroot. -func (d *Dirent) Freeze() { - d.mu.Lock() - defer d.mu.Unlock() - if d.frozen { - // Already frozen. - return - } - d.frozen = true - - // Take a reference when freezing. - for _, w := range d.children { - if child := w.Get(); child != nil { - // NOTE: We would normally drop the reference here. But - // instead we're hanging on to it. - ch := child.(*Dirent) - ch.Freeze() - } - } - - // Drop all expired weak references. - d.flush() -} - // descendantOf returns true if the receiver dirent is equal to, or a // descendant of, the argument dirent. // @@ -524,11 +484,6 @@ func (d *Dirent) walk(ctx context.Context, root *Dirent, name string, walkMayUnl w.Drop() } - // Are we allowed to do the lookup? - if d.frozen && !d.Inode.IsVirtual() { - return nil, syscall.ENOENT - } - // Slow path: load the InodeOperations into memory. Since this is a hot path and the lookup may be // expensive, if possible release the lock and re-acquire it. if walkMayUnlock { @@ -659,11 +614,6 @@ func (d *Dirent) Create(ctx context.Context, root *Dirent, name string, flags Fi return nil, syscall.EEXIST } - // Are we frozen? - if d.frozen && !d.Inode.IsVirtual() { - return nil, syscall.ENOENT - } - // Try the create. We need to trust the file system to return EEXIST (or something // that will translate to EEXIST) if name already exists. file, err := d.Inode.Create(ctx, d, name, flags, perms) @@ -727,11 +677,6 @@ func (d *Dirent) genericCreate(ctx context.Context, root *Dirent, name string, c return syscall.EEXIST } - // Are we frozen? - if d.frozen && !d.Inode.IsVirtual() { - return syscall.ENOENT - } - // Remove any negative Dirent. We've already asserted above with d.exists // that the only thing remaining here can be a negative Dirent. if w, ok := d.children[name]; ok { @@ -862,49 +807,6 @@ func (d *Dirent) GetDotAttrs(root *Dirent) (DentAttr, DentAttr) { return dot, dot } -// readdirFrozen returns readdir results based solely on the frozen children. -func (d *Dirent) readdirFrozen(root *Dirent, offset int64, dirCtx *DirCtx) (int64, error) { - // Collect attrs for "." and "..". - attrs := make(map[string]DentAttr) - names := []string{".", ".."} - attrs["."], attrs[".."] = d.GetDotAttrs(root) - - // Get info from all children. - d.mu.Lock() - defer d.mu.Unlock() - for name, w := range d.children { - if child := w.Get(); child != nil { - defer child.DecRef() - - // Skip negative children. - if child.(*Dirent).IsNegative() { - continue - } - - sattr := child.(*Dirent).Inode.StableAttr - attrs[name] = DentAttr{ - Type: sattr.Type, - InodeID: sattr.InodeID, - } - names = append(names, name) - } - } - - sort.Strings(names) - - if int(offset) >= len(names) { - return offset, nil - } - names = names[int(offset):] - for _, name := range names { - if err := dirCtx.DirEmit(name, attrs[name]); err != nil { - return offset, err - } - offset++ - } - return offset, nil -} - // DirIterator is an open directory containing directory entries that can be read. type DirIterator interface { // IterateDir emits directory entries by calling dirCtx.EmitDir, beginning @@ -964,10 +866,6 @@ func direntReaddir(ctx context.Context, d *Dirent, it DirIterator, root *Dirent, return offset, nil } - if d.frozen { - return d.readdirFrozen(root, offset, dirCtx) - } - // Collect attrs for "." and "..". dot, dotdot := d.GetDotAttrs(root) @@ -1068,11 +966,6 @@ func (d *Dirent) mount(ctx context.Context, inode *Inode) (newChild *Dirent, err return nil, syserror.EINVAL } - // Are we frozen? - if d.parent.frozen && !d.parent.Inode.IsVirtual() { - return nil, syserror.ENOENT - } - // Dirent that'll replace d. // // Note that NewDirent returns with one reference taken; the reference @@ -1101,11 +994,6 @@ func (d *Dirent) unmount(ctx context.Context, replacement *Dirent) error { return syserror.ENOENT } - // Are we frozen? - if d.parent.frozen && !d.parent.Inode.IsVirtual() { - return syserror.ENOENT - } - // Remount our former child in its place. // // As replacement used to be our child, it must already have the right @@ -1135,11 +1023,6 @@ func (d *Dirent) Remove(ctx context.Context, root *Dirent, name string, dirPath unlock := d.lockDirectory() defer unlock() - // Are we frozen? - if d.frozen && !d.Inode.IsVirtual() { - return syscall.ENOENT - } - // Try to walk to the node. child, err := d.walk(ctx, root, name, false /* may unlock */) if err != nil { @@ -1201,11 +1084,6 @@ func (d *Dirent) RemoveDirectory(ctx context.Context, root *Dirent, name string) unlock := d.lockDirectory() defer unlock() - // Are we frozen? - if d.frozen && !d.Inode.IsVirtual() { - return syscall.ENOENT - } - // Check for dots. if name == "." { // Rejected as the last component by rmdir(2). @@ -1438,8 +1316,8 @@ func lockForRename(oldParent *Dirent, oldName string, newParent *Dirent, newName }, nil } -func checkSticky(ctx context.Context, dir *Dirent, victim *Dirent) error { - uattr, err := dir.Inode.UnstableAttr(ctx) +func (d *Dirent) checkSticky(ctx context.Context, victim *Dirent) error { + uattr, err := d.Inode.UnstableAttr(ctx) if err != nil { return syserror.EPERM } @@ -1465,30 +1343,33 @@ func checkSticky(ctx context.Context, dir *Dirent, victim *Dirent) error { return syserror.EPERM } -// MayDelete determines whether `name`, a child of `dir`, can be deleted or +// MayDelete determines whether `name`, a child of `d`, can be deleted or // renamed by `ctx`. // // Compare Linux kernel fs/namei.c:may_delete. -func MayDelete(ctx context.Context, root, dir *Dirent, name string) error { - if err := dir.Inode.CheckPermission(ctx, PermMask{Write: true, Execute: true}); err != nil { +func (d *Dirent) MayDelete(ctx context.Context, root *Dirent, name string) error { + if err := d.Inode.CheckPermission(ctx, PermMask{Write: true, Execute: true}); err != nil { return err } - victim, err := dir.Walk(ctx, root, name) + unlock := d.lockDirectory() + defer unlock() + + victim, err := d.walk(ctx, root, name, true /* may unlock */) if err != nil { return err } defer victim.DecRef() - return mayDelete(ctx, dir, victim) + return d.mayDelete(ctx, victim) } // mayDelete determines whether `victim`, a child of `dir`, can be deleted or // renamed by `ctx`. // // Preconditions: `dir` is writable and executable by `ctx`. -func mayDelete(ctx context.Context, dir, victim *Dirent) error { - if err := checkSticky(ctx, dir, victim); err != nil { +func (d *Dirent) mayDelete(ctx context.Context, victim *Dirent) error { + if err := d.checkSticky(ctx, victim); err != nil { return err } @@ -1516,15 +1397,6 @@ func Rename(ctx context.Context, root *Dirent, oldParent *Dirent, oldName string return err } - // Are we frozen? - // TODO(jamieliu): Is this the right errno? - if oldParent.frozen && !oldParent.Inode.IsVirtual() { - return syscall.ENOENT - } - if newParent.frozen && !newParent.Inode.IsVirtual() { - return syscall.ENOENT - } - // Do we have general permission to remove from oldParent and // create/replace in newParent? if err := oldParent.Inode.CheckPermission(ctx, PermMask{Write: true, Execute: true}); err != nil { @@ -1542,7 +1414,7 @@ func Rename(ctx context.Context, root *Dirent, oldParent *Dirent, oldName string defer renamed.DecRef() // Check that the renamed dirent is deletable. - if err := mayDelete(ctx, oldParent, renamed); err != nil { + if err := oldParent.mayDelete(ctx, renamed); err != nil { return err } @@ -1580,7 +1452,7 @@ func Rename(ctx context.Context, root *Dirent, oldParent *Dirent, oldName string // across the Rename, so must call DecRef manually (no defer). // Check that we can delete replaced. - if err := mayDelete(ctx, newParent, replaced); err != nil { + if err := newParent.mayDelete(ctx, replaced); err != nil { replaced.DecRef() return err } diff --git a/pkg/sentry/fs/dirent_cache.go b/pkg/sentry/fs/dirent_cache.go index 25514ace4..33de32c69 100644 --- a/pkg/sentry/fs/dirent_cache.go +++ b/pkg/sentry/fs/dirent_cache.go @@ -101,8 +101,6 @@ func (c *DirentCache) remove(d *Dirent) { panic(fmt.Sprintf("trying to remove %v, which is not in the dirent cache", d)) } c.list.Remove(d) - d.SetPrev(nil) - d.SetNext(nil) d.DecRef() c.currentSize-- if c.limit != nil { diff --git a/pkg/sentry/fs/fdpipe/pipe_test.go b/pkg/sentry/fs/fdpipe/pipe_test.go index 5aff0cc95..a0082ecca 100644 --- a/pkg/sentry/fs/fdpipe/pipe_test.go +++ b/pkg/sentry/fs/fdpipe/pipe_test.go @@ -119,7 +119,7 @@ func TestNewPipe(t *testing.T) { continue } if flags := p.flags; test.flags != flags { - t.Errorf("%s: got file flags %s, want %s", test.desc, flags, test.flags) + t.Errorf("%s: got file flags %v, want %v", test.desc, flags, test.flags) continue } if len(test.readAheadBuffer) != len(p.readAheadBuffer) { @@ -136,7 +136,7 @@ func TestNewPipe(t *testing.T) { continue } if !fdnotifier.HasFD(int32(f.FD())) { - t.Errorf("%s: pipe fd %d is not registered for events", test.desc, f.FD) + t.Errorf("%s: pipe fd %d is not registered for events", test.desc, f.FD()) } } } diff --git a/pkg/sentry/fs/file.go b/pkg/sentry/fs/file.go index 78100e448..2a278fbe3 100644 --- a/pkg/sentry/fs/file.go +++ b/pkg/sentry/fs/file.go @@ -44,7 +44,7 @@ var ( RecordWaitTime = false reads = metric.MustCreateNewUint64Metric("/fs/reads", false /* sync */, "Number of file reads.") - readWait = metric.MustCreateNewUint64Metric("/fs/read_wait", false /* sync */, "Time waiting on file reads, in nanoseconds.") + readWait = metric.MustCreateNewUint64NanosecondsMetric("/fs/read_wait", false /* sync */, "Time waiting on file reads, in nanoseconds.") ) // IncrementWait increments the given wait time metric, if enabled. @@ -310,7 +310,6 @@ func (f *File) Writev(ctx context.Context, src usermem.IOSequence) (int64, error if !f.mu.Lock(ctx) { return 0, syserror.ErrInterrupted } - unlockAppendMu := f.Dirent.Inode.lockAppendMu(f.Flags().Append) // Handle append mode. if f.Flags().Append { @@ -355,7 +354,6 @@ func (f *File) Pwritev(ctx context.Context, src usermem.IOSequence, offset int64 // offset." unlockAppendMu := f.Dirent.Inode.lockAppendMu(f.Flags().Append) defer unlockAppendMu() - if f.Flags().Append { if err := f.offsetForAppend(ctx, &offset); err != nil { return 0, err @@ -374,9 +372,10 @@ func (f *File) Pwritev(ctx context.Context, src usermem.IOSequence, offset int64 return f.FileOperations.Write(ctx, f, src, offset) } -// offsetForAppend sets the given offset to the end of the file. +// offsetForAppend atomically sets the given offset to the end of the file. // -// Precondition: the file.Dirent.Inode.appendMu mutex should be held for writing. +// Precondition: the file.Dirent.Inode.appendMu mutex should be held for +// writing. func (f *File) offsetForAppend(ctx context.Context, offset *int64) error { uattr, err := f.Dirent.Inode.UnstableAttr(ctx) if err != nil { @@ -386,7 +385,7 @@ func (f *File) offsetForAppend(ctx context.Context, offset *int64) error { } // Update the offset. - *offset = uattr.Size + atomic.StoreInt64(offset, uattr.Size) return nil } diff --git a/pkg/sentry/fs/file_overlay_test.go b/pkg/sentry/fs/file_overlay_test.go index a76d87e3a..1971cc680 100644 --- a/pkg/sentry/fs/file_overlay_test.go +++ b/pkg/sentry/fs/file_overlay_test.go @@ -175,90 +175,6 @@ func TestReaddirRevalidation(t *testing.T) { } } -// TestReaddirOverlayFrozen tests that calling Readdir on an overlay file with -// a frozen dirent tree does not make Readdir calls to the underlying files. -// This is a regression test for b/114808269. -func TestReaddirOverlayFrozen(t *testing.T) { - ctx := contexttest.Context(t) - - // Create an overlay with two directories, each with two files. - upper := newTestRamfsDir(ctx, []dirContent{{name: "upper-file1"}, {name: "upper-file2"}}, nil) - lower := newTestRamfsDir(ctx, []dirContent{{name: "lower-file1"}, {name: "lower-file2"}}, nil) - overlayInode := fs.NewTestOverlayDir(ctx, upper, lower, false) - - // Set that overlay as the root. - root := fs.NewDirent(ctx, overlayInode, "root") - ctx = &rootContext{ - Context: ctx, - root: root, - } - - // Check that calling Readdir on the root now returns all 4 files (2 - // from each layer in the overlay). - rootFile, err := root.Inode.GetFile(ctx, root, fs.FileFlags{Read: true}) - if err != nil { - t.Fatalf("root.Inode.GetFile failed: %v", err) - } - defer rootFile.DecRef() - ser := &fs.CollectEntriesSerializer{} - if err := rootFile.Readdir(ctx, ser); err != nil { - t.Fatalf("rootFile.Readdir failed: %v", err) - } - if got, want := ser.Order, []string{".", "..", "lower-file1", "lower-file2", "upper-file1", "upper-file2"}; !reflect.DeepEqual(got, want) { - t.Errorf("Readdir got names %v, want %v", got, want) - } - - // Readdir should have been called on upper and lower. - upperDir := upper.InodeOperations.(*dir) - lowerDir := lower.InodeOperations.(*dir) - if !upperDir.ReaddirCalled { - t.Errorf("upperDir.ReaddirCalled got %v, want true", upperDir.ReaddirCalled) - } - if !lowerDir.ReaddirCalled { - t.Errorf("lowerDir.ReaddirCalled got %v, want true", lowerDir.ReaddirCalled) - } - - // Reset. - upperDir.ReaddirCalled = false - lowerDir.ReaddirCalled = false - - // Take references on "upper-file1" and "lower-file1", pinning them in - // the dirent tree. - for _, name := range []string{"upper-file1", "lower-file1"} { - if _, err := root.Walk(ctx, root, name); err != nil { - t.Fatalf("root.Walk(%q) failed: %v", name, err) - } - // Don't drop a reference on the returned dirent so that it - // will stay in the tree. - } - - // Freeze the dirent tree. - root.Freeze() - - // Seek back to the beginning of the file. - if _, err := rootFile.Seek(ctx, fs.SeekSet, 0); err != nil { - t.Fatalf("error seeking to beginning of directory: %v", err) - } - - // Calling Readdir on the root now will return only the pinned - // children. - ser = &fs.CollectEntriesSerializer{} - if err := rootFile.Readdir(ctx, ser); err != nil { - t.Fatalf("rootFile.Readdir failed: %v", err) - } - if got, want := ser.Order, []string{".", "..", "lower-file1", "upper-file1"}; !reflect.DeepEqual(got, want) { - t.Errorf("Readdir got names %v, want %v", got, want) - } - - // Readdir should NOT have been called on upper or lower. - if upperDir.ReaddirCalled { - t.Errorf("upperDir.ReaddirCalled got %v, want false", upperDir.ReaddirCalled) - } - if lowerDir.ReaddirCalled { - t.Errorf("lowerDir.ReaddirCalled got %v, want false", lowerDir.ReaddirCalled) - } -} - type rootContext struct { context.Context root *fs.Dirent diff --git a/pkg/sentry/fs/fs.go b/pkg/sentry/fs/fs.go index bdba6efe5..d2dbff268 100644 --- a/pkg/sentry/fs/fs.go +++ b/pkg/sentry/fs/fs.go @@ -42,9 +42,10 @@ // Dirent.dirMu // Dirent.mu // DirentCache.mu -// Locks in InodeOperations implementations or overlayEntry // Inode.Watches.mu (see `Inotify` for other lock ordering) // MountSource.mu +// Inode.appendMu +// Locks in InodeOperations implementations or overlayEntry // // If multiple Dirent or MountSource locks must be taken, locks in the parent must be // taken before locks in their children. diff --git a/pkg/sentry/fs/fsutil/frame_ref_set.go b/pkg/sentry/fs/fsutil/frame_ref_set.go index 6564fd0c6..dd6f5aba6 100644 --- a/pkg/sentry/fs/fsutil/frame_ref_set.go +++ b/pkg/sentry/fs/fsutil/frame_ref_set.go @@ -18,6 +18,7 @@ import ( "math" "gvisor.dev/gvisor/pkg/sentry/platform" + "gvisor.dev/gvisor/pkg/sentry/usage" ) // FrameRefSetFunctions implements segment.Functions for FrameRefSet. @@ -49,3 +50,42 @@ func (FrameRefSetFunctions) Merge(_ platform.FileRange, val1 uint64, _ platform. func (FrameRefSetFunctions) Split(_ platform.FileRange, val uint64, _ uint64) (uint64, uint64) { return val, val } + +// IncRefAndAccount adds a reference on the range fr. All newly inserted segments +// are accounted as host page cache memory mappings. +func (refs *FrameRefSet) IncRefAndAccount(fr platform.FileRange) { + seg, gap := refs.Find(fr.Start) + for { + switch { + case seg.Ok() && seg.Start() < fr.End: + seg = refs.Isolate(seg, fr) + seg.SetValue(seg.Value() + 1) + seg, gap = seg.NextNonEmpty() + case gap.Ok() && gap.Start() < fr.End: + newRange := gap.Range().Intersect(fr) + usage.MemoryAccounting.Inc(newRange.Length(), usage.Mapped) + seg, gap = refs.InsertWithoutMerging(gap, newRange, 1).NextNonEmpty() + default: + refs.MergeAdjacent(fr) + return + } + } +} + +// DecRefAndAccount removes a reference on the range fr and untracks segments +// that are removed from memory accounting. +func (refs *FrameRefSet) DecRefAndAccount(fr platform.FileRange) { + seg := refs.FindSegment(fr.Start) + + for seg.Ok() && seg.Start() < fr.End { + seg = refs.Isolate(seg, fr) + if old := seg.Value(); old == 1 { + usage.MemoryAccounting.Dec(seg.Range().Length(), usage.Mapped) + seg = refs.Remove(seg).NextSegment() + } else { + seg.SetValue(old - 1) + seg = seg.NextSegment() + } + } + refs.MergeAdjacent(fr) +} diff --git a/pkg/sentry/fs/fsutil/inode.go b/pkg/sentry/fs/fsutil/inode.go index daecc4ffe..1922ff08c 100644 --- a/pkg/sentry/fs/fsutil/inode.go +++ b/pkg/sentry/fs/fsutil/inode.go @@ -259,8 +259,8 @@ func (i *InodeSimpleExtendedAttributes) ListXattr(context.Context, *fs.Inode, ui // RemoveXattr implements fs.InodeOperations.RemoveXattr. func (i *InodeSimpleExtendedAttributes) RemoveXattr(_ context.Context, _ *fs.Inode, name string) error { - i.mu.RLock() - defer i.mu.RUnlock() + i.mu.Lock() + defer i.mu.Unlock() if _, ok := i.xattrs[name]; ok { delete(i.xattrs, name) return nil diff --git a/pkg/sentry/fs/g3doc/.gitignore b/pkg/sentry/fs/g3doc/.gitignore new file mode 100644 index 000000000..2d19fc766 --- /dev/null +++ b/pkg/sentry/fs/g3doc/.gitignore @@ -0,0 +1 @@ +*.html diff --git a/pkg/sentry/fs/g3doc/fuse.md b/pkg/sentry/fs/g3doc/fuse.md new file mode 100644 index 000000000..2ca84dd74 --- /dev/null +++ b/pkg/sentry/fs/g3doc/fuse.md @@ -0,0 +1,263 @@ +# Foreword + +This document describes an on-going project to support FUSE filesystems within +the sentry. This is intended to become the final documentation for this +subsystem, and is therefore written in the past tense. However FUSE support is +currently incomplete and the document will be updated as things progress. + +# FUSE: Filesystem in Userspace + +The sentry supports dispatching filesystem operations to a FUSE server, allowing +FUSE filesystem to be used with a sandbox. + +## Overview + +FUSE has two main components: + +1. A client kernel driver (canonically `fuse.ko` in Linux), which forwards + filesystem operations (usually initiated by syscalls) to the server. + +2. A server, which is a userspace daemon that implements the actual filesystem. + +The sentry implements the client component, which allows a server daemon running +within the sandbox to implement a filesystem within the sandbox. + +A FUSE filesystem is initialized with `mount(2)`, typically with the help of a +utility like `fusermount(1)`. Various mount options exist for establishing +ownership and access permissions on the filesystem, but the most important mount +option is a file descriptor used to establish communication between the client +and server. + +The FUSE device FD is obtained by opening `/dev/fuse`. During regular operation, +the client and server use the FUSE protocol described in `fuse(4)` to service +filesystem operations. See the "Protocol" section below for more information +about this protocol. The core of the sentry support for FUSE is the client-side +implementation of this protocol. + +## FUSE in the Sentry + +The sentry's FUSE client targets VFS2 and has the following components: + +- An implementation of `/dev/fuse`. + +- A VFS2 filesystem for mapping syscalls to FUSE ops. Since we're targeting + VFS2, one point of contention may be the lack of inodes in VFS2. We can + tentatively implement a kernfs-based filesystem to bridge the gap in APIs. + The kernfs base functionality can serve the role of the Linux inode cache + and, the filesystem can map VFS2 syscalls to kernfs inode operations; see + the `kernfs.Inode` interface. + +The FUSE protocol lends itself well to marshaling with `go_marshal`. The various +request and response packets can be defined in the ABI package and converted to +and from the wire format using `go_marshal`. + +### Design Goals + +- While filesystem performance is always important, the sentry's FUSE support + is primarily concerned with compatibility, with performance as a secondary + concern. + +- Avoiding deadlocks from a hung server daemon. + +- Consider the potential for denial of service from a malicious server daemon. + Protecting itself from userspace is already a design goal for the sentry, + but needs additional consideration for FUSE. Normally, an operating system + doesn't rely on userspace to make progress with filesystem operations. Since + this changes with FUSE, it opens up the possibility of creating a chain of + dependencies controlled by userspace, which could affect an entire sandbox. + For example: a FUSE op can block a syscall, which could be holding a + subsystem lock, which can then block another task goroutine. + +### Milestones + +Below are some broad goals to aim for while implementing FUSE in the sentry. +Many FUSE ops can be grouped into broad categories of functionality, and most +ops can be implemented in parallel. + +#### Minimal client that can mount a trivial FUSE filesystem. + +- 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`. + +#### Read-only mount with basic file operations + +- Implement the majority of file, directory and file descriptor FUSE ops. For + this milestone, we can skip uncommon or complex operations like mmap, mknod, + file locking, poll, and extended attributes. We can stub these out along + with any ops that modify the filesystem. The exact list of required ops are + to be determined, but the goal is to mount a real filesystem as read-only, + and be able to read contents from the filesystem in the sentry. + +#### Full read-write support + +- Implement the remaining FUSE ops and decide if we can omit rarely used + operations like ioctl. + +# Appendix + +## FUSE Protocol + +The FUSE protocol is a request-response protocol. All requests are initiated by +the client. The wire-format for the protocol is raw C structs serialized to +memory. + +All FUSE requests begin with the following request header: + +```c +struct fuse_in_header { + uint32_t len; // Length of the request, including this header. + uint32_t opcode; // Requested operation. + uint64_t unique; // A unique identifier for this request. + uint64_t nodeid; // ID of the filesystem object being operated on. + uint32_t uid; // UID of the requesting process. + uint32_t gid; // GID of the requesting process. + uint32_t pid; // PID of the requesting process. + uint32_t padding; +}; +``` + +The request is then followed by a payload specific to the `opcode`. + +All responses begin with this response header: + +```c +struct fuse_out_header { + uint32_t len; // Length of the response, including this header. + int32_t error; // Status of the request, 0 if success. + uint64_t unique; // The unique identifier from the corresponding request. +}; +``` + +The response payload also depends on the request `opcode`. If `error != 0`, the +response payload must be empty. + +### Operations + +The following is a list of all FUSE operations used in `fuse_in_header.opcode` +as of Linux v4.4, and a brief description of their purpose. These are defined in +`uapi/linux/fuse.h`. Many of these have a corresponding request and response +payload struct; `fuse(4)` has details for some of these. We also note how these +operations map to the sentry virtual filesystem. + +#### FUSE meta-operations + +These operations are specific to FUSE and don't have a corresponding action in a +generic filesystem. + +- `FUSE_INIT`: This operation initializes a new FUSE filesystem, and is the + first message sent by the client after mount. This is used for version and + feature negotiation. This is related to `mount(2)`. +- `FUSE_DESTROY`: Teardown a FUSE filesystem, related to `unmount(2)`. +- `FUSE_INTERRUPT`: Interrupts an in-flight operation, specified by the + `fuse_in_header.unique` value provided in the corresponding request header. + The client can send at most one of these per request, and will enter an + uninterruptible wait for a reply. The server is expected to reply promptly. +- `FUSE_FORGET`: A hint to the server that server should evict the indicate + node from any caches. This is wired up to `(struct + super_operations).evict_inode` in Linux, which is in turned hooked as the + inode cache shrinker which is typically triggered by system memory pressure. +- `FUSE_BATCH_FORGET`: Batch version of `FUSE_FORGET`. + +#### Filesystem Syscalls + +These FUSE ops map directly to an equivalent filesystem syscall, or family of +syscalls. The relevant syscalls have a similar name to the operation, unless +otherwise noted. + +Node creation: + +- `FUSE_MKNOD` +- `FUSE_MKDIR` +- `FUSE_CREATE`: This is equivalent to `open(2)` and `creat(2)`, which + atomically creates and opens a node. + +Node attributes and extended attributes: + +- `FUSE_GETATTR` +- `FUSE_SETATTR` +- `FUSE_SETXATTR` +- `FUSE_GETXATTR` +- `FUSE_LISTXATTR` +- `FUSE_REMOVEXATTR` + +Node link manipulation: + +- `FUSE_READLINK` +- `FUSE_LINK` +- `FUSE_SYMLINK` +- `FUSE_UNLINK` + +Directory operations: + +- `FUSE_RMDIR` +- `FUSE_RENAME` +- `FUSE_RENAME2` +- `FUSE_OPENDIR`: `open(2)` for directories. +- `FUSE_RELEASEDIR`: `close(2)` for directories. +- `FUSE_READDIR` +- `FUSE_READDIRPLUS` +- `FUSE_FSYNCDIR`: `fsync(2)` for directories. +- `FUSE_LOOKUP`: Establishes a unique identifier for a FS node. This is + reminiscent of `VirtualFilesystem.GetDentryAt` in that it resolves a path + component to a node. However the returned identifier is opaque to the + client. The server must remember this mapping, as this is how the client + will reference the node in the future. + +File operations: + +- `FUSE_OPEN`: `open(2)` for files. +- `FUSE_RELEASE`: `close(2)` for files. +- `FUSE_FSYNC` +- `FUSE_FALLOCATE` +- `FUSE_SETUPMAPPING`: Creates a memory map on a file for `mmap(2)`. +- `FUSE_REMOVEMAPPING`: Removes a memory map for `munmap(2)`. + +File locking: + +- `FUSE_GETLK` +- `FUSE_SETLK` +- `FUSE_SETLKW` +- `FUSE_COPY_FILE_RANGE` + +File descriptor operations: + +- `FUSE_IOCTL` +- `FUSE_POLL` +- `FUSE_LSEEK` + +Filesystem operations: + +- `FUSE_STATFS` + +#### Permissions + +- `FUSE_ACCESS` is used to check if a node is accessible, as part of many + syscall implementations. Maps to `vfs.FilesystemImpl.AccessAt` in the + sentry. + +#### I/O Operations + +These ops are used to read and write file pages. They're used to implement both +I/O syscalls like `read(2)`, `write(2)` and `mmap(2)`. + +- `FUSE_READ` +- `FUSE_WRITE` + +#### Miscellaneous + +- `FUSE_FLUSH`: Used by the client to indicate when a file descriptor is + closed. Distinct from `FUSE_FSYNC`, which corresponds to an `fsync(2)` + syscall from the user. Maps to `vfs.FileDescriptorImpl.Release` in the + sentry. +- `FUSE_BMAP`: Old address space API for block defrag. Probably not needed. +- `FUSE_NOTIFY_REPLY`: [TODO: what does this do?] + +# References + +- [fuse(4) Linux manual page](https://www.man7.org/linux/man-pages/man4/fuse.4.html) +- [Linux kernel FUSE documentation](https://www.kernel.org/doc/html/latest/filesystems/fuse.html) +- [The reference implementation of the Linux FUSE (Filesystem in Userspace) + interface](https://github.com/libfuse/libfuse) +- [The kernel interface of FUSE](https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/include/uapi/linux/fuse.h) diff --git a/pkg/sentry/fs/gofer/attr.go b/pkg/sentry/fs/gofer/attr.go index 6db4b762d..d481baf77 100644 --- a/pkg/sentry/fs/gofer/attr.go +++ b/pkg/sentry/fs/gofer/attr.go @@ -75,10 +75,18 @@ func owner(mounter fs.FileOwner, valid p9.AttrMask, pattr p9.Attr) fs.FileOwner // task's EUID/EGID. owner := mounter if valid.UID { - owner.UID = auth.KUID(pattr.UID) + if pattr.UID.Ok() { + owner.UID = auth.KUID(pattr.UID) + } else { + owner.UID = auth.KUID(auth.OverflowUID) + } } if valid.GID { - owner.GID = auth.KGID(pattr.GID) + if pattr.GID.Ok() { + owner.GID = auth.KGID(pattr.GID) + } else { + owner.GID = auth.KGID(auth.OverflowGID) + } } return owner } diff --git a/pkg/sentry/fs/gofer/file.go b/pkg/sentry/fs/gofer/file.go index 23296f246..b2fcab127 100644 --- a/pkg/sentry/fs/gofer/file.go +++ b/pkg/sentry/fs/gofer/file.go @@ -37,9 +37,9 @@ var ( opens9P = metric.MustCreateNewUint64Metric("/gofer/opens_9p", false /* sync */, "Number of times a 9P file was opened from a gofer.") opensHost = metric.MustCreateNewUint64Metric("/gofer/opens_host", false /* sync */, "Number of times a host file was opened from a gofer.") reads9P = metric.MustCreateNewUint64Metric("/gofer/reads_9p", false /* sync */, "Number of 9P file reads from a gofer.") - readWait9P = metric.MustCreateNewUint64Metric("/gofer/read_wait_9p", false /* sync */, "Time waiting on 9P file reads from a gofer, in nanoseconds.") + readWait9P = metric.MustCreateNewUint64NanosecondsMetric("/gofer/read_wait_9p", false /* sync */, "Time waiting on 9P file reads from a gofer, in nanoseconds.") readsHost = metric.MustCreateNewUint64Metric("/gofer/reads_host", false /* sync */, "Number of host file reads from a gofer.") - readWaitHost = metric.MustCreateNewUint64Metric("/gofer/read_wait_host", false /* sync */, "Time waiting on host file reads from a gofer, in nanoseconds.") + readWaitHost = metric.MustCreateNewUint64NanosecondsMetric("/gofer/read_wait_host", false /* sync */, "Time waiting on host file reads from a gofer, in nanoseconds.") ) // fileOperations implements fs.FileOperations for a remote file system. diff --git a/pkg/sentry/fs/gofer/file_state.go b/pkg/sentry/fs/gofer/file_state.go index ff96b28ba..edd6576aa 100644 --- a/pkg/sentry/fs/gofer/file_state.go +++ b/pkg/sentry/fs/gofer/file_state.go @@ -34,7 +34,6 @@ func (f *fileOperations) afterLoad() { flags := f.flags flags.Truncate = false - // TODO(b/38173783): Context is not plumbed to save/restore. f.handles, err = f.inodeOperations.fileState.getHandles(context.Background(), flags, f.inodeOperations.cachingInodeOps) if err != nil { return fmt.Errorf("failed to re-open handle: %v", err) diff --git a/pkg/sentry/fs/gofer/fs.go b/pkg/sentry/fs/gofer/fs.go index 9d41fcbdb..8ae2d78d7 100644 --- a/pkg/sentry/fs/gofer/fs.go +++ b/pkg/sentry/fs/gofer/fs.go @@ -60,8 +60,7 @@ const ( limitHostFDTranslationKey = "limit_host_fd_translation" // overlayfsStaleRead if present closes cached readonly file after the first - // write. This is done to workaround a limitation of overlayfs in kernels - // before 4.19 where open FDs are not updated after the file is copied up. + // write. This is done to workaround a limitation of Linux overlayfs. overlayfsStaleRead = "overlayfs_stale_read" ) diff --git a/pkg/sentry/fs/gofer/handles.go b/pkg/sentry/fs/gofer/handles.go index 9f7c3e89f..fc14249be 100644 --- a/pkg/sentry/fs/gofer/handles.go +++ b/pkg/sentry/fs/gofer/handles.go @@ -57,7 +57,6 @@ func (h *handles) DecRef() { } } } - // FIXME(b/38173783): Context is not plumbed here. if err := h.File.close(context.Background()); err != nil { log.Warningf("error closing p9 file: %v", err) } diff --git a/pkg/sentry/fs/gofer/inode.go b/pkg/sentry/fs/gofer/inode.go index 1c934981b..a016c896e 100644 --- a/pkg/sentry/fs/gofer/inode.go +++ b/pkg/sentry/fs/gofer/inode.go @@ -273,7 +273,7 @@ func (i *inodeFileState) recreateReadHandles(ctx context.Context, writer *handle // operations on the old will see the new data. Then, make the new handle take // ownereship of the old FD and mark the old readHandle to not close the FD // when done. - if err := syscall.Dup3(h.Host.FD(), i.readHandles.Host.FD(), 0); err != nil { + if err := syscall.Dup3(h.Host.FD(), i.readHandles.Host.FD(), syscall.O_CLOEXEC); err != nil { return err } @@ -710,13 +710,10 @@ func init() { } // AddLink implements InodeOperations.AddLink, but is currently a noop. -// FIXME(b/63117438): Remove this from InodeOperations altogether. func (*inodeOperations) AddLink() {} // DropLink implements InodeOperations.DropLink, but is currently a noop. -// FIXME(b/63117438): Remove this from InodeOperations altogether. func (*inodeOperations) DropLink() {} // NotifyStatusChange implements fs.InodeOperations.NotifyStatusChange. -// FIXME(b/63117438): Remove this from InodeOperations altogether. func (i *inodeOperations) NotifyStatusChange(ctx context.Context) {} diff --git a/pkg/sentry/fs/gofer/inode_state.go b/pkg/sentry/fs/gofer/inode_state.go index 238f7804c..a3402e343 100644 --- a/pkg/sentry/fs/gofer/inode_state.go +++ b/pkg/sentry/fs/gofer/inode_state.go @@ -123,7 +123,6 @@ func (i *inodeFileState) afterLoad() { // beforeSave. return fmt.Errorf("failed to find path for inode number %d. Device %s contains %s", i.sattr.InodeID, i.s.connID, fs.InodeMappings(i.s.inodeMappings)) } - // TODO(b/38173783): Context is not plumbed to save/restore. ctx := &dummyClockContext{context.Background()} _, i.file, err = i.s.attach.walk(ctx, splitAbsolutePath(name)) diff --git a/pkg/sentry/fs/gofer/path.go b/pkg/sentry/fs/gofer/path.go index a35c3a23d..cf9800100 100644 --- a/pkg/sentry/fs/gofer/path.go +++ b/pkg/sentry/fs/gofer/path.go @@ -16,7 +16,6 @@ package gofer import ( "fmt" - "syscall" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/log" @@ -68,7 +67,7 @@ func (i *inodeOperations) Lookup(ctx context.Context, dir *fs.Inode, name string // Get a p9.File for name. qids, newFile, mask, p9attr, err := i.fileState.file.walkGetAttr(ctx, []string{name}) if err != nil { - if err == syscall.ENOENT { + if err == syserror.ENOENT { if cp.cacheNegativeDirents() { // Return a negative Dirent. It will stay cached until something // is created over it. @@ -207,7 +206,7 @@ func (i *inodeOperations) CreateHardLink(ctx context.Context, inode *fs.Inode, t targetOpts, ok := target.InodeOperations.(*inodeOperations) if !ok { - return syscall.EXDEV + return syserror.EXDEV } if err := i.fileState.file.link(ctx, &targetOpts.fileState.file, newName); err != nil { @@ -251,7 +250,7 @@ func (i *inodeOperations) Bind(ctx context.Context, dir *fs.Inode, name string, } if i.session().overrides == nil { - return nil, syscall.EOPNOTSUPP + return nil, syserror.EOPNOTSUPP } // Stabilize the override map while creation is in progress. @@ -280,7 +279,7 @@ func (i *inodeOperations) CreateFifo(ctx context.Context, dir *fs.Inode, name st // N.B. FIFOs use major/minor numbers 0. if _, err := i.fileState.file.mknod(ctx, name, mode, 0, 0, p9.UID(owner.UID), p9.GID(owner.GID)); err != nil { - if i.session().overrides == nil || err != syscall.EPERM { + if i.session().overrides == nil || err != syserror.EPERM { return err } // If gofer doesn't support mknod, check if we can create an internal fifo. @@ -427,17 +426,16 @@ func (i *inodeOperations) Rename(ctx context.Context, inode *fs.Inode, oldParent return syserror.ENAMETOOLONG } - // Unwrap the new parent to a *inodeOperations. - newParentInodeOperations, ok := newParent.InodeOperations.(*inodeOperations) - if !ok { - return syscall.EXDEV + // Don't allow renames across different mounts. + if newParent.MountSource != oldParent.MountSource { + return syserror.EXDEV } + // Unwrap the new parent to a *inodeOperations. + newParentInodeOperations := newParent.InodeOperations.(*inodeOperations) + // Unwrap the old parent to a *inodeOperations. - oldParentInodeOperations, ok := oldParent.InodeOperations.(*inodeOperations) - if !ok { - return syscall.EXDEV - } + oldParentInodeOperations := oldParent.InodeOperations.(*inodeOperations) // Do the rename. if err := i.fileState.file.rename(ctx, newParentInodeOperations.fileState.file, newName); err != nil { diff --git a/pkg/sentry/fs/gofer/session.go b/pkg/sentry/fs/gofer/session.go index f6b3ef178..b5efc86f2 100644 --- a/pkg/sentry/fs/gofer/session.go +++ b/pkg/sentry/fs/gofer/session.go @@ -190,9 +190,9 @@ type session struct { // be socket/pipe files. This allows unix domain sockets and named pipes to // be used with paths that belong to a gofer. // - // TODO(gvisor.dev/issue/1200): there are few possible races with someone - // stat'ing the file and another deleting it concurrently, where the file - // will not be reported as socket file. + // There are a few possible races with someone stat'ing the file and another + // deleting it concurrently, where the file will not be reported as socket + // file. overrides *overrideMaps `state:"wait"` } diff --git a/pkg/sentry/fs/gofer/session_state.go b/pkg/sentry/fs/gofer/session_state.go index 111da59f9..2d398b753 100644 --- a/pkg/sentry/fs/gofer/session_state.go +++ b/pkg/sentry/fs/gofer/session_state.go @@ -104,7 +104,6 @@ func (s *session) afterLoad() { // If private unix sockets are enabled, create and fill the session's endpoint // maps. if opts.privateunixsocket { - // TODO(b/38173783): Context is not plumbed to save/restore. ctx := &dummyClockContext{context.Background()} if err = s.restoreEndpointMaps(ctx); err != nil { diff --git a/pkg/sentry/fs/gofer/socket.go b/pkg/sentry/fs/gofer/socket.go index 10ba2f5f0..40f2c1cad 100644 --- a/pkg/sentry/fs/gofer/socket.go +++ b/pkg/sentry/fs/gofer/socket.go @@ -47,6 +47,8 @@ func (i *inodeOperations) BoundEndpoint(inode *fs.Inode, path string) transport. return &endpoint{inode, i.fileState.file.file, path} } +// LINT.IfChange + // endpoint is a Gofer-backed transport.BoundEndpoint. // // An endpoint's lifetime is the time between when InodeOperations.BoundEndpoint() @@ -146,3 +148,5 @@ func (e *endpoint) Release() { func (e *endpoint) Passcred() bool { return false } + +// LINT.ThenChange(../../fsimpl/gofer/socket.go) diff --git a/pkg/sentry/fs/gofer/util.go b/pkg/sentry/fs/gofer/util.go index 2d8d3a2ea..47a6c69bf 100644 --- a/pkg/sentry/fs/gofer/util.go +++ b/pkg/sentry/fs/gofer/util.go @@ -20,17 +20,29 @@ import ( "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/p9" "gvisor.dev/gvisor/pkg/sentry/fs" + ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" ) func utimes(ctx context.Context, file contextFile, ts fs.TimeSpec) error { if ts.ATimeOmit && ts.MTimeOmit { return nil } + + // Replace requests to use the "system time" with the current time to + // ensure that timestamps remain consistent with the remote + // filesystem. + now := ktime.NowFromContext(ctx) + if ts.ATimeSetSystemTime { + ts.ATime = now + } + if ts.MTimeSetSystemTime { + ts.MTime = now + } mask := p9.SetAttrMask{ ATime: !ts.ATimeOmit, - ATimeNotSystemTime: !ts.ATimeSetSystemTime, + ATimeNotSystemTime: true, MTime: !ts.MTimeOmit, - MTimeNotSystemTime: !ts.MTimeSetSystemTime, + MTimeNotSystemTime: true, } as, ans := ts.ATime.Unix() ms, mns := ts.MTime.Unix() diff --git a/pkg/sentry/fs/host/BUILD b/pkg/sentry/fs/host/BUILD index 21003ea45..aabce6cc9 100644 --- a/pkg/sentry/fs/host/BUILD +++ b/pkg/sentry/fs/host/BUILD @@ -10,7 +10,7 @@ go_library( "descriptor_state.go", "device.go", "file.go", - "fs.go", + "host.go", "inode.go", "inode_state.go", "ioctl_unsafe.go", @@ -62,18 +62,15 @@ go_test( size = "small", srcs = [ "descriptor_test.go", - "fs_test.go", "inode_test.go", "socket_test.go", "wait_test.go", ], library = ":host", deps = [ - "//pkg/context", "//pkg/fd", "//pkg/fdnotifier", "//pkg/sentry/contexttest", - "//pkg/sentry/fs", "//pkg/sentry/kernel/time", "//pkg/sentry/socket", "//pkg/sentry/socket/unix/transport", diff --git a/pkg/sentry/fs/host/control.go b/pkg/sentry/fs/host/control.go index 1658979fc..39299b7e4 100644 --- a/pkg/sentry/fs/host/control.go +++ b/pkg/sentry/fs/host/control.go @@ -23,6 +23,8 @@ import ( "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" ) +// LINT.IfChange + type scmRights struct { fds []int } @@ -76,7 +78,7 @@ func fdsToFiles(ctx context.Context, fds []int) []*fs.File { } // Create the file backed by hostFD. - file, err := NewFile(ctx, fd, fs.FileOwnerFromContext(ctx)) + file, err := NewFile(ctx, fd) if err != nil { ctx.Warningf("Error creating file from host FD: %v", err) break @@ -91,3 +93,5 @@ func fdsToFiles(ctx context.Context, fds []int) []*fs.File { } return files } + +// LINT.ThenChange(../../fsimpl/host/control.go) diff --git a/pkg/sentry/fs/host/descriptor.go b/pkg/sentry/fs/host/descriptor.go index 2a4d1b291..cfdce6a74 100644 --- a/pkg/sentry/fs/host/descriptor.go +++ b/pkg/sentry/fs/host/descriptor.go @@ -16,7 +16,6 @@ package host import ( "fmt" - "path" "syscall" "gvisor.dev/gvisor/pkg/fdnotifier" @@ -28,12 +27,9 @@ import ( // // +stateify savable type descriptor struct { - // donated is true if the host fd was donated by another process. - donated bool - // If origFD >= 0, it is the host fd that this file was originally created // from, which must be available at time of restore. The FD can be closed - // after descriptor is created. Only set if donated is true. + // after descriptor is created. origFD int // wouldBlock is true if value (below) points to a file that can @@ -41,15 +37,13 @@ type descriptor struct { wouldBlock bool // value is the wrapped host fd. It is never saved or restored - // directly. How it is restored depends on whether it was - // donated and the fs.MountSource it was originally - // opened/created from. + // directly. value int `state:"nosave"` } // newDescriptor returns a wrapped host file descriptor. On success, // the descriptor is registered for event notifications with queue. -func newDescriptor(fd int, donated bool, saveable bool, wouldBlock bool, queue *waiter.Queue) (*descriptor, error) { +func newDescriptor(fd int, saveable bool, wouldBlock bool, queue *waiter.Queue) (*descriptor, error) { ownedFD := fd origFD := -1 if saveable { @@ -69,7 +63,6 @@ func newDescriptor(fd int, donated bool, saveable bool, wouldBlock bool, queue * } } return &descriptor{ - donated: donated, origFD: origFD, wouldBlock: wouldBlock, value: ownedFD, @@ -77,25 +70,11 @@ func newDescriptor(fd int, donated bool, saveable bool, wouldBlock bool, queue * } // initAfterLoad initializes the value of the descriptor after Load. -func (d *descriptor) initAfterLoad(mo *superOperations, id uint64, queue *waiter.Queue) error { - if d.donated { - var err error - d.value, err = syscall.Dup(d.origFD) - if err != nil { - return fmt.Errorf("failed to dup restored fd %d: %v", d.origFD, err) - } - } else { - name, ok := mo.inodeMappings[id] - if !ok { - return fmt.Errorf("failed to find path for inode number %d", id) - } - fullpath := path.Join(mo.root, name) - - var err error - d.value, err = open(nil, fullpath) - if err != nil { - return fmt.Errorf("failed to open %q: %v", fullpath, err) - } +func (d *descriptor) initAfterLoad(id uint64, queue *waiter.Queue) error { + var err error + d.value, err = syscall.Dup(d.origFD) + if err != nil { + return fmt.Errorf("failed to dup restored fd %d: %v", d.origFD, err) } if d.wouldBlock { if err := syscall.SetNonblock(d.value, true); err != nil { diff --git a/pkg/sentry/fs/host/descriptor_state.go b/pkg/sentry/fs/host/descriptor_state.go index 8167390a9..e880582ab 100644 --- a/pkg/sentry/fs/host/descriptor_state.go +++ b/pkg/sentry/fs/host/descriptor_state.go @@ -16,7 +16,7 @@ package host // beforeSave is invoked by stateify. func (d *descriptor) beforeSave() { - if d.donated && d.origFD < 0 { + if d.origFD < 0 { panic("donated file descriptor cannot be saved") } } diff --git a/pkg/sentry/fs/host/descriptor_test.go b/pkg/sentry/fs/host/descriptor_test.go index 4205981f5..d8e4605b6 100644 --- a/pkg/sentry/fs/host/descriptor_test.go +++ b/pkg/sentry/fs/host/descriptor_test.go @@ -47,10 +47,10 @@ func TestDescriptorRelease(t *testing.T) { // FD ownership is transferred to the descritor. queue := &waiter.Queue{} - d, err := newDescriptor(fd, false /* donated*/, tc.saveable, tc.wouldBlock, queue) + d, err := newDescriptor(fd, tc.saveable, tc.wouldBlock, queue) if err != nil { syscall.Close(fd) - t.Fatalf("newDescriptor(%d, %t, false, %t, queue) failed, err: %v", fd, tc.saveable, tc.wouldBlock, err) + t.Fatalf("newDescriptor(%d, %t, %t, queue) failed, err: %v", fd, tc.saveable, tc.wouldBlock, err) } if tc.saveable { if d.origFD < 0 { diff --git a/pkg/sentry/fs/host/file.go b/pkg/sentry/fs/host/file.go index e08f56d04..3e48b8b2c 100644 --- a/pkg/sentry/fs/host/file.go +++ b/pkg/sentry/fs/host/file.go @@ -60,8 +60,8 @@ var _ fs.FileOperations = (*fileOperations)(nil) // The returned File cannot be saved, since there is no guarantee that the same // FD will exist or represent the same file at time of restore. If such a // guarantee does exist, use ImportFile instead. -func NewFile(ctx context.Context, fd int, mounter fs.FileOwner) (*fs.File, error) { - return newFileFromDonatedFD(ctx, fd, mounter, false, false) +func NewFile(ctx context.Context, fd int) (*fs.File, error) { + return newFileFromDonatedFD(ctx, fd, false, false) } // ImportFile creates a new File backed by the provided host file descriptor. @@ -71,13 +71,13 @@ func NewFile(ctx context.Context, fd int, mounter fs.FileOwner) (*fs.File, error // If the returned file is saved, it will be restored by re-importing the FD // originally passed to ImportFile. It is the restorer's responsibility to // ensure that the FD represents the same file. -func ImportFile(ctx context.Context, fd int, mounter fs.FileOwner, isTTY bool) (*fs.File, error) { - return newFileFromDonatedFD(ctx, fd, mounter, true, isTTY) +func ImportFile(ctx context.Context, fd int, isTTY bool) (*fs.File, error) { + return newFileFromDonatedFD(ctx, fd, true, isTTY) } // newFileFromDonatedFD returns an fs.File from a donated FD. If the FD is // saveable, then saveable is true. -func newFileFromDonatedFD(ctx context.Context, donated int, mounter fs.FileOwner, saveable, isTTY bool) (*fs.File, error) { +func newFileFromDonatedFD(ctx context.Context, donated int, saveable, isTTY bool) (*fs.File, error) { var s syscall.Stat_t if err := syscall.Fstat(donated, &s); err != nil { return nil, err @@ -101,8 +101,8 @@ func newFileFromDonatedFD(ctx context.Context, donated int, mounter fs.FileOwner }) return s, nil default: - msrc := newMountSource(ctx, "/", mounter, &Filesystem{}, fs.MountSourceFlags{}, false /* dontTranslateOwnership */) - inode, err := newInode(ctx, msrc, donated, saveable, true /* donated */) + msrc := fs.NewNonCachingMountSource(ctx, &filesystem{}, fs.MountSourceFlags{}) + inode, err := newInode(ctx, msrc, donated, saveable) if err != nil { return nil, err } diff --git a/pkg/sentry/fs/host/fs.go b/pkg/sentry/fs/host/fs.go deleted file mode 100644 index d3e8e3a36..000000000 --- a/pkg/sentry/fs/host/fs.go +++ /dev/null @@ -1,339 +0,0 @@ -// Copyright 2018 The gVisor Authors. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -// Package host implements an fs.Filesystem for files backed by host -// file descriptors. -package host - -import ( - "fmt" - "path" - "path/filepath" - "strconv" - "strings" - - "gvisor.dev/gvisor/pkg/context" - "gvisor.dev/gvisor/pkg/log" - "gvisor.dev/gvisor/pkg/sentry/fs" -) - -// FilesystemName is the name under which Filesystem is registered. -const FilesystemName = "whitelistfs" - -const ( - // whitelistKey is the mount option containing a comma-separated list - // of host paths to whitelist. - whitelistKey = "whitelist" - - // rootPathKey is the mount option containing the root path of the - // mount. - rootPathKey = "root" - - // dontTranslateOwnershipKey is the key to superOperations.dontTranslateOwnership. - dontTranslateOwnershipKey = "dont_translate_ownership" -) - -// maxTraversals determines link traversals in building the whitelist. -const maxTraversals = 10 - -// Filesystem is a pseudo file system that is only available during the setup -// to lock down the configurations. This filesystem should only be mounted at root. -// -// Think twice before exposing this to applications. -// -// +stateify savable -type Filesystem struct { - // whitelist is a set of host paths to whitelist. - paths []string -} - -var _ fs.Filesystem = (*Filesystem)(nil) - -// Name is the identifier of this file system. -func (*Filesystem) Name() string { - return FilesystemName -} - -// AllowUserMount prohibits users from using mount(2) with this file system. -func (*Filesystem) AllowUserMount() bool { - return false -} - -// AllowUserList allows this filesystem to be listed in /proc/filesystems. -func (*Filesystem) AllowUserList() bool { - return true -} - -// Flags returns that there is nothing special about this file system. -func (*Filesystem) Flags() fs.FilesystemFlags { - return 0 -} - -// Mount returns an fs.Inode exposing the host file system. It is intended to be locked -// down in PreExec below. -func (f *Filesystem) Mount(ctx context.Context, _ string, flags fs.MountSourceFlags, data string, _ interface{}) (*fs.Inode, error) { - // Parse generic comma-separated key=value options. - options := fs.GenericMountSourceOptions(data) - - // Grab the whitelist if one was specified. - // TODO(edahlgren/mpratt/hzy): require another option "testonly" in order to allow - // no whitelist. - if wl, ok := options[whitelistKey]; ok { - f.paths = strings.Split(wl, "|") - delete(options, whitelistKey) - } - - // If the rootPath was set, use it. Othewise default to the root of the - // host fs. - rootPath := "/" - if rp, ok := options[rootPathKey]; ok { - rootPath = rp - delete(options, rootPathKey) - - // We must relativize the whitelisted paths to the new root. - for i, p := range f.paths { - rel, err := filepath.Rel(rootPath, p) - if err != nil { - return nil, fmt.Errorf("whitelist path %q must be a child of root path %q", p, rootPath) - } - f.paths[i] = path.Join("/", rel) - } - } - fd, err := open(nil, rootPath) - if err != nil { - return nil, fmt.Errorf("failed to find root: %v", err) - } - - var dontTranslateOwnership bool - if v, ok := options[dontTranslateOwnershipKey]; ok { - b, err := strconv.ParseBool(v) - if err != nil { - return nil, fmt.Errorf("invalid value for %q: %v", dontTranslateOwnershipKey, err) - } - dontTranslateOwnership = b - delete(options, dontTranslateOwnershipKey) - } - - // Fail if the caller passed us more options than we know about. - if len(options) > 0 { - return nil, fmt.Errorf("unsupported mount options: %v", options) - } - - // The mounting EUID/EGID will be cached by this file system. This will - // be used to assign ownership to files that we own. - owner := fs.FileOwnerFromContext(ctx) - - // Construct the host file system mount and inode. - msrc := newMountSource(ctx, rootPath, owner, f, flags, dontTranslateOwnership) - return newInode(ctx, msrc, fd, false /* saveable */, false /* donated */) -} - -// InstallWhitelist locks down the MountNamespace to only the currently installed -// Dirents and the given paths. -func (f *Filesystem) InstallWhitelist(ctx context.Context, m *fs.MountNamespace) error { - return installWhitelist(ctx, m, f.paths) -} - -func installWhitelist(ctx context.Context, m *fs.MountNamespace, paths []string) error { - if len(paths) == 0 || (len(paths) == 1 && paths[0] == "") { - // Warning will be logged during filter installation if the empty - // whitelist matters (allows for host file access). - return nil - } - - // Done tracks entries already added. - done := make(map[string]bool) - root := m.Root() - defer root.DecRef() - - for i := 0; i < len(paths); i++ { - // Make sure the path is absolute. This is a sanity check. - if !path.IsAbs(paths[i]) { - return fmt.Errorf("path %q is not absolute", paths[i]) - } - - // We need to add all the intermediate paths, in case one of - // them is a symlink that needs to be resolved. - for j := 1; j <= len(paths[i]); j++ { - if j < len(paths[i]) && paths[i][j] != '/' { - continue - } - current := paths[i][:j] - - // Lookup the given component in the tree. - remainingTraversals := uint(maxTraversals) - d, err := m.FindLink(ctx, root, nil, current, &remainingTraversals) - if err != nil { - log.Warningf("populate failed for %q: %v", current, err) - continue - } - - // It's critical that this DecRef happens after the - // freeze below. This ensures that the dentry is in - // place to be frozen. Otherwise, we freeze without - // these entries. - defer d.DecRef() - - // Expand the last component if necessary. - if current == paths[i] { - // Is it a directory or symlink? - sattr := d.Inode.StableAttr - if fs.IsDir(sattr) { - for name := range childDentAttrs(ctx, d) { - paths = append(paths, path.Join(current, name)) - } - } - if fs.IsSymlink(sattr) { - // Only expand symlinks once. The - // folder structure may contain - // recursive symlinks and we don't want - // to end up infinitely expanding this - // symlink. This is safe because this - // is the last component. If a later - // path wants to symlink something - // beneath this symlink that will still - // be handled by the FindLink above. - if done[current] { - continue - } - - s, err := d.Inode.Readlink(ctx) - if err != nil { - log.Warningf("readlink failed for %q: %v", current, err) - continue - } - if path.IsAbs(s) { - paths = append(paths, s) - } else { - target := path.Join(path.Dir(current), s) - paths = append(paths, target) - } - } - } - - // Only report this one once even though we may look - // it up more than once. If we whitelist /a/b,/a then - // /a will be "done" when it is looked up for /a/b, - // however we still need to expand all of its contents - // when whitelisting /a. - if !done[current] { - log.Debugf("whitelisted: %s", current) - } - done[current] = true - } - } - - // Freeze the mount tree in place. This prevents any new paths from - // being opened and any old ones from being removed. If we do provide - // tmpfs mounts, we'll want to freeze/thaw those separately. - m.Freeze() - return nil -} - -func childDentAttrs(ctx context.Context, d *fs.Dirent) map[string]fs.DentAttr { - dirname, _ := d.FullName(nil /* root */) - dir, err := d.Inode.GetFile(ctx, d, fs.FileFlags{Read: true}) - if err != nil { - log.Warningf("failed to open directory %q: %v", dirname, err) - return nil - } - dir.DecRef() - var stubSerializer fs.CollectEntriesSerializer - if err := dir.Readdir(ctx, &stubSerializer); err != nil { - log.Warningf("failed to iterate on host directory %q: %v", dirname, err) - return nil - } - delete(stubSerializer.Entries, ".") - delete(stubSerializer.Entries, "..") - return stubSerializer.Entries -} - -// newMountSource constructs a new host fs.MountSource -// relative to a root path. The root should match the mount point. -func newMountSource(ctx context.Context, root string, mounter fs.FileOwner, filesystem fs.Filesystem, flags fs.MountSourceFlags, dontTranslateOwnership bool) *fs.MountSource { - return fs.NewMountSource(ctx, &superOperations{ - root: root, - inodeMappings: make(map[uint64]string), - mounter: mounter, - dontTranslateOwnership: dontTranslateOwnership, - }, filesystem, flags) -} - -// superOperations implements fs.MountSourceOperations. -// -// +stateify savable -type superOperations struct { - fs.SimpleMountSourceOperations - - // root is the path of the mount point. All inode mappings - // are relative to this root. - root string - - // inodeMappings contains mappings of fs.Inodes associated - // with this MountSource to paths under root. - inodeMappings map[uint64]string - - // mounter is the cached EUID/EGID that mounted this file system. - mounter fs.FileOwner - - // dontTranslateOwnership indicates whether to not translate file - // ownership. - // - // By default, files/directories owned by the sandbox uses UID/GID - // of the mounter. For files/directories that are not owned by the - // sandbox, file UID/GID is translated to a UID/GID which cannot - // be mapped in the sandboxed application's user namespace. The - // UID/GID will look like the nobody UID/GID (65534) but is not - // strictly owned by the user "nobody". - // - // If whitelistfs is a lower filesystem in an overlay, set - // dont_translate_ownership=true in mount options. - dontTranslateOwnership bool -} - -var _ fs.MountSourceOperations = (*superOperations)(nil) - -// ResetInodeMappings implements fs.MountSourceOperations.ResetInodeMappings. -func (m *superOperations) ResetInodeMappings() { - m.inodeMappings = make(map[uint64]string) -} - -// SaveInodeMapping implements fs.MountSourceOperations.SaveInodeMapping. -func (m *superOperations) SaveInodeMapping(inode *fs.Inode, path string) { - // This is very unintuitive. We *CANNOT* trust the inode's StableAttrs, - // because overlay copyUp may have changed them out from under us. - // So much for "immutable". - sattr := inode.InodeOperations.(*inodeOperations).fileState.sattr - m.inodeMappings[sattr.InodeID] = path -} - -// Keep implements fs.MountSourceOperations.Keep. -// -// TODO(b/72455313,b/77596690): It is possible to change the permissions on a -// host file while it is in the dirent cache (say from RO to RW), but it is not -// possible to re-open the file with more relaxed permissions, since the host -// FD is already open and stored in the inode. -// -// Using the dirent LRU cache increases the odds that this bug is encountered. -// Since host file access is relatively fast anyways, we disable the LRU cache -// for host fs files. Once we can properly deal with permissions changes and -// re-opening host files, we should revisit whether or not to make use of the -// LRU cache. -func (*superOperations) Keep(*fs.Dirent) bool { - return false -} - -func init() { - fs.RegisterFilesystem(&Filesystem{}) -} diff --git a/pkg/sentry/fs/host/fs_test.go b/pkg/sentry/fs/host/fs_test.go deleted file mode 100644 index 3111d2df9..000000000 --- a/pkg/sentry/fs/host/fs_test.go +++ /dev/null @@ -1,380 +0,0 @@ -// Copyright 2018 The gVisor Authors. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -package host - -import ( - "fmt" - "io/ioutil" - "os" - "path" - "reflect" - "sort" - "testing" - - "gvisor.dev/gvisor/pkg/context" - "gvisor.dev/gvisor/pkg/sentry/contexttest" - "gvisor.dev/gvisor/pkg/sentry/fs" -) - -// newTestMountNamespace creates a MountNamespace with a ramfs root. -// It returns the host folder created, which should be removed when done. -func newTestMountNamespace(t *testing.T) (*fs.MountNamespace, string, error) { - p, err := ioutil.TempDir("", "root") - if err != nil { - return nil, "", err - } - - fd, err := open(nil, p) - if err != nil { - os.RemoveAll(p) - return nil, "", err - } - ctx := contexttest.Context(t) - root, err := newInode(ctx, newMountSource(ctx, p, fs.RootOwner, &Filesystem{}, fs.MountSourceFlags{}, false), fd, false, false) - if err != nil { - os.RemoveAll(p) - return nil, "", err - } - mm, err := fs.NewMountNamespace(ctx, root) - if err != nil { - os.RemoveAll(p) - return nil, "", err - } - return mm, p, nil -} - -// createTestDirs populates the root with some test files and directories. -// /a/a1.txt -// /a/a2.txt -// /b/b1.txt -// /b/c/c1.txt -// /symlinks/normal.txt -// /symlinks/to_normal.txt -> /symlinks/normal.txt -// /symlinks/recursive -> /symlinks -func createTestDirs(ctx context.Context, t *testing.T, m *fs.MountNamespace) error { - r := m.Root() - defer r.DecRef() - - if err := r.CreateDirectory(ctx, r, "a", fs.FilePermsFromMode(0777)); err != nil { - return err - } - - a, err := r.Walk(ctx, r, "a") - if err != nil { - return err - } - defer a.DecRef() - - a1, err := a.Create(ctx, r, "a1.txt", fs.FileFlags{Read: true, Write: true}, fs.FilePermsFromMode(0666)) - if err != nil { - return err - } - a1.DecRef() - - a2, err := a.Create(ctx, r, "a2.txt", fs.FileFlags{Read: true, Write: true}, fs.FilePermsFromMode(0666)) - if err != nil { - return err - } - a2.DecRef() - - if err := r.CreateDirectory(ctx, r, "b", fs.FilePermsFromMode(0777)); err != nil { - return err - } - - b, err := r.Walk(ctx, r, "b") - if err != nil { - return err - } - defer b.DecRef() - - b1, err := b.Create(ctx, r, "b1.txt", fs.FileFlags{Read: true, Write: true}, fs.FilePermsFromMode(0666)) - if err != nil { - return err - } - b1.DecRef() - - if err := b.CreateDirectory(ctx, r, "c", fs.FilePermsFromMode(0777)); err != nil { - return err - } - - c, err := b.Walk(ctx, r, "c") - if err != nil { - return err - } - defer c.DecRef() - - c1, err := c.Create(ctx, r, "c1.txt", fs.FileFlags{Read: true, Write: true}, fs.FilePermsFromMode(0666)) - if err != nil { - return err - } - c1.DecRef() - - if err := r.CreateDirectory(ctx, r, "symlinks", fs.FilePermsFromMode(0777)); err != nil { - return err - } - - symlinks, err := r.Walk(ctx, r, "symlinks") - if err != nil { - return err - } - defer symlinks.DecRef() - - normal, err := symlinks.Create(ctx, r, "normal.txt", fs.FileFlags{Read: true, Write: true}, fs.FilePermsFromMode(0666)) - if err != nil { - return err - } - normal.DecRef() - - if err := symlinks.CreateLink(ctx, r, "/symlinks/normal.txt", "to_normal.txt"); err != nil { - return err - } - - return symlinks.CreateLink(ctx, r, "/symlinks", "recursive") -} - -// allPaths returns a slice of all paths of entries visible in the rootfs. -func allPaths(ctx context.Context, t *testing.T, m *fs.MountNamespace, base string) ([]string, error) { - var paths []string - root := m.Root() - defer root.DecRef() - - maxTraversals := uint(1) - d, err := m.FindLink(ctx, root, nil, base, &maxTraversals) - if err != nil { - t.Logf("FindLink failed for %q", base) - return paths, err - } - defer d.DecRef() - - if fs.IsDir(d.Inode.StableAttr) { - dir, err := d.Inode.GetFile(ctx, d, fs.FileFlags{Read: true}) - if err != nil { - return nil, fmt.Errorf("failed to open directory %q: %v", base, err) - } - iter, ok := dir.FileOperations.(fs.DirIterator) - if !ok { - return nil, fmt.Errorf("cannot directly iterate on host directory %q", base) - } - dirCtx := &fs.DirCtx{ - Serializer: noopDentrySerializer{}, - } - if _, err := fs.DirentReaddir(ctx, d, iter, root, dirCtx, 0); err != nil { - return nil, err - } - for name := range dirCtx.DentAttrs() { - if name == "." || name == ".." { - continue - } - - fullName := path.Join(base, name) - paths = append(paths, fullName) - - // Recurse. - subpaths, err := allPaths(ctx, t, m, fullName) - if err != nil { - return paths, err - } - paths = append(paths, subpaths...) - } - } - - return paths, nil -} - -type noopDentrySerializer struct{} - -func (noopDentrySerializer) CopyOut(string, fs.DentAttr) error { - return nil -} -func (noopDentrySerializer) Written() int { - return 4096 -} - -// pathsEqual returns true if the two string slices contain the same entries. -func pathsEqual(got, want []string) bool { - sort.Strings(got) - sort.Strings(want) - - if len(got) != len(want) { - return false - } - - for i := range got { - if got[i] != want[i] { - return false - } - } - - return true -} - -func TestWhitelist(t *testing.T) { - for _, test := range []struct { - // description of the test. - desc string - // paths are the paths to whitelist - paths []string - // want are all of the directory entries that should be - // visible (nothing beyond this set should be visible). - want []string - }{ - { - desc: "root", - paths: []string{"/"}, - want: []string{"/a", "/a/a1.txt", "/a/a2.txt", "/b", "/b/b1.txt", "/b/c", "/b/c/c1.txt", "/symlinks", "/symlinks/normal.txt", "/symlinks/to_normal.txt", "/symlinks/recursive"}, - }, - { - desc: "top-level directories", - paths: []string{"/a", "/b"}, - want: []string{"/a", "/a/a1.txt", "/a/a2.txt", "/b", "/b/b1.txt", "/b/c", "/b/c/c1.txt"}, - }, - { - desc: "nested directories (1/2)", - paths: []string{"/b", "/b/c"}, - want: []string{"/b", "/b/b1.txt", "/b/c", "/b/c/c1.txt"}, - }, - { - desc: "nested directories (2/2)", - paths: []string{"/b/c", "/b"}, - want: []string{"/b", "/b/b1.txt", "/b/c", "/b/c/c1.txt"}, - }, - { - desc: "single file", - paths: []string{"/b/c/c1.txt"}, - want: []string{"/b", "/b/c", "/b/c/c1.txt"}, - }, - { - desc: "single file and directory", - paths: []string{"/a/a1.txt", "/b/c"}, - want: []string{"/a", "/a/a1.txt", "/b", "/b/c", "/b/c/c1.txt"}, - }, - { - desc: "symlink", - paths: []string{"/symlinks/to_normal.txt"}, - want: []string{"/symlinks", "/symlinks/normal.txt", "/symlinks/to_normal.txt"}, - }, - { - desc: "recursive symlink", - paths: []string{"/symlinks/recursive/normal.txt"}, - want: []string{"/symlinks", "/symlinks/normal.txt", "/symlinks/recursive"}, - }, - } { - t.Run(test.desc, func(t *testing.T) { - m, p, err := newTestMountNamespace(t) - if err != nil { - t.Errorf("Failed to create MountNamespace: %v", err) - } - defer os.RemoveAll(p) - - ctx := withRoot(contexttest.RootContext(t), m.Root()) - if err := createTestDirs(ctx, t, m); err != nil { - t.Errorf("Failed to create test dirs: %v", err) - } - - if err := installWhitelist(ctx, m, test.paths); err != nil { - t.Errorf("installWhitelist(%v) err got %v want nil", test.paths, err) - } - - got, err := allPaths(ctx, t, m, "/") - if err != nil { - t.Fatalf("Failed to lookup paths (whitelisted: %v): %v", test.paths, err) - } - - if !pathsEqual(got, test.want) { - t.Errorf("For paths %v got %v want %v", test.paths, got, test.want) - } - }) - } -} - -func TestRootPath(t *testing.T) { - // Create a temp dir, which will be the root of our mounted fs. - rootPath, err := ioutil.TempDir(os.TempDir(), "root") - if err != nil { - t.Fatalf("TempDir failed: %v", err) - } - defer os.RemoveAll(rootPath) - - // Create two files inside the new root, one which will be whitelisted - // and one not. - whitelisted, err := ioutil.TempFile(rootPath, "white") - if err != nil { - t.Fatalf("TempFile failed: %v", err) - } - if _, err := ioutil.TempFile(rootPath, "black"); err != nil { - t.Fatalf("TempFile failed: %v", err) - } - - // Create a mount with a root path and single whitelisted file. - hostFS := &Filesystem{} - ctx := contexttest.Context(t) - data := fmt.Sprintf("%s=%s,%s=%s", rootPathKey, rootPath, whitelistKey, whitelisted.Name()) - inode, err := hostFS.Mount(ctx, "", fs.MountSourceFlags{}, data, nil) - if err != nil { - t.Fatalf("Mount failed: %v", err) - } - mm, err := fs.NewMountNamespace(ctx, inode) - if err != nil { - t.Fatalf("NewMountNamespace failed: %v", err) - } - if err := hostFS.InstallWhitelist(ctx, mm); err != nil { - t.Fatalf("InstallWhitelist failed: %v", err) - } - - // Get the contents of the root directory. - rootDir := mm.Root() - rctx := withRoot(ctx, rootDir) - f, err := rootDir.Inode.GetFile(rctx, rootDir, fs.FileFlags{}) - if err != nil { - t.Fatalf("GetFile failed: %v", err) - } - c := &fs.CollectEntriesSerializer{} - if err := f.Readdir(rctx, c); err != nil { - t.Fatalf("Readdir failed: %v", err) - } - - // We should have only our whitelisted file, plus the dots. - want := []string{path.Base(whitelisted.Name()), ".", ".."} - got := c.Order - sort.Strings(want) - sort.Strings(got) - if !reflect.DeepEqual(got, want) { - t.Errorf("Readdir got %v, wanted %v", got, want) - } -} - -type rootContext struct { - context.Context - root *fs.Dirent -} - -// withRoot returns a copy of ctx with the given root. -func withRoot(ctx context.Context, root *fs.Dirent) context.Context { - return &rootContext{ - Context: ctx, - root: root, - } -} - -// Value implements Context.Value. -func (rc rootContext) Value(key interface{}) interface{} { - switch key { - case fs.CtxRoot: - rc.root.IncRef() - return rc.root - default: - return rc.Context.Value(key) - } -} diff --git a/pkg/sentry/fs/host/host.go b/pkg/sentry/fs/host/host.go new file mode 100644 index 000000000..081ba1dd8 --- /dev/null +++ b/pkg/sentry/fs/host/host.go @@ -0,0 +1,59 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// Package host supports file descriptors imported directly. +package host + +import ( + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/syserror" +) + +// filesystem is a host filesystem. +// +// +stateify savable +type filesystem struct{} + +func init() { + fs.RegisterFilesystem(&filesystem{}) +} + +// FilesystemName is the name under which the filesystem is registered. +const FilesystemName = "host" + +// Name is the name of the filesystem. +func (*filesystem) Name() string { + return FilesystemName +} + +// Mount returns an error. Mounting hostfs is not allowed. +func (*filesystem) Mount(ctx context.Context, device string, flags fs.MountSourceFlags, data string, dataObj interface{}) (*fs.Inode, error) { + return nil, syserror.EPERM +} + +// AllowUserMount prohibits users from using mount(2) with this file system. +func (*filesystem) AllowUserMount() bool { + return false +} + +// AllowUserList prohibits this filesystem to be listed in /proc/filesystems. +func (*filesystem) AllowUserList() bool { + return false +} + +// Flags returns that there is nothing special about this file system. +func (*filesystem) Flags() fs.FilesystemFlags { + return 0 +} diff --git a/pkg/sentry/fs/host/inode.go b/pkg/sentry/fs/host/inode.go index 6fa39caab..62f1246aa 100644 --- a/pkg/sentry/fs/host/inode.go +++ b/pkg/sentry/fs/host/inode.go @@ -17,12 +17,10 @@ package host import ( "syscall" - "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/fd" "gvisor.dev/gvisor/pkg/safemem" "gvisor.dev/gvisor/pkg/secio" - "gvisor.dev/gvisor/pkg/sentry/device" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fs/fsutil" "gvisor.dev/gvisor/pkg/sentry/memmap" @@ -69,9 +67,6 @@ type inodeOperations struct { // // +stateify savable type inodeFileState struct { - // Common file system state. - mops *superOperations `state:"wait"` - // descriptor is the backing host FD. descriptor *descriptor `state:"wait"` @@ -160,7 +155,7 @@ func (i *inodeFileState) unstableAttr(ctx context.Context) (fs.UnstableAttr, err if err := syscall.Fstat(i.FD(), &s); err != nil { return fs.UnstableAttr{}, err } - return unstableAttr(i.mops, &s), nil + return unstableAttr(&s), nil } // Allocate implements fsutil.CachedFileObject.Allocate. @@ -172,7 +167,7 @@ func (i *inodeFileState) Allocate(_ context.Context, offset, length int64) error var _ fs.InodeOperations = (*inodeOperations)(nil) // newInode returns a new fs.Inode backed by the host FD. -func newInode(ctx context.Context, msrc *fs.MountSource, fd int, saveable bool, donated bool) (*fs.Inode, error) { +func newInode(ctx context.Context, msrc *fs.MountSource, fd int, saveable bool) (*fs.Inode, error) { // Retrieve metadata. var s syscall.Stat_t err := syscall.Fstat(fd, &s) @@ -181,24 +176,17 @@ func newInode(ctx context.Context, msrc *fs.MountSource, fd int, saveable bool, } fileState := &inodeFileState{ - mops: msrc.MountSourceOperations.(*superOperations), sattr: stableAttr(&s), } // Initialize the wrapped host file descriptor. - fileState.descriptor, err = newDescriptor( - fd, - donated, - saveable, - wouldBlock(&s), - &fileState.queue, - ) + fileState.descriptor, err = newDescriptor(fd, saveable, wouldBlock(&s), &fileState.queue) if err != nil { return nil, err } // Build the fs.InodeOperations. - uattr := unstableAttr(msrc.MountSourceOperations.(*superOperations), &s) + uattr := unstableAttr(&s) iops := &inodeOperations{ fileState: fileState, cachingInodeOps: fsutil.NewCachingInodeOperations(ctx, fileState, uattr, fsutil.CachingInodeOperationsOptions{ @@ -232,54 +220,23 @@ func (i *inodeOperations) Release(context.Context) { // Lookup implements fs.InodeOperations.Lookup. func (i *inodeOperations) Lookup(ctx context.Context, dir *fs.Inode, name string) (*fs.Dirent, error) { - // Get a new FD relative to i at name. - fd, err := open(i, name) - if err != nil { - if err == syserror.ENOENT { - return nil, syserror.ENOENT - } - return nil, err - } - - inode, err := newInode(ctx, dir.MountSource, fd, false /* saveable */, false /* donated */) - if err != nil { - return nil, err - } - - // Return the fs.Dirent. - return fs.NewDirent(ctx, inode, name), nil + return nil, syserror.ENOENT } // Create implements fs.InodeOperations.Create. func (i *inodeOperations) Create(ctx context.Context, dir *fs.Inode, name string, flags fs.FileFlags, perm fs.FilePermissions) (*fs.File, error) { - // Create a file relative to i at name. - // - // N.B. We always open this file O_RDWR regardless of flags because a - // future GetFile might want more access. Open allows this regardless - // of perm. - fd, err := openAt(i, name, syscall.O_RDWR|syscall.O_CREAT|syscall.O_EXCL, perm.LinuxMode()) - if err != nil { - return nil, err - } - - inode, err := newInode(ctx, dir.MountSource, fd, false /* saveable */, false /* donated */) - if err != nil { - return nil, err - } + return nil, syserror.EPERM - d := fs.NewDirent(ctx, inode, name) - defer d.DecRef() - return inode.GetFile(ctx, d, flags) } // CreateDirectory implements fs.InodeOperations.CreateDirectory. func (i *inodeOperations) CreateDirectory(ctx context.Context, dir *fs.Inode, name string, perm fs.FilePermissions) error { - return syscall.Mkdirat(i.fileState.FD(), name, uint32(perm.LinuxMode())) + return syserror.EPERM } // CreateLink implements fs.InodeOperations.CreateLink. func (i *inodeOperations) CreateLink(ctx context.Context, dir *fs.Inode, oldname string, newname string) error { - return createLink(i.fileState.FD(), oldname, newname) + return syserror.EPERM } // CreateHardLink implements fs.InodeOperations.CreateHardLink. @@ -294,25 +251,17 @@ func (*inodeOperations) CreateFifo(context.Context, *fs.Inode, string, fs.FilePe // Remove implements fs.InodeOperations.Remove. func (i *inodeOperations) Remove(ctx context.Context, dir *fs.Inode, name string) error { - return unlinkAt(i.fileState.FD(), name, false /* dir */) + return syserror.EPERM } // RemoveDirectory implements fs.InodeOperations.RemoveDirectory. func (i *inodeOperations) RemoveDirectory(ctx context.Context, dir *fs.Inode, name string) error { - return unlinkAt(i.fileState.FD(), name, true /* dir */) + return syserror.EPERM } // Rename implements fs.InodeOperations.Rename. func (i *inodeOperations) Rename(ctx context.Context, inode *fs.Inode, oldParent *fs.Inode, oldName string, newParent *fs.Inode, newName string, replacement bool) error { - op, ok := oldParent.InodeOperations.(*inodeOperations) - if !ok { - return syscall.EXDEV - } - np, ok := newParent.InodeOperations.(*inodeOperations) - if !ok { - return syscall.EXDEV - } - return syscall.Renameat(op.fileState.FD(), oldName, np.fileState.FD(), newName) + return syserror.EPERM } // Bind implements fs.InodeOperations.Bind. @@ -448,82 +397,17 @@ func (i *inodeOperations) StatFS(context.Context) (fs.Info, error) { } // AddLink implements fs.InodeOperations.AddLink. -// FIXME(b/63117438): Remove this from InodeOperations altogether. func (i *inodeOperations) AddLink() {} // DropLink implements fs.InodeOperations.DropLink. -// FIXME(b/63117438): Remove this from InodeOperations altogether. func (i *inodeOperations) DropLink() {} // NotifyStatusChange implements fs.InodeOperations.NotifyStatusChange. -// FIXME(b/63117438): Remove this from InodeOperations altogether. func (i *inodeOperations) NotifyStatusChange(ctx context.Context) {} // readdirAll returns all of the directory entries in i. func (i *inodeOperations) readdirAll(d *dirInfo) (map[string]fs.DentAttr, error) { - i.readdirMu.Lock() - defer i.readdirMu.Unlock() - - fd := i.fileState.FD() - - // syscall.ReadDirent will use getdents, which will seek the file past - // the last directory entry. To read the directory entries a second - // time, we need to seek back to the beginning. - if _, err := syscall.Seek(fd, 0, 0); err != nil { - if err == syscall.ESPIPE { - // All directories should be seekable. If this file - // isn't seekable, it is not a directory and we should - // return that more sane error. - err = syscall.ENOTDIR - } - return nil, err - } - - names := make([]string, 0, 100) - for { - // Refill the buffer if necessary - if d.bufp >= d.nbuf { - d.bufp = 0 - // ReadDirent will just do a sys_getdents64 to the kernel. - n, err := syscall.ReadDirent(fd, d.buf) - if err != nil { - return nil, err - } - if n == 0 { - break // EOF - } - d.nbuf = n - } - - var nb int - // Parse the dirent buffer we just get and return the directory names along - // with the number of bytes consumed in the buffer. - nb, _, names = syscall.ParseDirent(d.buf[d.bufp:d.nbuf], -1, names) - d.bufp += nb - } - - entries := make(map[string]fs.DentAttr) - for _, filename := range names { - // Lookup the type and host device and inode. - stat, lerr := fstatat(fd, filename, linux.AT_SYMLINK_NOFOLLOW) - if lerr == syscall.ENOENT { - // File disappeared between readdir and lstat. - // Just treat it as if it didn't exist. - continue - } - - // There was a serious problem, we should probably report it. - if lerr != nil { - return nil, lerr - } - - entries[filename] = fs.DentAttr{ - Type: nodeType(&stat), - InodeID: hostFileDevice.Map(device.MultiDeviceKey{ - Device: stat.Dev, - Inode: stat.Ino, - }), - } - } - return entries, nil + // We only support non-directory file descriptors that have been + // imported, so just claim that this isn't a directory, even if it is. + return nil, syscall.ENOTDIR } diff --git a/pkg/sentry/fs/host/inode_state.go b/pkg/sentry/fs/host/inode_state.go index 299e0e0b0..1adbd4562 100644 --- a/pkg/sentry/fs/host/inode_state.go +++ b/pkg/sentry/fs/host/inode_state.go @@ -18,29 +18,14 @@ import ( "fmt" "syscall" - "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/device" "gvisor.dev/gvisor/pkg/sentry/fs" ) -// beforeSave is invoked by stateify. -func (i *inodeFileState) beforeSave() { - if !i.queue.IsEmpty() { - panic("event queue must be empty") - } - if !i.descriptor.donated && i.sattr.Type == fs.RegularFile { - uattr, err := i.unstableAttr(context.Background()) - if err != nil { - panic(fs.ErrSaveRejection{fmt.Errorf("failed to get unstable atttribute of %s: %v", i.mops.inodeMappings[i.sattr.InodeID], err)}) - } - i.savedUAttr = &uattr - } -} - // afterLoad is invoked by stateify. func (i *inodeFileState) afterLoad() { // Initialize the descriptor value. - if err := i.descriptor.initAfterLoad(i.mops, i.sattr.InodeID, &i.queue); err != nil { + if err := i.descriptor.initAfterLoad(i.sattr.InodeID, &i.queue); err != nil { panic(fmt.Sprintf("failed to load value of descriptor: %v", err)) } @@ -61,19 +46,4 @@ func (i *inodeFileState) afterLoad() { // change across save and restore, error out. panic(fs.ErrCorruption{fmt.Errorf("host %s conflict in host device mappings: %s", key, hostFileDevice)}) } - - if !i.descriptor.donated && i.sattr.Type == fs.RegularFile { - env, ok := fs.CurrentRestoreEnvironment() - if !ok { - panic("missing restore environment") - } - uattr := unstableAttr(i.mops, &s) - if env.ValidateFileSize && uattr.Size != i.savedUAttr.Size { - panic(fs.ErrCorruption{fmt.Errorf("file size has changed for %s: previously %d, now %d", i.mops.inodeMappings[i.sattr.InodeID], i.savedUAttr.Size, uattr.Size)}) - } - if env.ValidateFileTimestamp && uattr.ModificationTime != i.savedUAttr.ModificationTime { - panic(fs.ErrCorruption{fmt.Errorf("file modification time has changed for %s: previously %v, now %v", i.mops.inodeMappings[i.sattr.InodeID], i.savedUAttr.ModificationTime, uattr.ModificationTime)}) - } - i.savedUAttr = nil - } } diff --git a/pkg/sentry/fs/host/inode_test.go b/pkg/sentry/fs/host/inode_test.go index 7221bc825..c507f57eb 100644 --- a/pkg/sentry/fs/host/inode_test.go +++ b/pkg/sentry/fs/host/inode_test.go @@ -15,79 +15,12 @@ package host import ( - "io/ioutil" - "os" - "path" "syscall" "testing" "gvisor.dev/gvisor/pkg/sentry/contexttest" - "gvisor.dev/gvisor/pkg/sentry/fs" ) -// TestMultipleReaddir verifies that multiple Readdir calls return the same -// thing if they use different dir contexts. -func TestMultipleReaddir(t *testing.T) { - p, err := ioutil.TempDir("", "readdir") - if err != nil { - t.Fatalf("Failed to create test dir: %v", err) - } - defer os.RemoveAll(p) - - f, err := os.Create(path.Join(p, "a.txt")) - if err != nil { - t.Fatalf("Failed to create a.txt: %v", err) - } - f.Close() - - f, err = os.Create(path.Join(p, "b.txt")) - if err != nil { - t.Fatalf("Failed to create b.txt: %v", err) - } - f.Close() - - fd, err := open(nil, p) - if err != nil { - t.Fatalf("Failed to open %q: %v", p, err) - } - ctx := contexttest.Context(t) - n, err := newInode(ctx, newMountSource(ctx, p, fs.RootOwner, &Filesystem{}, fs.MountSourceFlags{}, false), fd, false, false) - if err != nil { - t.Fatalf("Failed to create inode: %v", err) - } - - dirent := fs.NewDirent(ctx, n, "readdir") - openFile, err := n.GetFile(ctx, dirent, fs.FileFlags{Read: true}) - if err != nil { - t.Fatalf("Failed to get file: %v", err) - } - defer openFile.DecRef() - - c1 := &fs.DirCtx{DirCursor: new(string)} - if _, err := openFile.FileOperations.(*fileOperations).IterateDir(ctx, dirent, c1, 0); err != nil { - t.Fatalf("First Readdir failed: %v", err) - } - - c2 := &fs.DirCtx{DirCursor: new(string)} - if _, err := openFile.FileOperations.(*fileOperations).IterateDir(ctx, dirent, c2, 0); err != nil { - t.Errorf("Second Readdir failed: %v", err) - } - - if _, ok := c1.DentAttrs()["a.txt"]; !ok { - t.Errorf("want a.txt in first Readdir, got %v", c1.DentAttrs()) - } - if _, ok := c1.DentAttrs()["b.txt"]; !ok { - t.Errorf("want b.txt in first Readdir, got %v", c1.DentAttrs()) - } - - if _, ok := c2.DentAttrs()["a.txt"]; !ok { - t.Errorf("want a.txt in second Readdir, got %v", c2.DentAttrs()) - } - if _, ok := c2.DentAttrs()["b.txt"]; !ok { - t.Errorf("want b.txt in second Readdir, got %v", c2.DentAttrs()) - } -} - // TestCloseFD verifies fds will be closed. func TestCloseFD(t *testing.T) { var p [2]int @@ -99,7 +32,7 @@ func TestCloseFD(t *testing.T) { // Use the write-end because we will detect if it's closed on the read end. ctx := contexttest.Context(t) - file, err := NewFile(ctx, p[1], fs.RootOwner) + file, err := NewFile(ctx, p[1]) if err != nil { t.Fatalf("Failed to create File: %v", err) } diff --git a/pkg/sentry/fs/host/ioctl_unsafe.go b/pkg/sentry/fs/host/ioctl_unsafe.go index 271582e54..150ac8e19 100644 --- a/pkg/sentry/fs/host/ioctl_unsafe.go +++ b/pkg/sentry/fs/host/ioctl_unsafe.go @@ -21,6 +21,8 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" ) +// LINT.IfChange + func ioctlGetTermios(fd int) (*linux.Termios, error) { var t linux.Termios _, _, errno := syscall.Syscall(syscall.SYS_IOCTL, uintptr(fd), linux.TCGETS, uintptr(unsafe.Pointer(&t))) @@ -54,3 +56,5 @@ func ioctlSetWinsize(fd int, w *linux.Winsize) error { } return nil } + +// LINT.ThenChange(../../fsimpl/host/ioctl_unsafe.go) diff --git a/pkg/sentry/fs/host/socket.go b/pkg/sentry/fs/host/socket.go index 06fc2d80a..b6e94583e 100644 --- a/pkg/sentry/fs/host/socket.go +++ b/pkg/sentry/fs/host/socket.go @@ -37,6 +37,8 @@ import ( "gvisor.dev/gvisor/pkg/waiter" ) +// LINT.IfChange + // maxSendBufferSize is the maximum host send buffer size allowed for endpoint. // // N.B. 8MB is the default maximum on Linux (2 * sysctl_wmem_max). @@ -388,3 +390,5 @@ func (c *ConnectedEndpoint) Release() { // CloseUnread implements transport.ConnectedEndpoint.CloseUnread. func (c *ConnectedEndpoint) CloseUnread() {} + +// LINT.ThenChange(../../fsimpl/host/socket.go) diff --git a/pkg/sentry/fs/host/socket_iovec.go b/pkg/sentry/fs/host/socket_iovec.go index af6955675..5c18dbd5e 100644 --- a/pkg/sentry/fs/host/socket_iovec.go +++ b/pkg/sentry/fs/host/socket_iovec.go @@ -21,6 +21,8 @@ import ( "gvisor.dev/gvisor/pkg/syserror" ) +// LINT.IfChange + // maxIovs is the maximum number of iovecs to pass to the host. var maxIovs = linux.UIO_MAXIOV @@ -111,3 +113,5 @@ func buildIovec(bufs [][]byte, maxlen int64, truncate bool) (length int64, iovec return total, iovecs, nil, err } + +// LINT.ThenChange(../../fsimpl/host/socket_iovec.go) diff --git a/pkg/sentry/fs/host/socket_test.go b/pkg/sentry/fs/host/socket_test.go index eb4afe520..affdbcacb 100644 --- a/pkg/sentry/fs/host/socket_test.go +++ b/pkg/sentry/fs/host/socket_test.go @@ -199,14 +199,14 @@ func TestListen(t *testing.T) { } func TestPasscred(t *testing.T) { - e := ConnectedEndpoint{} + e := &ConnectedEndpoint{} if got, want := e.Passcred(), false; got != want { t.Errorf("Got %#v.Passcred() = %t, want = %t", e, got, want) } } func TestGetLocalAddress(t *testing.T) { - e := ConnectedEndpoint{path: "foo"} + e := &ConnectedEndpoint{path: "foo"} want := tcpip.FullAddress{Addr: tcpip.Address("foo")} if got, err := e.GetLocalAddress(); err != nil || got != want { t.Errorf("Got %#v.GetLocalAddress() = %#v, %v, want = %#v, %v", e, got, err, want, nil) @@ -214,7 +214,7 @@ func TestGetLocalAddress(t *testing.T) { } func TestQueuedSize(t *testing.T) { - e := ConnectedEndpoint{} + e := &ConnectedEndpoint{} tests := []struct { name string f func() int64 diff --git a/pkg/sentry/fs/host/socket_unsafe.go b/pkg/sentry/fs/host/socket_unsafe.go index f3bbed7ea..5d4f312cf 100644 --- a/pkg/sentry/fs/host/socket_unsafe.go +++ b/pkg/sentry/fs/host/socket_unsafe.go @@ -19,6 +19,8 @@ import ( "unsafe" ) +// LINT.IfChange + // fdReadVec receives from fd to bufs. // // If the total length of bufs is > maxlen, fdReadVec will do a partial read @@ -99,3 +101,5 @@ func fdWriteVec(fd int, bufs [][]byte, maxlen int64, truncate bool) (int64, int6 return int64(n), length, err } + +// LINT.ThenChange(../../fsimpl/host/socket_unsafe.go) diff --git a/pkg/sentry/fs/host/tty.go b/pkg/sentry/fs/host/tty.go index 3f218b4a7..cb91355ab 100644 --- a/pkg/sentry/fs/host/tty.go +++ b/pkg/sentry/fs/host/tty.go @@ -26,6 +26,8 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) +// LINT.IfChange + // TTYFileOperations implements fs.FileOperations for a host file descriptor // that wraps a TTY FD. // @@ -43,6 +45,7 @@ type TTYFileOperations struct { // connected to this TTY. fgProcessGroup *kernel.ProcessGroup + // termios contains the terminal attributes for this TTY. termios linux.KernelTermios } @@ -357,3 +360,5 @@ func (t *TTYFileOperations) checkChange(ctx context.Context, sig linux.Signal) e _ = pg.SendSignal(kernel.SignalInfoPriv(sig)) return kernel.ERESTARTSYS } + +// LINT.ThenChange(../../fsimpl/host/tty.go) diff --git a/pkg/sentry/fs/host/util.go b/pkg/sentry/fs/host/util.go index e37e687c6..1b0356930 100644 --- a/pkg/sentry/fs/host/util.go +++ b/pkg/sentry/fs/host/util.go @@ -16,7 +16,6 @@ package host import ( "os" - "path" "syscall" "gvisor.dev/gvisor/pkg/abi/linux" @@ -28,45 +27,6 @@ import ( "gvisor.dev/gvisor/pkg/syserror" ) -func open(parent *inodeOperations, name string) (int, error) { - if parent == nil && !path.IsAbs(name) { - return -1, syserror.EINVAL - } - name = path.Clean(name) - - // Don't follow through symlinks. - flags := syscall.O_NOFOLLOW - - if fd, err := openAt(parent, name, flags|syscall.O_RDWR, 0); err == nil { - return fd, nil - } - // Retry as read-only. - if fd, err := openAt(parent, name, flags|syscall.O_RDONLY, 0); err == nil { - return fd, nil - } - - // Retry as write-only. - if fd, err := openAt(parent, name, flags|syscall.O_WRONLY, 0); err == nil { - return fd, nil - } - - // Retry as a symlink, by including O_PATH as an option. - fd, err := openAt(parent, name, linux.O_PATH|flags, 0) - if err == nil { - return fd, nil - } - - // Everything failed. - return -1, err -} - -func openAt(parent *inodeOperations, name string, flags int, perm linux.FileMode) (int, error) { - if parent == nil { - return syscall.Open(name, flags, uint32(perm)) - } - return syscall.Openat(parent.fileState.FD(), name, flags, uint32(perm)) -} - func nodeType(s *syscall.Stat_t) fs.InodeType { switch x := (s.Mode & syscall.S_IFMT); x { case syscall.S_IFLNK: @@ -107,51 +67,19 @@ func stableAttr(s *syscall.Stat_t) fs.StableAttr { } } -func owner(mo *superOperations, s *syscall.Stat_t) fs.FileOwner { - // User requested no translation, just return actual owner. - if mo.dontTranslateOwnership { - return fs.FileOwner{auth.KUID(s.Uid), auth.KGID(s.Gid)} +func owner(s *syscall.Stat_t) fs.FileOwner { + return fs.FileOwner{ + UID: auth.KUID(s.Uid), + GID: auth.KGID(s.Gid), } - - // Show only IDs relevant to the sandboxed task. I.e. if we not own the - // file, no sandboxed task can own the file. In that case, we - // use OverflowID for UID, implying that the IDs are not mapped in the - // "root" user namespace. - // - // E.g. - // sandbox's host EUID/EGID is 1/1. - // some_dir's host UID/GID is 2/1. - // Task that mounted this fs has virtualized EUID/EGID 5/5. - // - // If you executed `ls -n` in the sandboxed task, it would show: - // drwxwrxwrx [...] 65534 5 [...] some_dir - - // Files are owned by OverflowID by default. - owner := fs.FileOwner{auth.KUID(auth.OverflowUID), auth.KGID(auth.OverflowGID)} - - // If we own file on host, let mounting task's initial EUID own - // the file. - if s.Uid == hostUID { - owner.UID = mo.mounter.UID - } - - // If our group matches file's group, make file's group match - // the mounting task's initial EGID. - for _, gid := range hostGIDs { - if s.Gid == gid { - owner.GID = mo.mounter.GID - break - } - } - return owner } -func unstableAttr(mo *superOperations, s *syscall.Stat_t) fs.UnstableAttr { +func unstableAttr(s *syscall.Stat_t) fs.UnstableAttr { return fs.UnstableAttr{ Size: s.Size, Usage: s.Blocks * 512, Perms: fs.FilePermsFromMode(linux.FileMode(s.Mode)), - Owner: owner(mo, s), + Owner: owner(s), AccessTime: ktime.FromUnix(s.Atim.Sec, s.Atim.Nsec), ModificationTime: ktime.FromUnix(s.Mtim.Sec, s.Mtim.Nsec), StatusChangeTime: ktime.FromUnix(s.Ctim.Sec, s.Ctim.Nsec), @@ -165,6 +93,8 @@ type dirInfo struct { bufp int // location of next record in buf. } +// LINT.IfChange + // isBlockError unwraps os errors and checks if they are caused by EAGAIN or // EWOULDBLOCK. This is so they can be transformed into syserror.ErrWouldBlock. func isBlockError(err error) bool { @@ -177,6 +107,8 @@ func isBlockError(err error) bool { return false } +// LINT.ThenChange(../../fsimpl/host/util.go) + func hostEffectiveKIDs() (uint32, []uint32, error) { gids, err := os.Getgroups() if err != nil { diff --git a/pkg/sentry/fs/host/util_unsafe.go b/pkg/sentry/fs/host/util_unsafe.go index 3ab36b088..23bd35d64 100644 --- a/pkg/sentry/fs/host/util_unsafe.go +++ b/pkg/sentry/fs/host/util_unsafe.go @@ -26,26 +26,6 @@ import ( // NulByte is a single NUL byte. It is passed to readlinkat as an empty string. var NulByte byte = '\x00' -func createLink(fd int, name string, linkName string) error { - namePtr, err := syscall.BytePtrFromString(name) - if err != nil { - return err - } - linkNamePtr, err := syscall.BytePtrFromString(linkName) - if err != nil { - return err - } - _, _, errno := syscall.Syscall( - syscall.SYS_SYMLINKAT, - uintptr(unsafe.Pointer(namePtr)), - uintptr(fd), - uintptr(unsafe.Pointer(linkNamePtr))) - if errno != 0 { - return errno - } - return nil -} - func readLink(fd int) (string, error) { // Buffer sizing copied from os.Readlink. for l := 128; ; l *= 2 { @@ -66,27 +46,6 @@ func readLink(fd int) (string, error) { } } -func unlinkAt(fd int, name string, dir bool) error { - namePtr, err := syscall.BytePtrFromString(name) - if err != nil { - return err - } - var flags uintptr - if dir { - flags = linux.AT_REMOVEDIR - } - _, _, errno := syscall.Syscall( - syscall.SYS_UNLINKAT, - uintptr(fd), - uintptr(unsafe.Pointer(namePtr)), - flags, - ) - if errno != 0 { - return errno - } - return nil -} - func timespecFromTimestamp(t ktime.Time, omit, setSysTime bool) syscall.Timespec { if omit { return syscall.Timespec{0, linux.UTIME_OMIT} diff --git a/pkg/sentry/fs/host/wait_test.go b/pkg/sentry/fs/host/wait_test.go index d49c3a635..ce397a5e3 100644 --- a/pkg/sentry/fs/host/wait_test.go +++ b/pkg/sentry/fs/host/wait_test.go @@ -20,7 +20,6 @@ import ( "time" "gvisor.dev/gvisor/pkg/sentry/contexttest" - "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/waiter" ) @@ -34,7 +33,7 @@ func TestWait(t *testing.T) { defer syscall.Close(fds[1]) ctx := contexttest.Context(t) - file, err := NewFile(ctx, fds[0], fs.RootOwner) + file, err := NewFile(ctx, fds[0]) if err != nil { syscall.Close(fds[0]) t.Fatalf("NewFile failed: %v", err) diff --git a/pkg/sentry/fs/inode.go b/pkg/sentry/fs/inode.go index 55fb71c16..a34fbc946 100644 --- a/pkg/sentry/fs/inode.go +++ b/pkg/sentry/fs/inode.go @@ -102,7 +102,6 @@ func (i *Inode) DecRef() { // destroy releases the Inode and releases the msrc reference taken. func (i *Inode) destroy() { - // FIXME(b/38173783): Context is not plumbed here. ctx := context.Background() if err := i.WriteOut(ctx); err != nil { // FIXME(b/65209558): Mark as warning again once noatime is @@ -397,8 +396,6 @@ func (i *Inode) Getlink(ctx context.Context) (*Dirent, error) { // AddLink calls i.InodeOperations.AddLink. func (i *Inode) AddLink() { if i.overlay != nil { - // FIXME(b/63117438): Remove this from InodeOperations altogether. - // // This interface is only used by ramfs to update metadata of // children. These filesystems should _never_ have overlay // Inodes cached as children. So explicitly disallow this diff --git a/pkg/sentry/fs/inode_overlay.go b/pkg/sentry/fs/inode_overlay.go index 5ada33a32..537c8d257 100644 --- a/pkg/sentry/fs/inode_overlay.go +++ b/pkg/sentry/fs/inode_overlay.go @@ -231,7 +231,8 @@ func overlayCreate(ctx context.Context, o *overlayEntry, parent *Dirent, name st upperFile.Dirent.Inode.IncRef() entry, err := newOverlayEntry(ctx, upperFile.Dirent.Inode, nil, false) if err != nil { - cleanupUpper(ctx, o.upper, name) + werr := fmt.Errorf("newOverlayEntry failed: %v", err) + cleanupUpper(ctx, o.upper, name, werr) return nil, err } diff --git a/pkg/sentry/fs/inotify.go b/pkg/sentry/fs/inotify.go index 928c90aa0..e3a715c1f 100644 --- a/pkg/sentry/fs/inotify.go +++ b/pkg/sentry/fs/inotify.go @@ -143,7 +143,10 @@ func (i *Inotify) Read(ctx context.Context, _ *File, dst usermem.IOSequence, _ i } var writeLen int64 - for event := i.events.Front(); event != nil; event = event.Next() { + for it := i.events.Front(); it != nil; { + event := it + it = it.Next() + // Does the buffer have enough remaining space to hold the event we're // about to write out? if dst.NumBytes() < int64(event.sizeOf()) { diff --git a/pkg/sentry/fs/mount_test.go b/pkg/sentry/fs/mount_test.go index e672a438c..a3d10770b 100644 --- a/pkg/sentry/fs/mount_test.go +++ b/pkg/sentry/fs/mount_test.go @@ -36,11 +36,12 @@ func mountPathsAre(root *Dirent, got []*Mount, want ...string) error { gotPaths := make(map[string]struct{}, len(got)) gotStr := make([]string, len(got)) for i, g := range got { - groot := g.Root() - name, _ := groot.FullName(root) - groot.DecRef() - gotStr[i] = name - gotPaths[name] = struct{}{} + if groot := g.Root(); groot != nil { + name, _ := groot.FullName(root) + groot.DecRef() + gotStr[i] = name + gotPaths[name] = struct{}{} + } } if len(got) != len(want) { return fmt.Errorf("mount paths are different, got: %q, want: %q", gotStr, want) diff --git a/pkg/sentry/fs/mounts.go b/pkg/sentry/fs/mounts.go index 574a2cc91..3f2bd0e87 100644 --- a/pkg/sentry/fs/mounts.go +++ b/pkg/sentry/fs/mounts.go @@ -17,13 +17,9 @@ package fs import ( "fmt" "math" - "path" - "strings" "syscall" - "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" - "gvisor.dev/gvisor/pkg/log" "gvisor.dev/gvisor/pkg/refs" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sync" @@ -100,10 +96,14 @@ func newUndoMount(d *Dirent) *Mount { } } -// Root returns the root dirent of this mount. Callers must call DecRef on the -// returned dirent. +// Root returns the root dirent of this mount. +// +// This may return nil if the mount has already been free. Callers must handle this +// case appropriately. If non-nil, callers must call DecRef on the returned *Dirent. func (m *Mount) Root() *Dirent { - m.root.IncRef() + if !m.root.TryIncRef() { + return nil + } return m.root } @@ -269,19 +269,6 @@ func (mns *MountNamespace) DecRef() { mns.DecRefWithDestructor(mns.destroy) } -// Freeze freezes the entire mount tree. -func (mns *MountNamespace) Freeze() { - mns.mu.Lock() - defer mns.mu.Unlock() - - // We only want to freeze Dirents with active references, not Dirents referenced - // by a mount's MountSource. - mns.flushMountSourceRefsLocked() - - // Freeze the entire shebang. - mns.root.Freeze() -} - // withMountLocked prevents further walks to `node`, because `node` is about to // be a mount point. func (mns *MountNamespace) withMountLocked(node *Dirent, fn func() error) error { @@ -634,71 +621,3 @@ func (mns *MountNamespace) SyncAll(ctx context.Context) { defer mns.mu.Unlock() mns.root.SyncAll(ctx) } - -// ResolveExecutablePath resolves the given executable name given a set of -// paths that might contain it. -func (mns *MountNamespace) ResolveExecutablePath(ctx context.Context, wd, name string, paths []string) (string, error) { - // Absolute paths can be used directly. - if path.IsAbs(name) { - return name, nil - } - - // Paths with '/' in them should be joined to the working directory, or - // to the root if working directory is not set. - if strings.IndexByte(name, '/') > 0 { - if wd == "" { - wd = "/" - } - if !path.IsAbs(wd) { - return "", fmt.Errorf("working directory %q must be absolute", wd) - } - return path.Join(wd, name), nil - } - - // Otherwise, We must lookup the name in the paths, starting from the - // calling context's root directory. - root := RootFromContext(ctx) - if root == nil { - // Caller has no root. Don't bother traversing anything. - return "", syserror.ENOENT - } - defer root.DecRef() - for _, p := range paths { - binPath := path.Join(p, name) - traversals := uint(linux.MaxSymlinkTraversals) - d, err := mns.FindInode(ctx, root, nil, binPath, &traversals) - if err == syserror.ENOENT || err == syserror.EACCES { - // Didn't find it here. - continue - } - if err != nil { - return "", err - } - defer d.DecRef() - - // Check that it is a regular file. - if !IsRegular(d.Inode.StableAttr) { - continue - } - - // Check whether we can read and execute the found file. - if err := d.Inode.CheckPermission(ctx, PermMask{Read: true, Execute: true}); err != nil { - log.Infof("Found executable at %q, but user cannot execute it: %v", binPath, err) - continue - } - return path.Join("/", p, name), nil - } - return "", syserror.ENOENT -} - -// GetPath returns the PATH as a slice of strings given the environment -// variables. -func GetPath(env []string) []string { - const prefix = "PATH=" - for _, e := range env { - if strings.HasPrefix(e, prefix) { - return strings.Split(strings.TrimPrefix(e, prefix), ":") - } - } - return nil -} diff --git a/pkg/sentry/fs/proc/BUILD b/pkg/sentry/fs/proc/BUILD index 280093c5e..77c2c5c0e 100644 --- a/pkg/sentry/fs/proc/BUILD +++ b/pkg/sentry/fs/proc/BUILD @@ -36,6 +36,7 @@ go_library( "//pkg/sentry/fs/proc/device", "//pkg/sentry/fs/proc/seqfile", "//pkg/sentry/fs/ramfs", + "//pkg/sentry/fsbridge", "//pkg/sentry/inet", "//pkg/sentry/kernel", "//pkg/sentry/kernel/auth", diff --git a/pkg/sentry/fs/proc/meminfo.go b/pkg/sentry/fs/proc/meminfo.go index 465b47da9..91617267d 100644 --- a/pkg/sentry/fs/proc/meminfo.go +++ b/pkg/sentry/fs/proc/meminfo.go @@ -58,12 +58,16 @@ func (d *meminfoData) ReadSeqFileData(ctx context.Context, h seqfile.SeqHandle) var buf bytes.Buffer fmt.Fprintf(&buf, "MemTotal: %8d kB\n", totalSize/1024) - memFree := (totalSize - totalUsage) / 1024 + memFree := totalSize - totalUsage + if memFree > totalSize { + // Underflow. + memFree = 0 + } // We use MemFree as MemAvailable because we don't swap. // TODO(rahat): When reclaim is implemented the value of MemAvailable // should change. - fmt.Fprintf(&buf, "MemFree: %8d kB\n", memFree) - fmt.Fprintf(&buf, "MemAvailable: %8d kB\n", memFree) + fmt.Fprintf(&buf, "MemFree: %8d kB\n", memFree/1024) + fmt.Fprintf(&buf, "MemAvailable: %8d kB\n", memFree/1024) fmt.Fprintf(&buf, "Buffers: 0 kB\n") // memory usage by block devices fmt.Fprintf(&buf, "Cached: %8d kB\n", (file+snapshot.Tmpfs)/1024) // Emulate a system with no swap, which disables inactivation of anon pages. diff --git a/pkg/sentry/fs/proc/mounts.go b/pkg/sentry/fs/proc/mounts.go index c10888100..1fc9c703c 100644 --- a/pkg/sentry/fs/proc/mounts.go +++ b/pkg/sentry/fs/proc/mounts.go @@ -60,13 +60,15 @@ func forEachMount(t *kernel.Task, fn func(string, *fs.Mount)) { }) for _, m := range ms { mroot := m.Root() + if mroot == nil { + continue // No longer valid. + } mountPath, desc := mroot.FullName(rootDir) mroot.DecRef() if !desc { // MountSources that are not descendants of the chroot jail are ignored. continue } - fn(mountPath, m) } } @@ -91,6 +93,12 @@ func (mif *mountInfoFile) ReadSeqFileData(ctx context.Context, handle seqfile.Se var buf bytes.Buffer forEachMount(mif.t, func(mountPath string, m *fs.Mount) { + mroot := m.Root() + if mroot == nil { + return // No longer valid. + } + defer mroot.DecRef() + // Format: // 36 35 98:0 /mnt1 /mnt2 rw,noatime master:1 - ext3 /dev/root rw,errors=continue // (1)(2)(3) (4) (5) (6) (7) (8) (9) (10) (11) @@ -107,9 +115,6 @@ func (mif *mountInfoFile) ReadSeqFileData(ctx context.Context, handle seqfile.Se // (3) Major:Minor device ID. We don't have a superblock, so we // just use the root inode device number. - mroot := m.Root() - defer mroot.DecRef() - sa := mroot.Inode.StableAttr fmt.Fprintf(&buf, "%d:%d ", sa.DeviceFileMajor, sa.DeviceFileMinor) @@ -165,7 +170,8 @@ func superBlockOpts(mountPath string, msrc *fs.MountSource) string { // NOTE(b/147673608): If the mount is a cgroup, we also need to include // the cgroup name in the options. For now we just read that from the // path. - // TODO(gvisor.dev/issues/190): Once gVisor has full cgroup support, we + // + // TODO(gvisor.dev/issue/190): Once gVisor has full cgroup support, we // should get this value from the cgroup itself, and not rely on the // path. if msrc.FilesystemType == "cgroup" { @@ -207,6 +213,9 @@ func (mf *mountsFile) ReadSeqFileData(ctx context.Context, handle seqfile.SeqHan // // The "needs dump"and fsck flags are always 0, which is allowed. root := m.Root() + if root == nil { + return // No longer valid. + } defer root.DecRef() flags := root.Inode.MountSource.Flags diff --git a/pkg/sentry/fs/proc/net.go b/pkg/sentry/fs/proc/net.go index 6f2775344..bd18177d4 100644 --- a/pkg/sentry/fs/proc/net.go +++ b/pkg/sentry/fs/proc/net.go @@ -40,44 +40,48 @@ import ( // LINT.IfChange -// newNet creates a new proc net entry. -func (p *proc) newNetDir(ctx context.Context, k *kernel.Kernel, msrc *fs.MountSource) *fs.Inode { +// newNetDir creates a new proc net entry. +func newNetDir(t *kernel.Task, msrc *fs.MountSource) *fs.Inode { + k := t.Kernel() + var contents map[string]*fs.Inode - if s := p.k.NetworkStack(); s != nil { + if s := t.NetworkNamespace().Stack(); s != nil { + // TODO(gvisor.dev/issue/1833): Make sure file contents reflect the task + // network namespace. contents = map[string]*fs.Inode{ - "dev": seqfile.NewSeqFileInode(ctx, &netDev{s: s}, msrc), - "snmp": seqfile.NewSeqFileInode(ctx, &netSnmp{s: s}, msrc), + "dev": seqfile.NewSeqFileInode(t, &netDev{s: s}, msrc), + "snmp": seqfile.NewSeqFileInode(t, &netSnmp{s: s}, msrc), // The following files are simple stubs until they are // implemented in netstack, if the file contains a // header the stub is just the header otherwise it is // an empty file. - "arp": newStaticProcInode(ctx, msrc, []byte("IP address HW type Flags HW address Mask Device\n")), + "arp": newStaticProcInode(t, msrc, []byte("IP address HW type Flags HW address Mask Device\n")), - "netlink": newStaticProcInode(ctx, msrc, []byte("sk Eth Pid Groups Rmem Wmem Dump Locks Drops Inode\n")), - "netstat": newStaticProcInode(ctx, msrc, []byte("TcpExt: SyncookiesSent SyncookiesRecv SyncookiesFailed EmbryonicRsts PruneCalled RcvPruned OfoPruned OutOfWindowIcmps LockDroppedIcmps ArpFilter TW TWRecycled TWKilled PAWSPassive PAWSActive PAWSEstab DelayedACKs DelayedACKLocked DelayedACKLost ListenOverflows ListenDrops TCPPrequeued TCPDirectCopyFromBacklog TCPDirectCopyFromPrequeue TCPPrequeueDropped TCPHPHits TCPHPHitsToUser TCPPureAcks TCPHPAcks TCPRenoRecovery TCPSackRecovery TCPSACKReneging TCPFACKReorder TCPSACKReorder TCPRenoReorder TCPTSReorder TCPFullUndo TCPPartialUndo TCPDSACKUndo TCPLossUndo TCPLostRetransmit TCPRenoFailures TCPSackFailures TCPLossFailures TCPFastRetrans TCPForwardRetrans TCPSlowStartRetrans TCPTimeouts TCPLossProbes TCPLossProbeRecovery TCPRenoRecoveryFail TCPSackRecoveryFail TCPSchedulerFailed TCPRcvCollapsed TCPDSACKOldSent TCPDSACKOfoSent TCPDSACKRecv TCPDSACKOfoRecv TCPAbortOnData TCPAbortOnClose TCPAbortOnMemory TCPAbortOnTimeout TCPAbortOnLinger TCPAbortFailed TCPMemoryPressures TCPSACKDiscard TCPDSACKIgnoredOld TCPDSACKIgnoredNoUndo TCPSpuriousRTOs TCPMD5NotFound TCPMD5Unexpected TCPMD5Failure TCPSackShifted TCPSackMerged TCPSackShiftFallback TCPBacklogDrop TCPMinTTLDrop TCPDeferAcceptDrop IPReversePathFilter TCPTimeWaitOverflow TCPReqQFullDoCookies TCPReqQFullDrop TCPRetransFail TCPRcvCoalesce TCPOFOQueue TCPOFODrop TCPOFOMerge TCPChallengeACK TCPSYNChallenge TCPFastOpenActive TCPFastOpenActiveFail TCPFastOpenPassive TCPFastOpenPassiveFail TCPFastOpenListenOverflow TCPFastOpenCookieReqd TCPSpuriousRtxHostQueues BusyPollRxPackets TCPAutoCorking TCPFromZeroWindowAdv TCPToZeroWindowAdv TCPWantZeroWindowAdv TCPSynRetrans TCPOrigDataSent TCPHystartTrainDetect TCPHystartTrainCwnd TCPHystartDelayDetect TCPHystartDelayCwnd TCPACKSkippedSynRecv TCPACKSkippedPAWS TCPACKSkippedSeq TCPACKSkippedFinWait2 TCPACKSkippedTimeWait TCPACKSkippedChallenge TCPWinProbe TCPKeepAlive TCPMTUPFail TCPMTUPSuccess\n")), - "packet": newStaticProcInode(ctx, msrc, []byte("sk RefCnt Type Proto Iface R Rmem User Inode\n")), - "protocols": newStaticProcInode(ctx, msrc, []byte("protocol size sockets memory press maxhdr slab module cl co di ac io in de sh ss gs se re sp bi br ha uh gp em\n")), + "netlink": newStaticProcInode(t, msrc, []byte("sk Eth Pid Groups Rmem Wmem Dump Locks Drops Inode\n")), + "netstat": newStaticProcInode(t, msrc, []byte("TcpExt: SyncookiesSent SyncookiesRecv SyncookiesFailed EmbryonicRsts PruneCalled RcvPruned OfoPruned OutOfWindowIcmps LockDroppedIcmps ArpFilter TW TWRecycled TWKilled PAWSPassive PAWSActive PAWSEstab DelayedACKs DelayedACKLocked DelayedACKLost ListenOverflows ListenDrops TCPPrequeued TCPDirectCopyFromBacklog TCPDirectCopyFromPrequeue TCPPrequeueDropped TCPHPHits TCPHPHitsToUser TCPPureAcks TCPHPAcks TCPRenoRecovery TCPSackRecovery TCPSACKReneging TCPFACKReorder TCPSACKReorder TCPRenoReorder TCPTSReorder TCPFullUndo TCPPartialUndo TCPDSACKUndo TCPLossUndo TCPLostRetransmit TCPRenoFailures TCPSackFailures TCPLossFailures TCPFastRetrans TCPForwardRetrans TCPSlowStartRetrans TCPTimeouts TCPLossProbes TCPLossProbeRecovery TCPRenoRecoveryFail TCPSackRecoveryFail TCPSchedulerFailed TCPRcvCollapsed TCPDSACKOldSent TCPDSACKOfoSent TCPDSACKRecv TCPDSACKOfoRecv TCPAbortOnData TCPAbortOnClose TCPAbortOnMemory TCPAbortOnTimeout TCPAbortOnLinger TCPAbortFailed TCPMemoryPressures TCPSACKDiscard TCPDSACKIgnoredOld TCPDSACKIgnoredNoUndo TCPSpuriousRTOs TCPMD5NotFound TCPMD5Unexpected TCPMD5Failure TCPSackShifted TCPSackMerged TCPSackShiftFallback TCPBacklogDrop TCPMinTTLDrop TCPDeferAcceptDrop IPReversePathFilter TCPTimeWaitOverflow TCPReqQFullDoCookies TCPReqQFullDrop TCPRetransFail TCPRcvCoalesce TCPOFOQueue TCPOFODrop TCPOFOMerge TCPChallengeACK TCPSYNChallenge TCPFastOpenActive TCPFastOpenActiveFail TCPFastOpenPassive TCPFastOpenPassiveFail TCPFastOpenListenOverflow TCPFastOpenCookieReqd TCPSpuriousRtxHostQueues BusyPollRxPackets TCPAutoCorking TCPFromZeroWindowAdv TCPToZeroWindowAdv TCPWantZeroWindowAdv TCPSynRetrans TCPOrigDataSent TCPHystartTrainDetect TCPHystartTrainCwnd TCPHystartDelayDetect TCPHystartDelayCwnd TCPACKSkippedSynRecv TCPACKSkippedPAWS TCPACKSkippedSeq TCPACKSkippedFinWait2 TCPACKSkippedTimeWait TCPACKSkippedChallenge TCPWinProbe TCPKeepAlive TCPMTUPFail TCPMTUPSuccess\n")), + "packet": newStaticProcInode(t, msrc, []byte("sk RefCnt Type Proto Iface R Rmem User Inode\n")), + "protocols": newStaticProcInode(t, msrc, []byte("protocol size sockets memory press maxhdr slab module cl co di ac io in de sh ss gs se re sp bi br ha uh gp em\n")), // Linux sets psched values to: nsec per usec, psched // tick in ns, 1000000, high res timer ticks per sec // (ClockGetres returns 1ns resolution). - "psched": newStaticProcInode(ctx, msrc, []byte(fmt.Sprintf("%08x %08x %08x %08x\n", uint64(time.Microsecond/time.Nanosecond), 64, 1000000, uint64(time.Second/time.Nanosecond)))), - "ptype": newStaticProcInode(ctx, msrc, []byte("Type Device Function\n")), - "route": seqfile.NewSeqFileInode(ctx, &netRoute{s: s}, msrc), - "tcp": seqfile.NewSeqFileInode(ctx, &netTCP{k: k}, msrc), - "udp": seqfile.NewSeqFileInode(ctx, &netUDP{k: k}, msrc), - "unix": seqfile.NewSeqFileInode(ctx, &netUnix{k: k}, msrc), + "psched": newStaticProcInode(t, msrc, []byte(fmt.Sprintf("%08x %08x %08x %08x\n", uint64(time.Microsecond/time.Nanosecond), 64, 1000000, uint64(time.Second/time.Nanosecond)))), + "ptype": newStaticProcInode(t, msrc, []byte("Type Device Function\n")), + "route": seqfile.NewSeqFileInode(t, &netRoute{s: s}, msrc), + "tcp": seqfile.NewSeqFileInode(t, &netTCP{k: k}, msrc), + "udp": seqfile.NewSeqFileInode(t, &netUDP{k: k}, msrc), + "unix": seqfile.NewSeqFileInode(t, &netUnix{k: k}, msrc), } if s.SupportsIPv6() { - contents["if_inet6"] = seqfile.NewSeqFileInode(ctx, &ifinet6{s: s}, msrc) - contents["ipv6_route"] = newStaticProcInode(ctx, msrc, []byte("")) - contents["tcp6"] = seqfile.NewSeqFileInode(ctx, &netTCP6{k: k}, msrc) - contents["udp6"] = newStaticProcInode(ctx, msrc, []byte(" sl local_address remote_address st tx_queue rx_queue tr tm->when retrnsmt uid timeout inode\n")) + contents["if_inet6"] = seqfile.NewSeqFileInode(t, &ifinet6{s: s}, msrc) + contents["ipv6_route"] = newStaticProcInode(t, msrc, []byte("")) + contents["tcp6"] = seqfile.NewSeqFileInode(t, &netTCP6{k: k}, msrc) + contents["udp6"] = newStaticProcInode(t, msrc, []byte(" sl local_address remote_address st tx_queue rx_queue tr tm->when retrnsmt uid timeout inode\n")) } } - d := ramfs.NewDir(ctx, contents, fs.RootOwner, fs.FilePermsFromMode(0555)) - return newProcInode(ctx, d, msrc, fs.SpecialDirectory, nil) + d := ramfs.NewDir(t, contents, fs.RootOwner, fs.FilePermsFromMode(0555)) + return newProcInode(t, d, msrc, fs.SpecialDirectory, t) } // ifinet6 implements seqfile.SeqSource for /proc/net/if_inet6. @@ -834,4 +838,4 @@ func (n *netUDP) ReadSeqFileData(ctx context.Context, h seqfile.SeqHandle) ([]se return data, 0 } -// LINT.ThenChange(../../fsimpl/proc/tasks_net.go) +// LINT.ThenChange(../../fsimpl/proc/task_net.go) diff --git a/pkg/sentry/fs/proc/proc.go b/pkg/sentry/fs/proc/proc.go index c8abb5052..c659224a7 100644 --- a/pkg/sentry/fs/proc/proc.go +++ b/pkg/sentry/fs/proc/proc.go @@ -70,6 +70,7 @@ func New(ctx context.Context, msrc *fs.MountSource, cgroupControllers map[string "loadavg": seqfile.NewSeqFileInode(ctx, &loadavgData{}, msrc), "meminfo": seqfile.NewSeqFileInode(ctx, &meminfoData{k}, msrc), "mounts": newProcInode(ctx, ramfs.NewSymlink(ctx, fs.RootOwner, "self/mounts"), msrc, fs.Symlink, nil), + "net": newProcInode(ctx, ramfs.NewSymlink(ctx, fs.RootOwner, "self/net"), msrc, fs.Symlink, nil), "self": newSelf(ctx, pidns, msrc), "stat": seqfile.NewSeqFileInode(ctx, &statData{k}, msrc), "thread-self": newThreadSelf(ctx, pidns, msrc), @@ -86,7 +87,6 @@ func New(ctx context.Context, msrc *fs.MountSource, cgroupControllers map[string } // Add more contents that need proc to be initialized. - p.AddChild(ctx, "net", p.newNetDir(ctx, k, msrc)) p.AddChild(ctx, "sys", p.newSysDir(ctx, msrc)) return newProcInode(ctx, p, msrc, fs.SpecialDirectory, nil), nil diff --git a/pkg/sentry/fs/proc/sys_net.go b/pkg/sentry/fs/proc/sys_net.go index 0772d4ae4..702fdd392 100644 --- a/pkg/sentry/fs/proc/sys_net.go +++ b/pkg/sentry/fs/proc/sys_net.go @@ -80,7 +80,7 @@ func newTCPMemInode(ctx context.Context, msrc *fs.MountSource, s inet.Stack, dir } // Truncate implements fs.InodeOperations.Truncate. -func (tcpMemInode) Truncate(context.Context, *fs.Inode, int64) error { +func (*tcpMemInode) Truncate(context.Context, *fs.Inode, int64) error { return nil } @@ -196,7 +196,7 @@ func newTCPSackInode(ctx context.Context, msrc *fs.MountSource, s inet.Stack) *f } // Truncate implements fs.InodeOperations.Truncate. -func (tcpSack) Truncate(context.Context, *fs.Inode, int64) error { +func (*tcpSack) Truncate(context.Context, *fs.Inode, int64) error { return nil } @@ -357,7 +357,9 @@ func (p *proc) newSysNetIPv4Dir(ctx context.Context, msrc *fs.MountSource, s ine func (p *proc) newSysNetDir(ctx context.Context, msrc *fs.MountSource) *fs.Inode { var contents map[string]*fs.Inode - if s := p.k.NetworkStack(); s != nil { + // TODO(gvisor.dev/issue/1833): Support for using the network stack in the + // network namespace of the calling process. + if s := p.k.RootNetworkNamespace().Stack(); s != nil { contents = map[string]*fs.Inode{ "ipv4": p.newSysNetIPv4Dir(ctx, msrc, s), "core": p.newSysNetCore(ctx, msrc, s), diff --git a/pkg/sentry/fs/proc/task.go b/pkg/sentry/fs/proc/task.go index ca020e11e..4bbe90198 100644 --- a/pkg/sentry/fs/proc/task.go +++ b/pkg/sentry/fs/proc/task.go @@ -28,6 +28,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/fs/proc/device" "gvisor.dev/gvisor/pkg/sentry/fs/proc/seqfile" "gvisor.dev/gvisor/pkg/sentry/fs/ramfs" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/limits" "gvisor.dev/gvisor/pkg/sentry/mm" @@ -56,14 +57,23 @@ func getTaskMM(t *kernel.Task) (*mm.MemoryManager, error) { return m, nil } +func checkTaskState(t *kernel.Task) error { + switch t.ExitState() { + case kernel.TaskExitZombie: + return syserror.EACCES + case kernel.TaskExitDead: + return syserror.ESRCH + } + return nil +} + // taskDir represents a task-level directory. // // +stateify savable type taskDir struct { ramfs.Dir - t *kernel.Task - pidns *kernel.PIDNamespace + t *kernel.Task } var _ fs.InodeOperations = (*taskDir)(nil) @@ -71,24 +81,27 @@ var _ fs.InodeOperations = (*taskDir)(nil) // newTaskDir creates a new proc task entry. func (p *proc) newTaskDir(t *kernel.Task, msrc *fs.MountSource, isThreadGroup bool) *fs.Inode { contents := map[string]*fs.Inode{ - "auxv": newAuxvec(t, msrc), - "cmdline": newExecArgInode(t, msrc, cmdlineExecArg), - "comm": newComm(t, msrc), - "environ": newExecArgInode(t, msrc, environExecArg), - "exe": newExe(t, msrc), - "fd": newFdDir(t, msrc), - "fdinfo": newFdInfoDir(t, msrc), - "gid_map": newGIDMap(t, msrc), - "io": newIO(t, msrc, isThreadGroup), - "maps": newMaps(t, msrc), - "mountinfo": seqfile.NewSeqFileInode(t, &mountInfoFile{t: t}, msrc), - "mounts": seqfile.NewSeqFileInode(t, &mountsFile{t: t}, msrc), - "ns": newNamespaceDir(t, msrc), - "smaps": newSmaps(t, msrc), - "stat": newTaskStat(t, msrc, isThreadGroup, p.pidns), - "statm": newStatm(t, msrc), - "status": newStatus(t, msrc, p.pidns), - "uid_map": newUIDMap(t, msrc), + "auxv": newAuxvec(t, msrc), + "cmdline": newExecArgInode(t, msrc, cmdlineExecArg), + "comm": newComm(t, msrc), + "environ": newExecArgInode(t, msrc, environExecArg), + "exe": newExe(t, msrc), + "fd": newFdDir(t, msrc), + "fdinfo": newFdInfoDir(t, msrc), + "gid_map": newGIDMap(t, msrc), + "io": newIO(t, msrc, isThreadGroup), + "maps": newMaps(t, msrc), + "mountinfo": seqfile.NewSeqFileInode(t, &mountInfoFile{t: t}, msrc), + "mounts": seqfile.NewSeqFileInode(t, &mountsFile{t: t}, msrc), + "net": newNetDir(t, msrc), + "ns": newNamespaceDir(t, msrc), + "oom_score": newOOMScore(t, msrc), + "oom_score_adj": newOOMScoreAdj(t, msrc), + "smaps": newSmaps(t, msrc), + "stat": newTaskStat(t, msrc, isThreadGroup, p.pidns), + "statm": newStatm(t, msrc), + "status": newStatus(t, msrc, p.pidns), + "uid_map": newUIDMap(t, msrc), } if isThreadGroup { contents["task"] = p.newSubtasks(t, msrc) @@ -249,12 +262,13 @@ func newExe(t *kernel.Task, msrc *fs.MountSource) *fs.Inode { return newProcInode(t, exeSymlink, msrc, fs.Symlink, t) } -func (e *exe) executable() (d *fs.Dirent, err error) { +func (e *exe) executable() (file fsbridge.File, err error) { + if err := checkTaskState(e.t); err != nil { + return nil, err + } e.t.WithMuLocked(func(t *kernel.Task) { mm := t.MemoryManager() if mm == nil { - // TODO(b/34851096): Check shouldn't allow Readlink once the - // Task is zombied. err = syserror.EACCES return } @@ -262,9 +276,9 @@ func (e *exe) executable() (d *fs.Dirent, err error) { // The MemoryManager may be destroyed, in which case // MemoryManager.destroy will simply set the executable to nil // (with locks held). - d = mm.Executable() - if d == nil { - err = syserror.ENOENT + file = mm.Executable() + if file == nil { + err = syserror.ESRCH } }) return @@ -283,15 +297,7 @@ func (e *exe) Readlink(ctx context.Context, inode *fs.Inode) (string, error) { } defer exec.DecRef() - root := fs.RootFromContext(ctx) - if root == nil { - // This doesn't correspond to anything in Linux because the vfs is - // global there. - return "", syserror.EINVAL - } - defer root.DecRef() - n, _ := exec.FullName(root) - return n, nil + return exec.PathnameWithDeleted(ctx), nil } // namespaceSymlink represents a symlink in the namespacefs, such as the files @@ -317,11 +323,22 @@ func newNamespaceSymlink(t *kernel.Task, msrc *fs.MountSource, name string) *fs. return newProcInode(t, n, msrc, fs.Symlink, t) } +// Readlink reads the symlink value. +func (n *namespaceSymlink) Readlink(ctx context.Context, inode *fs.Inode) (string, error) { + if err := checkTaskState(n.t); err != nil { + return "", err + } + return n.Symlink.Readlink(ctx, inode) +} + // Getlink implements fs.InodeOperations.Getlink. func (n *namespaceSymlink) Getlink(ctx context.Context, inode *fs.Inode) (*fs.Dirent, error) { if !kernel.ContextCanTrace(ctx, n.t, false) { return nil, syserror.EACCES } + if err := checkTaskState(n.t); err != nil { + return nil, err + } // Create a new regular file to fake the namespace file. iops := fsutil.NewNoReadWriteFileInode(ctx, fs.RootOwner, fs.FilePermsFromMode(0777), linux.PROC_SUPER_MAGIC) @@ -803,4 +820,95 @@ func (f *auxvecFile) Read(ctx context.Context, _ *fs.File, dst usermem.IOSequenc return int64(n), err } +// newOOMScore returns a oom_score file. It is a stub that always returns 0. +// TODO(gvisor.dev/issue/1967) +func newOOMScore(t *kernel.Task, msrc *fs.MountSource) *fs.Inode { + return newStaticProcInode(t, msrc, []byte("0\n")) +} + +// oomScoreAdj is a file containing the oom_score adjustment for a task. +// +// +stateify savable +type oomScoreAdj struct { + fsutil.SimpleFileInode + + t *kernel.Task +} + +// +stateify savable +type oomScoreAdjFile struct { + fsutil.FileGenericSeek `state:"nosave"` + fsutil.FileNoIoctl `state:"nosave"` + fsutil.FileNoMMap `state:"nosave"` + fsutil.FileNoSplice `state:"nosave"` + fsutil.FileNoopFlush `state:"nosave"` + fsutil.FileNoopFsync `state:"nosave"` + fsutil.FileNoopRelease `state:"nosave"` + fsutil.FileNotDirReaddir `state:"nosave"` + fsutil.FileUseInodeUnstableAttr `state:"nosave"` + waiter.AlwaysReady `state:"nosave"` + + t *kernel.Task +} + +// newOOMScoreAdj returns a oom_score_adj file. +func newOOMScoreAdj(t *kernel.Task, msrc *fs.MountSource) *fs.Inode { + i := &oomScoreAdj{ + SimpleFileInode: *fsutil.NewSimpleFileInode(t, fs.RootOwner, fs.FilePermsFromMode(0644), linux.PROC_SUPER_MAGIC), + t: t, + } + return newProcInode(t, i, msrc, fs.SpecialFile, t) +} + +// Truncate implements fs.InodeOperations.Truncate. Truncate is called when +// O_TRUNC is specified for any kind of existing Dirent but is not called via +// (f)truncate for proc files. +func (*oomScoreAdj) Truncate(context.Context, *fs.Inode, int64) error { + return nil +} + +// GetFile implements fs.InodeOperations.GetFile. +func (o *oomScoreAdj) GetFile(ctx context.Context, dirent *fs.Dirent, flags fs.FileFlags) (*fs.File, error) { + return fs.NewFile(ctx, dirent, flags, &oomScoreAdjFile{t: o.t}), nil +} + +// Read implements fs.FileOperations.Read. +func (f *oomScoreAdjFile) Read(ctx context.Context, _ *fs.File, dst usermem.IOSequence, offset int64) (int64, error) { + if f.t.ExitState() == kernel.TaskExitDead { + return 0, syserror.ESRCH + } + var buf bytes.Buffer + fmt.Fprintf(&buf, "%d\n", f.t.OOMScoreAdj()) + if offset >= int64(buf.Len()) { + return 0, io.EOF + } + n, err := dst.CopyOut(ctx, buf.Bytes()[offset:]) + return int64(n), err +} + +// Write implements fs.FileOperations.Write. +func (f *oomScoreAdjFile) Write(ctx context.Context, _ *fs.File, src usermem.IOSequence, offset int64) (int64, error) { + if src.NumBytes() == 0 { + return 0, nil + } + + // Limit input size so as not to impact performance if input size is large. + src = src.TakeFirst(usermem.PageSize - 1) + + var v int32 + n, err := usermem.CopyInt32StringInVec(ctx, src.IO, src.Addrs, &v, src.Opts) + if err != nil { + return 0, err + } + + if f.t.ExitState() == kernel.TaskExitDead { + return 0, syserror.ESRCH + } + if err := f.t.SetOOMScoreAdj(v); err != nil { + return 0, err + } + + return n, nil +} + // LINT.ThenChange(../../fsimpl/proc/task.go|../../fsimpl/proc/task_files.go) diff --git a/pkg/sentry/fs/tmpfs/fs.go b/pkg/sentry/fs/tmpfs/fs.go index d5be56c3f..bc117ca6a 100644 --- a/pkg/sentry/fs/tmpfs/fs.go +++ b/pkg/sentry/fs/tmpfs/fs.go @@ -44,9 +44,6 @@ const ( // lookup. cacheRevalidate = "revalidate" - // TODO(edahlgren/mpratt): support a tmpfs size limit. - // size = "size" - // Permissions that exceed modeMask will be rejected. modeMask = 01777 diff --git a/pkg/sentry/fs/tmpfs/inode_file.go b/pkg/sentry/fs/tmpfs/inode_file.go index 25abbc151..1dc75291d 100644 --- a/pkg/sentry/fs/tmpfs/inode_file.go +++ b/pkg/sentry/fs/tmpfs/inode_file.go @@ -39,7 +39,7 @@ var ( opensRO = metric.MustCreateNewUint64Metric("/in_memory_file/opens_ro", false /* sync */, "Number of times an in-memory file was opened in read-only mode.") opensW = metric.MustCreateNewUint64Metric("/in_memory_file/opens_w", false /* sync */, "Number of times an in-memory file was opened in write mode.") reads = metric.MustCreateNewUint64Metric("/in_memory_file/reads", false /* sync */, "Number of in-memory file reads.") - readWait = metric.MustCreateNewUint64Metric("/in_memory_file/read_wait", false /* sync */, "Time waiting on in-memory file reads, in nanoseconds.") + readWait = metric.MustCreateNewUint64NanosecondsMetric("/in_memory_file/read_wait", false /* sync */, "Time waiting on in-memory file reads, in nanoseconds.") ) // fileInodeOperations implements fs.InodeOperations for a regular tmpfs file. diff --git a/pkg/sentry/fs/tmpfs/tmpfs.go b/pkg/sentry/fs/tmpfs/tmpfs.go index 3c2b583ae..b095312fe 100644 --- a/pkg/sentry/fs/tmpfs/tmpfs.go +++ b/pkg/sentry/fs/tmpfs/tmpfs.go @@ -39,14 +39,13 @@ var fsInfo = fs.Info{ // rename implements fs.InodeOperations.Rename for tmpfs nodes. func rename(ctx context.Context, oldParent *fs.Inode, oldName string, newParent *fs.Inode, newName string, replacement bool) error { - op, ok := oldParent.InodeOperations.(*Dir) - if !ok { - return syserror.EXDEV - } - np, ok := newParent.InodeOperations.(*Dir) - if !ok { + // Don't allow renames across different mounts. + if newParent.MountSource != oldParent.MountSource { return syserror.EXDEV } + + op := oldParent.InodeOperations.(*Dir) + np := newParent.InodeOperations.(*Dir) return ramfs.Rename(ctx, op.ramfsDir, oldName, np.ramfsDir, newName, replacement) } diff --git a/pkg/sentry/fs/tty/line_discipline.go b/pkg/sentry/fs/tty/line_discipline.go index 12b1c6097..2e9dd2d55 100644 --- a/pkg/sentry/fs/tty/line_discipline.go +++ b/pkg/sentry/fs/tty/line_discipline.go @@ -27,6 +27,8 @@ import ( "gvisor.dev/gvisor/pkg/waiter" ) +// LINT.IfChange + const ( // canonMaxBytes is the number of bytes that fit into a single line of // terminal input in canonical mode. This corresponds to N_TTY_BUF_SIZE @@ -443,3 +445,5 @@ func (l *lineDiscipline) peek(b []byte) int { } return size } + +// LINT.ThenChange(../../fsimpl/devpts/line_discipline.go) diff --git a/pkg/sentry/fs/tty/master.go b/pkg/sentry/fs/tty/master.go index f62da49bd..fe07fa929 100644 --- a/pkg/sentry/fs/tty/master.go +++ b/pkg/sentry/fs/tty/master.go @@ -26,6 +26,8 @@ import ( "gvisor.dev/gvisor/pkg/waiter" ) +// LINT.IfChange + // masterInodeOperations are the fs.InodeOperations for the master end of the // Terminal (ptmx file). // @@ -232,3 +234,5 @@ func maybeEmitUnimplementedEvent(ctx context.Context, cmd uint32) { unimpl.EmitUnimplementedEvent(ctx) } } + +// LINT.ThenChange(../../fsimpl/devpts/master.go) diff --git a/pkg/sentry/fs/tty/queue.go b/pkg/sentry/fs/tty/queue.go index 1ca79c0b2..ceabb9b1e 100644 --- a/pkg/sentry/fs/tty/queue.go +++ b/pkg/sentry/fs/tty/queue.go @@ -25,6 +25,8 @@ import ( "gvisor.dev/gvisor/pkg/waiter" ) +// LINT.IfChange + // waitBufMaxBytes is the maximum size of a wait buffer. It is based on // TTYB_DEFAULT_MEM_LIMIT. const waitBufMaxBytes = 131072 @@ -234,3 +236,5 @@ func (q *queue) waitBufAppend(b []byte) { q.waitBuf = append(q.waitBuf, b) q.waitBufLen += uint64(len(b)) } + +// LINT.ThenChange(../../fsimpl/devpts/queue.go) diff --git a/pkg/sentry/fs/tty/slave.go b/pkg/sentry/fs/tty/slave.go index 6a2dbc576..9871f6fc6 100644 --- a/pkg/sentry/fs/tty/slave.go +++ b/pkg/sentry/fs/tty/slave.go @@ -25,6 +25,8 @@ import ( "gvisor.dev/gvisor/pkg/waiter" ) +// LINT.IfChange + // slaveInodeOperations are the fs.InodeOperations for the slave end of the // Terminal (pts file). // @@ -172,3 +174,5 @@ func (sf *slaveFileOperations) Ioctl(ctx context.Context, _ *fs.File, io usermem return 0, syserror.ENOTTY } } + +// LINT.ThenChange(../../fsimpl/devpts/slave.go) diff --git a/pkg/sentry/fs/tty/terminal.go b/pkg/sentry/fs/tty/terminal.go index 5883f26db..ddcccf4da 100644 --- a/pkg/sentry/fs/tty/terminal.go +++ b/pkg/sentry/fs/tty/terminal.go @@ -23,6 +23,8 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) +// LINT.IfChange + // Terminal is a pseudoterminal. // // +stateify savable @@ -126,3 +128,5 @@ func (tm *Terminal) tty(isMaster bool) *kernel.TTY { } return tm.slaveKTTY } + +// LINT.ThenChange(../../fsimpl/devpts/terminal.go) diff --git a/pkg/sentry/fs/user/BUILD b/pkg/sentry/fs/user/BUILD new file mode 100644 index 000000000..bd5dac373 --- /dev/null +++ b/pkg/sentry/fs/user/BUILD @@ -0,0 +1,39 @@ +load("//tools:defs.bzl", "go_library", "go_test") + +package(licenses = ["notice"]) + +go_library( + name = "user", + srcs = [ + "path.go", + "user.go", + ], + visibility = ["//pkg/sentry:internal"], + deps = [ + "//pkg/abi/linux", + "//pkg/context", + "//pkg/fspath", + "//pkg/log", + "//pkg/sentry/fs", + "//pkg/sentry/kernel/auth", + "//pkg/sentry/vfs", + "//pkg/syserror", + "//pkg/usermem", + ], +) + +go_test( + name = "user_test", + size = "small", + srcs = ["user_test.go"], + library = ":user", + deps = [ + "//pkg/abi/linux", + "//pkg/context", + "//pkg/sentry/fs", + "//pkg/sentry/fs/tmpfs", + "//pkg/sentry/kernel/auth", + "//pkg/sentry/kernel/contexttest", + "//pkg/usermem", + ], +) diff --git a/pkg/sentry/fs/user/path.go b/pkg/sentry/fs/user/path.go new file mode 100644 index 000000000..fbd4547a7 --- /dev/null +++ b/pkg/sentry/fs/user/path.go @@ -0,0 +1,169 @@ +// 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 user + +import ( + "fmt" + "path" + "strings" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/log" + "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" +) + +// ResolveExecutablePath resolves the given executable name given the working +// dir and environment. +func ResolveExecutablePath(ctx context.Context, creds *auth.Credentials, mns *fs.MountNamespace, envv []string, wd, name string) (string, error) { + // Absolute paths can be used directly. + if path.IsAbs(name) { + return name, nil + } + + // Paths with '/' in them should be joined to the working directory, or + // to the root if working directory is not set. + if strings.IndexByte(name, '/') > 0 { + if wd == "" { + wd = "/" + } + if !path.IsAbs(wd) { + return "", fmt.Errorf("working directory %q must be absolute", wd) + } + return path.Join(wd, name), nil + } + + // Otherwise, We must lookup the name in the paths, starting from the + // calling context's root directory. + paths := getPath(envv) + + root := fs.RootFromContext(ctx) + if root == nil { + // Caller has no root. Don't bother traversing anything. + return "", syserror.ENOENT + } + defer root.DecRef() + for _, p := range paths { + if !path.IsAbs(p) { + // Relative paths aren't safe, no one should be using them. + log.Warningf("Skipping relative path %q in $PATH", p) + continue + } + + binPath := path.Join(p, name) + traversals := uint(linux.MaxSymlinkTraversals) + d, err := mns.FindInode(ctx, root, nil, binPath, &traversals) + if err == syserror.ENOENT || err == syserror.EACCES { + // Didn't find it here. + continue + } + if err != nil { + return "", err + } + defer d.DecRef() + + // Check that it is a regular file. + if !fs.IsRegular(d.Inode.StableAttr) { + continue + } + + // Check whether we can read and execute the found file. + if err := d.Inode.CheckPermission(ctx, fs.PermMask{Read: true, Execute: true}); err != nil { + log.Infof("Found executable at %q, but user cannot execute it: %v", binPath, err) + continue + } + return path.Join("/", p, name), nil + } + + // Couldn't find it. + return "", syserror.ENOENT +} + +// ResolveExecutablePathVFS2 resolves the given executable name given the +// working dir and environment. +func ResolveExecutablePathVFS2(ctx context.Context, creds *auth.Credentials, mns *vfs.MountNamespace, envv []string, wd, name string) (string, error) { + // Absolute paths can be used directly. + if path.IsAbs(name) { + return name, nil + } + + // Paths with '/' in them should be joined to the working directory, or + // to the root if working directory is not set. + if strings.IndexByte(name, '/') > 0 { + if wd == "" { + wd = "/" + } + if !path.IsAbs(wd) { + return "", fmt.Errorf("working directory %q must be absolute", wd) + } + return path.Join(wd, name), nil + } + + // Otherwise, We must lookup the name in the paths, starting from the + // calling context's root directory. + paths := getPath(envv) + + root := mns.Root() + defer root.DecRef() + for _, p := range paths { + if !path.IsAbs(p) { + // Relative paths aren't safe, no one should be using them. + log.Warningf("Skipping relative path %q in $PATH", p) + continue + } + + binPath := path.Join(p, name) + pop := &vfs.PathOperation{ + Root: root, + Start: root, + Path: fspath.Parse(binPath), + FollowFinalSymlink: true, + } + opts := &vfs.OpenOptions{ + FileExec: true, + Flags: linux.O_RDONLY, + } + dentry, err := root.Mount().Filesystem().VirtualFilesystem().OpenAt(ctx, creds, pop, opts) + if err == syserror.ENOENT || err == syserror.EACCES { + // Didn't find it here. + continue + } + if err != nil { + return "", err + } + dentry.DecRef() + + return binPath, nil + } + + // Couldn't find it. + return "", syserror.ENOENT +} + +// getPath returns the PATH as a slice of strings given the environment +// variables. +func getPath(env []string) []string { + const prefix = "PATH=" + for _, e := range env { + if strings.HasPrefix(e, prefix) { + return strings.Split(strings.TrimPrefix(e, prefix), ":") + } + } + return nil +} diff --git a/pkg/sentry/fs/user/user.go b/pkg/sentry/fs/user/user.go new file mode 100644 index 000000000..f4d525523 --- /dev/null +++ b/pkg/sentry/fs/user/user.go @@ -0,0 +1,239 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// Package user contains methods for resolving filesystem paths based on the +// user and their environment. +package user + +import ( + "bufio" + "fmt" + "io" + "strconv" + "strings" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/usermem" +) + +type fileReader struct { + // Ctx is the context for the file reader. + Ctx context.Context + + // File is the file to read from. + File *fs.File +} + +// Read implements io.Reader.Read. +func (r *fileReader) Read(buf []byte) (int, error) { + n, err := r.File.Readv(r.Ctx, usermem.BytesIOSequence(buf)) + return int(n), err +} + +// getExecUserHome returns the home directory of the executing user read from +// /etc/passwd as read from the container filesystem. +func getExecUserHome(ctx context.Context, rootMns *fs.MountNamespace, uid auth.KUID) (string, error) { + // The default user home directory to return if no user matching the user + // if found in the /etc/passwd found in the image. + const defaultHome = "/" + + // Open the /etc/passwd file from the dirent via the root mount namespace. + mnsRoot := rootMns.Root() + maxTraversals := uint(linux.MaxSymlinkTraversals) + dirent, err := rootMns.FindInode(ctx, mnsRoot, nil, "/etc/passwd", &maxTraversals) + if err != nil { + // NOTE: Ignore errors opening the passwd file. If the passwd file + // doesn't exist we will return the default home directory. + return defaultHome, nil + } + defer dirent.DecRef() + + // Check read permissions on the file. + if err := dirent.Inode.CheckPermission(ctx, fs.PermMask{Read: true}); err != nil { + // NOTE: Ignore permissions errors here and return default root dir. + return defaultHome, nil + } + + // Only open regular files. We don't open other files like named pipes as + // they may block and might present some attack surface to the container. + // Note that runc does not seem to do this kind of checking. + if !fs.IsRegular(dirent.Inode.StableAttr) { + return defaultHome, nil + } + + f, err := dirent.Inode.GetFile(ctx, dirent, fs.FileFlags{Read: true, Directory: false}) + if err != nil { + return "", err + } + defer f.DecRef() + + r := &fileReader{ + Ctx: ctx, + File: f, + } + + return findHomeInPasswd(uint32(uid), r, defaultHome) +} + +type fileReaderVFS2 struct { + ctx context.Context + fd *vfs.FileDescription +} + +func (r *fileReaderVFS2) Read(buf []byte) (int, error) { + n, err := r.fd.Read(r.ctx, usermem.BytesIOSequence(buf), vfs.ReadOptions{}) + return int(n), err +} + +func getExecUserHomeVFS2(ctx context.Context, mns *vfs.MountNamespace, uid auth.KUID) (string, error) { + const defaultHome = "/" + + root := mns.Root() + defer root.DecRef() + + creds := auth.CredentialsFromContext(ctx) + + target := &vfs.PathOperation{ + Root: root, + Start: root, + Path: fspath.Parse("/etc/passwd"), + } + + opts := &vfs.OpenOptions{ + Flags: linux.O_RDONLY, + } + + fd, err := root.Mount().Filesystem().VirtualFilesystem().OpenAt(ctx, creds, target, opts) + if err != nil { + return defaultHome, nil + } + defer fd.DecRef() + + r := &fileReaderVFS2{ + ctx: ctx, + fd: fd, + } + + homeDir, err := findHomeInPasswd(uint32(uid), r, defaultHome) + if err != nil { + return "", err + } + + return homeDir, nil +} + +// MaybeAddExecUserHome returns a new slice with the HOME enviroment variable +// set if the slice does not already contain it, otherwise it returns the +// original slice unmodified. +func MaybeAddExecUserHome(ctx context.Context, mns *fs.MountNamespace, uid auth.KUID, envv []string) ([]string, error) { + // Check if the envv already contains HOME. + for _, env := range envv { + if strings.HasPrefix(env, "HOME=") { + // We have it. Return the original slice unmodified. + return envv, nil + } + } + + // Read /etc/passwd for the user's HOME directory and set the HOME + // environment variable as required by POSIX if it is not overridden by + // the user. + homeDir, err := getExecUserHome(ctx, mns, uid) + if err != nil { + return nil, fmt.Errorf("error reading exec user: %v", err) + } + + return append(envv, "HOME="+homeDir), nil +} + +// MaybeAddExecUserHomeVFS2 returns a new slice with the HOME enviroment +// variable set if the slice does not already contain it, otherwise it returns +// the original slice unmodified. +func MaybeAddExecUserHomeVFS2(ctx context.Context, vmns *vfs.MountNamespace, uid auth.KUID, envv []string) ([]string, error) { + // Check if the envv already contains HOME. + for _, env := range envv { + if strings.HasPrefix(env, "HOME=") { + // We have it. Return the original slice unmodified. + return envv, nil + } + } + + // Read /etc/passwd for the user's HOME directory and set the HOME + // environment variable as required by POSIX if it is not overridden by + // the user. + homeDir, err := getExecUserHomeVFS2(ctx, vmns, uid) + if err != nil { + return nil, fmt.Errorf("error reading exec user: %v", err) + } + return append(envv, "HOME="+homeDir), nil +} + +// findHomeInPasswd parses a passwd file and returns the given user's home +// directory. This function does it's best to replicate the runc's behavior. +func findHomeInPasswd(uid uint32, passwd io.Reader, defaultHome string) (string, error) { + s := bufio.NewScanner(passwd) + + for s.Scan() { + if err := s.Err(); err != nil { + return "", err + } + + line := strings.TrimSpace(s.Text()) + if line == "" { + continue + } + + // Pull out part of passwd entry. Loosely parse the passwd entry as some + // passwd files could be poorly written and for compatibility with runc. + // + // Per 'man 5 passwd' + // /etc/passwd contains one line for each user account, with seven + // fields delimited by colons (“:”). These fields are: + // + // - login name + // - optional encrypted password + // - numerical user ID + // - numerical group ID + // - user name or comment field + // - user home directory + // - optional user command interpreter + parts := strings.Split(line, ":") + + found := false + homeDir := "" + for i, p := range parts { + switch i { + case 2: + parsedUID, err := strconv.ParseUint(p, 10, 32) + if err == nil && parsedUID == uint64(uid) { + found = true + } + case 5: + homeDir = p + } + } + if found { + // NOTE: If the uid is present but the home directory is not + // present in the /etc/passwd entry we return an empty string. This + // is, for better or worse, what runc does. + return homeDir, nil + } + } + + return defaultHome, nil +} diff --git a/pkg/sentry/fs/user/user_test.go b/pkg/sentry/fs/user/user_test.go new file mode 100644 index 000000000..7d8e9ac7c --- /dev/null +++ b/pkg/sentry/fs/user/user_test.go @@ -0,0 +1,198 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package user + +import ( + "fmt" + "strings" + "testing" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/fs/tmpfs" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/kernel/contexttest" + "gvisor.dev/gvisor/pkg/usermem" +) + +// createEtcPasswd creates /etc/passwd with the given contents and mode. If +// mode is empty, then no file will be created. If mode is not a regular file +// mode, then contents is ignored. +func createEtcPasswd(ctx context.Context, root *fs.Dirent, contents string, mode linux.FileMode) error { + if err := root.CreateDirectory(ctx, root, "etc", fs.FilePermsFromMode(0755)); err != nil { + return err + } + etc, err := root.Walk(ctx, root, "etc") + if err != nil { + return err + } + defer etc.DecRef() + switch mode.FileType() { + case 0: + // Don't create anything. + return nil + case linux.S_IFREG: + passwd, err := etc.Create(ctx, root, "passwd", fs.FileFlags{Write: true}, fs.FilePermsFromMode(mode)) + if err != nil { + return err + } + defer passwd.DecRef() + if _, err := passwd.Writev(ctx, usermem.BytesIOSequence([]byte(contents))); err != nil { + return err + } + return nil + case linux.S_IFDIR: + return etc.CreateDirectory(ctx, root, "passwd", fs.FilePermsFromMode(mode)) + case linux.S_IFIFO: + return etc.CreateFifo(ctx, root, "passwd", fs.FilePermsFromMode(mode)) + default: + return fmt.Errorf("unknown file type %x", mode.FileType()) + } +} + +// TestGetExecUserHome tests the getExecUserHome function. +func TestGetExecUserHome(t *testing.T) { + tests := map[string]struct { + uid auth.KUID + passwdContents string + passwdMode linux.FileMode + expected string + }{ + "success": { + uid: 1000, + passwdContents: "adin::1000:1111::/home/adin:/bin/sh", + passwdMode: linux.S_IFREG | 0666, + expected: "/home/adin", + }, + "no_perms": { + uid: 1000, + passwdContents: "adin::1000:1111::/home/adin:/bin/sh", + passwdMode: linux.S_IFREG, + expected: "/", + }, + "no_passwd": { + uid: 1000, + expected: "/", + }, + "directory": { + uid: 1000, + passwdMode: linux.S_IFDIR | 0666, + expected: "/", + }, + // Currently we don't allow named pipes. + "named_pipe": { + uid: 1000, + passwdMode: linux.S_IFIFO | 0666, + expected: "/", + }, + } + + for name, tc := range tests { + t.Run(name, func(t *testing.T) { + ctx := contexttest.Context(t) + msrc := fs.NewPseudoMountSource(ctx) + rootInode := tmpfs.NewDir(ctx, nil, fs.RootOwner, fs.FilePermsFromMode(0777), msrc) + + mns, err := fs.NewMountNamespace(ctx, rootInode) + if err != nil { + t.Fatalf("NewMountNamespace failed: %v", err) + } + defer mns.DecRef() + root := mns.Root() + defer root.DecRef() + ctx = fs.WithRoot(ctx, root) + + if err := createEtcPasswd(ctx, root, tc.passwdContents, tc.passwdMode); err != nil { + t.Fatalf("createEtcPasswd failed: %v", err) + } + + got, err := getExecUserHome(ctx, mns, tc.uid) + if err != nil { + t.Fatalf("failed to get user home: %v", err) + } + + if got != tc.expected { + t.Fatalf("expected %v, got: %v", tc.expected, got) + } + }) + } +} + +// TestFindHomeInPasswd tests the findHomeInPasswd function's passwd file parsing. +func TestFindHomeInPasswd(t *testing.T) { + tests := map[string]struct { + uid uint32 + passwd string + expected string + def string + }{ + "empty": { + uid: 1000, + passwd: "", + expected: "/", + def: "/", + }, + "whitespace": { + uid: 1000, + passwd: " ", + expected: "/", + def: "/", + }, + "full": { + uid: 1000, + passwd: "adin::1000:1111::/home/adin:/bin/sh", + expected: "/home/adin", + def: "/", + }, + // For better or worse, this is how runc works. + "partial": { + uid: 1000, + passwd: "adin::1000:1111:", + expected: "", + def: "/", + }, + "multiple": { + uid: 1001, + passwd: "adin::1000:1111::/home/adin:/bin/sh\nian::1001:1111::/home/ian:/bin/sh", + expected: "/home/ian", + def: "/", + }, + "duplicate": { + uid: 1000, + passwd: "adin::1000:1111::/home/adin:/bin/sh\nian::1000:1111::/home/ian:/bin/sh", + expected: "/home/adin", + def: "/", + }, + "empty_lines": { + uid: 1001, + passwd: "adin::1000:1111::/home/adin:/bin/sh\n\n\nian::1001:1111::/home/ian:/bin/sh", + expected: "/home/ian", + def: "/", + }, + } + + for name, tc := range tests { + t.Run(name, func(t *testing.T) { + got, err := findHomeInPasswd(tc.uid, strings.NewReader(tc.passwd), tc.def) + if err != nil { + t.Fatalf("error parsing passwd: %v", err) + } + if tc.expected != got { + t.Fatalf("expected %v, got: %v", tc.expected, got) + } + }) + } +} diff --git a/pkg/sentry/fsbridge/BUILD b/pkg/sentry/fsbridge/BUILD new file mode 100644 index 000000000..6c798f0bd --- /dev/null +++ b/pkg/sentry/fsbridge/BUILD @@ -0,0 +1,24 @@ +load("//tools:defs.bzl", "go_library") + +licenses(["notice"]) + +go_library( + name = "fsbridge", + srcs = [ + "bridge.go", + "fs.go", + "vfs.go", + ], + visibility = ["//pkg/sentry:internal"], + deps = [ + "//pkg/abi/linux", + "//pkg/context", + "//pkg/fspath", + "//pkg/sentry/fs", + "//pkg/sentry/kernel/auth", + "//pkg/sentry/memmap", + "//pkg/sentry/vfs", + "//pkg/syserror", + "//pkg/usermem", + ], +) diff --git a/pkg/sentry/fsbridge/bridge.go b/pkg/sentry/fsbridge/bridge.go new file mode 100644 index 000000000..8e7590721 --- /dev/null +++ b/pkg/sentry/fsbridge/bridge.go @@ -0,0 +1,54 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// Package fsbridge provides common interfaces to bridge between VFS1 and VFS2 +// files. +package fsbridge + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/memmap" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/usermem" +) + +// File provides a common interface to bridge between VFS1 and VFS2 files. +type File interface { + // PathnameWithDeleted returns an absolute pathname to vd, consistent with + // Linux's d_path(). In particular, if vd.Dentry() has been disowned, + // PathnameWithDeleted appends " (deleted)" to the returned pathname. + PathnameWithDeleted(ctx context.Context) string + + // ReadFull read all contents from the file. + ReadFull(ctx context.Context, dst usermem.IOSequence, offset int64) (int64, error) + + // ConfigureMMap mutates opts to implement mmap(2) for the file. + ConfigureMMap(context.Context, *memmap.MMapOpts) error + + // Type returns the file type, e.g. linux.S_IFREG. + Type(context.Context) (linux.FileMode, error) + + // IncRef increments reference. + IncRef() + + // DecRef decrements reference. + DecRef() +} + +// Lookup provides a common interface to open files. +type Lookup interface { + // OpenPath opens a file. + OpenPath(ctx context.Context, path string, opts vfs.OpenOptions, remainingTraversals *uint, resolveFinal bool) (File, error) +} diff --git a/pkg/sentry/fsbridge/fs.go b/pkg/sentry/fsbridge/fs.go new file mode 100644 index 000000000..093ce1fb3 --- /dev/null +++ b/pkg/sentry/fsbridge/fs.go @@ -0,0 +1,181 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package fsbridge + +import ( + "io" + "strings" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/memmap" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +// fsFile implements File interface over fs.File. +// +// +stateify savable +type fsFile struct { + file *fs.File +} + +var _ File = (*fsFile)(nil) + +// NewFSFile creates a new File over fs.File. +func NewFSFile(file *fs.File) File { + return &fsFile{file: file} +} + +// PathnameWithDeleted implements File. +func (f *fsFile) PathnameWithDeleted(ctx context.Context) string { + root := fs.RootFromContext(ctx) + if root == nil { + // This doesn't correspond to anything in Linux because the vfs is + // global there. + return "" + } + defer root.DecRef() + + name, _ := f.file.Dirent.FullName(root) + return name +} + +// ReadFull implements File. +func (f *fsFile) ReadFull(ctx context.Context, dst usermem.IOSequence, offset int64) (int64, error) { + var total int64 + for dst.NumBytes() > 0 { + n, err := f.file.Preadv(ctx, dst, offset+total) + total += n + if err == io.EOF && total != 0 { + return total, io.ErrUnexpectedEOF + } else if err != nil { + return total, err + } + dst = dst.DropFirst64(n) + } + return total, nil +} + +// ConfigureMMap implements File. +func (f *fsFile) ConfigureMMap(ctx context.Context, opts *memmap.MMapOpts) error { + return f.file.ConfigureMMap(ctx, opts) +} + +// Type implements File. +func (f *fsFile) Type(context.Context) (linux.FileMode, error) { + return linux.FileMode(f.file.Dirent.Inode.StableAttr.Type.LinuxType()), nil +} + +// IncRef implements File. +func (f *fsFile) IncRef() { + f.file.IncRef() +} + +// DecRef implements File. +func (f *fsFile) DecRef() { + f.file.DecRef() +} + +// fsLookup implements Lookup interface using fs.File. +// +// +stateify savable +type fsLookup struct { + mntns *fs.MountNamespace + + root *fs.Dirent + workingDir *fs.Dirent +} + +var _ Lookup = (*fsLookup)(nil) + +// NewFSLookup creates a new Lookup using VFS1. +func NewFSLookup(mntns *fs.MountNamespace, root, workingDir *fs.Dirent) Lookup { + return &fsLookup{ + mntns: mntns, + root: root, + workingDir: workingDir, + } +} + +// OpenPath implements Lookup. +func (l *fsLookup) OpenPath(ctx context.Context, path string, opts vfs.OpenOptions, remainingTraversals *uint, resolveFinal bool) (File, error) { + var d *fs.Dirent + var err error + if resolveFinal { + d, err = l.mntns.FindInode(ctx, l.root, l.workingDir, path, remainingTraversals) + } else { + d, err = l.mntns.FindLink(ctx, l.root, l.workingDir, path, remainingTraversals) + } + if err != nil { + return nil, err + } + defer d.DecRef() + + if !resolveFinal && fs.IsSymlink(d.Inode.StableAttr) { + return nil, syserror.ELOOP + } + + fsPerm := openOptionsToPermMask(&opts) + if err := d.Inode.CheckPermission(ctx, fsPerm); err != nil { + return nil, err + } + + // If they claim it's a directory, then make sure. + if strings.HasSuffix(path, "/") { + if d.Inode.StableAttr.Type != fs.Directory { + return nil, syserror.ENOTDIR + } + } + + if opts.FileExec && d.Inode.StableAttr.Type != fs.RegularFile { + ctx.Infof("%q is not a regular file: %v", path, d.Inode.StableAttr.Type) + return nil, syserror.EACCES + } + + f, err := d.Inode.GetFile(ctx, d, flagsToFileFlags(opts.Flags)) + if err != nil { + return nil, err + } + + return &fsFile{file: f}, nil +} + +func openOptionsToPermMask(opts *vfs.OpenOptions) fs.PermMask { + mode := opts.Flags & linux.O_ACCMODE + return fs.PermMask{ + Read: mode == linux.O_RDONLY || mode == linux.O_RDWR, + Write: mode == linux.O_WRONLY || mode == linux.O_RDWR, + Execute: opts.FileExec, + } +} + +func flagsToFileFlags(flags uint32) fs.FileFlags { + return fs.FileFlags{ + Direct: flags&linux.O_DIRECT != 0, + DSync: flags&(linux.O_DSYNC|linux.O_SYNC) != 0, + Sync: flags&linux.O_SYNC != 0, + NonBlocking: flags&linux.O_NONBLOCK != 0, + Read: (flags & linux.O_ACCMODE) != linux.O_WRONLY, + Write: (flags & linux.O_ACCMODE) != linux.O_RDONLY, + Append: flags&linux.O_APPEND != 0, + Directory: flags&linux.O_DIRECTORY != 0, + Async: flags&linux.O_ASYNC != 0, + LargeFile: flags&linux.O_LARGEFILE != 0, + Truncate: flags&linux.O_TRUNC != 0, + } +} diff --git a/pkg/sentry/fsbridge/vfs.go b/pkg/sentry/fsbridge/vfs.go new file mode 100644 index 000000000..89168220a --- /dev/null +++ b/pkg/sentry/fsbridge/vfs.go @@ -0,0 +1,142 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package fsbridge + +import ( + "io" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/memmap" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/usermem" +) + +// VFSFile implements File interface over vfs.FileDescription. +// +// +stateify savable +type VFSFile struct { + file *vfs.FileDescription +} + +var _ File = (*VFSFile)(nil) + +// NewVFSFile creates a new File over fs.File. +func NewVFSFile(file *vfs.FileDescription) File { + return &VFSFile{file: file} +} + +// PathnameWithDeleted implements File. +func (f *VFSFile) PathnameWithDeleted(ctx context.Context) string { + root := vfs.RootFromContext(ctx) + defer root.DecRef() + + vfsObj := f.file.VirtualDentry().Mount().Filesystem().VirtualFilesystem() + name, _ := vfsObj.PathnameWithDeleted(ctx, root, f.file.VirtualDentry()) + return name +} + +// ReadFull implements File. +func (f *VFSFile) ReadFull(ctx context.Context, dst usermem.IOSequence, offset int64) (int64, error) { + var total int64 + for dst.NumBytes() > 0 { + n, err := f.file.PRead(ctx, dst, offset+total, vfs.ReadOptions{}) + total += n + if err == io.EOF && total != 0 { + return total, io.ErrUnexpectedEOF + } else if err != nil { + return total, err + } + dst = dst.DropFirst64(n) + } + return total, nil +} + +// ConfigureMMap implements File. +func (f *VFSFile) ConfigureMMap(ctx context.Context, opts *memmap.MMapOpts) error { + return f.file.ConfigureMMap(ctx, opts) +} + +// Type implements File. +func (f *VFSFile) Type(ctx context.Context) (linux.FileMode, error) { + stat, err := f.file.Stat(ctx, vfs.StatOptions{}) + if err != nil { + return 0, err + } + return linux.FileMode(stat.Mode).FileType(), nil +} + +// IncRef implements File. +func (f *VFSFile) IncRef() { + f.file.IncRef() +} + +// DecRef implements File. +func (f *VFSFile) DecRef() { + f.file.DecRef() +} + +// FileDescription returns the FileDescription represented by f. It does not +// take an additional reference on the returned FileDescription. +func (f *VFSFile) FileDescription() *vfs.FileDescription { + return f.file +} + +// fsLookup implements Lookup interface using fs.File. +// +// +stateify savable +type vfsLookup struct { + mntns *vfs.MountNamespace + + root vfs.VirtualDentry + workingDir vfs.VirtualDentry +} + +var _ Lookup = (*vfsLookup)(nil) + +// NewVFSLookup creates a new Lookup using VFS2. +func NewVFSLookup(mntns *vfs.MountNamespace, root, workingDir vfs.VirtualDentry) Lookup { + return &vfsLookup{ + mntns: mntns, + root: root, + workingDir: workingDir, + } +} + +// OpenPath implements Lookup. +// +// remainingTraversals is not configurable in VFS2, all callers are using the +// default anyways. +func (l *vfsLookup) OpenPath(ctx context.Context, pathname string, opts vfs.OpenOptions, _ *uint, resolveFinal bool) (File, error) { + vfsObj := l.mntns.Root().Mount().Filesystem().VirtualFilesystem() + creds := auth.CredentialsFromContext(ctx) + path := fspath.Parse(pathname) + pop := &vfs.PathOperation{ + Root: l.root, + Start: l.workingDir, + Path: path, + FollowFinalSymlink: resolveFinal, + } + if path.Absolute { + pop.Start = l.root + } + fd, err := vfsObj.OpenAt(ctx, creds, pop, &opts) + if err != nil { + return nil, err + } + return &VFSFile{file: fd}, nil +} diff --git a/pkg/sentry/fsimpl/devpts/BUILD b/pkg/sentry/fsimpl/devpts/BUILD new file mode 100644 index 000000000..585764223 --- /dev/null +++ b/pkg/sentry/fsimpl/devpts/BUILD @@ -0,0 +1,43 @@ +load("//tools:defs.bzl", "go_library", "go_test") + +licenses(["notice"]) + +go_library( + name = "devpts", + srcs = [ + "devpts.go", + "line_discipline.go", + "master.go", + "queue.go", + "slave.go", + "terminal.go", + ], + visibility = ["//pkg/sentry:internal"], + deps = [ + "//pkg/abi/linux", + "//pkg/context", + "//pkg/safemem", + "//pkg/sentry/arch", + "//pkg/sentry/fsimpl/kernfs", + "//pkg/sentry/kernel", + "//pkg/sentry/kernel/auth", + "//pkg/sentry/unimpl", + "//pkg/sentry/vfs", + "//pkg/sync", + "//pkg/syserror", + "//pkg/usermem", + "//pkg/waiter", + ], +) + +go_test( + name = "devpts_test", + size = "small", + srcs = ["devpts_test.go"], + library = ":devpts", + deps = [ + "//pkg/abi/linux", + "//pkg/sentry/contexttest", + "//pkg/usermem", + ], +) diff --git a/pkg/sentry/fsimpl/devpts/devpts.go b/pkg/sentry/fsimpl/devpts/devpts.go new file mode 100644 index 000000000..c03c65445 --- /dev/null +++ b/pkg/sentry/fsimpl/devpts/devpts.go @@ -0,0 +1,231 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// Package devpts provides a filesystem implementation that behaves like +// devpts. +package devpts + +import ( + "fmt" + "math" + "sort" + "strconv" + "sync" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" +) + +// Name is the filesystem name. +const Name = "devpts" + +// FilesystemType implements vfs.FilesystemType. +type FilesystemType struct{} + +// Name implements vfs.FilesystemType.Name. +func (FilesystemType) Name() string { + return Name +} + +var _ vfs.FilesystemType = (*FilesystemType)(nil) + +// GetFilesystem implements vfs.FilesystemType.GetFilesystem. +func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { + // No data allowed. + if opts.Data != "" { + return nil, nil, syserror.EINVAL + } + + fs, root, err := fstype.newFilesystem(vfsObj, creds) + if err != nil { + return nil, nil, err + } + return fs.Filesystem.VFSFilesystem(), root.VFSDentry(), nil +} + +type filesystem struct { + kernfs.Filesystem + + devMinor uint32 +} + +// newFilesystem creates a new devpts filesystem with root directory and ptmx +// master inode. It returns the filesystem and root Dentry. +func (fstype FilesystemType) newFilesystem(vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials) (*filesystem, *kernfs.Dentry, error) { + devMinor, err := vfsObj.GetAnonBlockDevMinor() + if err != nil { + return nil, nil, err + } + + fs := &filesystem{ + devMinor: devMinor, + } + fs.Filesystem.VFSFilesystem().Init(vfsObj, fstype, fs) + + // Construct the root directory. This is always inode id 1. + root := &rootInode{ + slaves: make(map[uint32]*slaveInode), + } + root.InodeAttrs.Init(creds, linux.UNNAMED_MAJOR, devMinor, 1, linux.ModeDirectory|0555) + root.OrderedChildren.Init(kernfs.OrderedChildrenOptions{}) + root.dentry.Init(root) + + // Construct the pts master inode and dentry. Linux always uses inode + // id 2 for ptmx. See fs/devpts/inode.c:mknod_ptmx. + master := &masterInode{ + root: root, + } + master.InodeAttrs.Init(creds, linux.UNNAMED_MAJOR, devMinor, 2, linux.ModeCharacterDevice|0666) + master.dentry.Init(master) + + // Add the master as a child of the root. + links := root.OrderedChildren.Populate(&root.dentry, map[string]*kernfs.Dentry{ + "ptmx": &master.dentry, + }) + root.IncLinks(links) + + return fs, &root.dentry, nil +} + +// Release implements vfs.FilesystemImpl.Release. +func (fs *filesystem) Release() { + fs.Filesystem.VFSFilesystem().VirtualFilesystem().PutAnonBlockDevMinor(fs.devMinor) + fs.Filesystem.Release() +} + +// rootInode is the root directory inode for the devpts mounts. +type rootInode struct { + kernfs.AlwaysValid + kernfs.InodeAttrs + kernfs.InodeDirectoryNoNewChildren + kernfs.InodeNotSymlink + kernfs.OrderedChildren + + // Keep a reference to this inode's dentry. + dentry kernfs.Dentry + + // master is the master pty inode. Immutable. + master *masterInode + + // root is the root directory inode for this filesystem. Immutable. + root *rootInode + + // mu protects the fields below. + mu sync.Mutex + + // slaves maps pty ids to slave inodes. + slaves map[uint32]*slaveInode + + // nextIdx is the next pty index to use. Must be accessed atomically. + // + // TODO(b/29356795): reuse indices when ptys are closed. + nextIdx uint32 +} + +var _ kernfs.Inode = (*rootInode)(nil) + +// allocateTerminal creates a new Terminal and installs a pts node for it. +func (i *rootInode) allocateTerminal(creds *auth.Credentials) (*Terminal, error) { + i.mu.Lock() + defer i.mu.Unlock() + if i.nextIdx == math.MaxUint32 { + return nil, syserror.ENOMEM + } + idx := i.nextIdx + i.nextIdx++ + + // Sanity check that slave with idx does not exist. + if _, ok := i.slaves[idx]; ok { + panic(fmt.Sprintf("pty index collision; index %d already exists", idx)) + } + + // Create the new terminal and slave. + t := newTerminal(idx) + slave := &slaveInode{ + root: i, + t: t, + } + // Linux always uses pty index + 3 as the inode id. See + // fs/devpts/inode.c:devpts_pty_new(). + slave.InodeAttrs.Init(creds, i.InodeAttrs.DevMajor(), i.InodeAttrs.DevMinor(), uint64(idx+3), linux.ModeCharacterDevice|0600) + slave.dentry.Init(slave) + i.slaves[idx] = slave + + return t, nil +} + +// masterClose is called when the master end of t is closed. +func (i *rootInode) masterClose(t *Terminal) { + i.mu.Lock() + defer i.mu.Unlock() + + // Sanity check that slave with idx exists. + if _, ok := i.slaves[t.n]; !ok { + panic(fmt.Sprintf("pty with index %d does not exist", t.n)) + } + delete(i.slaves, t.n) +} + +// Open implements kernfs.Inode.Open. +func (i *rootInode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), vfsd, &i.OrderedChildren, &opts) + if err != nil { + return nil, err + } + return fd.VFSFileDescription(), nil +} + +// Lookup implements kernfs.Inode.Lookup. +func (i *rootInode) Lookup(ctx context.Context, name string) (*vfs.Dentry, error) { + idx, err := strconv.ParseUint(name, 10, 32) + if err != nil { + return nil, syserror.ENOENT + } + i.mu.Lock() + defer i.mu.Unlock() + if si, ok := i.slaves[uint32(idx)]; ok { + si.dentry.IncRef() + return si.dentry.VFSDentry(), nil + + } + return nil, syserror.ENOENT +} + +// IterDirents implements kernfs.Inode.IterDirents. +func (i *rootInode) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback, offset, relOffset int64) (int64, error) { + i.mu.Lock() + defer i.mu.Unlock() + ids := make([]int, 0, len(i.slaves)) + for id := range i.slaves { + ids = append(ids, int(id)) + } + sort.Ints(ids) + for _, id := range ids[relOffset:] { + dirent := vfs.Dirent{ + Name: strconv.FormatUint(uint64(id), 10), + Type: linux.DT_CHR, + Ino: i.slaves[uint32(id)].InodeAttrs.Ino(), + NextOff: offset + 1, + } + if err := cb.Handle(dirent); err != nil { + return offset, err + } + offset++ + } + return offset, nil +} diff --git a/pkg/sentry/fsimpl/devpts/devpts_test.go b/pkg/sentry/fsimpl/devpts/devpts_test.go new file mode 100644 index 000000000..b7c149047 --- /dev/null +++ b/pkg/sentry/fsimpl/devpts/devpts_test.go @@ -0,0 +1,56 @@ +// Copyright 2018 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package devpts + +import ( + "testing" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/contexttest" + "gvisor.dev/gvisor/pkg/usermem" +) + +func TestSimpleMasterToSlave(t *testing.T) { + ld := newLineDiscipline(linux.DefaultSlaveTermios) + ctx := contexttest.Context(t) + inBytes := []byte("hello, tty\n") + src := usermem.BytesIOSequence(inBytes) + outBytes := make([]byte, 32) + dst := usermem.BytesIOSequence(outBytes) + + // Write to the input queue. + nw, err := ld.inputQueueWrite(ctx, src) + if err != nil { + t.Fatalf("error writing to input queue: %v", err) + } + if nw != int64(len(inBytes)) { + t.Fatalf("wrote wrong length: got %d, want %d", nw, len(inBytes)) + } + + // Read from the input queue. + nr, err := ld.inputQueueRead(ctx, dst) + if err != nil { + t.Fatalf("error reading from input queue: %v", err) + } + if nr != int64(len(inBytes)) { + t.Fatalf("read wrong length: got %d, want %d", nr, len(inBytes)) + } + + outStr := string(outBytes[:nr]) + inStr := string(inBytes) + if outStr != inStr { + t.Fatalf("written and read strings do not match: got %q, want %q", outStr, inStr) + } +} diff --git a/pkg/sentry/fsimpl/devpts/line_discipline.go b/pkg/sentry/fsimpl/devpts/line_discipline.go new file mode 100644 index 000000000..f7bc325d1 --- /dev/null +++ b/pkg/sentry/fsimpl/devpts/line_discipline.go @@ -0,0 +1,445 @@ +// Copyright 2018 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package devpts + +import ( + "bytes" + "unicode/utf8" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" + "gvisor.dev/gvisor/pkg/waiter" +) + +const ( + // canonMaxBytes is the number of bytes that fit into a single line of + // terminal input in canonical mode. This corresponds to N_TTY_BUF_SIZE + // in include/linux/tty.h. + canonMaxBytes = 4096 + + // nonCanonMaxBytes is the maximum number of bytes that can be read at + // a time in noncanonical mode. + nonCanonMaxBytes = canonMaxBytes - 1 + + spacesPerTab = 8 +) + +// lineDiscipline dictates how input and output are handled between the +// pseudoterminal (pty) master and slave. It can be configured to alter I/O, +// modify control characters (e.g. Ctrl-C for SIGINT), etc. The following man +// pages are good resources for how to affect the line discipline: +// +// * termios(3) +// * tty_ioctl(4) +// +// This file corresponds most closely to drivers/tty/n_tty.c. +// +// lineDiscipline has a simple structure but supports a multitude of options +// (see the above man pages). It consists of two queues of bytes: one from the +// terminal master to slave (the input queue) and one from slave to master (the +// output queue). When bytes are written to one end of the pty, the line +// discipline reads the bytes, modifies them or takes special action if +// required, and enqueues them to be read by the other end of the pty: +// +// input from terminal +-------------+ input to process (e.g. bash) +// +------------------------>| input queue |---------------------------+ +// | (inputQueueWrite) +-------------+ (inputQueueRead) | +// | | +// | v +// masterFD slaveFD +// ^ | +// | | +// | output to terminal +--------------+ output from process | +// +------------------------| output queue |<--------------------------+ +// (outputQueueRead) +--------------+ (outputQueueWrite) +// +// Lock order: +// termiosMu +// inQueue.mu +// outQueue.mu +// +// +stateify savable +type lineDiscipline struct { + // sizeMu protects size. + sizeMu sync.Mutex `state:"nosave"` + + // size is the terminal size (width and height). + size linux.WindowSize + + // inQueue is the input queue of the terminal. + inQueue queue + + // outQueue is the output queue of the terminal. + outQueue queue + + // termiosMu protects termios. + termiosMu sync.RWMutex `state:"nosave"` + + // termios is the terminal configuration used by the lineDiscipline. + termios linux.KernelTermios + + // column is the location in a row of the cursor. This is important for + // handling certain special characters like backspace. + column int + + // masterWaiter is used to wait on the master end of the TTY. + masterWaiter waiter.Queue `state:"zerovalue"` + + // slaveWaiter is used to wait on the slave end of the TTY. + slaveWaiter waiter.Queue `state:"zerovalue"` +} + +func newLineDiscipline(termios linux.KernelTermios) *lineDiscipline { + ld := lineDiscipline{termios: termios} + ld.inQueue.transformer = &inputQueueTransformer{} + ld.outQueue.transformer = &outputQueueTransformer{} + return &ld +} + +// getTermios gets the linux.Termios for the tty. +func (l *lineDiscipline) getTermios(ctx context.Context, io usermem.IO, args arch.SyscallArguments) (uintptr, error) { + l.termiosMu.RLock() + defer l.termiosMu.RUnlock() + // We must copy a Termios struct, not KernelTermios. + t := l.termios.ToTermios() + _, err := usermem.CopyObjectOut(ctx, io, args[2].Pointer(), t, usermem.IOOpts{ + AddressSpaceActive: true, + }) + return 0, err +} + +// setTermios sets a linux.Termios for the tty. +func (l *lineDiscipline) setTermios(ctx context.Context, io usermem.IO, args arch.SyscallArguments) (uintptr, error) { + l.termiosMu.Lock() + defer l.termiosMu.Unlock() + oldCanonEnabled := l.termios.LEnabled(linux.ICANON) + // We must copy a Termios struct, not KernelTermios. + var t linux.Termios + _, err := usermem.CopyObjectIn(ctx, io, args[2].Pointer(), &t, usermem.IOOpts{ + AddressSpaceActive: true, + }) + l.termios.FromTermios(t) + + // If canonical mode is turned off, move bytes from inQueue's wait + // buffer to its read buffer. Anything already in the read buffer is + // now readable. + if oldCanonEnabled && !l.termios.LEnabled(linux.ICANON) { + l.inQueue.mu.Lock() + l.inQueue.pushWaitBufLocked(l) + l.inQueue.readable = true + l.inQueue.mu.Unlock() + l.slaveWaiter.Notify(waiter.EventIn) + } + + return 0, err +} + +func (l *lineDiscipline) windowSize(ctx context.Context, io usermem.IO, args arch.SyscallArguments) error { + l.sizeMu.Lock() + defer l.sizeMu.Unlock() + _, err := usermem.CopyObjectOut(ctx, io, args[2].Pointer(), l.size, usermem.IOOpts{ + AddressSpaceActive: true, + }) + return err +} + +func (l *lineDiscipline) setWindowSize(ctx context.Context, io usermem.IO, args arch.SyscallArguments) error { + l.sizeMu.Lock() + defer l.sizeMu.Unlock() + _, err := usermem.CopyObjectIn(ctx, io, args[2].Pointer(), &l.size, usermem.IOOpts{ + AddressSpaceActive: true, + }) + return err +} + +func (l *lineDiscipline) masterReadiness() waiter.EventMask { + // We don't have to lock a termios because the default master termios + // is immutable. + return l.inQueue.writeReadiness(&linux.MasterTermios) | l.outQueue.readReadiness(&linux.MasterTermios) +} + +func (l *lineDiscipline) slaveReadiness() waiter.EventMask { + l.termiosMu.RLock() + defer l.termiosMu.RUnlock() + return l.outQueue.writeReadiness(&l.termios) | l.inQueue.readReadiness(&l.termios) +} + +func (l *lineDiscipline) inputQueueReadSize(ctx context.Context, io usermem.IO, args arch.SyscallArguments) error { + return l.inQueue.readableSize(ctx, io, args) +} + +func (l *lineDiscipline) inputQueueRead(ctx context.Context, dst usermem.IOSequence) (int64, error) { + l.termiosMu.RLock() + defer l.termiosMu.RUnlock() + n, pushed, err := l.inQueue.read(ctx, dst, l) + if err != nil { + return 0, err + } + if n > 0 { + l.masterWaiter.Notify(waiter.EventOut) + if pushed { + l.slaveWaiter.Notify(waiter.EventIn) + } + return n, nil + } + return 0, syserror.ErrWouldBlock +} + +func (l *lineDiscipline) inputQueueWrite(ctx context.Context, src usermem.IOSequence) (int64, error) { + l.termiosMu.RLock() + defer l.termiosMu.RUnlock() + n, err := l.inQueue.write(ctx, src, l) + if err != nil { + return 0, err + } + if n > 0 { + l.slaveWaiter.Notify(waiter.EventIn) + return n, nil + } + return 0, syserror.ErrWouldBlock +} + +func (l *lineDiscipline) outputQueueReadSize(ctx context.Context, io usermem.IO, args arch.SyscallArguments) error { + return l.outQueue.readableSize(ctx, io, args) +} + +func (l *lineDiscipline) outputQueueRead(ctx context.Context, dst usermem.IOSequence) (int64, error) { + l.termiosMu.RLock() + defer l.termiosMu.RUnlock() + n, pushed, err := l.outQueue.read(ctx, dst, l) + if err != nil { + return 0, err + } + if n > 0 { + l.slaveWaiter.Notify(waiter.EventOut) + if pushed { + l.masterWaiter.Notify(waiter.EventIn) + } + return n, nil + } + return 0, syserror.ErrWouldBlock +} + +func (l *lineDiscipline) outputQueueWrite(ctx context.Context, src usermem.IOSequence) (int64, error) { + l.termiosMu.RLock() + defer l.termiosMu.RUnlock() + n, err := l.outQueue.write(ctx, src, l) + if err != nil { + return 0, err + } + if n > 0 { + l.masterWaiter.Notify(waiter.EventIn) + return n, nil + } + return 0, syserror.ErrWouldBlock +} + +// transformer is a helper interface to make it easier to stateify queue. +type transformer interface { + // transform functions require queue's mutex to be held. + transform(*lineDiscipline, *queue, []byte) int +} + +// outputQueueTransformer implements transformer. It performs line discipline +// transformations on the output queue. +// +// +stateify savable +type outputQueueTransformer struct{} + +// transform does output processing for one end of the pty. See +// drivers/tty/n_tty.c:do_output_char for an analogous kernel function. +// +// Preconditions: +// * l.termiosMu must be held for reading. +// * q.mu must be held. +func (*outputQueueTransformer) transform(l *lineDiscipline, q *queue, buf []byte) int { + // transformOutput is effectively always in noncanonical mode, as the + // master termios never has ICANON set. + + if !l.termios.OEnabled(linux.OPOST) { + q.readBuf = append(q.readBuf, buf...) + if len(q.readBuf) > 0 { + q.readable = true + } + return len(buf) + } + + var ret int + for len(buf) > 0 { + size := l.peek(buf) + cBytes := append([]byte{}, buf[:size]...) + ret += size + buf = buf[size:] + // We're guaranteed that cBytes has at least one element. + switch cBytes[0] { + case '\n': + if l.termios.OEnabled(linux.ONLRET) { + l.column = 0 + } + if l.termios.OEnabled(linux.ONLCR) { + q.readBuf = append(q.readBuf, '\r', '\n') + continue + } + case '\r': + if l.termios.OEnabled(linux.ONOCR) && l.column == 0 { + continue + } + if l.termios.OEnabled(linux.OCRNL) { + cBytes[0] = '\n' + if l.termios.OEnabled(linux.ONLRET) { + l.column = 0 + } + break + } + l.column = 0 + case '\t': + spaces := spacesPerTab - l.column%spacesPerTab + if l.termios.OutputFlags&linux.TABDLY == linux.XTABS { + l.column += spaces + q.readBuf = append(q.readBuf, bytes.Repeat([]byte{' '}, spacesPerTab)...) + continue + } + l.column += spaces + case '\b': + if l.column > 0 { + l.column-- + } + default: + l.column++ + } + q.readBuf = append(q.readBuf, cBytes...) + } + if len(q.readBuf) > 0 { + q.readable = true + } + return ret +} + +// inputQueueTransformer implements transformer. It performs line discipline +// transformations on the input queue. +// +// +stateify savable +type inputQueueTransformer struct{} + +// transform does input processing for one end of the pty. Characters read are +// transformed according to flags set in the termios struct. See +// drivers/tty/n_tty.c:n_tty_receive_char_special for an analogous kernel +// function. +// +// Preconditions: +// * l.termiosMu must be held for reading. +// * q.mu must be held. +func (*inputQueueTransformer) transform(l *lineDiscipline, q *queue, buf []byte) int { + // If there's a line waiting to be read in canonical mode, don't write + // anything else to the read buffer. + if l.termios.LEnabled(linux.ICANON) && q.readable { + return 0 + } + + maxBytes := nonCanonMaxBytes + if l.termios.LEnabled(linux.ICANON) { + maxBytes = canonMaxBytes + } + + var ret int + for len(buf) > 0 && len(q.readBuf) < canonMaxBytes { + size := l.peek(buf) + cBytes := append([]byte{}, buf[:size]...) + // We're guaranteed that cBytes has at least one element. + switch cBytes[0] { + case '\r': + if l.termios.IEnabled(linux.IGNCR) { + buf = buf[size:] + ret += size + continue + } + if l.termios.IEnabled(linux.ICRNL) { + cBytes[0] = '\n' + } + case '\n': + if l.termios.IEnabled(linux.INLCR) { + cBytes[0] = '\r' + } + } + + // In canonical mode, we discard non-terminating characters + // after the first 4095. + if l.shouldDiscard(q, cBytes) { + buf = buf[size:] + ret += size + continue + } + + // Stop if the buffer would be overfilled. + if len(q.readBuf)+size > maxBytes { + break + } + buf = buf[size:] + ret += size + + // If we get EOF, make the buffer available for reading. + if l.termios.LEnabled(linux.ICANON) && l.termios.IsEOF(cBytes[0]) { + q.readable = true + break + } + + q.readBuf = append(q.readBuf, cBytes...) + + // Anything written to the readBuf will have to be echoed. + if l.termios.LEnabled(linux.ECHO) { + l.outQueue.writeBytes(cBytes, l) + l.masterWaiter.Notify(waiter.EventIn) + } + + // If we finish a line, make it available for reading. + if l.termios.LEnabled(linux.ICANON) && l.termios.IsTerminating(cBytes) { + q.readable = true + break + } + } + + // In noncanonical mode, everything is readable. + if !l.termios.LEnabled(linux.ICANON) && len(q.readBuf) > 0 { + q.readable = true + } + + return ret +} + +// shouldDiscard returns whether c should be discarded. In canonical mode, if +// too many bytes are enqueued, we keep reading input and discarding it until +// we find a terminating character. Signal/echo processing still occurs. +// +// Precondition: +// * l.termiosMu must be held for reading. +// * q.mu must be held. +func (l *lineDiscipline) shouldDiscard(q *queue, cBytes []byte) bool { + return l.termios.LEnabled(linux.ICANON) && len(q.readBuf)+len(cBytes) >= canonMaxBytes && !l.termios.IsTerminating(cBytes) +} + +// peek returns the size in bytes of the next character to process. As long as +// b isn't empty, peek returns a value of at least 1. +func (l *lineDiscipline) peek(b []byte) int { + size := 1 + // If UTF-8 support is enabled, runes might be multiple bytes. + if l.termios.IEnabled(linux.IUTF8) { + _, size = utf8.DecodeRune(b) + } + return size +} diff --git a/pkg/sentry/fsimpl/devpts/master.go b/pkg/sentry/fsimpl/devpts/master.go new file mode 100644 index 000000000..7a7ce5d81 --- /dev/null +++ b/pkg/sentry/fsimpl/devpts/master.go @@ -0,0 +1,222 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package devpts + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/unimpl" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" + "gvisor.dev/gvisor/pkg/waiter" +) + +// masterInode is the inode for the master end of the Terminal. +type masterInode struct { + kernfs.InodeAttrs + kernfs.InodeNoopRefCount + kernfs.InodeNotDirectory + kernfs.InodeNotSymlink + + // Keep a reference to this inode's dentry. + dentry kernfs.Dentry + + // root is the devpts root inode. + root *rootInode +} + +var _ kernfs.Inode = (*masterInode)(nil) + +// Open implements kernfs.Inode.Open. +func (mi *masterInode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + t, err := mi.root.allocateTerminal(rp.Credentials()) + if err != nil { + return nil, err + } + + mi.IncRef() + fd := &masterFileDescription{ + inode: mi, + t: t, + } + if err := fd.vfsfd.Init(fd, opts.Flags, rp.Mount(), vfsd, &vfs.FileDescriptionOptions{}); err != nil { + mi.DecRef() + return nil, err + } + return &fd.vfsfd, nil +} + +// Stat implements kernfs.Inode.Stat. +func (mi *masterInode) Stat(vfsfs *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) { + statx, err := mi.InodeAttrs.Stat(vfsfs, opts) + if err != nil { + return linux.Statx{}, err + } + statx.Blksize = 1024 + statx.RdevMajor = linux.TTYAUX_MAJOR + statx.RdevMinor = linux.PTMX_MINOR + return statx, nil +} + +// SetStat implements kernfs.Inode.SetStat +func (mi *masterInode) SetStat(ctx context.Context, vfsfs *vfs.Filesystem, creds *auth.Credentials, opts vfs.SetStatOptions) error { + if opts.Stat.Mask&linux.STATX_SIZE != 0 { + return syserror.EINVAL + } + return mi.InodeAttrs.SetStat(ctx, vfsfs, creds, opts) +} + +type masterFileDescription struct { + vfsfd vfs.FileDescription + vfs.FileDescriptionDefaultImpl + + inode *masterInode + t *Terminal +} + +var _ vfs.FileDescriptionImpl = (*masterFileDescription)(nil) + +// Release implements vfs.FileDescriptionImpl.Release. +func (mfd *masterFileDescription) Release() { + mfd.inode.root.masterClose(mfd.t) + mfd.inode.DecRef() +} + +// EventRegister implements waiter.Waitable.EventRegister. +func (mfd *masterFileDescription) EventRegister(e *waiter.Entry, mask waiter.EventMask) { + mfd.t.ld.masterWaiter.EventRegister(e, mask) +} + +// EventUnregister implements waiter.Waitable.EventUnregister. +func (mfd *masterFileDescription) EventUnregister(e *waiter.Entry) { + mfd.t.ld.masterWaiter.EventUnregister(e) +} + +// Readiness implements waiter.Waitable.Readiness. +func (mfd *masterFileDescription) Readiness(mask waiter.EventMask) waiter.EventMask { + return mfd.t.ld.masterReadiness() +} + +// Read implements vfs.FileDescriptionImpl.Read. +func (mfd *masterFileDescription) Read(ctx context.Context, dst usermem.IOSequence, _ vfs.ReadOptions) (int64, error) { + return mfd.t.ld.outputQueueRead(ctx, dst) +} + +// Write implements vfs.FileDescriptionImpl.Write. +func (mfd *masterFileDescription) Write(ctx context.Context, src usermem.IOSequence, _ vfs.WriteOptions) (int64, error) { + return mfd.t.ld.inputQueueWrite(ctx, src) +} + +// Ioctl implements vfs.FileDescriptionImpl.Ioctl. +func (mfd *masterFileDescription) Ioctl(ctx context.Context, io usermem.IO, args arch.SyscallArguments) (uintptr, error) { + switch cmd := args[1].Uint(); cmd { + case linux.FIONREAD: // linux.FIONREAD == linux.TIOCINQ + // Get the number of bytes in the output queue read buffer. + return 0, mfd.t.ld.outputQueueReadSize(ctx, io, args) + case linux.TCGETS: + // N.B. TCGETS on the master actually returns the configuration + // of the slave end. + return mfd.t.ld.getTermios(ctx, io, args) + case linux.TCSETS: + // N.B. TCSETS on the master actually affects the configuration + // of the slave end. + return mfd.t.ld.setTermios(ctx, io, args) + case linux.TCSETSW: + // TODO(b/29356795): This should drain the output queue first. + return mfd.t.ld.setTermios(ctx, io, args) + case linux.TIOCGPTN: + _, err := usermem.CopyObjectOut(ctx, io, args[2].Pointer(), uint32(mfd.t.n), usermem.IOOpts{ + AddressSpaceActive: true, + }) + return 0, err + case linux.TIOCSPTLCK: + // TODO(b/29356795): Implement pty locking. For now just pretend we do. + return 0, nil + case linux.TIOCGWINSZ: + return 0, mfd.t.ld.windowSize(ctx, io, args) + case linux.TIOCSWINSZ: + return 0, mfd.t.ld.setWindowSize(ctx, io, args) + case linux.TIOCSCTTY: + // Make the given terminal the controlling terminal of the + // calling process. + return 0, mfd.t.setControllingTTY(ctx, io, args, true /* isMaster */) + case linux.TIOCNOTTY: + // Release this process's controlling terminal. + return 0, mfd.t.releaseControllingTTY(ctx, io, args, true /* isMaster */) + case linux.TIOCGPGRP: + // Get the foreground process group. + return mfd.t.foregroundProcessGroup(ctx, io, args, true /* isMaster */) + case linux.TIOCSPGRP: + // Set the foreground process group. + return mfd.t.setForegroundProcessGroup(ctx, io, args, true /* isMaster */) + default: + maybeEmitUnimplementedEvent(ctx, cmd) + return 0, syserror.ENOTTY + } +} + +// SetStat implements vfs.FileDescriptionImpl.SetStat. +func (mfd *masterFileDescription) SetStat(ctx context.Context, opts vfs.SetStatOptions) error { + creds := auth.CredentialsFromContext(ctx) + fs := mfd.vfsfd.VirtualDentry().Mount().Filesystem() + return mfd.inode.SetStat(ctx, fs, creds, opts) +} + +// Stat implements vfs.FileDescriptionImpl.Stat. +func (mfd *masterFileDescription) Stat(ctx context.Context, opts vfs.StatOptions) (linux.Statx, error) { + fs := mfd.vfsfd.VirtualDentry().Mount().Filesystem() + return mfd.inode.Stat(fs, opts) +} + +// maybeEmitUnimplementedEvent emits unimplemented event if cmd is valid. +func maybeEmitUnimplementedEvent(ctx context.Context, cmd uint32) { + switch cmd { + case linux.TCGETS, + linux.TCSETS, + linux.TCSETSW, + linux.TCSETSF, + linux.TIOCGWINSZ, + linux.TIOCSWINSZ, + linux.TIOCSETD, + linux.TIOCSBRK, + linux.TIOCCBRK, + linux.TCSBRK, + linux.TCSBRKP, + linux.TIOCSTI, + linux.TIOCCONS, + linux.FIONBIO, + linux.TIOCEXCL, + linux.TIOCNXCL, + linux.TIOCGEXCL, + linux.TIOCGSID, + linux.TIOCGETD, + linux.TIOCVHANGUP, + linux.TIOCGDEV, + linux.TIOCMGET, + linux.TIOCMSET, + linux.TIOCMBIC, + linux.TIOCMBIS, + linux.TIOCGICOUNT, + linux.TCFLSH, + linux.TIOCSSERIAL, + linux.TIOCGPTPEER: + + unimpl.EmitUnimplementedEvent(ctx) + } +} diff --git a/pkg/sentry/fsimpl/devpts/queue.go b/pkg/sentry/fsimpl/devpts/queue.go new file mode 100644 index 000000000..dffb4232c --- /dev/null +++ b/pkg/sentry/fsimpl/devpts/queue.go @@ -0,0 +1,236 @@ +// Copyright 2018 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package devpts + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/safemem" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" + "gvisor.dev/gvisor/pkg/waiter" +) + +// waitBufMaxBytes is the maximum size of a wait buffer. It is based on +// TTYB_DEFAULT_MEM_LIMIT. +const waitBufMaxBytes = 131072 + +// queue represents one of the input or output queues between a pty master and +// slave. Bytes written to a queue are added to the read buffer until it is +// full, at which point they are written to the wait buffer. Bytes are +// processed (i.e. undergo termios transformations) as they are added to the +// read buffer. The read buffer is readable when its length is nonzero and +// readable is true. +// +// +stateify savable +type queue struct { + // mu protects everything in queue. + mu sync.Mutex `state:"nosave"` + + // readBuf is buffer of data ready to be read when readable is true. + // This data has been processed. + readBuf []byte + + // waitBuf contains data that can't fit into readBuf. It is put here + // until it can be loaded into the read buffer. waitBuf contains data + // that hasn't been processed. + waitBuf [][]byte + waitBufLen uint64 + + // readable indicates whether the read buffer can be read from. In + // canonical mode, there can be an unterminated line in the read buffer, + // so readable must be checked. + readable bool + + // transform is the the queue's function for transforming bytes + // entering the queue. For example, transform might convert all '\r's + // entering the queue to '\n's. + transformer +} + +// readReadiness returns whether q is ready to be read from. +func (q *queue) readReadiness(t *linux.KernelTermios) waiter.EventMask { + q.mu.Lock() + defer q.mu.Unlock() + if len(q.readBuf) > 0 && q.readable { + return waiter.EventIn + } + return waiter.EventMask(0) +} + +// writeReadiness returns whether q is ready to be written to. +func (q *queue) writeReadiness(t *linux.KernelTermios) waiter.EventMask { + q.mu.Lock() + defer q.mu.Unlock() + if q.waitBufLen < waitBufMaxBytes { + return waiter.EventOut + } + return waiter.EventMask(0) +} + +// readableSize writes the number of readable bytes to userspace. +func (q *queue) readableSize(ctx context.Context, io usermem.IO, args arch.SyscallArguments) error { + q.mu.Lock() + defer q.mu.Unlock() + var size int32 + if q.readable { + size = int32(len(q.readBuf)) + } + + _, err := usermem.CopyObjectOut(ctx, io, args[2].Pointer(), size, usermem.IOOpts{ + AddressSpaceActive: true, + }) + return err + +} + +// read reads from q to userspace. It returns the number of bytes read as well +// as whether the read caused more readable data to become available (whether +// data was pushed from the wait buffer to the read buffer). +// +// Preconditions: +// * l.termiosMu must be held for reading. +func (q *queue) read(ctx context.Context, dst usermem.IOSequence, l *lineDiscipline) (int64, bool, error) { + q.mu.Lock() + defer q.mu.Unlock() + + if !q.readable { + return 0, false, syserror.ErrWouldBlock + } + + if dst.NumBytes() > canonMaxBytes { + dst = dst.TakeFirst(canonMaxBytes) + } + + n, err := dst.CopyOutFrom(ctx, safemem.ReaderFunc(func(dst safemem.BlockSeq) (uint64, error) { + src := safemem.BlockSeqOf(safemem.BlockFromSafeSlice(q.readBuf)) + n, err := safemem.CopySeq(dst, src) + if err != nil { + return 0, err + } + q.readBuf = q.readBuf[n:] + + // If we read everything, this queue is no longer readable. + if len(q.readBuf) == 0 { + q.readable = false + } + + return n, nil + })) + if err != nil { + return 0, false, err + } + + // Move data from the queue's wait buffer to its read buffer. + nPushed := q.pushWaitBufLocked(l) + + return int64(n), nPushed > 0, nil +} + +// write writes to q from userspace. +// +// Preconditions: +// * l.termiosMu must be held for reading. +func (q *queue) write(ctx context.Context, src usermem.IOSequence, l *lineDiscipline) (int64, error) { + q.mu.Lock() + defer q.mu.Unlock() + + // Copy data into the wait buffer. + n, err := src.CopyInTo(ctx, safemem.WriterFunc(func(src safemem.BlockSeq) (uint64, error) { + copyLen := src.NumBytes() + room := waitBufMaxBytes - q.waitBufLen + // If out of room, return EAGAIN. + if room == 0 && copyLen > 0 { + return 0, syserror.ErrWouldBlock + } + // Cap the size of the wait buffer. + if copyLen > room { + copyLen = room + src = src.TakeFirst64(room) + } + buf := make([]byte, copyLen) + + // Copy the data into the wait buffer. + dst := safemem.BlockSeqOf(safemem.BlockFromSafeSlice(buf)) + n, err := safemem.CopySeq(dst, src) + if err != nil { + return 0, err + } + q.waitBufAppend(buf) + + return n, nil + })) + if err != nil { + return 0, err + } + + // Push data from the wait to the read buffer. + q.pushWaitBufLocked(l) + + return n, nil +} + +// writeBytes writes to q from b. +// +// Preconditions: +// * l.termiosMu must be held for reading. +func (q *queue) writeBytes(b []byte, l *lineDiscipline) { + q.mu.Lock() + defer q.mu.Unlock() + + // Write to the wait buffer. + q.waitBufAppend(b) + q.pushWaitBufLocked(l) +} + +// pushWaitBufLocked fills the queue's read buffer with data from the wait +// buffer. +// +// Preconditions: +// * l.termiosMu must be held for reading. +// * q.mu must be locked. +func (q *queue) pushWaitBufLocked(l *lineDiscipline) int { + if q.waitBufLen == 0 { + return 0 + } + + // Move data from the wait to the read buffer. + var total int + var i int + for i = 0; i < len(q.waitBuf); i++ { + n := q.transform(l, q, q.waitBuf[i]) + total += n + if n != len(q.waitBuf[i]) { + // The read buffer filled up without consuming the + // entire buffer. + q.waitBuf[i] = q.waitBuf[i][n:] + break + } + } + + // Update wait buffer based on consumed data. + q.waitBuf = q.waitBuf[i:] + q.waitBufLen -= uint64(total) + + return total +} + +// Precondition: q.mu must be locked. +func (q *queue) waitBufAppend(b []byte) { + q.waitBuf = append(q.waitBuf, b) + q.waitBufLen += uint64(len(b)) +} diff --git a/pkg/sentry/fsimpl/devpts/slave.go b/pkg/sentry/fsimpl/devpts/slave.go new file mode 100644 index 000000000..526cd406c --- /dev/null +++ b/pkg/sentry/fsimpl/devpts/slave.go @@ -0,0 +1,182 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package devpts + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" + "gvisor.dev/gvisor/pkg/waiter" +) + +// slaveInode is the inode for the slave end of the Terminal. +type slaveInode struct { + kernfs.InodeAttrs + kernfs.InodeNoopRefCount + kernfs.InodeNotDirectory + kernfs.InodeNotSymlink + + // Keep a reference to this inode's dentry. + dentry kernfs.Dentry + + // root is the devpts root inode. + root *rootInode + + // t is the connected Terminal. + t *Terminal +} + +var _ kernfs.Inode = (*slaveInode)(nil) + +// Open implements kernfs.Inode.Open. +func (si *slaveInode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + si.IncRef() + fd := &slaveFileDescription{ + inode: si, + } + if err := fd.vfsfd.Init(fd, opts.Flags, rp.Mount(), vfsd, &vfs.FileDescriptionOptions{}); err != nil { + si.DecRef() + return nil, err + } + return &fd.vfsfd, nil + +} + +// Valid implements kernfs.Inode.Valid. +func (si *slaveInode) Valid(context.Context) bool { + // Return valid if the slave still exists. + si.root.mu.Lock() + defer si.root.mu.Unlock() + _, ok := si.root.slaves[si.t.n] + return ok +} + +// Stat implements kernfs.Inode.Stat. +func (si *slaveInode) Stat(vfsfs *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) { + statx, err := si.InodeAttrs.Stat(vfsfs, opts) + if err != nil { + return linux.Statx{}, err + } + statx.Blksize = 1024 + statx.RdevMajor = linux.UNIX98_PTY_SLAVE_MAJOR + statx.RdevMinor = si.t.n + return statx, nil +} + +// SetStat implements kernfs.Inode.SetStat +func (si *slaveInode) SetStat(ctx context.Context, vfsfs *vfs.Filesystem, creds *auth.Credentials, opts vfs.SetStatOptions) error { + if opts.Stat.Mask&linux.STATX_SIZE != 0 { + return syserror.EINVAL + } + return si.InodeAttrs.SetStat(ctx, vfsfs, creds, opts) +} + +type slaveFileDescription struct { + vfsfd vfs.FileDescription + vfs.FileDescriptionDefaultImpl + + inode *slaveInode +} + +var _ vfs.FileDescriptionImpl = (*slaveFileDescription)(nil) + +// Release implements fs.FileOperations.Release. +func (sfd *slaveFileDescription) Release() { + sfd.inode.DecRef() +} + +// EventRegister implements waiter.Waitable.EventRegister. +func (sfd *slaveFileDescription) EventRegister(e *waiter.Entry, mask waiter.EventMask) { + sfd.inode.t.ld.slaveWaiter.EventRegister(e, mask) +} + +// EventUnregister implements waiter.Waitable.EventUnregister. +func (sfd *slaveFileDescription) EventUnregister(e *waiter.Entry) { + sfd.inode.t.ld.slaveWaiter.EventUnregister(e) +} + +// Readiness implements waiter.Waitable.Readiness. +func (sfd *slaveFileDescription) Readiness(mask waiter.EventMask) waiter.EventMask { + return sfd.inode.t.ld.slaveReadiness() +} + +// Read implements vfs.FileDescriptionImpl.Read. +func (sfd *slaveFileDescription) Read(ctx context.Context, dst usermem.IOSequence, _ vfs.ReadOptions) (int64, error) { + return sfd.inode.t.ld.inputQueueRead(ctx, dst) +} + +// Write implements vfs.FileDescriptionImpl.Write. +func (sfd *slaveFileDescription) Write(ctx context.Context, src usermem.IOSequence, _ vfs.WriteOptions) (int64, error) { + return sfd.inode.t.ld.outputQueueWrite(ctx, src) +} + +// Ioctl implements vfs.FileDescripionImpl.Ioctl. +func (sfd *slaveFileDescription) Ioctl(ctx context.Context, io usermem.IO, args arch.SyscallArguments) (uintptr, error) { + switch cmd := args[1].Uint(); cmd { + case linux.FIONREAD: // linux.FIONREAD == linux.TIOCINQ + // Get the number of bytes in the input queue read buffer. + return 0, sfd.inode.t.ld.inputQueueReadSize(ctx, io, args) + case linux.TCGETS: + return sfd.inode.t.ld.getTermios(ctx, io, args) + case linux.TCSETS: + return sfd.inode.t.ld.setTermios(ctx, io, args) + case linux.TCSETSW: + // TODO(b/29356795): This should drain the output queue first. + return sfd.inode.t.ld.setTermios(ctx, io, args) + case linux.TIOCGPTN: + _, err := usermem.CopyObjectOut(ctx, io, args[2].Pointer(), uint32(sfd.inode.t.n), usermem.IOOpts{ + AddressSpaceActive: true, + }) + return 0, err + case linux.TIOCGWINSZ: + return 0, sfd.inode.t.ld.windowSize(ctx, io, args) + case linux.TIOCSWINSZ: + return 0, sfd.inode.t.ld.setWindowSize(ctx, io, args) + case linux.TIOCSCTTY: + // Make the given terminal the controlling terminal of the + // calling process. + return 0, sfd.inode.t.setControllingTTY(ctx, io, args, false /* isMaster */) + case linux.TIOCNOTTY: + // Release this process's controlling terminal. + return 0, sfd.inode.t.releaseControllingTTY(ctx, io, args, false /* isMaster */) + case linux.TIOCGPGRP: + // Get the foreground process group. + return sfd.inode.t.foregroundProcessGroup(ctx, io, args, false /* isMaster */) + case linux.TIOCSPGRP: + // Set the foreground process group. + return sfd.inode.t.setForegroundProcessGroup(ctx, io, args, false /* isMaster */) + default: + maybeEmitUnimplementedEvent(ctx, cmd) + return 0, syserror.ENOTTY + } +} + +// SetStat implements vfs.FileDescriptionImpl.SetStat. +func (sfd *slaveFileDescription) SetStat(ctx context.Context, opts vfs.SetStatOptions) error { + creds := auth.CredentialsFromContext(ctx) + fs := sfd.vfsfd.VirtualDentry().Mount().Filesystem() + return sfd.inode.SetStat(ctx, fs, creds, opts) +} + +// Stat implements vfs.FileDescriptionImpl.Stat. +func (sfd *slaveFileDescription) Stat(ctx context.Context, opts vfs.StatOptions) (linux.Statx, error) { + fs := sfd.vfsfd.VirtualDentry().Mount().Filesystem() + return sfd.inode.Stat(fs, opts) +} diff --git a/pkg/sentry/fsimpl/devpts/terminal.go b/pkg/sentry/fsimpl/devpts/terminal.go new file mode 100644 index 000000000..7d2781c54 --- /dev/null +++ b/pkg/sentry/fsimpl/devpts/terminal.go @@ -0,0 +1,120 @@ +// Copyright 2018 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package devpts + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/usermem" +) + +// Terminal is a pseudoterminal. +// +// +stateify savable +type Terminal struct { + // n is the terminal index. It is immutable. + n uint32 + + // ld is the line discipline of the terminal. It is immutable. + ld *lineDiscipline + + // masterKTTY contains the controlling process of the master end of + // this terminal. This field is immutable. + masterKTTY *kernel.TTY + + // slaveKTTY contains the controlling process of the slave end of this + // terminal. This field is immutable. + slaveKTTY *kernel.TTY +} + +func newTerminal(n uint32) *Terminal { + termios := linux.DefaultSlaveTermios + t := Terminal{ + n: n, + ld: newLineDiscipline(termios), + masterKTTY: &kernel.TTY{Index: n}, + slaveKTTY: &kernel.TTY{Index: n}, + } + return &t +} + +// setControllingTTY makes tm the controlling terminal of the calling thread +// group. +func (tm *Terminal) setControllingTTY(ctx context.Context, io usermem.IO, args arch.SyscallArguments, isMaster bool) error { + task := kernel.TaskFromContext(ctx) + if task == nil { + panic("setControllingTTY must be called from a task context") + } + + return task.ThreadGroup().SetControllingTTY(tm.tty(isMaster), args[2].Int()) +} + +// releaseControllingTTY removes tm as the controlling terminal of the calling +// thread group. +func (tm *Terminal) releaseControllingTTY(ctx context.Context, io usermem.IO, args arch.SyscallArguments, isMaster bool) error { + task := kernel.TaskFromContext(ctx) + if task == nil { + panic("releaseControllingTTY must be called from a task context") + } + + return task.ThreadGroup().ReleaseControllingTTY(tm.tty(isMaster)) +} + +// foregroundProcessGroup gets the process group ID of tm's foreground process. +func (tm *Terminal) foregroundProcessGroup(ctx context.Context, io usermem.IO, args arch.SyscallArguments, isMaster bool) (uintptr, error) { + task := kernel.TaskFromContext(ctx) + if task == nil { + panic("foregroundProcessGroup must be called from a task context") + } + + ret, err := task.ThreadGroup().ForegroundProcessGroup(tm.tty(isMaster)) + if err != nil { + return 0, err + } + + // Write it out to *arg. + _, err = usermem.CopyObjectOut(ctx, io, args[2].Pointer(), int32(ret), usermem.IOOpts{ + AddressSpaceActive: true, + }) + return 0, err +} + +// foregroundProcessGroup sets tm's foreground process. +func (tm *Terminal) setForegroundProcessGroup(ctx context.Context, io usermem.IO, args arch.SyscallArguments, isMaster bool) (uintptr, error) { + task := kernel.TaskFromContext(ctx) + if task == nil { + panic("setForegroundProcessGroup must be called from a task context") + } + + // Read in the process group ID. + var pgid int32 + if _, err := usermem.CopyObjectIn(ctx, io, args[2].Pointer(), &pgid, usermem.IOOpts{ + AddressSpaceActive: true, + }); err != nil { + return 0, err + } + + ret, err := task.ThreadGroup().SetForegroundProcessGroup(tm.tty(isMaster), kernel.ProcessGroupID(pgid)) + return uintptr(ret), err +} + +func (tm *Terminal) tty(isMaster bool) *kernel.TTY { + if isMaster { + return tm.masterKTTY + } + return tm.slaveKTTY +} diff --git a/pkg/sentry/fsimpl/devtmpfs/devtmpfs.go b/pkg/sentry/fsimpl/devtmpfs/devtmpfs.go index e03a0c665..142ee53b0 100644 --- a/pkg/sentry/fsimpl/devtmpfs/devtmpfs.go +++ b/pkg/sentry/fsimpl/devtmpfs/devtmpfs.go @@ -28,6 +28,9 @@ import ( "gvisor.dev/gvisor/pkg/sync" ) +// Name is the default filesystem name. +const Name = "devtmpfs" + // FilesystemType implements vfs.FilesystemType. type FilesystemType struct { initOnce sync.Once @@ -39,6 +42,11 @@ type FilesystemType struct { root *vfs.Dentry } +// Name implements vfs.FilesystemType.Name. +func (*FilesystemType) Name() string { + return Name +} + // GetFilesystem implements vfs.FilesystemType.GetFilesystem. func (fst *FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { fst.initOnce.Do(func() { @@ -107,6 +115,7 @@ func (a *Accessor) wrapContext(ctx context.Context) *accessorContext { func (ac *accessorContext) Value(key interface{}) interface{} { switch key { case vfs.CtxMountNamespace: + ac.a.mntns.IncRef() return ac.a.mntns case vfs.CtxRoot: ac.a.root.IncRef() @@ -154,16 +163,25 @@ func (a *Accessor) CreateDeviceFile(ctx context.Context, pathname string, kind v func (a *Accessor) UserspaceInit(ctx context.Context) error { actx := a.wrapContext(ctx) - // systemd: src/shared/dev-setup.c:dev_setup() + // Initialize symlinks. for _, symlink := range []struct { source string target string }{ - // /proc/kcore is not implemented. + // systemd: src/shared/dev-setup.c:dev_setup() {source: "fd", target: "/proc/self/fd"}, {source: "stdin", target: "/proc/self/fd/0"}, {source: "stdout", target: "/proc/self/fd/1"}, {source: "stderr", target: "/proc/self/fd/2"}, + // /proc/kcore is not implemented. + + // Linux implements /dev/ptmx as a device node, but advises + // container implementations to create /dev/ptmx as a symlink + // to pts/ptmx (Documentation/filesystems/devpts.txt). Systemd + // follows this advice (src/nspawn/nspawn.c:setup_pts()), while + // LXC tries to create a bind mount and falls back to a symlink + // (src/lxc/conf.c:lxc_setup_devpts()). + {source: "ptmx", target: "pts/ptmx"}, } { if err := a.vfsObj.SymlinkAt(actx, a.creds, a.pathOperationAt(symlink.source), symlink.target); err != nil { return fmt.Errorf("failed to create symlink %q => %q: %v", symlink.source, symlink.target, err) diff --git a/pkg/sentry/fsimpl/devtmpfs/devtmpfs_test.go b/pkg/sentry/fsimpl/devtmpfs/devtmpfs_test.go index 73308a2b5..b6d52c015 100644 --- a/pkg/sentry/fsimpl/devtmpfs/devtmpfs_test.go +++ b/pkg/sentry/fsimpl/devtmpfs/devtmpfs_test.go @@ -29,7 +29,10 @@ func TestDevtmpfs(t *testing.T) { ctx := contexttest.Context(t) creds := auth.CredentialsFromContext(ctx) - vfsObj := vfs.New() + vfsObj := &vfs.VirtualFilesystem{} + if err := vfsObj.Init(); err != nil { + t.Fatalf("VFS init: %v", err) + } // Register tmpfs just so that we can have a root filesystem that isn't // devtmpfs. vfsObj.MustRegisterFilesystemType("tmpfs", tmpfs.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ diff --git a/pkg/sentry/fsimpl/eventfd/BUILD b/pkg/sentry/fsimpl/eventfd/BUILD new file mode 100644 index 000000000..ea167d38c --- /dev/null +++ b/pkg/sentry/fsimpl/eventfd/BUILD @@ -0,0 +1,33 @@ +load("//tools:defs.bzl", "go_library", "go_test") + +licenses(["notice"]) + +go_library( + name = "eventfd", + srcs = ["eventfd.go"], + visibility = ["//pkg/sentry:internal"], + deps = [ + "//pkg/abi/linux", + "//pkg/context", + "//pkg/fdnotifier", + "//pkg/log", + "//pkg/sentry/vfs", + "//pkg/syserror", + "//pkg/usermem", + "//pkg/waiter", + ], +) + +go_test( + name = "eventfd_test", + size = "small", + srcs = ["eventfd_test.go"], + library = ":eventfd", + deps = [ + "//pkg/abi/linux", + "//pkg/sentry/contexttest", + "//pkg/sentry/vfs", + "//pkg/usermem", + "//pkg/waiter", + ], +) diff --git a/pkg/sentry/fsimpl/eventfd/eventfd.go b/pkg/sentry/fsimpl/eventfd/eventfd.go new file mode 100644 index 000000000..c573d7935 --- /dev/null +++ b/pkg/sentry/fsimpl/eventfd/eventfd.go @@ -0,0 +1,284 @@ +// 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 eventfd implements event fds. +package eventfd + +import ( + "math" + "sync" + "syscall" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/fdnotifier" + "gvisor.dev/gvisor/pkg/log" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" + "gvisor.dev/gvisor/pkg/waiter" +) + +// EventFileDescription implements 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. +type EventFileDescription struct { + vfsfd vfs.FileDescription + vfs.FileDescriptionDefaultImpl + vfs.DentryMetadataFileDescriptionImpl + + // queue is used to notify interested parties when the event object + // becomes readable or writable. + queue waiter.Queue `state:"zerovalue"` + + // mu protects the fields below. + mu sync.Mutex `state:"nosave"` + + // val is the current value of the event counter. + val uint64 + + // semMode specifies whether the event is in "semaphore" mode. + semMode bool + + // hostfd indicates whether this eventfd is passed through to the host. + hostfd int +} + +var _ vfs.FileDescriptionImpl = (*EventFileDescription)(nil) + +// New creates a new event fd. +func New(vfsObj *vfs.VirtualFilesystem, initVal uint64, semMode bool, flags uint32) (*vfs.FileDescription, error) { + vd := vfsObj.NewAnonVirtualDentry("[eventfd]") + defer vd.DecRef() + efd := &EventFileDescription{ + val: initVal, + semMode: semMode, + hostfd: -1, + } + if err := efd.vfsfd.Init(efd, flags, vd.Mount(), vd.Dentry(), &vfs.FileDescriptionOptions{ + UseDentryMetadata: true, + DenyPRead: true, + DenyPWrite: true, + }); err != nil { + return nil, err + } + return &efd.vfsfd, nil +} + +// HostFD returns the host eventfd associated with this event. +func (efd *EventFileDescription) HostFD() (int, error) { + efd.mu.Lock() + defer efd.mu.Unlock() + if efd.hostfd >= 0 { + return efd.hostfd, nil + } + + flags := linux.EFD_NONBLOCK + if efd.semMode { + flags |= linux.EFD_SEMAPHORE + } + + fd, _, errno := syscall.Syscall(syscall.SYS_EVENTFD2, uintptr(efd.val), uintptr(flags), 0) + if errno != 0 { + return -1, errno + } + + if err := fdnotifier.AddFD(int32(fd), &efd.queue); err != nil { + if closeErr := syscall.Close(int(fd)); closeErr != nil { + log.Warningf("close(%d) eventfd failed: %v", fd, closeErr) + } + return -1, err + } + + efd.hostfd = int(fd) + return efd.hostfd, nil +} + +// Release implements FileDescriptionImpl.Release() +func (efd *EventFileDescription) Release() { + efd.mu.Lock() + defer efd.mu.Unlock() + if efd.hostfd >= 0 { + fdnotifier.RemoveFD(int32(efd.hostfd)) + if closeErr := syscall.Close(int(efd.hostfd)); closeErr != nil { + log.Warningf("close(%d) eventfd failed: %v", efd.hostfd, closeErr) + } + efd.hostfd = -1 + } +} + +// Read implements FileDescriptionImpl.Read. +func (efd *EventFileDescription) Read(ctx context.Context, dst usermem.IOSequence, _ vfs.ReadOptions) (int64, error) { + if dst.NumBytes() < 8 { + return 0, syscall.EINVAL + } + if err := efd.read(ctx, dst); err != nil { + return 0, err + } + return 8, nil +} + +// Write implements FileDescriptionImpl.Write. +func (efd *EventFileDescription) Write(ctx context.Context, src usermem.IOSequence, _ vfs.WriteOptions) (int64, error) { + if src.NumBytes() < 8 { + return 0, syscall.EINVAL + } + if err := efd.write(ctx, src); err != nil { + return 0, err + } + return 8, nil +} + +// Preconditions: Must be called with efd.mu locked. +func (efd *EventFileDescription) hostReadLocked(ctx context.Context, dst usermem.IOSequence) error { + var buf [8]byte + if _, err := syscall.Read(efd.hostfd, buf[:]); err != nil { + if err == syscall.EWOULDBLOCK { + return syserror.ErrWouldBlock + } + return err + } + _, err := dst.CopyOut(ctx, buf[:]) + return err +} + +func (efd *EventFileDescription) read(ctx context.Context, dst usermem.IOSequence) error { + efd.mu.Lock() + if efd.hostfd >= 0 { + defer efd.mu.Unlock() + return efd.hostReadLocked(ctx, dst) + } + + // We can't complete the read if the value is currently zero. + if efd.val == 0 { + efd.mu.Unlock() + return syserror.ErrWouldBlock + } + + // Update the value based on the mode the event is operating in. + var val uint64 + if efd.semMode { + val = 1 + // Consistent with Linux, this is done even if writing to memory fails. + efd.val-- + } else { + val = efd.val + efd.val = 0 + } + + efd.mu.Unlock() + + // Notify writers. We do this even if we were already writable because + // it is possible that a writer is waiting to write the maximum value + // to the event. + efd.queue.Notify(waiter.EventOut) + + var buf [8]byte + usermem.ByteOrder.PutUint64(buf[:], val) + _, err := dst.CopyOut(ctx, buf[:]) + return err +} + +// Preconditions: Must be called with efd.mu locked. +func (efd *EventFileDescription) hostWriteLocked(val uint64) error { + var buf [8]byte + usermem.ByteOrder.PutUint64(buf[:], val) + _, err := syscall.Write(efd.hostfd, buf[:]) + if err == syscall.EWOULDBLOCK { + return syserror.ErrWouldBlock + } + return err +} + +func (efd *EventFileDescription) write(ctx context.Context, src usermem.IOSequence) error { + var buf [8]byte + if _, err := src.CopyIn(ctx, buf[:]); err != nil { + return err + } + val := usermem.ByteOrder.Uint64(buf[:]) + + return efd.Signal(val) +} + +// Signal is an internal function to signal the event fd. +func (efd *EventFileDescription) Signal(val uint64) error { + if val == math.MaxUint64 { + return syscall.EINVAL + } + + efd.mu.Lock() + + if efd.hostfd >= 0 { + defer efd.mu.Unlock() + return efd.hostWriteLocked(val) + } + + // We only allow writes that won't cause the value to go over the max + // uint64 minus 1. + if val > math.MaxUint64-1-efd.val { + efd.mu.Unlock() + return syserror.ErrWouldBlock + } + + efd.val += val + efd.mu.Unlock() + + // Always trigger a notification. + efd.queue.Notify(waiter.EventIn) + + return nil +} + +// Readiness implements waiter.Waitable.Readiness. +func (efd *EventFileDescription) Readiness(mask waiter.EventMask) waiter.EventMask { + efd.mu.Lock() + defer efd.mu.Unlock() + + if efd.hostfd >= 0 { + return fdnotifier.NonBlockingPoll(int32(efd.hostfd), mask) + } + + ready := waiter.EventMask(0) + if efd.val > 0 { + ready |= waiter.EventIn + } + + if efd.val < math.MaxUint64-1 { + ready |= waiter.EventOut + } + + return mask & ready +} + +// EventRegister implements waiter.Waitable.EventRegister. +func (efd *EventFileDescription) EventRegister(entry *waiter.Entry, mask waiter.EventMask) { + efd.queue.EventRegister(entry, mask) + + efd.mu.Lock() + defer efd.mu.Unlock() + if efd.hostfd >= 0 { + fdnotifier.UpdateFD(int32(efd.hostfd)) + } +} + +// EventUnregister implements waiter.Waitable.EventUnregister. +func (efd *EventFileDescription) EventUnregister(entry *waiter.Entry) { + efd.queue.EventUnregister(entry) + + efd.mu.Lock() + defer efd.mu.Unlock() + if efd.hostfd >= 0 { + fdnotifier.UpdateFD(int32(efd.hostfd)) + } +} diff --git a/pkg/sentry/fsimpl/eventfd/eventfd_test.go b/pkg/sentry/fsimpl/eventfd/eventfd_test.go new file mode 100644 index 000000000..20e3adffc --- /dev/null +++ b/pkg/sentry/fsimpl/eventfd/eventfd_test.go @@ -0,0 +1,97 @@ +// Copyright 2018 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package eventfd + +import ( + "testing" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/contexttest" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/usermem" + "gvisor.dev/gvisor/pkg/waiter" +) + +func TestEventFD(t *testing.T) { + initVals := []uint64{ + 0, + // Using a non-zero initial value verifies that writing to an + // eventfd signals when the eventfd's counter was already + // non-zero. + 343, + } + + for _, initVal := range initVals { + ctx := contexttest.Context(t) + vfsObj := &vfs.VirtualFilesystem{} + if err := vfsObj.Init(); err != nil { + t.Fatalf("VFS init: %v", err) + } + + // Make a new eventfd that is writable. + eventfd, err := New(vfsObj, initVal, false, linux.O_RDWR) + if err != nil { + t.Fatalf("New() failed: %v", err) + } + defer eventfd.DecRef() + + // Register a callback for a write event. + w, ch := waiter.NewChannelEntry(nil) + eventfd.EventRegister(&w, waiter.EventIn) + defer eventfd.EventUnregister(&w) + + data := []byte("00000124") + // Create and submit a write request. + n, err := eventfd.Write(ctx, usermem.BytesIOSequence(data), vfs.WriteOptions{}) + if err != nil { + t.Fatal(err) + } + if n != 8 { + t.Errorf("eventfd.write wrote %d bytes, not full int64", n) + } + + // Check if the callback fired due to the write event. + select { + case <-ch: + default: + t.Errorf("Didn't get notified of EventIn after write") + } + } +} + +func TestEventFDStat(t *testing.T) { + ctx := contexttest.Context(t) + vfsObj := &vfs.VirtualFilesystem{} + if err := vfsObj.Init(); err != nil { + t.Fatalf("VFS init: %v", err) + } + + // Make a new eventfd that is writable. + eventfd, err := New(vfsObj, 0, false, linux.O_RDWR) + if err != nil { + t.Fatalf("New() failed: %v", err) + } + defer eventfd.DecRef() + + statx, err := eventfd.Stat(ctx, vfs.StatOptions{ + Mask: linux.STATX_BASIC_STATS, + }) + if err != nil { + t.Fatalf("eventfd.Stat failed: %v", err) + } + if statx.Size != 0 { + t.Errorf("eventfd size should be 0") + } +} diff --git a/pkg/sentry/fsimpl/ext/BUILD b/pkg/sentry/fsimpl/ext/BUILD index 6f78f478f..ff861d0fe 100644 --- a/pkg/sentry/fsimpl/ext/BUILD +++ b/pkg/sentry/fsimpl/ext/BUILD @@ -15,6 +15,17 @@ go_template_instance( }, ) +go_template_instance( + name = "fstree", + out = "fstree.go", + package = "ext", + prefix = "generic", + template = "//pkg/sentry/vfs/genericfstree:generic_fstree", + types = { + "Dentry": "dentry", + }, +) + go_library( name = "ext", srcs = [ @@ -26,6 +37,7 @@ go_library( "extent_file.go", "file_description.go", "filesystem.go", + "fstree.go", "inode.go", "regular_file.go", "symlink.go", @@ -45,6 +57,7 @@ go_library( "//pkg/sentry/fsimpl/ext/disklayout", "//pkg/sentry/kernel/auth", "//pkg/sentry/memmap", + "//pkg/sentry/socket/unix/transport", "//pkg/sentry/syscalls/linux", "//pkg/sentry/vfs", "//pkg/sync", @@ -80,8 +93,8 @@ go_test( "//pkg/sentry/kernel/auth", "//pkg/sentry/vfs", "//pkg/syserror", + "//pkg/test/testutil", "//pkg/usermem", - "//runsc/testutil", "@com_github_google_go-cmp//cmp:go_default_library", "@com_github_google_go-cmp//cmp/cmpopts:go_default_library", ], diff --git a/pkg/sentry/fsimpl/ext/benchmark/benchmark_test.go b/pkg/sentry/fsimpl/ext/benchmark/benchmark_test.go index 2015a8871..89caee3df 100644 --- a/pkg/sentry/fsimpl/ext/benchmark/benchmark_test.go +++ b/pkg/sentry/fsimpl/ext/benchmark/benchmark_test.go @@ -52,7 +52,10 @@ func setUp(b *testing.B, imagePath string) (context.Context, *vfs.VirtualFilesys creds := auth.CredentialsFromContext(ctx) // Create VFS. - vfsObj := vfs.New() + vfsObj := &vfs.VirtualFilesystem{} + if err := vfsObj.Init(); err != nil { + return nil, nil, nil, nil, err + } vfsObj.MustRegisterFilesystemType("extfs", ext.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ AllowUserMount: true, }) diff --git a/pkg/sentry/fsimpl/ext/dentry.go b/pkg/sentry/fsimpl/ext/dentry.go index a080cb189..6bd1a9fc6 100644 --- a/pkg/sentry/fsimpl/ext/dentry.go +++ b/pkg/sentry/fsimpl/ext/dentry.go @@ -22,6 +22,10 @@ import ( type dentry struct { vfsd vfs.Dentry + // Protected by filesystem.mu. + parent *dentry + name string + // inode is the inode represented by this dentry. Multiple Dentries may // share a single non-directory Inode (with hard links). inode is // immutable. @@ -56,3 +60,15 @@ func (d *dentry) DecRef() { // inode.decRef(). d.inode.decRef() } + +// InotifyWithParent implements vfs.DentryImpl.InotifyWithParent. +// +// TODO(gvisor.dev/issue/1479): Implement inotify. +func (d *dentry) InotifyWithParent(events uint32, cookie uint32, et vfs.EventType) {} + +// Watches implements vfs.DentryImpl.Watches. +// +// TODO(gvisor.dev/issue/1479): Implement inotify. +func (d *dentry) Watches() *vfs.Watches { + return nil +} diff --git a/pkg/sentry/fsimpl/ext/directory.go b/pkg/sentry/fsimpl/ext/directory.go index ebb72b75e..12b875c8f 100644 --- a/pkg/sentry/fsimpl/ext/directory.go +++ b/pkg/sentry/fsimpl/ext/directory.go @@ -21,7 +21,6 @@ import ( "gvisor.dev/gvisor/pkg/log" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fsimpl/ext/disklayout" - "gvisor.dev/gvisor/pkg/sentry/memmap" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/syserror" @@ -31,6 +30,10 @@ import ( type directory struct { inode inode + // childCache maps filenames to dentries for children for which dentries + // have been instantiated. childCache is protected by filesystem.mu. + childCache map[string]*dentry + // mu serializes the changes to childList. // Lock Order (outermost locks must be taken first): // directory.mu @@ -50,9 +53,13 @@ type directory struct { childMap map[string]*dirent } -// newDirectroy is the directory constructor. -func newDirectroy(inode inode, newDirent bool) (*directory, error) { - file := &directory{inode: inode, childMap: make(map[string]*dirent)} +// newDirectory is the directory constructor. +func newDirectory(inode inode, newDirent bool) (*directory, error) { + file := &directory{ + inode: inode, + childCache: make(map[string]*dentry), + childMap: make(map[string]*dirent), + } file.inode.impl = file // Initialize childList by reading dirents from the underlying file. @@ -188,14 +195,14 @@ func (fd *directoryFD) IterDirents(ctx context.Context, cb vfs.IterDirentsCallba childType = fs.ToInodeType(childInode.diskInode.Mode().FileType()) } - if !cb.Handle(vfs.Dirent{ + if err := cb.Handle(vfs.Dirent{ Name: child.diskDirent.FileName(), Type: fs.ToDirentType(childType), Ino: uint64(child.diskDirent.Inode()), NextOff: fd.off + 1, - }) { + }); err != nil { dir.childList.InsertBefore(child, fd.iter) - return nil + return err } fd.off++ } @@ -299,9 +306,3 @@ func (fd *directoryFD) Seek(ctx context.Context, offset int64, whence int32) (in fd.off = offset return offset, nil } - -// ConfigureMMap implements vfs.FileDescriptionImpl.ConfigureMMap. -func (fd *directoryFD) ConfigureMMap(ctx context.Context, opts *memmap.MMapOpts) error { - // mmap(2) specifies that EACCESS should be returned for non-regular file fds. - return syserror.EACCES -} diff --git a/pkg/sentry/fsimpl/ext/ext.go b/pkg/sentry/fsimpl/ext/ext.go index 373d23b74..dac6effbf 100644 --- a/pkg/sentry/fsimpl/ext/ext.go +++ b/pkg/sentry/fsimpl/ext/ext.go @@ -30,6 +30,9 @@ import ( "gvisor.dev/gvisor/pkg/syserror" ) +// Name is the name of this filesystem. +const Name = "ext" + // FilesystemType implements vfs.FilesystemType. type FilesystemType struct{} @@ -91,42 +94,61 @@ func isCompatible(sb disklayout.SuperBlock) bool { return true } +// Name implements vfs.FilesystemType.Name. +func (FilesystemType) Name() string { + return Name +} + // GetFilesystem implements vfs.FilesystemType.GetFilesystem. -func (FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { +func (fsType FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { // TODO(b/134676337): Ensure that the user is mounting readonly. If not, // EACCESS should be returned according to mount(2). Filesystem independent // flags (like readonly) are currently not available in pkg/sentry/vfs. + devMinor, err := vfsObj.GetAnonBlockDevMinor() + if err != nil { + return nil, nil, err + } + dev, err := getDeviceFd(source, opts) if err != nil { return nil, nil, err } - fs := filesystem{dev: dev, inodeCache: make(map[uint32]*inode)} - fs.vfsfs.Init(vfsObj, &fs) + fs := filesystem{ + dev: dev, + inodeCache: make(map[uint32]*inode), + devMinor: devMinor, + } + fs.vfsfs.Init(vfsObj, &fsType, &fs) fs.sb, err = readSuperBlock(dev) if err != nil { + fs.vfsfs.DecRef() return nil, nil, err } if fs.sb.Magic() != linux.EXT_SUPER_MAGIC { // mount(2) specifies that EINVAL should be returned if the superblock is // invalid. + fs.vfsfs.DecRef() return nil, nil, syserror.EINVAL } // Refuse to mount if the filesystem is incompatible. if !isCompatible(fs.sb) { + fs.vfsfs.DecRef() return nil, nil, syserror.EINVAL } fs.bgs, err = readBlockGroups(dev, fs.sb) if err != nil { + fs.vfsfs.DecRef() return nil, nil, err } rootInode, err := fs.getOrCreateInodeLocked(disklayout.RootDirInode) if err != nil { + fs.vfsfs.DecRef() return nil, nil, err } rootInode.incRef() diff --git a/pkg/sentry/fsimpl/ext/ext_test.go b/pkg/sentry/fsimpl/ext/ext_test.go index 05f992826..64e9a579f 100644 --- a/pkg/sentry/fsimpl/ext/ext_test.go +++ b/pkg/sentry/fsimpl/ext/ext_test.go @@ -32,9 +32,8 @@ import ( "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/test/testutil" "gvisor.dev/gvisor/pkg/usermem" - - "gvisor.dev/gvisor/runsc/testutil" ) const ( @@ -65,7 +64,10 @@ func setUp(t *testing.T, imagePath string) (context.Context, *vfs.VirtualFilesys creds := auth.CredentialsFromContext(ctx) // Create VFS. - vfsObj := vfs.New() + vfsObj := &vfs.VirtualFilesystem{} + if err := vfsObj.Init(); err != nil { + t.Fatalf("VFS init: %v", err) + } vfsObj.MustRegisterFilesystemType("extfs", FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ AllowUserMount: true, }) @@ -496,9 +498,9 @@ func newIterDirentCb() *iterDirentsCb { } // Handle implements vfs.IterDirentsCallback.Handle. -func (cb *iterDirentsCb) Handle(dirent vfs.Dirent) bool { +func (cb *iterDirentsCb) Handle(dirent vfs.Dirent) error { cb.dirents = append(cb.dirents, dirent) - return true + return nil } // TestIterDirents tests the FileDescriptionImpl.IterDirents functionality. diff --git a/pkg/sentry/fsimpl/ext/filesystem.go b/pkg/sentry/fsimpl/ext/filesystem.go index e05429d41..557963e03 100644 --- a/pkg/sentry/fsimpl/ext/filesystem.go +++ b/pkg/sentry/fsimpl/ext/filesystem.go @@ -22,6 +22,8 @@ import ( "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/fspath" "gvisor.dev/gvisor/pkg/sentry/fsimpl/ext/disklayout" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/syserror" @@ -62,6 +64,10 @@ type filesystem struct { // bgs represents all the block group descriptors for the filesystem. // Immutable after initialization. bgs []disklayout.BlockGroup + + // devMinor is this filesystem's device minor number. Immutable after + // initialization. + devMinor uint32 } // Compiles only if filesystem implements vfs.FilesystemImpl. @@ -87,14 +93,33 @@ func stepLocked(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, inode *inode, write boo } for { - nextVFSD, err := rp.ResolveComponent(vfsd) - if err != nil { - return nil, nil, err + name := rp.Component() + if name == "." { + rp.Advance() + return vfsd, inode, nil } - if nextVFSD == nil { - // Since the Dentry tree is not the sole source of truth for extfs, if it's - // not in the Dentry tree, it might need to be pulled from disk. - childDirent, ok := inode.impl.(*directory).childMap[rp.Component()] + d := vfsd.Impl().(*dentry) + if name == ".." { + isRoot, err := rp.CheckRoot(vfsd) + if err != nil { + return nil, nil, err + } + if isRoot || d.parent == nil { + rp.Advance() + return vfsd, inode, nil + } + if err := rp.CheckMount(&d.parent.vfsd); err != nil { + return nil, nil, err + } + rp.Advance() + return &d.parent.vfsd, d.parent.inode, nil + } + + dir := inode.impl.(*directory) + child, ok := dir.childCache[name] + if !ok { + // We may need to instantiate a new dentry for this child. + childDirent, ok := dir.childMap[name] if !ok { // The underlying inode does not exist on disk. return nil, nil, syserror.ENOENT @@ -113,21 +138,22 @@ func stepLocked(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, inode *inode, write boo } // incRef because this is being added to the dentry tree. childInode.incRef() - child := newDentry(childInode) - vfsd.InsertChild(&child.vfsd, rp.Component()) - - // Continue as usual now that nextVFSD is not nil. - nextVFSD = &child.vfsd + child = newDentry(childInode) + child.parent = d + child.name = name + dir.childCache[name] = child } - nextInode := nextVFSD.Impl().(*dentry).inode - if nextInode.isSymlink() && rp.ShouldFollowSymlink() { - if err := rp.HandleSymlink(inode.impl.(*symlink).target); err != nil { + if err := rp.CheckMount(&child.vfsd); err != nil { + return nil, nil, err + } + if child.inode.isSymlink() && rp.ShouldFollowSymlink() { + if err := rp.HandleSymlink(child.inode.impl.(*symlink).target); err != nil { return nil, nil, err } continue } rp.Advance() - return nextVFSD, nextInode, nil + return &child.vfsd, child.inode, nil } } @@ -255,6 +281,15 @@ func (fs *filesystem) statTo(stat *linux.Statfs) { // TODO(b/134676337): Set Statfs.Flags and Statfs.FSID. } +// AccessAt implements vfs.Filesystem.Impl.AccessAt. +func (fs *filesystem) AccessAt(ctx context.Context, rp *vfs.ResolvingPath, creds *auth.Credentials, ats vfs.AccessTypes) error { + _, inode, err := fs.walk(rp, false) + if err != nil { + return err + } + return inode.checkPermissions(rp.Credentials(), ats) +} + // GetDentryAt implements vfs.FilesystemImpl.GetDentryAt. func (fs *filesystem) GetDentryAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.GetDentryOptions) (*vfs.Dentry, error) { vfsd, inode, err := fs.walk(rp, false) @@ -335,7 +370,9 @@ func (fs *filesystem) StatFSAt(ctx context.Context, rp *vfs.ResolvingPath) (linu } // Release implements vfs.FilesystemImpl.Release. -func (fs *filesystem) Release() {} +func (fs *filesystem) Release() { + fs.vfsfs.VirtualFilesystem().PutAnonBlockDevMinor(fs.devMinor) +} // Sync implements vfs.FilesystemImpl.Sync. func (fs *filesystem) Sync(ctx context.Context) error { @@ -453,8 +490,22 @@ func (fs *filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error return syserror.EROFS } +// BoundEndpointAt implements FilesystemImpl.BoundEndpointAt. +func (fs *filesystem) BoundEndpointAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.BoundEndpointOptions) (transport.BoundEndpoint, error) { + _, inode, err := fs.walk(rp, false) + if err != nil { + return nil, err + } + if err := inode.checkPermissions(rp.Credentials(), vfs.MayWrite); err != nil { + return nil, err + } + + // TODO(b/134676337): Support sockets. + return nil, syserror.ECONNREFUSED +} + // ListxattrAt implements vfs.FilesystemImpl.ListxattrAt. -func (fs *filesystem) ListxattrAt(ctx context.Context, rp *vfs.ResolvingPath) ([]string, error) { +func (fs *filesystem) ListxattrAt(ctx context.Context, rp *vfs.ResolvingPath, size uint64) ([]string, error) { _, _, err := fs.walk(rp, false) if err != nil { return nil, err @@ -463,7 +514,7 @@ func (fs *filesystem) ListxattrAt(ctx context.Context, rp *vfs.ResolvingPath) ([ } // GetxattrAt implements vfs.FilesystemImpl.GetxattrAt. -func (fs *filesystem) GetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) (string, error) { +func (fs *filesystem) GetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.GetxattrOptions) (string, error) { _, _, err := fs.walk(rp, false) if err != nil { return "", err @@ -493,5 +544,5 @@ func (fs *filesystem) RemovexattrAt(ctx context.Context, rp *vfs.ResolvingPath, func (fs *filesystem) PrependPath(ctx context.Context, vfsroot, vd vfs.VirtualDentry, b *fspath.Builder) error { fs.mu.RLock() defer fs.mu.RUnlock() - return vfs.GenericPrependPath(vfsroot, vd, b) + return genericPrependPath(vfsroot, vd.Mount(), vd.Dentry().Impl().(*dentry), b) } diff --git a/pkg/sentry/fsimpl/ext/inode.go b/pkg/sentry/fsimpl/ext/inode.go index 6962083f5..485f86f4b 100644 --- a/pkg/sentry/fsimpl/ext/inode.go +++ b/pkg/sentry/fsimpl/ext/inode.go @@ -136,7 +136,7 @@ func newInode(fs *filesystem, inodeNum uint32) (*inode, error) { } return &f.inode, nil case linux.ModeDirectory: - f, err := newDirectroy(inode, fs.sb.IncompatibleFeatures().DirentFileType) + f, err := newDirectory(inode, fs.sb.IncompatibleFeatures().DirentFileType) if err != nil { return nil, err } @@ -186,7 +186,7 @@ func (in *inode) open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts *vfs.OpenOpt } func (in *inode) checkPermissions(creds *auth.Credentials, ats vfs.AccessTypes) error { - return vfs.GenericCheckPermissions(creds, ats, in.isDir(), uint16(in.diskInode.Mode()), in.diskInode.UID(), in.diskInode.GID()) + return vfs.GenericCheckPermissions(creds, ats, in.diskInode.Mode(), in.diskInode.UID(), in.diskInode.GID()) } // statTo writes the statx fields to the output parameter. @@ -204,6 +204,8 @@ func (in *inode) statTo(stat *linux.Statx) { stat.Atime = in.diskInode.AccessTime().StatxTimestamp() stat.Ctime = in.diskInode.ChangeTime().StatxTimestamp() stat.Mtime = in.diskInode.ModificationTime().StatxTimestamp() + stat.DevMajor = linux.UNNAMED_MAJOR + stat.DevMinor = in.fs.devMinor // TODO(b/134676337): Set stat.Blocks which is the number of 512 byte blocks // (including metadata blocks) required to represent this file. } diff --git a/pkg/sentry/fsimpl/gofer/BUILD b/pkg/sentry/fsimpl/gofer/BUILD index 4ba76a1e8..f5f35a3bc 100644 --- a/pkg/sentry/fsimpl/gofer/BUILD +++ b/pkg/sentry/fsimpl/gofer/BUILD @@ -1,4 +1,4 @@ -load("//tools:defs.bzl", "go_library") +load("//tools:defs.bzl", "go_library", "go_test") load("//tools/go_generics:defs.bzl", "go_template_instance") licenses(["notice"]) @@ -15,18 +15,30 @@ go_template_instance( }, ) +go_template_instance( + name = "fstree", + out = "fstree.go", + package = "gofer", + prefix = "generic", + template = "//pkg/sentry/vfs/genericfstree:generic_fstree", + types = { + "Dentry": "dentry", + }, +) + go_library( name = "gofer", srcs = [ "dentry_list.go", "directory.go", "filesystem.go", + "fstree.go", "gofer.go", "handle.go", - "handle_unsafe.go", + "host_named_pipe.go", "p9file.go", - "pagemath.go", "regular_file.go", + "socket.go", "special_file.go", "symlink.go", "time.go", @@ -36,20 +48,41 @@ go_library( "//pkg/abi/linux", "//pkg/context", "//pkg/fd", + "//pkg/fdnotifier", "//pkg/fspath", "//pkg/log", "//pkg/p9", "//pkg/safemem", "//pkg/sentry/fs/fsutil", + "//pkg/sentry/fsimpl/host", + "//pkg/sentry/hostfd", + "//pkg/sentry/kernel", "//pkg/sentry/kernel/auth", + "//pkg/sentry/kernel/pipe", "//pkg/sentry/kernel/time", "//pkg/sentry/memmap", "//pkg/sentry/pgalloc", "//pkg/sentry/platform", + "//pkg/sentry/socket/control", + "//pkg/sentry/socket/unix", + "//pkg/sentry/socket/unix/transport", "//pkg/sentry/usage", "//pkg/sentry/vfs", + "//pkg/syserr", "//pkg/syserror", "//pkg/unet", "//pkg/usermem", + "//pkg/waiter", + "@org_golang_x_sys//unix:go_default_library", + ], +) + +go_test( + name = "gofer_test", + srcs = ["gofer_test.go"], + library = ":gofer", + deps = [ + "//pkg/p9", + "//pkg/sentry/contexttest", ], ) diff --git a/pkg/sentry/fsimpl/gofer/directory.go b/pkg/sentry/fsimpl/gofer/directory.go index 6d4ebc2bf..b98218753 100644 --- a/pkg/sentry/fsimpl/gofer/directory.go +++ b/pkg/sentry/fsimpl/gofer/directory.go @@ -15,27 +15,100 @@ package gofer import ( + "fmt" "sync" "sync/atomic" "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/p9" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/kernel/pipe" + "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" ) func (d *dentry) isDir() bool { return d.fileType() == linux.S_IFDIR } -// Preconditions: d.dirMu must be locked. d.isDir(). fs.opts.interop != -// InteropModeShared. -func (d *dentry) cacheNegativeChildLocked(name string) { - if d.negativeChildren == nil { - d.negativeChildren = make(map[string]struct{}) +// Preconditions: filesystem.renameMu must be locked. d.dirMu must be locked. +// d.isDir(). child must be a newly-created dentry that has never had a parent. +func (d *dentry) cacheNewChildLocked(child *dentry, name string) { + d.IncRef() // reference held by child on its parent + child.parent = d + child.name = name + if d.children == nil { + d.children = make(map[string]*dentry) } - d.negativeChildren[name] = struct{}{} + d.children[name] = child +} + +// Preconditions: d.dirMu must be locked. d.isDir(). +func (d *dentry) cacheNegativeLookupLocked(name string) { + // Don't cache negative lookups if InteropModeShared is in effect (since + // this makes remote lookup unavoidable), or if d.isSynthetic() (in which + // case the only files in the directory are those for which a dentry exists + // in d.children). Instead, just delete any previously-cached dentry. + if d.fs.opts.interop == InteropModeShared || d.isSynthetic() { + delete(d.children, name) + return + } + if d.children == nil { + d.children = make(map[string]*dentry) + } + d.children[name] = nil +} + +type createSyntheticOpts struct { + name string + mode linux.FileMode + kuid auth.KUID + kgid auth.KGID + + // The endpoint for a synthetic socket. endpoint should be nil if the file + // being created is not a socket. + endpoint transport.BoundEndpoint + + // pipe should be nil if the file being created is not a pipe. + pipe *pipe.VFSPipe +} + +// createSyntheticChildLocked creates a synthetic file with the given name +// in d. +// +// Preconditions: d.dirMu must be locked. d.isDir(). d does not already contain +// a child with the given name. +func (d *dentry) createSyntheticChildLocked(opts *createSyntheticOpts) { + d2 := &dentry{ + refs: 1, // held by d + fs: d.fs, + mode: uint32(opts.mode), + uid: uint32(opts.kuid), + gid: uint32(opts.kgid), + blockSize: usermem.PageSize, // arbitrary + handle: handle{ + fd: -1, + }, + nlink: uint32(2), + } + switch opts.mode.FileType() { + case linux.S_IFDIR: + // Nothing else needs to be done. + case linux.S_IFSOCK: + d2.endpoint = opts.endpoint + case linux.S_IFIFO: + d2.pipe = opts.pipe + default: + panic(fmt.Sprintf("failed to create synthetic file of unrecognized type: %v", opts.mode.FileType())) + } + d2.pf.dentry = d2 + d2.vfsd.Init(d2) + + d.cacheNewChildLocked(d2, opts.name) + d.syntheticChildren++ } type directoryFD struct { @@ -56,17 +129,22 @@ func (fd *directoryFD) IterDirents(ctx context.Context, cb vfs.IterDirentsCallba fd.mu.Lock() defer fd.mu.Unlock() + d := fd.dentry() if fd.dirents == nil { - ds, err := fd.dentry().getDirents(ctx) + ds, err := d.getDirents(ctx) if err != nil { return err } fd.dirents = ds } + if d.cachedMetadataAuthoritative() { + d.touchAtime(fd.vfsfd.Mount()) + } + for fd.off < int64(len(fd.dirents)) { - if !cb.Handle(fd.dirents[fd.off]) { - return nil + if err := cb.Handle(fd.dirents[fd.off]); err != nil { + return err } fd.off++ } @@ -75,23 +153,21 @@ func (fd *directoryFD) IterDirents(ctx context.Context, cb vfs.IterDirentsCallba // Preconditions: d.isDir(). There exists at least one directoryFD representing d. func (d *dentry) getDirents(ctx context.Context) ([]vfs.Dirent, error) { - // 9P2000.L's readdir does not specify behavior in the presence of - // concurrent mutation of an iterated directory, so implementations may - // duplicate or omit entries in this case, which violates POSIX semantics. - // Thus we read all directory entries while holding d.dirMu to exclude - // directory mutations. (Note that it is impossible for the client to - // exclude concurrent mutation from other remote filesystem users. Since - // there is no way to detect if the server has incorrectly omitted - // directory entries, we simply assume that the server is well-behaved - // under InteropModeShared.) This is inconsistent with Linux (which appears - // to assume that directory fids have the correct semantics, and translates - // struct file_operations::readdir calls directly to readdir RPCs), but is - // consistent with VFS1. - // - // NOTE(b/135560623): In particular, some gofer implementations may not - // retain state between calls to Readdir, so may not provide a coherent - // directory stream across in the presence of mutation. + // NOTE(b/135560623): 9P2000.L's readdir does not specify behavior in the + // presence of concurrent mutation of an iterated directory, so + // implementations may duplicate or omit entries in this case, which + // violates POSIX semantics. Thus we read all directory entries while + // holding d.dirMu to exclude directory mutations. (Note that it is + // impossible for the client to exclude concurrent mutation from other + // remote filesystem users. Since there is no way to detect if the server + // has incorrectly omitted directory entries, we simply assume that the + // server is well-behaved under InteropModeShared.) This is inconsistent + // with Linux (which appears to assume that directory fids have the correct + // semantics, and translates struct file_operations::readdir calls directly + // to readdir RPCs), but is consistent with VFS1. + // filesystem.renameMu is needed for d.parent, and must be locked before + // dentry.dirMu. d.fs.renameMu.RLock() defer d.fs.renameMu.RUnlock() d.dirMu.Lock() @@ -102,7 +178,7 @@ func (d *dentry) getDirents(ctx context.Context) ([]vfs.Dirent, error) { // It's not clear if 9P2000.L's readdir is expected to return "." and "..", // so we generate them here. - parent := d.vfsd.ParentOrSelf().Impl().(*dentry) + parent := genericParentOrSelf(d) dirents := []vfs.Dirent{ { Name: ".", @@ -117,50 +193,81 @@ func (d *dentry) getDirents(ctx context.Context) ([]vfs.Dirent, error) { NextOff: 2, }, } - off := uint64(0) - const count = 64 * 1024 // for consistency with the vfs1 client - d.handleMu.RLock() - defer d.handleMu.RUnlock() - if !d.handleReadable { - // This should not be possible because a readable handle should have - // been opened when the calling directoryFD was opened. - panic("gofer.dentry.getDirents called without a readable handle") - } - for { - p9ds, err := d.handle.file.readdir(ctx, off, count) - if err != nil { - return nil, err + var realChildren map[string]struct{} + if !d.isSynthetic() { + if d.syntheticChildren != 0 && d.fs.opts.interop == InteropModeShared { + // Record the set of children d actually has so that we don't emit + // duplicate entries for synthetic children. + realChildren = make(map[string]struct{}) } - if len(p9ds) == 0 { - // Cache dirents for future directoryFDs if permitted. - if d.fs.opts.interop != InteropModeShared { - d.dirents = dirents + off := uint64(0) + const count = 64 * 1024 // for consistency with the vfs1 client + d.handleMu.RLock() + if !d.handleReadable { + // This should not be possible because a readable handle should + // have been opened when the calling directoryFD was opened. + d.handleMu.RUnlock() + panic("gofer.dentry.getDirents called without a readable handle") + } + for { + p9ds, err := d.handle.file.readdir(ctx, off, count) + if err != nil { + d.handleMu.RUnlock() + return nil, err + } + if len(p9ds) == 0 { + d.handleMu.RUnlock() + break } - return dirents, nil + for _, p9d := range p9ds { + if p9d.Name == "." || p9d.Name == ".." { + continue + } + dirent := vfs.Dirent{ + Name: p9d.Name, + Ino: p9d.QID.Path, + NextOff: int64(len(dirents) + 1), + } + // p9 does not expose 9P2000.U's DMDEVICE, DMNAMEDPIPE, or + // DMSOCKET. + switch p9d.Type { + case p9.TypeSymlink: + dirent.Type = linux.DT_LNK + case p9.TypeDir: + dirent.Type = linux.DT_DIR + default: + dirent.Type = linux.DT_REG + } + dirents = append(dirents, dirent) + if realChildren != nil { + realChildren[p9d.Name] = struct{}{} + } + } + off = p9ds[len(p9ds)-1].Offset } - for _, p9d := range p9ds { - if p9d.Name == "." || p9d.Name == ".." { + } + // Emit entries for synthetic children. + if d.syntheticChildren != 0 { + for _, child := range d.children { + if child == nil || !child.isSynthetic() { continue } - dirent := vfs.Dirent{ - Name: p9d.Name, - Ino: p9d.QID.Path, - NextOff: int64(len(dirents) + 1), - } - // p9 does not expose 9P2000.U's DMDEVICE, DMNAMEDPIPE, or - // DMSOCKET. - switch p9d.Type { - case p9.TypeSymlink: - dirent.Type = linux.DT_LNK - case p9.TypeDir: - dirent.Type = linux.DT_DIR - default: - dirent.Type = linux.DT_REG + if _, ok := realChildren[child.name]; ok { + continue } - dirents = append(dirents, dirent) + dirents = append(dirents, vfs.Dirent{ + Name: child.name, + Type: uint8(atomic.LoadUint32(&child.mode) >> 12), + Ino: child.ino, + NextOff: int64(len(dirents) + 1), + }) } - off = p9ds[len(p9ds)-1].Offset } + // Cache dirents for future directoryFDs if permitted. + if d.cachedMetadataAuthoritative() { + d.dirents = dirents + } + return dirents, nil } // Seek implements vfs.FileDescriptionImpl.Seek. diff --git a/pkg/sentry/fsimpl/gofer/filesystem.go b/pkg/sentry/fsimpl/gofer/filesystem.go index 138adb9f7..36e0e1856 100644 --- a/pkg/sentry/fsimpl/gofer/filesystem.go +++ b/pkg/sentry/fsimpl/gofer/filesystem.go @@ -21,20 +21,28 @@ import ( "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/fspath" "gvisor.dev/gvisor/pkg/p9" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/host" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/kernel/pipe" + "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" ) // Sync implements vfs.FilesystemImpl.Sync. func (fs *filesystem) Sync(ctx context.Context) error { - // Snapshot current dentries and special files. + // Snapshot current syncable dentries and special files. fs.syncMu.Lock() - ds := make([]*dentry, 0, len(fs.dentries)) - for d := range fs.dentries { + ds := make([]*dentry, 0, len(fs.syncableDentries)) + for d := range fs.syncableDentries { + d.IncRef() ds = append(ds, d) } sffds := make([]*specialFileFD, 0, len(fs.specialFileFDs)) for sffd := range fs.specialFileFDs { + sffd.vfsfd.IncRef() sffds = append(sffds, sffd) } fs.syncMu.Unlock() @@ -45,9 +53,6 @@ func (fs *filesystem) Sync(ctx context.Context) error { // Sync regular files. for _, d := range ds { - if !d.TryIncRef() { - continue - } err := d.syncSharedHandle(ctx) d.DecRef() if err != nil && retErr == nil { @@ -58,9 +63,6 @@ func (fs *filesystem) Sync(ctx context.Context) error { // Sync special files, which may be writable but do not use dentry shared // handles (so they won't be synced by the above). for _, sffd := range sffds { - if !sffd.vfsfd.TryIncRef() { - continue - } err := sffd.Sync(ctx) sffd.vfsfd.DecRef() if err != nil && retErr == nil { @@ -112,13 +114,15 @@ func putDentrySlice(ds *[]*dentry) { // to *ds. // // Preconditions: fs.renameMu must be locked. d.dirMu must be locked. -// !rp.Done(). If fs.opts.interop == InteropModeShared, then d's cached -// metadata must be up to date. +// !rp.Done(). If !d.cachedMetadataAuthoritative(), then d's cached metadata +// must be up to date. +// +// Postconditions: The returned dentry's cached metadata is up to date. func (fs *filesystem) stepLocked(ctx context.Context, rp *vfs.ResolvingPath, d *dentry, ds **[]*dentry) (*dentry, error) { if !d.isDir() { return nil, syserror.ENOTDIR } - if err := d.checkPermissions(rp.Credentials(), vfs.MayExec, true); err != nil { + if err := d.checkPermissions(rp.Credentials(), vfs.MayExec); err != nil { return nil, err } afterSymlink: @@ -128,39 +132,42 @@ afterSymlink: return d, nil } if name == ".." { - parentVFSD, err := rp.ResolveParent(&d.vfsd) - if err != nil { + if isRoot, err := rp.CheckRoot(&d.vfsd); err != nil { return nil, err + } else if isRoot || d.parent == nil { + rp.Advance() + return d, nil } - parent := parentVFSD.Impl().(*dentry) - if fs.opts.interop == InteropModeShared { - // We must assume that parentVFSD is correct, because if d has been - // moved elsewhere in the remote filesystem so that its parent has - // changed, we have no way of determining its new parent's location - // in the filesystem. Get updated metadata for parentVFSD. - _, attrMask, attr, err := parent.file.getAttr(ctx, dentryAttrMask()) + // We must assume that d.parent is correct, because if d has been moved + // elsewhere in the remote filesystem so that its parent has changed, + // we have no way of determining its new parent's location in the + // filesystem. + // + // Call rp.CheckMount() before updating d.parent's metadata, since if + // we traverse to another mount then d.parent's metadata is irrelevant. + if err := rp.CheckMount(&d.parent.vfsd); err != nil { + return nil, err + } + if d != d.parent && !d.cachedMetadataAuthoritative() { + _, attrMask, attr, err := d.parent.file.getAttr(ctx, dentryAttrMask()) if err != nil { return nil, err } - parent.updateFromP9Attrs(attrMask, &attr) + d.parent.updateFromP9Attrs(attrMask, &attr) } rp.Advance() - return parent, nil + return d.parent, nil } - childVFSD, err := rp.ResolveChild(&d.vfsd, name) - if err != nil { - return nil, err - } - // FIXME(jamieliu): Linux performs revalidation before mount lookup - // (fs/namei.c:lookup_fast() => __d_lookup_rcu(), d_revalidate(), - // __follow_mount_rcu()). - child, err := fs.revalidateChildLocked(ctx, rp.VirtualFilesystem(), d, name, childVFSD, ds) + child, err := fs.getChildLocked(ctx, rp.VirtualFilesystem(), d, name, ds) if err != nil { return nil, err } if child == nil { return nil, syserror.ENOENT } + if err := rp.CheckMount(&child.vfsd); err != nil { + return nil, err + } if child.isSymlink() && rp.ShouldFollowSymlink() { target, err := child.readlink(ctx, rp.Mount()) if err != nil { @@ -175,38 +182,37 @@ afterSymlink: return child, nil } -// revalidateChildLocked must be called after a call to parent.vfsd.Child(name) -// or vfs.ResolvingPath.ResolveChild(name) returns childVFSD (which may be -// nil) to verify that the returned child (or lack thereof) is correct. If no file -// exists at name, revalidateChildLocked returns (nil, nil). +// getChildLocked returns a dentry representing the child of parent with the +// given name. If no such child exists, getChildLocked returns (nil, nil). // // Preconditions: fs.renameMu must be locked. parent.dirMu must be locked. // parent.isDir(). name is not "." or "..". // -// Postconditions: If revalidateChildLocked returns a non-nil dentry, its -// cached metadata is up to date. -func (fs *filesystem) revalidateChildLocked(ctx context.Context, vfsObj *vfs.VirtualFilesystem, parent *dentry, name string, childVFSD *vfs.Dentry, ds **[]*dentry) (*dentry, error) { - if childVFSD != nil && fs.opts.interop != InteropModeShared { - // We have a cached dentry that is assumed to be correct. - return childVFSD.Impl().(*dentry), nil - } - // We either don't have a cached dentry or need to verify that it's still - // correct, either of which requires a remote lookup. Check if this name is - // valid before performing the lookup. +// Postconditions: If getChildLocked returns a non-nil dentry, its cached +// metadata is up to date. +func (fs *filesystem) getChildLocked(ctx context.Context, vfsObj *vfs.VirtualFilesystem, parent *dentry, name string, ds **[]*dentry) (*dentry, error) { if len(name) > maxFilenameLen { return nil, syserror.ENAMETOOLONG } - // Check if we've already cached this lookup with a negative result. - if _, ok := parent.negativeChildren[name]; ok { - return nil, nil + child, ok := parent.children[name] + if (ok && fs.opts.interop != InteropModeShared) || parent.isSynthetic() { + // Whether child is nil or not, it is cached information that is + // assumed to be correct. + return child, nil } - // Perform the remote lookup. + // We either don't have cached information or need to verify that it's + // still correct, either of which requires a remote lookup. Check if this + // name is valid before performing the lookup. + return fs.revalidateChildLocked(ctx, vfsObj, parent, name, child, ds) +} + +// Preconditions: As for getChildLocked. !parent.isSynthetic(). +func (fs *filesystem) revalidateChildLocked(ctx context.Context, vfsObj *vfs.VirtualFilesystem, parent *dentry, name string, child *dentry, ds **[]*dentry) (*dentry, error) { qid, file, attrMask, attr, err := parent.file.walkGetAttrOne(ctx, name) if err != nil && err != syserror.ENOENT { return nil, err } - if childVFSD != nil { - child := childVFSD.Impl().(*dentry) + if child != nil { if !file.isNil() && qid.Path == child.ino { // The file at this path hasn't changed. Just update cached // metadata. @@ -214,29 +220,44 @@ func (fs *filesystem) revalidateChildLocked(ctx context.Context, vfsObj *vfs.Vir child.updateFromP9Attrs(attrMask, &attr) return child, nil } - // The file at this path has changed or no longer exists. Remove - // the stale dentry from the tree, and re-evaluate its caching - // status (i.e. if it has 0 references, drop it). - vfsObj.ForceDeleteDentry(childVFSD) + if file.isNil() && child.isSynthetic() { + // We have a synthetic file, and no remote file has arisen to + // replace it. + return child, nil + } + // The file at this path has changed or no longer exists. Mark the + // dentry invalidated, and re-evaluate its caching status (i.e. if it + // has 0 references, drop it). Wait to update parent.children until we + // know what to replace the existing dentry with (i.e. one of the + // returns below), to avoid a redundant map access. + vfsObj.InvalidateDentry(&child.vfsd) + if child.isSynthetic() { + // Normally we don't mark invalidated dentries as deleted since + // they may still exist (but at a different path), and also for + // consistency with Linux. However, synthetic files are guaranteed + // to become unreachable if their dentries are invalidated, so + // treat their invalidation as deletion. + child.setDeleted() + parent.syntheticChildren-- + child.decRefLocked() + parent.dirents = nil + } *ds = appendDentry(*ds, child) - childVFSD = nil } if file.isNil() { // No file exists at this path now. Cache the negative lookup if // allowed. - if fs.opts.interop != InteropModeShared { - parent.cacheNegativeChildLocked(name) - } + parent.cacheNegativeLookupLocked(name) return nil, nil } // Create a new dentry representing the file. - child, err := fs.newDentry(ctx, file, qid, attrMask, &attr) + child, err = fs.newDentry(ctx, file, qid, attrMask, &attr) if err != nil { file.close(ctx) + delete(parent.children, name) return nil, err } - parent.IncRef() // reference held by child on its parent - parent.vfsd.InsertChild(&child.vfsd, name) + parent.cacheNewChildLocked(child, name) // For now, child has 0 references, so our caller should call // child.checkCachingLocked(). *ds = appendDentry(*ds, child) @@ -248,8 +269,9 @@ func (fs *filesystem) revalidateChildLocked(ctx context.Context, vfsObj *vfs.Vir // rp.Start().Impl().(*dentry)). It does not check that the returned directory // is searchable by the provider of rp. // -// Preconditions: fs.renameMu must be locked. !rp.Done(). If fs.opts.interop == -// InteropModeShared, then d's cached metadata must be up to date. +// Preconditions: fs.renameMu must be locked. !rp.Done(). If +// !d.cachedMetadataAuthoritative(), then d's cached metadata must be up to +// date. func (fs *filesystem) walkParentDirLocked(ctx context.Context, rp *vfs.ResolvingPath, d *dentry, ds **[]*dentry) (*dentry, error) { for !rp.Final() { d.dirMu.Lock() @@ -271,7 +293,7 @@ func (fs *filesystem) walkParentDirLocked(ctx context.Context, rp *vfs.Resolving // Preconditions: fs.renameMu must be locked. func (fs *filesystem) resolveLocked(ctx context.Context, rp *vfs.ResolvingPath, ds **[]*dentry) (*dentry, error) { d := rp.Start().Impl().(*dentry) - if fs.opts.interop == InteropModeShared { + if !d.cachedMetadataAuthoritative() { // Get updated metadata for rp.Start() as required by fs.stepLocked(). if err := d.updateFromGetattr(ctx); err != nil { return nil, err @@ -293,16 +315,17 @@ func (fs *filesystem) resolveLocked(ctx context.Context, rp *vfs.ResolvingPath, } // doCreateAt checks that creating a file at rp is permitted, then invokes -// create to do so. +// createInRemoteDir (if the parent directory is a real remote directory) or +// createInSyntheticDir (if the parent directory is synthetic) to do so. // // Preconditions: !rp.Done(). For the final path component in rp, // !rp.ShouldFollowSymlink(). -func (fs *filesystem) doCreateAt(ctx context.Context, rp *vfs.ResolvingPath, dir bool, create func(parent *dentry, name string) error) error { +func (fs *filesystem) doCreateAt(ctx context.Context, rp *vfs.ResolvingPath, dir bool, createInRemoteDir func(parent *dentry, name string) error, createInSyntheticDir func(parent *dentry, name string) error) error { var ds *[]*dentry fs.renameMu.RLock() defer fs.renameMuRUnlockAndCheckCaching(&ds) start := rp.Start().Impl().(*dentry) - if fs.opts.interop == InteropModeShared { + if !start.cachedMetadataAuthoritative() { // Get updated metadata for start as required by // fs.walkParentDirLocked(). if err := start.updateFromGetattr(ctx); err != nil { @@ -313,12 +336,9 @@ func (fs *filesystem) doCreateAt(ctx context.Context, rp *vfs.ResolvingPath, dir if err != nil { return err } - if err := parent.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true); err != nil { + if err := parent.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec); err != nil { return err } - if parent.isDeleted() { - return syserror.ENOENT - } name := rp.Component() if name == "." || name == ".." { return syserror.EEXIST @@ -329,6 +349,9 @@ func (fs *filesystem) doCreateAt(ctx context.Context, rp *vfs.ResolvingPath, dir if !dir && rp.MustBeDir() { return syserror.ENOENT } + if parent.isDeleted() { + return syserror.ENOENT + } mnt := rp.Mount() if err := mnt.CheckBeginWrite(); err != nil { return err @@ -336,6 +359,20 @@ func (fs *filesystem) doCreateAt(ctx context.Context, rp *vfs.ResolvingPath, dir defer mnt.EndWrite() parent.dirMu.Lock() defer parent.dirMu.Unlock() + if parent.isSynthetic() { + if child := parent.children[name]; child != nil { + return syserror.EEXIST + } + if createInSyntheticDir == nil { + return syserror.EPERM + } + if err := createInSyntheticDir(parent, name); err != nil { + return err + } + parent.touchCMtime() + parent.dirents = nil + return nil + } if fs.opts.interop == InteropModeShared { // The existence of a dentry at name would be inconclusive because the // file it represents may have been deleted from the remote filesystem, @@ -344,18 +381,21 @@ func (fs *filesystem) doCreateAt(ctx context.Context, rp *vfs.ResolvingPath, dir // will fail with EEXIST like we would have. If the RPC succeeds, and a // stale dentry exists, the dentry will fail revalidation next time // it's used. - return create(parent, name) + return createInRemoteDir(parent, name) } - if parent.vfsd.Child(name) != nil { + if child := parent.children[name]; child != nil { return syserror.EEXIST } // No cached dentry exists; however, there might still be an existing file // at name. As above, we attempt the file creation RPC anyway. - if err := create(parent, name); err != nil { + if err := createInRemoteDir(parent, name); err != nil { return err } - parent.touchCMtime(ctx) - delete(parent.negativeChildren, name) + if child, ok := parent.children[name]; ok && child == nil { + // Delete the now-stale negative dentry. + delete(parent.children, name) + } + parent.touchCMtime() parent.dirents = nil return nil } @@ -366,7 +406,7 @@ func (fs *filesystem) unlinkAt(ctx context.Context, rp *vfs.ResolvingPath, dir b fs.renameMu.RLock() defer fs.renameMuRUnlockAndCheckCaching(&ds) start := rp.Start().Impl().(*dentry) - if fs.opts.interop == InteropModeShared { + if !start.cachedMetadataAuthoritative() { // Get updated metadata for start as required by // fs.walkParentDirLocked(). if err := start.updateFromGetattr(ctx); err != nil { @@ -377,7 +417,7 @@ func (fs *filesystem) unlinkAt(ctx context.Context, rp *vfs.ResolvingPath, dir b if err != nil { return err } - if err := parent.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true); err != nil { + if err := parent.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec); err != nil { return err } if err := rp.Mount().CheckBeginWrite(); err != nil { @@ -400,66 +440,110 @@ func (fs *filesystem) unlinkAt(ctx context.Context, rp *vfs.ResolvingPath, dir b } vfsObj := rp.VirtualFilesystem() mntns := vfs.MountNamespaceFromContext(ctx) + defer mntns.DecRef() parent.dirMu.Lock() defer parent.dirMu.Unlock() - childVFSD := parent.vfsd.Child(name) - var child *dentry + child, ok := parent.children[name] + if ok && child == nil { + return syserror.ENOENT + } // We only need a dentry representing the file at name if it can be a mount - // point. If childVFSD is nil, then it can't be a mount point. If childVFSD - // is non-nil but stale, the actual file can't be a mount point either; we + // point. If child is nil, then it can't be a mount point. If child is + // non-nil but stale, the actual file can't be a mount point either; we // detect this case by just speculatively calling PrepareDeleteDentry and // only revalidating the dentry if that fails (indicating that the existing // dentry is a mount point). - if childVFSD != nil { - child = childVFSD.Impl().(*dentry) - if err := vfsObj.PrepareDeleteDentry(mntns, childVFSD); err != nil { - child, err = fs.revalidateChildLocked(ctx, vfsObj, parent, name, childVFSD, &ds) + if child != nil { + child.dirMu.Lock() + defer child.dirMu.Unlock() + if err := vfsObj.PrepareDeleteDentry(mntns, &child.vfsd); err != nil { + if parent.cachedMetadataAuthoritative() { + return err + } + child, err = fs.revalidateChildLocked(ctx, vfsObj, parent, name, child, &ds) if err != nil { return err } if child != nil { - childVFSD = &child.vfsd - if err := vfsObj.PrepareDeleteDentry(mntns, childVFSD); err != nil { + if err := vfsObj.PrepareDeleteDentry(mntns, &child.vfsd); err != nil { return err } - } else { - childVFSD = nil } } - } else if _, ok := parent.negativeChildren[name]; ok { - return syserror.ENOENT } flags := uint32(0) + // If a dentry exists, use it for best-effort checks on its deletability. if dir { - if child != nil && !child.isDir() { - return syserror.ENOTDIR + if child != nil { + // child must be an empty directory. + if child.syntheticChildren != 0 { + // This is definitely not an empty directory, irrespective of + // fs.opts.interop. + vfsObj.AbortDeleteDentry(&child.vfsd) + return syserror.ENOTEMPTY + } + // If InteropModeShared is in effect and the first call to + // PrepareDeleteDentry above succeeded, then child wasn't + // revalidated (so we can't expect its file type to be correct) and + // individually revalidating its children (to confirm that they + // still exist) would be a waste of time. + if child.cachedMetadataAuthoritative() { + if !child.isDir() { + vfsObj.AbortDeleteDentry(&child.vfsd) + return syserror.ENOTDIR + } + for _, grandchild := range child.children { + if grandchild != nil { + vfsObj.AbortDeleteDentry(&child.vfsd) + return syserror.ENOTEMPTY + } + } + } } flags = linux.AT_REMOVEDIR } else { + // child must be a non-directory file. if child != nil && child.isDir() { + vfsObj.AbortDeleteDentry(&child.vfsd) return syserror.EISDIR } if rp.MustBeDir() { + if child != nil { + vfsObj.AbortDeleteDentry(&child.vfsd) + } return syserror.ENOTDIR } } - err = parent.file.unlinkAt(ctx, name, flags) - if err != nil { - if childVFSD != nil { - vfsObj.AbortDeleteDentry(childVFSD) + if parent.isSynthetic() { + if child == nil { + return syserror.ENOENT + } + } else { + err = parent.file.unlinkAt(ctx, name, flags) + if err != nil { + if child != nil { + vfsObj.AbortDeleteDentry(&child.vfsd) + } + return err } - return err - } - if fs.opts.interop != InteropModeShared { - parent.touchCMtime(ctx) - parent.cacheNegativeChildLocked(name) - parent.dirents = nil } if child != nil { + vfsObj.CommitDeleteDentry(&child.vfsd) child.setDeleted() - vfsObj.CommitDeleteDentry(childVFSD) + if child.isSynthetic() { + parent.syntheticChildren-- + child.decRefLocked() + } ds = appendDentry(ds, child) } + parent.cacheNegativeLookupLocked(name) + if parent.cachedMetadataAuthoritative() { + parent.dirents = nil + parent.touchCMtime() + if dir { + parent.decLinks() + } + } return nil } @@ -498,6 +582,18 @@ func (fs *filesystem) renameMuUnlockAndCheckCaching(ds **[]*dentry) { putDentrySlice(*ds) } +// AccessAt implements vfs.Filesystem.Impl.AccessAt. +func (fs *filesystem) AccessAt(ctx context.Context, rp *vfs.ResolvingPath, creds *auth.Credentials, ats vfs.AccessTypes) error { + var ds *[]*dentry + fs.renameMu.RLock() + defer fs.renameMuRUnlockAndCheckCaching(&ds) + d, err := fs.resolveLocked(ctx, rp, &ds) + if err != nil { + return err + } + return d.checkPermissions(creds, ats) +} + // GetDentryAt implements vfs.FilesystemImpl.GetDentryAt. func (fs *filesystem) GetDentryAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.GetDentryOptions) (*vfs.Dentry, error) { var ds *[]*dentry @@ -511,7 +607,7 @@ func (fs *filesystem) GetDentryAt(ctx context.Context, rp *vfs.ResolvingPath, op if !d.isDir() { return nil, syserror.ENOTDIR } - if err := d.checkPermissions(rp.Credentials(), vfs.MayExec, true); err != nil { + if err := d.checkPermissions(rp.Credentials(), vfs.MayExec); err != nil { return nil, err } } @@ -525,7 +621,7 @@ func (fs *filesystem) GetParentDentryAt(ctx context.Context, rp *vfs.ResolvingPa fs.renameMu.RLock() defer fs.renameMuRUnlockAndCheckCaching(&ds) start := rp.Start().Impl().(*dentry) - if fs.opts.interop == InteropModeShared { + if !start.cachedMetadataAuthoritative() { // Get updated metadata for start as required by // fs.walkParentDirLocked(). if err := start.updateFromGetattr(ctx); err != nil { @@ -548,15 +644,42 @@ func (fs *filesystem) LinkAt(ctx context.Context, rp *vfs.ResolvingPath, vd vfs. } // 9P2000.L supports hard links, but we don't. return syserror.EPERM - }) + }, nil) } // MkdirAt implements vfs.FilesystemImpl.MkdirAt. func (fs *filesystem) MkdirAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.MkdirOptions) error { + creds := rp.Credentials() return fs.doCreateAt(ctx, rp, true /* dir */, func(parent *dentry, name string) error { - creds := rp.Credentials() - _, err := parent.file.mkdir(ctx, name, (p9.FileMode)(opts.Mode), (p9.UID)(creds.EffectiveKUID), (p9.GID)(creds.EffectiveKGID)) - return err + if _, err := parent.file.mkdir(ctx, name, (p9.FileMode)(opts.Mode), (p9.UID)(creds.EffectiveKUID), (p9.GID)(creds.EffectiveKGID)); err != nil { + if !opts.ForSyntheticMountpoint || err == syserror.EEXIST { + return err + } + ctx.Infof("Failed to create remote directory %q: %v; falling back to synthetic directory", name, err) + parent.createSyntheticChildLocked(&createSyntheticOpts{ + name: name, + mode: linux.S_IFDIR | opts.Mode, + kuid: creds.EffectiveKUID, + kgid: creds.EffectiveKGID, + }) + } + if fs.opts.interop != InteropModeShared { + parent.incLinks() + } + return nil + }, func(parent *dentry, name string) error { + if !opts.ForSyntheticMountpoint { + // Can't create non-synthetic files in synthetic directories. + return syserror.EPERM + } + parent.createSyntheticChildLocked(&createSyntheticOpts{ + name: name, + mode: linux.S_IFDIR | opts.Mode, + kuid: creds.EffectiveKUID, + kgid: creds.EffectiveKGID, + }) + parent.incLinks() + return nil }) } @@ -565,8 +688,32 @@ func (fs *filesystem) MknodAt(ctx context.Context, rp *vfs.ResolvingPath, opts v return fs.doCreateAt(ctx, rp, false /* dir */, func(parent *dentry, name string) error { creds := rp.Credentials() _, err := parent.file.mknod(ctx, name, (p9.FileMode)(opts.Mode), opts.DevMajor, opts.DevMinor, (p9.UID)(creds.EffectiveKUID), (p9.GID)(creds.EffectiveKGID)) + // If the gofer does not allow creating a socket or pipe, create a + // synthetic one, i.e. one that is kept entirely in memory. + if err == syserror.EPERM { + switch opts.Mode.FileType() { + case linux.S_IFSOCK: + parent.createSyntheticChildLocked(&createSyntheticOpts{ + name: name, + mode: opts.Mode, + kuid: creds.EffectiveKUID, + kgid: creds.EffectiveKGID, + endpoint: opts.Endpoint, + }) + return nil + case linux.S_IFIFO: + parent.createSyntheticChildLocked(&createSyntheticOpts{ + name: name, + mode: opts.Mode, + kuid: creds.EffectiveKUID, + kgid: creds.EffectiveKGID, + pipe: pipe.NewVFSPipe(true /* isNamed */, pipe.DefaultPipeSize, usermem.PageSize), + }) + return nil + } + } return err - }) + }, nil) } // OpenAt implements vfs.FilesystemImpl.OpenAt. @@ -586,7 +733,7 @@ func (fs *filesystem) OpenAt(ctx context.Context, rp *vfs.ResolvingPath, opts vf defer fs.renameMuRUnlockAndCheckCaching(&ds) start := rp.Start().Impl().(*dentry) - if fs.opts.interop == InteropModeShared { + if !start.cachedMetadataAuthoritative() { // Get updated metadata for start as required by fs.stepLocked(). if err := start.updateFromGetattr(ctx); err != nil { return nil, err @@ -602,14 +749,18 @@ afterTrailingSymlink: return nil, err } // Check for search permission in the parent directory. - if err := parent.checkPermissions(rp.Credentials(), vfs.MayExec, true); err != nil { + if err := parent.checkPermissions(rp.Credentials(), vfs.MayExec); err != nil { return nil, err } // Determine whether or not we need to create a file. parent.dirMu.Lock() child, err := fs.stepLocked(ctx, rp, parent, &ds) if err == syserror.ENOENT && mayCreate { - fd, err := parent.createAndOpenChildLocked(ctx, rp, &opts) + if parent.isSynthetic() { + parent.dirMu.Unlock() + return nil, syserror.EPERM + } + fd, err := parent.createAndOpenChildLocked(ctx, rp, &opts, &ds) parent.dirMu.Unlock() return fd, err } @@ -639,24 +790,25 @@ afterTrailingSymlink: // Preconditions: fs.renameMu must be locked. func (d *dentry) openLocked(ctx context.Context, rp *vfs.ResolvingPath, opts *vfs.OpenOptions) (*vfs.FileDescription, error) { ats := vfs.AccessTypesForOpenFlags(opts) - if err := d.checkPermissions(rp.Credentials(), ats, d.isDir()); err != nil { + if err := d.checkPermissions(rp.Credentials(), ats); err != nil { return nil, err } mnt := rp.Mount() - filetype := d.fileType() - switch { - case filetype == linux.S_IFREG && !d.fs.opts.regularFilesUseSpecialFileFD: - if err := d.ensureSharedHandle(ctx, ats&vfs.MayRead != 0, ats&vfs.MayWrite != 0, opts.Flags&linux.O_TRUNC != 0); err != nil { - return nil, err - } - fd := ®ularFileFD{} - if err := fd.vfsfd.Init(fd, opts.Flags, mnt, &d.vfsd, &vfs.FileDescriptionOptions{ - AllowDirectIO: true, - }); err != nil { - return nil, err + switch d.fileType() { + case linux.S_IFREG: + if !d.fs.opts.regularFilesUseSpecialFileFD { + if err := d.ensureSharedHandle(ctx, ats&vfs.MayRead != 0, ats&vfs.MayWrite != 0, opts.Flags&linux.O_TRUNC != 0); err != nil { + return nil, err + } + fd := ®ularFileFD{} + if err := fd.vfsfd.Init(fd, opts.Flags, mnt, &d.vfsd, &vfs.FileDescriptionOptions{ + AllowDirectIO: true, + }); err != nil { + return nil, err + } + return &fd.vfsfd, nil } - return &fd.vfsfd, nil - case filetype == linux.S_IFDIR: + case linux.S_IFDIR: // Can't open directories with O_CREAT. if opts.Flags&linux.O_CREAT != 0 { return nil, syserror.EISDIR @@ -668,39 +820,100 @@ func (d *dentry) openLocked(ctx context.Context, rp *vfs.ResolvingPath, opts *vf if opts.Flags&linux.O_DIRECT != 0 { return nil, syserror.EINVAL } - if err := d.ensureSharedHandle(ctx, ats&vfs.MayRead != 0, false /* write */, false /* trunc */); err != nil { - return nil, err + if !d.isSynthetic() { + if err := d.ensureSharedHandle(ctx, ats&vfs.MayRead != 0, false /* write */, false /* trunc */); err != nil { + return nil, err + } } fd := &directoryFD{} if err := fd.vfsfd.Init(fd, opts.Flags, mnt, &d.vfsd, &vfs.FileDescriptionOptions{}); err != nil { return nil, err } return &fd.vfsfd, nil - case filetype == linux.S_IFLNK: + case linux.S_IFLNK: // Can't open symlinks without O_PATH (which is unimplemented). return nil, syserror.ELOOP - default: - if opts.Flags&linux.O_DIRECT != 0 { - return nil, syserror.EINVAL + case linux.S_IFSOCK: + if d.isSynthetic() { + return nil, syserror.ENXIO } - h, err := openHandle(ctx, d.file, ats&vfs.MayRead != 0, ats&vfs.MayWrite != 0, opts.Flags&linux.O_TRUNC != 0) - if err != nil { - return nil, err + if d.fs.iopts.OpenSocketsByConnecting { + return d.connectSocketLocked(ctx, opts) } - fd := &specialFileFD{ - handle: h, + case linux.S_IFIFO: + if d.isSynthetic() { + return d.pipe.Open(ctx, mnt, &d.vfsd, opts.Flags) } - if err := fd.vfsfd.Init(fd, opts.Flags, mnt, &d.vfsd, &vfs.FileDescriptionOptions{}); err != nil { + } + return d.openSpecialFileLocked(ctx, mnt, opts) +} + +func (d *dentry) connectSocketLocked(ctx context.Context, opts *vfs.OpenOptions) (*vfs.FileDescription, error) { + if opts.Flags&linux.O_DIRECT != 0 { + return nil, syserror.EINVAL + } + fdObj, err := d.file.connect(ctx, p9.AnonymousSocket) + if err != nil { + return nil, err + } + fd, err := host.NewFD(ctx, kernel.KernelFromContext(ctx).HostMount(), fdObj.FD(), &host.NewFDOptions{ + HaveFlags: true, + Flags: opts.Flags, + }) + if err != nil { + fdObj.Close() + return nil, err + } + fdObj.Release() + return fd, nil +} + +func (d *dentry) openSpecialFileLocked(ctx context.Context, mnt *vfs.Mount, opts *vfs.OpenOptions) (*vfs.FileDescription, error) { + ats := vfs.AccessTypesForOpenFlags(opts) + if opts.Flags&linux.O_DIRECT != 0 { + return nil, syserror.EINVAL + } + // We assume that the server silently inserts O_NONBLOCK in the open flags + // for all named pipes (because all existing gofers do this). + // + // NOTE(b/133875563): This makes named pipe opens racy, because the + // mechanisms for translating nonblocking to blocking opens can only detect + // the instantaneous presence of a peer holding the other end of the pipe + // open, not whether the pipe was *previously* opened by a peer that has + // since closed its end. + isBlockingOpenOfNamedPipe := d.fileType() == linux.S_IFIFO && opts.Flags&linux.O_NONBLOCK == 0 +retry: + h, err := openHandle(ctx, d.file, ats.MayRead(), ats.MayWrite(), opts.Flags&linux.O_TRUNC != 0) + if err != nil { + if isBlockingOpenOfNamedPipe && ats == vfs.MayWrite && err == syserror.ENXIO { + // An attempt to open a named pipe with O_WRONLY|O_NONBLOCK fails + // with ENXIO if opening the same named pipe with O_WRONLY would + // block because there are no readers of the pipe. + if err := sleepBetweenNamedPipeOpenChecks(ctx); err != nil { + return nil, err + } + goto retry + } + return nil, err + } + if isBlockingOpenOfNamedPipe && ats == vfs.MayRead && h.fd >= 0 { + if err := blockUntilNonblockingPipeHasWriter(ctx, h.fd); err != nil { h.close(ctx) return nil, err } - return &fd.vfsfd, nil } + fd, err := newSpecialFileFD(h, mnt, d, opts.Flags) + if err != nil { + h.close(ctx) + return nil, err + } + return &fd.vfsfd, nil } // Preconditions: d.fs.renameMu must be locked. d.dirMu must be locked. -func (d *dentry) createAndOpenChildLocked(ctx context.Context, rp *vfs.ResolvingPath, opts *vfs.OpenOptions) (*vfs.FileDescription, error) { - if err := d.checkPermissions(rp.Credentials(), vfs.MayWrite, true); err != nil { +// !d.isSynthetic(). +func (d *dentry) createAndOpenChildLocked(ctx context.Context, rp *vfs.ResolvingPath, opts *vfs.OpenOptions, ds **[]*dentry) (*vfs.FileDescription, error) { + if err := d.checkPermissions(rp.Credentials(), vfs.MayWrite); err != nil { return nil, err } if d.isDeleted() { @@ -721,7 +934,11 @@ func (d *dentry) createAndOpenChildLocked(ctx context.Context, rp *vfs.Resolving } creds := rp.Credentials() name := rp.Component() - fdobj, openFile, createQID, _, err := dirfile.create(ctx, name, (p9.OpenFlags)(opts.Flags), (p9.FileMode)(opts.Mode), (p9.UID)(creds.EffectiveKUID), (p9.GID)(creds.EffectiveKGID)) + // Filter file creation flags and O_LARGEFILE out; the create RPC already + // has the semantics of O_CREAT|O_EXCL, while some servers will choke on + // O_LARGEFILE. + createFlags := p9.OpenFlags(opts.Flags &^ (linux.O_CREAT | linux.O_EXCL | linux.O_NOCTTY | linux.O_TRUNC | linux.O_LARGEFILE)) + fdobj, openFile, createQID, _, err := dirfile.create(ctx, name, createFlags, (p9.FileMode)(opts.Mode), (p9.UID)(creds.EffectiveKUID), (p9.GID)(creds.EffectiveKGID)) if err != nil { dirfile.close(ctx) return nil, err @@ -729,7 +946,7 @@ func (d *dentry) createAndOpenChildLocked(ctx context.Context, rp *vfs.Resolving // Then we need to walk to the file we just created to get a non-open fid // representing it, and to get its metadata. This must use d.file since, as // explained above, dirfile was invalidated by dirfile.Create(). - walkQID, nonOpenFile, attrMask, attr, err := d.file.walkGetAttrOne(ctx, name) + _, nonOpenFile, attrMask, attr, err := d.file.walkGetAttrOne(ctx, name) if err != nil { openFile.close(ctx) if fdobj != nil { @@ -737,17 +954,6 @@ func (d *dentry) createAndOpenChildLocked(ctx context.Context, rp *vfs.Resolving } return nil, err } - // Sanity-check that we walked to the file we created. - if createQID.Path != walkQID.Path { - // Probably due to concurrent remote filesystem mutation? - ctx.Warningf("gofer.dentry.createAndOpenChildLocked: created file has QID %v before walk, QID %v after (interop=%v)", createQID, walkQID, d.fs.opts.interop) - nonOpenFile.close(ctx) - openFile.close(ctx) - if fdobj != nil { - fdobj.Close() - } - return nil, syserror.EAGAIN - } // Construct the new dentry. child, err := d.fs.newDentry(ctx, nonOpenFile, createQID, attrMask, &attr) @@ -759,6 +965,7 @@ func (d *dentry) createAndOpenChildLocked(ctx context.Context, rp *vfs.Resolving } return nil, err } + *ds = appendDentry(*ds, child) // Incorporate the fid that was opened by lcreate. useRegularFileFD := child.fileType() == linux.S_IFREG && !d.fs.opts.regularFilesUseSpecialFileFD if useRegularFileFD { @@ -771,22 +978,15 @@ func (d *dentry) createAndOpenChildLocked(ctx context.Context, rp *vfs.Resolving child.handleWritable = vfs.MayWriteFileWithOpenFlags(opts.Flags) child.handleMu.Unlock() } - // Take a reference on the new dentry to be held by the new file - // description. (This reference also means that the new dentry is not - // eligible for caching yet, so we don't need to append to a dentry slice.) - child.refs = 1 // Insert the dentry into the tree. - d.IncRef() // reference held by child on its parent d - d.vfsd.InsertChild(&child.vfsd, name) - if d.fs.opts.interop != InteropModeShared { - d.touchCMtime(ctx) - delete(d.negativeChildren, name) + d.cacheNewChildLocked(child, name) + if d.cachedMetadataAuthoritative() { + d.touchCMtime() d.dirents = nil } // Finally, construct a file description representing the created file. var childVFSFD *vfs.FileDescription - mnt.IncRef() if useRegularFileFD { fd := ®ularFileFD{} if err := fd.vfsfd.Init(fd, opts.Flags, mnt, &child.vfsd, &vfs.FileDescriptionOptions{ @@ -796,17 +996,16 @@ func (d *dentry) createAndOpenChildLocked(ctx context.Context, rp *vfs.Resolving } childVFSFD = &fd.vfsfd } else { - fd := &specialFileFD{ - handle: handle{ - file: openFile, - fd: -1, - }, + h := handle{ + file: openFile, + fd: -1, } if fdobj != nil { - fd.handle.fd = int32(fdobj.Release()) + h.fd = int32(fdobj.Release()) } - if err := fd.vfsfd.Init(fd, opts.Flags, mnt, &child.vfsd, &vfs.FileDescriptionOptions{}); err != nil { - fd.handle.close(ctx) + fd, err := newSpecialFileFD(h, mnt, child, opts.Flags) + if err != nil { + h.close(ctx) return nil, err } childVFSFD = &fd.vfsfd @@ -857,12 +1056,12 @@ func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa defer mnt.EndWrite() oldParent := oldParentVD.Dentry().Impl().(*dentry) - if fs.opts.interop == InteropModeShared { + if !oldParent.cachedMetadataAuthoritative() { if err := oldParent.updateFromGetattr(ctx); err != nil { return err } } - if err := oldParent.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true); err != nil { + if err := oldParent.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec); err != nil { return err } vfsObj := rp.VirtualFilesystem() @@ -870,7 +1069,7 @@ func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa // directory, we need to check for write permission on it. oldParent.dirMu.Lock() defer oldParent.dirMu.Unlock() - renamed, err := fs.revalidateChildLocked(ctx, vfsObj, oldParent, oldName, oldParent.vfsd.Child(oldName), &ds) + renamed, err := fs.getChildLocked(ctx, vfsObj, oldParent, oldName, &ds) if err != nil { return err } @@ -878,11 +1077,11 @@ func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa return syserror.ENOENT } if renamed.isDir() { - if renamed == newParent || renamed.vfsd.IsAncestorOf(&newParent.vfsd) { + if renamed == newParent || genericIsAncestorDentry(renamed, newParent) { return syserror.EINVAL } if oldParent != newParent { - if err := renamed.checkPermissions(rp.Credentials(), vfs.MayWrite, true); err != nil { + if err := renamed.checkPermissions(rp.Credentials(), vfs.MayWrite); err != nil { return err } } @@ -893,7 +1092,7 @@ func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa } if oldParent != newParent { - if err := newParent.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true); err != nil { + if err := newParent.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec); err != nil { return err } newParent.dirMu.Lock() @@ -902,52 +1101,99 @@ func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa if newParent.isDeleted() { return syserror.ENOENT } - replacedVFSD := newParent.vfsd.Child(newName) - var replaced *dentry - // This is similar to unlinkAt, except: - // - // - We revalidate the replaced dentry unconditionally for simplicity. - // - // - If rp.MustBeDir(), then we need a dentry representing the replaced - // file regardless to confirm that it's a directory. - if replacedVFSD != nil || rp.MustBeDir() { - replaced, err = fs.revalidateChildLocked(ctx, vfsObj, newParent, newName, replacedVFSD, &ds) - if err != nil { - return err - } - if replaced != nil { - if replaced.isDir() { - if !renamed.isDir() { - return syserror.EISDIR - } - } else { - if rp.MustBeDir() || renamed.isDir() { - return syserror.ENOTDIR - } + replaced, err := fs.getChildLocked(ctx, rp.VirtualFilesystem(), newParent, newName, &ds) + if err != nil { + return err + } + var replacedVFSD *vfs.Dentry + if replaced != nil { + replacedVFSD = &replaced.vfsd + if replaced.isDir() { + if !renamed.isDir() { + return syserror.EISDIR } - replacedVFSD = &replaced.vfsd } else { - replacedVFSD = nil + if rp.MustBeDir() || renamed.isDir() { + return syserror.ENOTDIR + } } } if oldParent == newParent && oldName == newName { return nil } - if err := vfsObj.PrepareRenameDentry(vfs.MountNamespaceFromContext(ctx), &renamed.vfsd, replacedVFSD); err != nil { + mntns := vfs.MountNamespaceFromContext(ctx) + defer mntns.DecRef() + if err := vfsObj.PrepareRenameDentry(mntns, &renamed.vfsd, replacedVFSD); err != nil { return err } - if err := renamed.file.rename(ctx, newParent.file, newName); err != nil { - vfsObj.AbortRenameDentry(&renamed.vfsd, replacedVFSD) - return err + + // Update the remote filesystem. + if !renamed.isSynthetic() { + if err := renamed.file.rename(ctx, newParent.file, newName); err != nil { + vfsObj.AbortRenameDentry(&renamed.vfsd, replacedVFSD) + return err + } + } else if replaced != nil && !replaced.isSynthetic() { + // We are replacing an existing real file with a synthetic one, so we + // need to unlink the former. + flags := uint32(0) + if replaced.isDir() { + flags = linux.AT_REMOVEDIR + } + if err := newParent.file.unlinkAt(ctx, newName, flags); err != nil { + vfsObj.AbortRenameDentry(&renamed.vfsd, replacedVFSD) + return err + } + } + + // Update the dentry tree. + vfsObj.CommitRenameReplaceDentry(&renamed.vfsd, replacedVFSD) + if replaced != nil { + replaced.setDeleted() + if replaced.isSynthetic() { + newParent.syntheticChildren-- + replaced.decRefLocked() + } + ds = appendDentry(ds, replaced) } - if fs.opts.interop != InteropModeShared { - oldParent.cacheNegativeChildLocked(oldName) + oldParent.cacheNegativeLookupLocked(oldName) + // We don't use newParent.cacheNewChildLocked() since we don't want to mess + // with reference counts and queue oldParent for checkCachingLocked if the + // parent isn't actually changing. + if oldParent != newParent { + ds = appendDentry(ds, oldParent) + newParent.IncRef() + if renamed.isSynthetic() { + oldParent.syntheticChildren-- + newParent.syntheticChildren++ + } + } + renamed.parent = newParent + renamed.name = newName + if newParent.children == nil { + newParent.children = make(map[string]*dentry) + } + newParent.children[newName] = renamed + + // Update metadata. + if renamed.cachedMetadataAuthoritative() { + renamed.touchCtime() + } + if oldParent.cachedMetadataAuthoritative() { oldParent.dirents = nil - delete(newParent.negativeChildren, newName) + oldParent.touchCMtime() + if renamed.isDir() { + oldParent.decLinks() + } + } + if newParent.cachedMetadataAuthoritative() { newParent.dirents = nil + newParent.touchCMtime() + if renamed.isDir() { + newParent.incLinks() + } } - vfsObj.CommitRenameReplaceDentry(&renamed.vfsd, &newParent.vfsd, newName, replacedVFSD) return nil } @@ -994,6 +1240,10 @@ func (fs *filesystem) StatFSAt(ctx context.Context, rp *vfs.ResolvingPath) (linu if err != nil { return linux.Statfs{}, err } + // If d is synthetic, invoke statfs on the first ancestor of d that isn't. + for d.isSynthetic() { + d = d.parent + } fsstat, err := d.file.statFS(ctx) if err != nil { return linux.Statfs{}, err @@ -1023,7 +1273,7 @@ func (fs *filesystem) SymlinkAt(ctx context.Context, rp *vfs.ResolvingPath, targ creds := rp.Credentials() _, err := parent.file.symlink(ctx, target, name, (p9.UID)(creds.EffectiveKUID), (p9.GID)(creds.EffectiveKGID)) return err - }) + }, nil) } // UnlinkAt implements vfs.FilesystemImpl.UnlinkAt. @@ -1031,8 +1281,34 @@ func (fs *filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error return fs.unlinkAt(ctx, rp, false /* dir */) } +// BoundEndpointAt implements FilesystemImpl.BoundEndpointAt. +func (fs *filesystem) BoundEndpointAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.BoundEndpointOptions) (transport.BoundEndpoint, error) { + var ds *[]*dentry + fs.renameMu.RLock() + defer fs.renameMuRUnlockAndCheckCaching(&ds) + d, err := fs.resolveLocked(ctx, rp, &ds) + if err != nil { + return nil, err + } + if err := d.checkPermissions(rp.Credentials(), vfs.MayWrite); err != nil { + return nil, err + } + if d.isSocket() { + if !d.isSynthetic() { + d.IncRef() + return &endpoint{ + dentry: d, + file: d.file.file, + path: opts.Addr, + }, nil + } + return d.endpoint, nil + } + return nil, syserror.ECONNREFUSED +} + // ListxattrAt implements vfs.FilesystemImpl.ListxattrAt. -func (fs *filesystem) ListxattrAt(ctx context.Context, rp *vfs.ResolvingPath) ([]string, error) { +func (fs *filesystem) ListxattrAt(ctx context.Context, rp *vfs.ResolvingPath, size uint64) ([]string, error) { var ds *[]*dentry fs.renameMu.RLock() defer fs.renameMuRUnlockAndCheckCaching(&ds) @@ -1040,11 +1316,11 @@ func (fs *filesystem) ListxattrAt(ctx context.Context, rp *vfs.ResolvingPath) ([ if err != nil { return nil, err } - return d.listxattr(ctx) + return d.listxattr(ctx, rp.Credentials(), size) } // GetxattrAt implements vfs.FilesystemImpl.GetxattrAt. -func (fs *filesystem) GetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) (string, error) { +func (fs *filesystem) GetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.GetxattrOptions) (string, error) { var ds *[]*dentry fs.renameMu.RLock() defer fs.renameMuRUnlockAndCheckCaching(&ds) @@ -1052,7 +1328,7 @@ func (fs *filesystem) GetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, nam if err != nil { return "", err } - return d.getxattr(ctx, name) + return d.getxattr(ctx, rp.Credentials(), &opts) } // SetxattrAt implements vfs.FilesystemImpl.SetxattrAt. @@ -1064,7 +1340,7 @@ func (fs *filesystem) SetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opt if err != nil { return err } - return d.setxattr(ctx, &opts) + return d.setxattr(ctx, rp.Credentials(), &opts) } // RemovexattrAt implements vfs.FilesystemImpl.RemovexattrAt. @@ -1076,12 +1352,12 @@ func (fs *filesystem) RemovexattrAt(ctx context.Context, rp *vfs.ResolvingPath, if err != nil { return err } - return d.removexattr(ctx, name) + return d.removexattr(ctx, rp.Credentials(), name) } // PrependPath implements vfs.FilesystemImpl.PrependPath. func (fs *filesystem) PrependPath(ctx context.Context, vfsroot, vd vfs.VirtualDentry, b *fspath.Builder) error { fs.renameMu.RLock() defer fs.renameMu.RUnlock() - return vfs.GenericPrependPath(vfsroot, vd, b) + return genericPrependPath(vfsroot, vd.Mount(), vd.Dentry().Impl().(*dentry), b) } diff --git a/pkg/sentry/fsimpl/gofer/gofer.go b/pkg/sentry/fsimpl/gofer/gofer.go index d0552bd99..3f3bd56f0 100644 --- a/pkg/sentry/fsimpl/gofer/gofer.go +++ b/pkg/sentry/fsimpl/gofer/gofer.go @@ -27,13 +27,15 @@ // dentry.handleMu // dentry.dataMu // -// Locking dentry.dirMu in multiple dentries requires holding -// filesystem.renameMu for writing. +// Locking dentry.dirMu in multiple dentries requires that either ancestor +// dentries are locked before descendant dentries, or that filesystem.renameMu +// is locked for writing. package gofer import ( "fmt" "strconv" + "strings" "sync" "sync/atomic" "syscall" @@ -44,14 +46,20 @@ import ( "gvisor.dev/gvisor/pkg/p9" "gvisor.dev/gvisor/pkg/sentry/fs/fsutil" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/kernel/pipe" + ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" "gvisor.dev/gvisor/pkg/sentry/memmap" "gvisor.dev/gvisor/pkg/sentry/pgalloc" + "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/unet" "gvisor.dev/gvisor/pkg/usermem" ) +// Name is the default filesystem name. +const Name = "9p" + // FilesystemType implements vfs.FilesystemType. type FilesystemType struct{} @@ -64,16 +72,17 @@ type filesystem struct { mfp pgalloc.MemoryFileProvider // Immutable options. - opts filesystemOptions + opts filesystemOptions + iopts InternalFilesystemOptions // client is the client used by this filesystem. client is immutable. client *p9.Client - // uid and gid are the effective KUID and KGID of the filesystem's creator, - // and are used as the owner and group for files that don't specify one. - // uid and gid are immutable. - uid auth.KUID - gid auth.KGID + // clock is a realtime clock used to set timestamps in file operations. + clock ktime.Clock + + // devMinor is the filesystem's minor device number. devMinor is immutable. + devMinor uint32 // renameMu serves two purposes: // @@ -94,11 +103,12 @@ type filesystem struct { cachedDentries dentryList cachedDentriesLen uint64 - // dentries contains all dentries in this filesystem. specialFileFDs - // contains all open specialFileFDs. These fields are protected by syncMu. - syncMu sync.Mutex - dentries map[*dentry]struct{} - specialFileFDs map[*specialFileFD]struct{} + // syncableDentries contains all dentries in this filesystem for which + // !dentry.file.isNil(). specialFileFDs contains all open specialFileFDs. + // These fields are protected by syncMu. + syncMu sync.Mutex + syncableDentries map[*dentry]struct{} + specialFileFDs map[*specialFileFD]struct{} } type filesystemOptions struct { @@ -106,6 +116,8 @@ type filesystemOptions struct { fd int aname string interop InteropMode // derived from the "cache" mount option + dfltuid auth.KUID + dfltgid auth.KGID msize uint32 version string @@ -127,9 +139,12 @@ type filesystemOptions struct { // If overlayfsStaleRead is true, O_RDONLY host FDs provided by the remote // filesystem may not be coherent with writable host FDs opened later, so - // mappings of the former must be replaced by mappings of the latter. This - // is usually only the case when the remote filesystem is an overlayfs - // mount on Linux < 4.19. + // all uses of the former must be replaced by uses of the latter. This is + // usually only the case when the remote filesystem is a Linux overlayfs + // mount. (Prior to Linux 4.18, patch series centered on commit + // d1d04ef8572b "ovl: stack file ops", both I/O and memory mappings were + // incoherent between pre-copy-up and post-copy-up FDs; after that patch + // series, only memory mappings are incoherent.) overlayfsStaleRead bool // If regularFilesUseSpecialFileFD is true, application FDs representing @@ -179,7 +194,8 @@ const ( // InteropModeShared is appropriate when there are users of the remote // filesystem that may mutate its state other than the client. // - // - The client must verify cached filesystem state before using it. + // - The client must verify ("revalidate") cached filesystem state before + // using it. // // - Client changes to filesystem state must be sent to the remote // filesystem synchronously. @@ -196,6 +212,34 @@ const ( InteropModeShared ) +// InternalFilesystemOptions may be passed as +// vfs.GetFilesystemOptions.InternalData to FilesystemType.GetFilesystem. +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 + // which servers can handle only a single client and report failure if that + // client disconnects. + LeakConnection bool + + // If OpenSocketsByConnecting is true, silently translate attempts to open + // files identifying as sockets to connect RPCs. + OpenSocketsByConnecting bool +} + +// _V9FS_DEFUID and _V9FS_DEFGID (from Linux's fs/9p/v9fs.h) are the default +// UIDs and GIDs used for files that do not provide a specific owner or group +// respectively. +const ( + // uint32(-2) doesn't work in Go. + _V9FS_DEFUID = auth.KUID(4294967294) + _V9FS_DEFGID = auth.KGID(4294967294) +) + +// Name implements vfs.FilesystemType.Name. +func (FilesystemType) Name() string { + return Name +} + // GetFilesystem implements vfs.FilesystemType.GetFilesystem. func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { mfp := pgalloc.MemoryFileProviderFromContext(ctx) @@ -276,6 +320,31 @@ func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virt } } + // Parse the default UID and GID. + fsopts.dfltuid = _V9FS_DEFUID + if dfltuidstr, ok := mopts["dfltuid"]; ok { + delete(mopts, "dfltuid") + dfltuid, err := strconv.ParseUint(dfltuidstr, 10, 32) + if err != nil { + ctx.Warningf("gofer.FilesystemType.GetFilesystem: invalid default UID: dfltuid=%s", dfltuidstr) + return nil, nil, syserror.EINVAL + } + // In Linux, dfltuid is interpreted as a UID and is converted to a KUID + // in the caller's user namespace, but goferfs isn't + // application-mountable. + fsopts.dfltuid = auth.KUID(dfltuid) + } + fsopts.dfltgid = _V9FS_DEFGID + if dfltgidstr, ok := mopts["dfltgid"]; ok { + delete(mopts, "dfltgid") + dfltgid, err := strconv.ParseUint(dfltgidstr, 10, 32) + if err != nil { + ctx.Warningf("gofer.FilesystemType.GetFilesystem: invalid default UID: dfltgid=%s", dfltgidstr) + return nil, nil, syserror.EINVAL + } + fsopts.dfltgid = auth.KGID(dfltgid) + } + // Parse the 9P message size. fsopts.msize = 1024 * 1024 // 1M, tested to give good enough performance up to 64M if msizestr, ok := mopts["msize"]; ok { @@ -329,6 +398,14 @@ func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virt return nil, nil, syserror.EINVAL } + // Handle internal options. + iopts, ok := opts.InternalData.(InternalFilesystemOptions) + if opts.InternalData != nil && !ok { + ctx.Warningf("gofer.FilesystemType.GetFilesystem: GetFilesystemOptions.InternalData has type %T, wanted gofer.InternalFilesystemOptions", opts.InternalData) + return nil, nil, syserror.EINVAL + } + // If !ok, iopts being the zero value is correct. + // Establish a connection with the server. conn, err := unet.NewSocket(fsopts.fd) if err != nil { @@ -362,16 +439,23 @@ func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virt } // Construct the filesystem object. + devMinor, err := vfsObj.GetAnonBlockDevMinor() + if err != nil { + attachFile.close(ctx) + client.Close() + return nil, nil, err + } fs := &filesystem{ - mfp: mfp, - opts: fsopts, - uid: creds.EffectiveKUID, - gid: creds.EffectiveKGID, - client: client, - dentries: make(map[*dentry]struct{}), - specialFileFDs: make(map[*specialFileFD]struct{}), + mfp: mfp, + opts: fsopts, + iopts: iopts, + client: client, + clock: ktime.RealtimeClockFromContext(ctx), + devMinor: devMinor, + syncableDentries: make(map[*dentry]struct{}), + specialFileFDs: make(map[*specialFileFD]struct{}), } - fs.vfsfs.Init(vfsObj, fs) + fs.vfsfs.Init(vfsObj, &fstype, fs) // Construct the root dentry. root, err := fs.newDentry(ctx, attachFile, qid, attrMask, &attr) @@ -395,7 +479,7 @@ func (fs *filesystem) Release() { mf := fs.mfp.MemoryFile() fs.syncMu.Lock() - for d := range fs.dentries { + for d := range fs.syncableDentries { d.handleMu.Lock() d.dataMu.Lock() if d.handleWritable { @@ -421,8 +505,12 @@ func (fs *filesystem) Release() { // fs. fs.syncMu.Unlock() - // Close the connection to the server. This implicitly clunks all fids. - fs.client.Close() + if !fs.iopts.LeakConnection { + // Close the connection to the server. This implicitly clunks all fids. + fs.client.Close() + } + + fs.vfsfs.VirtualFilesystem().PutAnonBlockDevMinor(fs.devMinor) } // dentry implements vfs.DentryImpl. @@ -430,16 +518,33 @@ type dentry struct { vfsd vfs.Dentry // refs is the reference count. Each dentry holds a reference on its - // parent, even if disowned. refs is accessed using atomic memory - // operations. + // parent, even if disowned. An additional reference is held on all + // synthetic dentries until they are unlinked or invalidated. When refs + // reaches 0, the dentry may be added to the cache or destroyed. If refs == + // -1, the dentry has already been destroyed. refs is accessed using atomic + // memory operations. refs int64 // fs is the owning filesystem. fs is immutable. fs *filesystem + // parent is this dentry's parent directory. Each dentry holds a reference + // on its parent. If this dentry is a filesystem root, parent is nil. + // parent is protected by filesystem.renameMu. + parent *dentry + + // name is the name of this dentry in its parent. If this dentry is a + // filesystem root, name is the empty string. name is protected by + // filesystem.renameMu. + name string + // We don't support hard links, so each dentry maps 1:1 to an inode. // file is the unopened p9.File that backs this dentry. file is immutable. + // + // If file.isNil(), this dentry represents a synthetic file, i.e. a file + // that does not exist on the remote filesystem. As of this writing, the + // only files that can be synthetic are sockets, pipes, and directories. file p9file // If deleted is non-zero, the file represented by this dentry has been @@ -454,15 +559,26 @@ type dentry struct { dirMu sync.Mutex - // If this dentry represents a directory, and InteropModeShared is not in - // effect, negativeChildren is a set of child names in this directory that - // are known not to exist. negativeChildren is protected by dirMu. - negativeChildren map[string]struct{} - - // If this dentry represents a directory, InteropModeShared is not in - // effect, and dirents is not nil, it is a cache of all entries in the - // directory, in the order they were returned by the server. dirents is - // protected by dirMu. + // If this dentry represents a directory, children contains: + // + // - Mappings of child filenames to dentries representing those children. + // + // - Mappings of child filenames that are known not to exist to nil + // dentries (only if InteropModeShared is not in effect and the directory + // is not synthetic). + // + // children is protected by dirMu. + children map[string]*dentry + + // If this dentry represents a directory, syntheticChildren is the number + // of child dentries for which dentry.isSynthetic() == true. + // syntheticChildren is protected by dirMu. + syntheticChildren int + + // If this dentry represents a directory, + // dentry.cachedMetadataAuthoritative() == true, and dirents is not nil, it + // is a cache of all entries in the directory, in the order they were + // returned by the server. dirents is protected by dirMu. dirents []vfs.Dirent // Cached metadata; protected by metadataMu and accessed using atomic @@ -482,6 +598,11 @@ type dentry struct { // locked to mutate it). size uint64 + // nlink counts the number of hard links to this dentry. It's updated and + // accessed using atomic operations. It's not protected by metadataMu like the + // other metadata fields. + nlink uint32 + mapsMu sync.Mutex // If this dentry represents a regular file, mappings tracks mappings of @@ -533,6 +654,14 @@ type dentry struct { // and target are protected by dataMu. haveTarget bool target string + + // If this dentry represents a synthetic socket file, endpoint is the + // transport endpoint bound to this file. + endpoint transport.BoundEndpoint + + // If this dentry represents a synthetic named pipe, pipe is the pipe + // endpoint bound to this file. + pipe *pipe.VFSPipe } // dentryAttrMask returns a p9.AttrMask enabling all attributes used by the @@ -554,6 +683,8 @@ func dentryAttrMask() p9.AttrMask { // initially has no references, but is not cached; it is the caller's // responsibility to set the dentry's reference count and/or call // dentry.checkCachingLocked() as appropriate. +// +// Preconditions: !file.isNil(). func (fs *filesystem) newDentry(ctx context.Context, file p9file, qid p9.QID, mask p9.AttrMask, attr *p9.Attr) (*dentry, error) { if !mask.Mode { ctx.Warningf("can't create gofer.dentry without file type") @@ -569,8 +700,8 @@ func (fs *filesystem) newDentry(ctx context.Context, file p9file, qid p9.QID, ma file: file, ino: qid.Path, mode: uint32(attr.Mode), - uid: uint32(fs.uid), - gid: uint32(fs.gid), + uid: uint32(fs.opts.dfltuid), + gid: uint32(fs.opts.dfltgid), blockSize: usermem.PageSize, handle: handle{ fd: -1, @@ -578,10 +709,10 @@ func (fs *filesystem) newDentry(ctx context.Context, file p9file, qid p9.QID, ma } d.pf.dentry = d if mask.UID { - d.uid = uint32(attr.UID) + d.uid = dentryUIDFromP9UID(attr.UID) } if mask.GID { - d.gid = uint32(attr.GID) + d.gid = dentryGIDFromP9GID(attr.GID) } if mask.Size { d.size = attr.Size @@ -601,14 +732,25 @@ func (fs *filesystem) newDentry(ctx context.Context, file p9file, qid p9.QID, ma if mask.BTime { d.btime = dentryTimestampFromP9(attr.BTimeSeconds, attr.BTimeNanoSeconds) } + if mask.NLink { + d.nlink = uint32(attr.NLink) + } d.vfsd.Init(d) fs.syncMu.Lock() - fs.dentries[d] = struct{}{} + fs.syncableDentries[d] = struct{}{} fs.syncMu.Unlock() return d, nil } +func (d *dentry) isSynthetic() bool { + return d.file.isNil() +} + +func (d *dentry) cachedMetadataAuthoritative() bool { + return d.fs.opts.interop != InteropModeShared || d.isSynthetic() +} + // updateFromP9Attrs is called to update d's metadata after an update from the // remote filesystem. func (d *dentry) updateFromP9Attrs(mask p9.AttrMask, attr *p9.Attr) { @@ -621,10 +763,10 @@ func (d *dentry) updateFromP9Attrs(mask p9.AttrMask, attr *p9.Attr) { atomic.StoreUint32(&d.mode, uint32(attr.Mode)) } if mask.UID { - atomic.StoreUint32(&d.uid, uint32(attr.UID)) + atomic.StoreUint32(&d.uid, dentryUIDFromP9UID(attr.UID)) } if mask.GID { - atomic.StoreUint32(&d.gid, uint32(attr.GID)) + atomic.StoreUint32(&d.gid, dentryGIDFromP9GID(attr.GID)) } // There is no P9_GETATTR_* bit for I/O block size. if attr.BlockSize != 0 { @@ -642,6 +784,9 @@ func (d *dentry) updateFromP9Attrs(mask p9.AttrMask, attr *p9.Attr) { if mask.BTime { atomic.StoreInt64(&d.btime, dentryTimestampFromP9(attr.BTimeSeconds, attr.BTimeNanoSeconds)) } + if mask.NLink { + atomic.StoreUint32(&d.nlink, uint32(attr.NLink)) + } if mask.Size { d.dataMu.Lock() atomic.StoreUint64(&d.size, attr.Size) @@ -650,6 +795,7 @@ func (d *dentry) updateFromP9Attrs(mask p9.AttrMask, attr *p9.Attr) { d.metadataMu.Unlock() } +// Preconditions: !d.isSynthetic() func (d *dentry) updateFromGetattr(ctx context.Context) error { // Use d.handle.file, which represents a 9P fid that has been opened, in // preference to d.file, which represents a 9P fid that has not. This may @@ -684,10 +830,7 @@ func (d *dentry) fileType() uint32 { func (d *dentry) statTo(stat *linux.Statx) { stat.Mask = linux.STATX_TYPE | linux.STATX_MODE | linux.STATX_NLINK | linux.STATX_UID | linux.STATX_GID | linux.STATX_ATIME | linux.STATX_MTIME | linux.STATX_CTIME | linux.STATX_INO | linux.STATX_SIZE | linux.STATX_BLOCKS | linux.STATX_BTIME stat.Blksize = atomic.LoadUint32(&d.blockSize) - stat.Nlink = 1 - if d.isDir() { - stat.Nlink = 2 - } + stat.Nlink = atomic.LoadUint32(&d.nlink) stat.UID = atomic.LoadUint32(&d.uid) stat.GID = atomic.LoadUint32(&d.gid) stat.Mode = uint16(atomic.LoadUint32(&d.mode)) @@ -700,7 +843,8 @@ func (d *dentry) statTo(stat *linux.Statx) { stat.Btime = statxTimestampFromDentry(atomic.LoadInt64(&d.btime)) stat.Ctime = statxTimestampFromDentry(atomic.LoadInt64(&d.ctime)) stat.Mtime = statxTimestampFromDentry(atomic.LoadInt64(&d.mtime)) - // TODO(jamieliu): device number + stat.DevMajor = linux.UNNAMED_MAJOR + stat.DevMinor = d.fs.devMinor } func (d *dentry) setStat(ctx context.Context, creds *auth.Credentials, stat *linux.Statx, mnt *vfs.Mount) error { @@ -710,7 +854,8 @@ func (d *dentry) setStat(ctx context.Context, creds *auth.Credentials, stat *lin if stat.Mask&^(linux.STATX_MODE|linux.STATX_UID|linux.STATX_GID|linux.STATX_ATIME|linux.STATX_MTIME|linux.STATX_SIZE) != 0 { return syserror.EPERM } - if err := vfs.CheckSetStat(creds, stat, uint16(atomic.LoadUint32(&d.mode))&^linux.S_IFMT, auth.KUID(atomic.LoadUint32(&d.uid)), auth.KGID(atomic.LoadUint32(&d.gid))); err != nil { + mode := linux.FileMode(atomic.LoadUint32(&d.mode)) + if err := vfs.CheckSetStat(ctx, creds, stat, mode, auth.KUID(atomic.LoadUint32(&d.uid)), auth.KGID(atomic.LoadUint32(&d.gid))); err != nil { return err } if err := mnt.CheckBeginWrite(); err != nil { @@ -719,53 +864,63 @@ func (d *dentry) setStat(ctx context.Context, creds *auth.Credentials, stat *lin defer mnt.EndWrite() setLocalAtime := false setLocalMtime := false - if d.fs.opts.interop != InteropModeShared { + if d.cachedMetadataAuthoritative() { // Timestamp updates will be handled locally. setLocalAtime = stat.Mask&linux.STATX_ATIME != 0 setLocalMtime = stat.Mask&linux.STATX_MTIME != 0 stat.Mask &^= linux.STATX_ATIME | linux.STATX_MTIME - if !setLocalMtime && (stat.Mask&linux.STATX_SIZE != 0) { - // Truncate updates mtime. - setLocalMtime = true - stat.Mtime.Nsec = linux.UTIME_NOW + + // Prepare for truncate. + if stat.Mask&linux.STATX_SIZE != 0 { + switch d.mode & linux.S_IFMT { + case linux.S_IFREG: + if !setLocalMtime { + // Truncate updates mtime. + setLocalMtime = true + stat.Mtime.Nsec = linux.UTIME_NOW + } + case linux.S_IFDIR: + return syserror.EISDIR + default: + return syserror.EINVAL + } } } d.metadataMu.Lock() defer d.metadataMu.Unlock() - if stat.Mask != 0 { - if err := d.file.setAttr(ctx, p9.SetAttrMask{ - Permissions: stat.Mask&linux.STATX_MODE != 0, - UID: stat.Mask&linux.STATX_UID != 0, - GID: stat.Mask&linux.STATX_GID != 0, - Size: stat.Mask&linux.STATX_SIZE != 0, - ATime: stat.Mask&linux.STATX_ATIME != 0, - MTime: stat.Mask&linux.STATX_MTIME != 0, - ATimeNotSystemTime: stat.Atime.Nsec != linux.UTIME_NOW, - MTimeNotSystemTime: stat.Mtime.Nsec != linux.UTIME_NOW, - }, p9.SetAttr{ - Permissions: p9.FileMode(stat.Mode), - UID: p9.UID(stat.UID), - GID: p9.GID(stat.GID), - Size: stat.Size, - ATimeSeconds: uint64(stat.Atime.Sec), - ATimeNanoSeconds: uint64(stat.Atime.Nsec), - MTimeSeconds: uint64(stat.Mtime.Sec), - MTimeNanoSeconds: uint64(stat.Mtime.Nsec), - }); err != nil { - return err + if !d.isSynthetic() { + if stat.Mask != 0 { + if err := d.file.setAttr(ctx, p9.SetAttrMask{ + Permissions: stat.Mask&linux.STATX_MODE != 0, + UID: stat.Mask&linux.STATX_UID != 0, + GID: stat.Mask&linux.STATX_GID != 0, + Size: stat.Mask&linux.STATX_SIZE != 0, + ATime: stat.Mask&linux.STATX_ATIME != 0, + MTime: stat.Mask&linux.STATX_MTIME != 0, + ATimeNotSystemTime: stat.Mask&linux.STATX_ATIME != 0 && stat.Atime.Nsec != linux.UTIME_NOW, + MTimeNotSystemTime: stat.Mask&linux.STATX_MTIME != 0 && stat.Mtime.Nsec != linux.UTIME_NOW, + }, p9.SetAttr{ + Permissions: p9.FileMode(stat.Mode), + UID: p9.UID(stat.UID), + GID: p9.GID(stat.GID), + Size: stat.Size, + ATimeSeconds: uint64(stat.Atime.Sec), + ATimeNanoSeconds: uint64(stat.Atime.Nsec), + MTimeSeconds: uint64(stat.Mtime.Sec), + MTimeNanoSeconds: uint64(stat.Mtime.Nsec), + }); err != nil { + return err + } + } + if d.fs.opts.interop == InteropModeShared { + // There's no point to updating d's metadata in this case since + // it'll be overwritten by revalidation before the next time it's + // used anyway. (InteropModeShared inhibits client caching of + // regular file data, so there's no cache to truncate either.) + return nil } } - if d.fs.opts.interop == InteropModeShared { - // There's no point to updating d's metadata in this case since it'll - // be overwritten by revalidation before the next time it's used - // anyway. (InteropModeShared inhibits client caching of regular file - // data, so there's no cache to truncate either.) - return nil - } - now, haveNow := nowFromContext(ctx) - if !haveNow { - ctx.Warningf("gofer.dentry.setStat: current time not available") - } + now := d.fs.clock.Now().Nanoseconds() if stat.Mask&linux.STATX_MODE != 0 { atomic.StoreUint32(&d.mode, d.fileType()|uint32(stat.Mode)) } @@ -777,25 +932,19 @@ func (d *dentry) setStat(ctx context.Context, creds *auth.Credentials, stat *lin } if setLocalAtime { if stat.Atime.Nsec == linux.UTIME_NOW { - if haveNow { - atomic.StoreInt64(&d.atime, now) - } + atomic.StoreInt64(&d.atime, now) } else { atomic.StoreInt64(&d.atime, dentryTimestampFromStatx(stat.Atime)) } } if setLocalMtime { if stat.Mtime.Nsec == linux.UTIME_NOW { - if haveNow { - atomic.StoreInt64(&d.mtime, now) - } + atomic.StoreInt64(&d.mtime, now) } else { atomic.StoreInt64(&d.mtime, dentryTimestampFromStatx(stat.Mtime)) } } - if haveNow { - atomic.StoreInt64(&d.ctime, now) - } + atomic.StoreInt64(&d.ctime, now) if stat.Mask&linux.STATX_SIZE != 0 { d.dataMu.Lock() oldSize := d.size @@ -807,8 +956,8 @@ func (d *dentry) setStat(ctx context.Context, creds *auth.Credentials, stat *lin // so we can't race with Write or another truncate.) d.dataMu.Unlock() if d.size < oldSize { - oldpgend := pageRoundUp(oldSize) - newpgend := pageRoundUp(d.size) + oldpgend, _ := usermem.PageRoundUp(oldSize) + newpgend, _ := usermem.PageRoundUp(d.size) if oldpgend != newpgend { d.mapsMu.Lock() d.mappings.Invalidate(memmap.MappableRange{newpgend, oldpgend}, memmap.InvalidateOpts{ @@ -832,8 +981,22 @@ func (d *dentry) setStat(ctx context.Context, creds *auth.Credentials, stat *lin return nil } -func (d *dentry) checkPermissions(creds *auth.Credentials, ats vfs.AccessTypes, isDir bool) error { - return vfs.GenericCheckPermissions(creds, ats, isDir, uint16(atomic.LoadUint32(&d.mode))&0777, auth.KUID(atomic.LoadUint32(&d.uid)), auth.KGID(atomic.LoadUint32(&d.gid))) +func (d *dentry) checkPermissions(creds *auth.Credentials, ats vfs.AccessTypes) error { + return vfs.GenericCheckPermissions(creds, ats, linux.FileMode(atomic.LoadUint32(&d.mode)), auth.KUID(atomic.LoadUint32(&d.uid)), auth.KGID(atomic.LoadUint32(&d.gid))) +} + +func dentryUIDFromP9UID(uid p9.UID) uint32 { + if !uid.Ok() { + return uint32(auth.OverflowUID) + } + return uint32(uid) +} + +func dentryGIDFromP9GID(gid p9.GID) uint32 { + if !gid.Ok() { + return uint32(auth.OverflowGID) + } + return uint32(gid) } // IncRef implements vfs.DentryImpl.IncRef. @@ -847,7 +1010,7 @@ func (d *dentry) IncRef() { func (d *dentry) TryIncRef() bool { for { refs := atomic.LoadInt64(&d.refs) - if refs == 0 { + if refs <= 0 { return false } if atomic.CompareAndSwapInt64(&d.refs, refs, refs+1) { @@ -867,16 +1030,44 @@ func (d *dentry) DecRef() { } } +// decRefLocked decrements d's reference count without calling +// d.checkCachingLocked, even if d's reference count reaches 0; callers are +// responsible for ensuring that d.checkCachingLocked will be called later. +func (d *dentry) decRefLocked() { + if refs := atomic.AddInt64(&d.refs, -1); refs < 0 { + panic("gofer.dentry.decRefLocked() called without holding a reference") + } +} + +// InotifyWithParent implements vfs.DentryImpl.InotifyWithParent. +// +// TODO(gvisor.dev/issue/1479): Implement inotify. +func (d *dentry) InotifyWithParent(events uint32, cookie uint32, et vfs.EventType) {} + +// Watches implements vfs.DentryImpl.Watches. +// +// TODO(gvisor.dev/issue/1479): Implement inotify. +func (d *dentry) Watches() *vfs.Watches { + return nil +} + // checkCachingLocked should be called after d's reference count becomes 0 or it // becomes disowned. // +// It may be called on a destroyed dentry. For example, +// renameMu[R]UnlockAndCheckCaching may call checkCachingLocked multiple times +// for the same dentry when the dentry is visited more than once in the same +// operation. One of the calls may destroy the dentry, so subsequent calls will +// do nothing. +// // Preconditions: d.fs.renameMu must be locked for writing. func (d *dentry) checkCachingLocked() { // Dentries with a non-zero reference count must be retained. (The only way // to obtain a reference on a dentry with zero references is via path // resolution, which requires renameMu, so if d.refs is zero then it will // remain zero while we hold renameMu for writing.) - if atomic.LoadInt64(&d.refs) != 0 { + refs := atomic.LoadInt64(&d.refs) + if refs > 0 { if d.cached { d.fs.cachedDentries.Remove(d) d.fs.cachedDentriesLen-- @@ -884,9 +1075,13 @@ func (d *dentry) checkCachingLocked() { } return } - // Non-child dentries with zero references are no longer reachable by path - // resolution and should be dropped immediately. - if d.vfsd.Parent() == nil || d.vfsd.IsDisowned() { + if refs == -1 { + // Dentry has already been destroyed. + return + } + // Deleted and invalidated dentries with zero references are no longer + // reachable by path resolution and should be dropped immediately. + if d.vfsd.IsDead() { if d.cached { d.fs.cachedDentries.Remove(d) d.fs.cachedDentriesLen-- @@ -911,34 +1106,45 @@ func (d *dentry) checkCachingLocked() { d.fs.cachedDentries.Remove(victim) d.fs.cachedDentriesLen-- victim.cached = false - // victim.refs may have become non-zero from an earlier path - // resolution since it was inserted into fs.cachedDentries; see - // dentry.incRefLocked(). Either way, we brought - // fs.cachedDentriesLen back down to fs.opts.maxCachedDentries, so - // we don't loop. + // victim.refs may have become non-zero from an earlier path resolution + // since it was inserted into fs.cachedDentries. if atomic.LoadInt64(&victim.refs) == 0 { - if victimParentVFSD := victim.vfsd.Parent(); victimParentVFSD != nil { - victimParent := victimParentVFSD.Impl().(*dentry) - victimParent.dirMu.Lock() - if !victim.vfsd.IsDisowned() { - // victim can't be a mount point (in any mount - // namespace), since VFS holds references on mount - // points. - d.fs.vfsfs.VirtualFilesystem().ForceDeleteDentry(&victim.vfsd) + if victim.parent != nil { + victim.parent.dirMu.Lock() + if !victim.vfsd.IsDead() { + // Note that victim can't be a mount point (in any mount + // namespace), since VFS holds references on mount points. + d.fs.vfsfs.VirtualFilesystem().InvalidateDentry(&victim.vfsd) + delete(victim.parent.children, victim.name) // We're only deleting the dentry, not the file it // represents, so we don't need to update // victimParent.dirents etc. } - victimParent.dirMu.Unlock() + victim.parent.dirMu.Unlock() } victim.destroyLocked() } + // Whether or not victim was destroyed, we brought fs.cachedDentriesLen + // back down to fs.opts.maxCachedDentries, so we don't loop. } } +// destroyLocked destroys the dentry. It may flushes dirty pages from cache, +// close p9 file and remove reference on parent dentry. +// // Preconditions: d.fs.renameMu must be locked for writing. d.refs == 0. d is // not a child dentry. func (d *dentry) destroyLocked() { + switch atomic.LoadInt64(&d.refs) { + case 0: + // Mark the dentry destroyed. + atomic.StoreInt64(&d.refs, -1) + case -1: + panic("dentry.destroyLocked() called on already destroyed dentry") + default: + panic("dentry.destroyLocked() called with references on the dentry") + } + ctx := context.Background() d.handleMu.Lock() if !d.handle.file.isNil() { @@ -958,17 +1164,20 @@ func (d *dentry) destroyLocked() { d.handle.close(ctx) } d.handleMu.Unlock() - d.file.close(ctx) - // Remove d from the set of all dentries. - d.fs.syncMu.Lock() - delete(d.fs.dentries, d) - d.fs.syncMu.Unlock() - // Drop the reference held by d on its parent. - if parentVFSD := d.vfsd.Parent(); parentVFSD != nil { - parent := parentVFSD.Impl().(*dentry) - // This is parent.DecRef() without recursive locking of d.fs.renameMu. - if refs := atomic.AddInt64(&parent.refs, -1); refs == 0 { - parent.checkCachingLocked() + + if !d.file.isNil() { + d.file.close(ctx) + d.file = p9file{} + // Remove d from the set of syncable dentries. + d.fs.syncMu.Lock() + delete(d.fs.syncableDentries, d) + d.fs.syncMu.Unlock() + } + // Drop the reference held by d on its parent without recursively locking + // d.fs.renameMu. + if d.parent != nil { + if refs := atomic.AddInt64(&d.parent.refs, -1); refs == 0 { + d.parent.checkCachingLocked() } else if refs < 0 { panic("gofer.dentry.DecRef() called without holding a reference") } @@ -983,24 +1192,65 @@ func (d *dentry) setDeleted() { atomic.StoreUint32(&d.deleted, 1) } -func (d *dentry) listxattr(ctx context.Context) ([]string, error) { - return nil, syserror.ENOTSUP +// We only support xattrs prefixed with "user." (see b/148380782). Currently, +// there is no need to expose any other xattrs through a gofer. +func (d *dentry) listxattr(ctx context.Context, creds *auth.Credentials, size uint64) ([]string, error) { + if d.file.isNil() { + return nil, nil + } + xattrMap, err := d.file.listXattr(ctx, size) + if err != nil { + return nil, err + } + xattrs := make([]string, 0, len(xattrMap)) + for x := range xattrMap { + if strings.HasPrefix(x, linux.XATTR_USER_PREFIX) { + xattrs = append(xattrs, x) + } + } + return xattrs, nil } -func (d *dentry) getxattr(ctx context.Context, name string) (string, error) { - // TODO(jamieliu): add vfs.GetxattrOptions.Size - return d.file.getXattr(ctx, name, linux.XATTR_SIZE_MAX) +func (d *dentry) getxattr(ctx context.Context, creds *auth.Credentials, opts *vfs.GetxattrOptions) (string, error) { + if d.file.isNil() { + return "", syserror.ENODATA + } + if err := d.checkPermissions(creds, vfs.MayRead); err != nil { + return "", err + } + if !strings.HasPrefix(opts.Name, linux.XATTR_USER_PREFIX) { + return "", syserror.EOPNOTSUPP + } + return d.file.getXattr(ctx, opts.Name, opts.Size) } -func (d *dentry) setxattr(ctx context.Context, opts *vfs.SetxattrOptions) error { +func (d *dentry) setxattr(ctx context.Context, creds *auth.Credentials, opts *vfs.SetxattrOptions) error { + if d.file.isNil() { + return syserror.EPERM + } + if err := d.checkPermissions(creds, vfs.MayWrite); err != nil { + return err + } + if !strings.HasPrefix(opts.Name, linux.XATTR_USER_PREFIX) { + return syserror.EOPNOTSUPP + } return d.file.setXattr(ctx, opts.Name, opts.Value, opts.Flags) } -func (d *dentry) removexattr(ctx context.Context, name string) error { - return syserror.ENOTSUP +func (d *dentry) removexattr(ctx context.Context, creds *auth.Credentials, name string) error { + if d.file.isNil() { + return syserror.EPERM + } + if err := d.checkPermissions(creds, vfs.MayWrite); err != nil { + return err + } + if !strings.HasPrefix(name, linux.XATTR_USER_PREFIX) { + return syserror.EOPNOTSUPP + } + return d.file.removeXattr(ctx, name) } -// Preconditions: d.isRegularFile() || d.isDirectory(). +// Preconditions: !d.isSynthetic(). d.isRegularFile() || d.isDirectory(). func (d *dentry) ensureSharedHandle(ctx context.Context, read, write, trunc bool) error { // O_TRUNC unconditionally requires us to obtain a new handle (opened with // O_TRUNC). @@ -1042,13 +1292,13 @@ func (d *dentry) ensureSharedHandle(ctx context.Context, read, write, trunc bool // using the old file descriptor, preventing us from safely // closing it. We could handle this by invalidating existing // memmap.Translations, but this is expensive. Instead, use - // dup2() to make the old file descriptor refer to the new file + // dup3 to make the old file descriptor refer to the new file // description, then close the new file descriptor (which is no // longer needed). Racing callers may use the old or new file // description, but this doesn't matter since they refer to the // same file (unless d.fs.opts.overlayfsStaleRead is true, // which we handle separately). - if err := syscall.Dup2(int(h.fd), int(d.handle.fd)); err != nil { + if err := syscall.Dup3(int(h.fd), int(d.handle.fd), syscall.O_CLOEXEC); err != nil { d.handleMu.Unlock() ctx.Warningf("gofer.dentry.ensureSharedHandle: failed to dup fd %d to fd %d: %v", h.fd, d.handle.fd, err) h.close(ctx) @@ -1091,6 +1341,26 @@ func (d *dentry) ensureSharedHandle(ctx context.Context, read, write, trunc bool return nil } +// incLinks increments link count. +// +// Preconditions: d.nlink != 0 && d.nlink < math.MaxUint32. +func (d *dentry) incLinks() { + v := atomic.AddUint32(&d.nlink, 1) + if v < 2 { + panic(fmt.Sprintf("dentry.nlink is invalid (was 0 or overflowed): %d", v)) + } +} + +// decLinks decrements link count. +// +// Preconditions: d.nlink > 1. +func (d *dentry) decLinks() { + v := atomic.AddUint32(&d.nlink, ^uint32(0)) + if v == 0 { + panic(fmt.Sprintf("dentry.nlink must be greater than 0: %d", v)) + } +} + // fileDescription is embedded by gofer implementations of // vfs.FileDescriptionImpl. type fileDescription struct { @@ -1109,7 +1379,8 @@ func (fd *fileDescription) dentry() *dentry { // Stat implements vfs.FileDescriptionImpl.Stat. func (fd *fileDescription) Stat(ctx context.Context, opts vfs.StatOptions) (linux.Statx, error) { d := fd.dentry() - if d.fs.opts.interop == InteropModeShared && opts.Mask&(linux.STATX_MODE|linux.STATX_UID|linux.STATX_GID|linux.STATX_ATIME|linux.STATX_MTIME|linux.STATX_CTIME|linux.STATX_SIZE|linux.STATX_BLOCKS|linux.STATX_BTIME) != 0 && opts.Sync != linux.AT_STATX_DONT_SYNC { + const validMask = uint32(linux.STATX_MODE | linux.STATX_UID | linux.STATX_GID | linux.STATX_ATIME | linux.STATX_MTIME | linux.STATX_CTIME | linux.STATX_SIZE | linux.STATX_BLOCKS | linux.STATX_BTIME) + if !d.cachedMetadataAuthoritative() && opts.Mask&validMask != 0 && opts.Sync != linux.AT_STATX_DONT_SYNC { // TODO(jamieliu): Use specialFileFD.handle.file for the getattr if // available? if err := d.updateFromGetattr(ctx); err != nil { @@ -1127,21 +1398,21 @@ func (fd *fileDescription) SetStat(ctx context.Context, opts vfs.SetStatOptions) } // Listxattr implements vfs.FileDescriptionImpl.Listxattr. -func (fd *fileDescription) Listxattr(ctx context.Context) ([]string, error) { - return fd.dentry().listxattr(ctx) +func (fd *fileDescription) Listxattr(ctx context.Context, size uint64) ([]string, error) { + return fd.dentry().listxattr(ctx, auth.CredentialsFromContext(ctx), size) } // Getxattr implements vfs.FileDescriptionImpl.Getxattr. -func (fd *fileDescription) Getxattr(ctx context.Context, name string) (string, error) { - return fd.dentry().getxattr(ctx, name) +func (fd *fileDescription) Getxattr(ctx context.Context, opts vfs.GetxattrOptions) (string, error) { + return fd.dentry().getxattr(ctx, auth.CredentialsFromContext(ctx), &opts) } // Setxattr implements vfs.FileDescriptionImpl.Setxattr. func (fd *fileDescription) Setxattr(ctx context.Context, opts vfs.SetxattrOptions) error { - return fd.dentry().setxattr(ctx, &opts) + return fd.dentry().setxattr(ctx, auth.CredentialsFromContext(ctx), &opts) } // Removexattr implements vfs.FileDescriptionImpl.Removexattr. func (fd *fileDescription) Removexattr(ctx context.Context, name string) error { - return fd.dentry().removexattr(ctx, name) + return fd.dentry().removexattr(ctx, auth.CredentialsFromContext(ctx), name) } diff --git a/pkg/sentry/fsimpl/gofer/gofer_test.go b/pkg/sentry/fsimpl/gofer/gofer_test.go new file mode 100644 index 000000000..adff39490 --- /dev/null +++ b/pkg/sentry/fsimpl/gofer/gofer_test.go @@ -0,0 +1,63 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package gofer + +import ( + "sync/atomic" + "testing" + + "gvisor.dev/gvisor/pkg/p9" + "gvisor.dev/gvisor/pkg/sentry/contexttest" +) + +func TestDestroyIdempotent(t *testing.T) { + fs := filesystem{ + syncableDentries: make(map[*dentry]struct{}), + opts: filesystemOptions{ + // Test relies on no dentry being held in the cache. + maxCachedDentries: 0, + }, + } + + ctx := contexttest.Context(t) + attr := &p9.Attr{ + Mode: p9.ModeRegular, + } + mask := p9.AttrMask{ + Mode: true, + Size: true, + } + parent, err := fs.newDentry(ctx, p9file{}, p9.QID{}, mask, attr) + if err != nil { + t.Fatalf("fs.newDentry(): %v", err) + } + + child, err := fs.newDentry(ctx, p9file{}, p9.QID{}, mask, attr) + if err != nil { + t.Fatalf("fs.newDentry(): %v", err) + } + parent.cacheNewChildLocked(child, "child") + + child.checkCachingLocked() + if got := atomic.LoadInt64(&child.refs); got != -1 { + t.Fatalf("child.refs=%d, want: -1", got) + } + // Parent will also be destroyed when child reference is removed. + if got := atomic.LoadInt64(&parent.refs); got != -1 { + t.Fatalf("parent.refs=%d, want: -1", got) + } + child.checkCachingLocked() + child.checkCachingLocked() +} diff --git a/pkg/sentry/fsimpl/gofer/handle.go b/pkg/sentry/fsimpl/gofer/handle.go index cfe66f797..724a3f1f7 100644 --- a/pkg/sentry/fsimpl/gofer/handle.go +++ b/pkg/sentry/fsimpl/gofer/handle.go @@ -20,6 +20,7 @@ import ( "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/p9" "gvisor.dev/gvisor/pkg/safemem" + "gvisor.dev/gvisor/pkg/sentry/hostfd" ) // handle represents a remote "open file descriptor", consisting of an opened @@ -77,7 +78,7 @@ func (h *handle) readToBlocksAt(ctx context.Context, dsts safemem.BlockSeq, offs } if h.fd >= 0 { ctx.UninterruptibleSleepStart(false) - n, err := hostPreadv(h.fd, dsts, int64(offset)) + n, err := hostfd.Preadv2(h.fd, dsts, int64(offset), 0 /* flags */) ctx.UninterruptibleSleepFinish(false) return n, err } @@ -103,7 +104,7 @@ func (h *handle) writeFromBlocksAt(ctx context.Context, srcs safemem.BlockSeq, o } if h.fd >= 0 { ctx.UninterruptibleSleepStart(false) - n, err := hostPwritev(h.fd, srcs, int64(offset)) + n, err := hostfd.Pwritev2(h.fd, srcs, int64(offset), 0 /* flags */) ctx.UninterruptibleSleepFinish(false) return n, err } diff --git a/pkg/sentry/fsimpl/gofer/handle_unsafe.go b/pkg/sentry/fsimpl/gofer/handle_unsafe.go deleted file mode 100644 index 19560ab26..000000000 --- a/pkg/sentry/fsimpl/gofer/handle_unsafe.go +++ /dev/null @@ -1,66 +0,0 @@ -// Copyright 2019 The gVisor Authors. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -package gofer - -import ( - "syscall" - "unsafe" - - "gvisor.dev/gvisor/pkg/safemem" -) - -// Preconditions: !dsts.IsEmpty(). -func hostPreadv(fd int32, dsts safemem.BlockSeq, off int64) (uint64, error) { - // No buffering is necessary regardless of safecopy; host syscalls will - // return EFAULT if appropriate, instead of raising SIGBUS. - if dsts.NumBlocks() == 1 { - // Use pread() instead of preadv() to avoid iovec allocation and - // copying. - dst := dsts.Head() - n, _, e := syscall.Syscall6(syscall.SYS_PREAD64, uintptr(fd), dst.Addr(), uintptr(dst.Len()), uintptr(off), 0, 0) - if e != 0 { - return 0, e - } - return uint64(n), nil - } - iovs := safemem.IovecsFromBlockSeq(dsts) - n, _, e := syscall.Syscall6(syscall.SYS_PREADV, uintptr(fd), uintptr((unsafe.Pointer)(&iovs[0])), uintptr(len(iovs)), uintptr(off), 0, 0) - if e != 0 { - return 0, e - } - return uint64(n), nil -} - -// Preconditions: !srcs.IsEmpty(). -func hostPwritev(fd int32, srcs safemem.BlockSeq, off int64) (uint64, error) { - // No buffering is necessary regardless of safecopy; host syscalls will - // return EFAULT if appropriate, instead of raising SIGBUS. - if srcs.NumBlocks() == 1 { - // Use pwrite() instead of pwritev() to avoid iovec allocation and - // copying. - src := srcs.Head() - n, _, e := syscall.Syscall6(syscall.SYS_PWRITE64, uintptr(fd), src.Addr(), uintptr(src.Len()), uintptr(off), 0, 0) - if e != 0 { - return 0, e - } - return uint64(n), nil - } - iovs := safemem.IovecsFromBlockSeq(srcs) - n, _, e := syscall.Syscall6(syscall.SYS_PWRITEV, uintptr(fd), uintptr((unsafe.Pointer)(&iovs[0])), uintptr(len(iovs)), uintptr(off), 0, 0) - if e != 0 { - return 0, e - } - return uint64(n), nil -} diff --git a/pkg/sentry/fsimpl/gofer/host_named_pipe.go b/pkg/sentry/fsimpl/gofer/host_named_pipe.go new file mode 100644 index 000000000..7294de7d6 --- /dev/null +++ b/pkg/sentry/fsimpl/gofer/host_named_pipe.go @@ -0,0 +1,97 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package gofer + +import ( + "fmt" + "sync" + "time" + + "golang.org/x/sys/unix" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/syserror" +) + +// Global pipe used by blockUntilNonblockingPipeHasWriter since we can't create +// pipes after sentry initialization due to syscall filters. +var ( + tempPipeMu sync.Mutex + tempPipeReadFD int + tempPipeWriteFD int + tempPipeBuf [1]byte +) + +func init() { + var pipeFDs [2]int + if err := unix.Pipe(pipeFDs[:]); err != nil { + panic(fmt.Sprintf("failed to create pipe for gofer.blockUntilNonblockingPipeHasWriter: %v", err)) + } + tempPipeReadFD = pipeFDs[0] + tempPipeWriteFD = pipeFDs[1] +} + +func blockUntilNonblockingPipeHasWriter(ctx context.Context, fd int32) error { + for { + ok, err := nonblockingPipeHasWriter(fd) + if err != nil { + return err + } + if ok { + return nil + } + if err := sleepBetweenNamedPipeOpenChecks(ctx); err != nil { + return err + } + } +} + +func nonblockingPipeHasWriter(fd int32) (bool, error) { + tempPipeMu.Lock() + defer tempPipeMu.Unlock() + // Copy 1 byte from fd into the temporary pipe. + n, err := unix.Tee(int(fd), tempPipeWriteFD, 1, unix.SPLICE_F_NONBLOCK) + if err == syserror.EAGAIN { + // The pipe represented by fd is empty, but has a writer. + return true, nil + } + if err != nil { + return false, err + } + if n == 0 { + // The pipe represented by fd is empty and has no writer. + return false, nil + } + // The pipe represented by fd is non-empty, so it either has, or has + // previously had, a writer. Remove the byte copied to the temporary pipe + // before returning. + if n, err := unix.Read(tempPipeReadFD, tempPipeBuf[:]); err != nil || n != 1 { + panic(fmt.Sprintf("failed to drain pipe for gofer.blockUntilNonblockingPipeHasWriter: got (%d, %v), wanted (1, nil)", n, err)) + } + return true, nil +} + +func sleepBetweenNamedPipeOpenChecks(ctx context.Context) error { + t := time.NewTimer(100 * time.Millisecond) + defer t.Stop() + cancel := ctx.SleepStart() + select { + case <-t.C: + ctx.SleepFinish(true) + return nil + case <-cancel: + ctx.SleepFinish(false) + return syserror.ErrInterrupted + } +} diff --git a/pkg/sentry/fsimpl/gofer/p9file.go b/pkg/sentry/fsimpl/gofer/p9file.go index 755ac2985..87f0b877f 100644 --- a/pkg/sentry/fsimpl/gofer/p9file.go +++ b/pkg/sentry/fsimpl/gofer/p9file.go @@ -85,6 +85,13 @@ func (f p9file) setAttr(ctx context.Context, valid p9.SetAttrMask, attr p9.SetAt return err } +func (f p9file) listXattr(ctx context.Context, size uint64) (map[string]struct{}, error) { + ctx.UninterruptibleSleepStart(false) + xattrs, err := f.file.ListXattr(size) + ctx.UninterruptibleSleepFinish(false) + return xattrs, err +} + func (f p9file) getXattr(ctx context.Context, name string, size uint64) (string, error) { ctx.UninterruptibleSleepStart(false) val, err := f.file.GetXattr(name, size) @@ -99,6 +106,13 @@ func (f p9file) setXattr(ctx context.Context, name, value string, flags uint32) return err } +func (f p9file) removeXattr(ctx context.Context, name string) error { + ctx.UninterruptibleSleepStart(false) + err := f.file.RemoveXattr(name) + ctx.UninterruptibleSleepFinish(false) + return err +} + func (f p9file) allocate(ctx context.Context, mode p9.AllocateMode, offset, length uint64) error { ctx.UninterruptibleSleepStart(false) err := f.file.Allocate(mode, offset, length) diff --git a/pkg/sentry/fsimpl/gofer/regular_file.go b/pkg/sentry/fsimpl/gofer/regular_file.go index 54c1031a7..0d10cf7ac 100644 --- a/pkg/sentry/fsimpl/gofer/regular_file.go +++ b/pkg/sentry/fsimpl/gofer/regular_file.go @@ -104,7 +104,7 @@ func (fd *regularFileFD) PRead(ctx context.Context, dst usermem.IOSequence, offs putDentryReadWriter(rw) if d.fs.opts.interop != InteropModeShared { // Compare Linux's mm/filemap.c:do_generic_file_read() => file_accessed(). - d.touchAtime(ctx, fd.vfsfd.Mount()) + d.touchAtime(fd.vfsfd.Mount()) } return n, err } @@ -126,6 +126,11 @@ func (fd *regularFileFD) PWrite(ctx context.Context, src usermem.IOSequence, off if opts.Flags != 0 { return 0, syserror.EOPNOTSUPP } + limit, err := vfs.CheckLimit(ctx, offset, src.NumBytes()) + if err != nil { + return 0, err + } + src = src.TakeFirst64(limit) d := fd.dentry() d.metadataMu.Lock() @@ -134,10 +139,7 @@ func (fd *regularFileFD) PWrite(ctx context.Context, src usermem.IOSequence, off // Compare Linux's mm/filemap.c:__generic_file_write_iter() => // file_update_time(). This is d.touchCMtime(), but without locking // d.metadataMu (recursively). - if now, ok := nowFromContext(ctx); ok { - atomic.StoreInt64(&d.mtime, now) - atomic.StoreInt64(&d.ctime, now) - } + d.touchCMtimeLocked() } if fd.vfsfd.StatusFlags()&linux.O_DIRECT != 0 { // Write dirty cached pages that will be touched by the write back to @@ -146,9 +148,9 @@ func (fd *regularFileFD) PWrite(ctx context.Context, src usermem.IOSequence, off return 0, err } // Remove touched pages from the cache. - pgstart := pageRoundDown(uint64(offset)) - pgend := pageRoundUp(uint64(offset + src.NumBytes())) - if pgend < pgstart { + pgstart := usermem.PageRoundDown(uint64(offset)) + pgend, ok := usermem.PageRoundUp(uint64(offset + src.NumBytes())) + if !ok { return 0, syserror.EINVAL } mr := memmap.MappableRange{pgstart, pgend} @@ -304,9 +306,10 @@ func (rw *dentryReadWriter) ReadToBlocks(dsts safemem.BlockSeq) (uint64, error) if fillCache { // Read into the cache, then re-enter the loop to read from the // cache. + gapEnd, _ := usermem.PageRoundUp(gapMR.End) reqMR := memmap.MappableRange{ - Start: pageRoundDown(gapMR.Start), - End: pageRoundUp(gapMR.End), + Start: usermem.PageRoundDown(gapMR.Start), + End: gapEnd, } optMR := gap.Range() err := rw.d.cache.Fill(rw.ctx, reqMR, maxFillRange(reqMR, optMR), mf, usage.PageCache, rw.d.handle.readToBlocksAt) @@ -361,8 +364,15 @@ func (rw *dentryReadWriter) WriteFromBlocks(srcs safemem.BlockSeq) (uint64, erro rw.d.handleMu.RLock() if (rw.d.handle.fd >= 0 && !rw.d.fs.opts.forcePageCache) || rw.d.fs.opts.interop == InteropModeShared || rw.direct { n, err := rw.d.handle.writeFromBlocksAt(rw.ctx, srcs, rw.off) - rw.d.handleMu.RUnlock() rw.off += n + rw.d.dataMu.Lock() + if rw.off > rw.d.size { + atomic.StoreUint64(&rw.d.size, rw.off) + // The remote file's size will implicitly be extended to the correct + // value when we write back to it. + } + rw.d.dataMu.Unlock() + rw.d.handleMu.RUnlock() return n, err } @@ -662,7 +672,7 @@ func (d *dentry) Translate(ctx context.Context, required, optional memmap.Mappab // Constrain translations to d.size (rounded up) to prevent translation to // pages that may be concurrently truncated. - pgend := pageRoundUp(d.size) + pgend, _ := usermem.PageRoundUp(d.size) var beyondEOF bool if required.End > pgend { if required.Start >= pgend { @@ -809,43 +819,15 @@ type dentryPlatformFile struct { // IncRef implements platform.File.IncRef. func (d *dentryPlatformFile) IncRef(fr platform.FileRange) { d.dataMu.Lock() - seg, gap := d.fdRefs.Find(fr.Start) - for { - switch { - case seg.Ok() && seg.Start() < fr.End: - seg = d.fdRefs.Isolate(seg, fr) - seg.SetValue(seg.Value() + 1) - seg, gap = seg.NextNonEmpty() - case gap.Ok() && gap.Start() < fr.End: - newRange := gap.Range().Intersect(fr) - usage.MemoryAccounting.Inc(newRange.Length(), usage.Mapped) - seg, gap = d.fdRefs.InsertWithoutMerging(gap, newRange, 1).NextNonEmpty() - default: - d.fdRefs.MergeAdjacent(fr) - d.dataMu.Unlock() - return - } - } + d.fdRefs.IncRefAndAccount(fr) + d.dataMu.Unlock() } // DecRef implements platform.File.DecRef. func (d *dentryPlatformFile) DecRef(fr platform.FileRange) { d.dataMu.Lock() - seg := d.fdRefs.FindSegment(fr.Start) - - for seg.Ok() && seg.Start() < fr.End { - seg = d.fdRefs.Isolate(seg, fr) - if old := seg.Value(); old == 1 { - usage.MemoryAccounting.Dec(seg.Range().Length(), usage.Mapped) - seg = d.fdRefs.Remove(seg).NextSegment() - } else { - seg.SetValue(old - 1) - seg = seg.NextSegment() - } - } - d.fdRefs.MergeAdjacent(fr) + d.fdRefs.DecRefAndAccount(fr) d.dataMu.Unlock() - } // MapInternal implements platform.File.MapInternal. diff --git a/pkg/sentry/fsimpl/gofer/socket.go b/pkg/sentry/fsimpl/gofer/socket.go new file mode 100644 index 000000000..d6dbe9092 --- /dev/null +++ b/pkg/sentry/fsimpl/gofer/socket.go @@ -0,0 +1,146 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package gofer + +import ( + "syscall" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/log" + "gvisor.dev/gvisor/pkg/p9" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/host" + "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" + "gvisor.dev/gvisor/pkg/syserr" + "gvisor.dev/gvisor/pkg/waiter" +) + +func (d *dentry) isSocket() bool { + return d.fileType() == linux.S_IFSOCK +} + +// endpoint is a Gofer-backed transport.BoundEndpoint. +// +// An endpoint's lifetime is the time between when filesystem.BoundEndpointAt() +// is called and either BoundEndpoint.BidirectionalConnect or +// BoundEndpoint.UnidirectionalConnect is called. +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 + + // path is the sentry path where this endpoint is bound. + path string +} + +func sockTypeToP9(t linux.SockType) (p9.ConnectFlags, bool) { + switch t { + case linux.SOCK_STREAM: + return p9.StreamSocket, true + case linux.SOCK_SEQPACKET: + return p9.SeqpacketSocket, true + case linux.SOCK_DGRAM: + return p9.DgramSocket, true + } + return 0, false +} + +// BidirectionalConnect implements ConnectableEndpoint.BidirectionalConnect. +func (e *endpoint) BidirectionalConnect(ctx context.Context, ce transport.ConnectingEndpoint, returnConnect func(transport.Receiver, transport.ConnectedEndpoint)) *syserr.Error { + cf, ok := sockTypeToP9(ce.Type()) + if !ok { + return syserr.ErrConnectionRefused + } + + // No lock ordering required as only the ConnectingEndpoint has a mutex. + ce.Lock() + + // Check connecting state. + if ce.Connected() { + ce.Unlock() + return syserr.ErrAlreadyConnected + } + if ce.Listening() { + ce.Unlock() + return syserr.ErrInvalidEndpointState + } + + c, err := e.newConnectedEndpoint(ctx, cf, ce.WaiterQueue()) + if err != nil { + ce.Unlock() + return err + } + + returnConnect(c, c) + ce.Unlock() + if err := c.Init(); err != nil { + return syserr.FromError(err) + } + + return nil +} + +// UnidirectionalConnect implements +// transport.BoundEndpoint.UnidirectionalConnect. +func (e *endpoint) UnidirectionalConnect(ctx context.Context) (transport.ConnectedEndpoint, *syserr.Error) { + c, err := e.newConnectedEndpoint(ctx, p9.DgramSocket, &waiter.Queue{}) + if err != nil { + return nil, err + } + + if err := c.Init(); err != nil { + return nil, syserr.FromError(err) + } + + // We don't need the receiver. + c.CloseRecv() + c.Release() + + return c, nil +} + +func (e *endpoint) newConnectedEndpoint(ctx context.Context, flags p9.ConnectFlags, queue *waiter.Queue) (*host.SCMConnectedEndpoint, *syserr.Error) { + hostFile, err := e.file.Connect(flags) + if err != nil { + return nil, syserr.ErrConnectionRefused + } + // Dup the fd so that the new endpoint can manage its lifetime. + hostFD, err := syscall.Dup(hostFile.FD()) + if err != nil { + log.Warningf("Could not dup host socket fd %d: %v", hostFile.FD(), err) + return nil, syserr.FromError(err) + } + // After duplicating, we no longer need hostFile. + hostFile.Close() + + c, serr := host.NewSCMEndpoint(ctx, hostFD, queue, e.path) + if serr != nil { + log.Warningf("Gofer returned invalid host socket for BidirectionalConnect; file %+v flags %+v: %v", e.file, flags, serr) + return nil, serr + } + return c, nil +} + +// Release implements transport.BoundEndpoint.Release. +func (e *endpoint) Release() { + e.dentry.DecRef() +} + +// Passcred implements transport.BoundEndpoint.Passcred. +func (e *endpoint) Passcred() bool { + return false +} diff --git a/pkg/sentry/fsimpl/gofer/special_file.go b/pkg/sentry/fsimpl/gofer/special_file.go index 08c691c47..ff6126b87 100644 --- a/pkg/sentry/fsimpl/gofer/special_file.go +++ b/pkg/sentry/fsimpl/gofer/special_file.go @@ -19,33 +19,69 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/fdnotifier" "gvisor.dev/gvisor/pkg/safemem" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" + "gvisor.dev/gvisor/pkg/waiter" ) -// specialFileFD implements vfs.FileDescriptionImpl for files other than -// regular files, directories, and symlinks: pipes, sockets, etc. It is also -// used for regular files when filesystemOptions.specialRegularFiles is in -// effect. specialFileFD differs from regularFileFD by using per-FD handles -// instead of shared per-dentry handles, and never buffering I/O. +// specialFileFD implements vfs.FileDescriptionImpl for pipes, sockets, device +// special files, and (when filesystemOptions.specialRegularFiles 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. type specialFileFD struct { fileDescription - // handle is immutable. + // handle is used for file I/O. handle is immutable. handle handle - // off is the file offset. off is protected by mu. (POSIX 2.9.7 only - // requires operations using the file offset to be atomic for regular files - // and symlinks; however, since specialFileFD may be used for regular - // files, we apply this atomicity unconditionally.) + // seekable is true if this file description represents a file for which + // file offset is significant, i.e. a regular file. seekable is immutable. + seekable bool + + // mayBlock is true if this file description represents a file for which + // queue may send I/O readiness events. mayBlock is immutable. + mayBlock bool + queue waiter.Queue + + // If seekable is true, off is the file offset. off is protected by mu. mu sync.Mutex off int64 } +func newSpecialFileFD(h handle, mnt *vfs.Mount, d *dentry, flags uint32) (*specialFileFD, error) { + ftype := d.fileType() + seekable := ftype == linux.S_IFREG + mayBlock := ftype == linux.S_IFIFO || ftype == linux.S_IFSOCK + fd := &specialFileFD{ + handle: h, + seekable: seekable, + mayBlock: mayBlock, + } + if mayBlock && h.fd >= 0 { + if err := fdnotifier.AddFD(h.fd, &fd.queue); err != nil { + return nil, err + } + } + if err := fd.vfsfd.Init(fd, flags, mnt, &d.vfsd, &vfs.FileDescriptionOptions{ + DenyPRead: !seekable, + DenyPWrite: !seekable, + }); err != nil { + if mayBlock && h.fd >= 0 { + fdnotifier.RemoveFD(h.fd) + } + return nil, err + } + return fd, nil +} + // Release implements vfs.FileDescriptionImpl.Release. func (fd *specialFileFD) Release() { + if fd.mayBlock && fd.handle.fd >= 0 { + fdnotifier.RemoveFD(fd.handle.fd) + } fd.handle.close(context.Background()) fs := fd.vfsfd.Mount().Filesystem().Impl().(*filesystem) fs.syncMu.Lock() @@ -61,9 +97,35 @@ func (fd *specialFileFD) OnClose(ctx context.Context) error { return fd.handle.file.flush(ctx) } +// Readiness implements waiter.Waitable.Readiness. +func (fd *specialFileFD) Readiness(mask waiter.EventMask) waiter.EventMask { + if fd.mayBlock { + return fdnotifier.NonBlockingPoll(fd.handle.fd, mask) + } + return fd.fileDescription.Readiness(mask) +} + +// EventRegister implements waiter.Waitable.EventRegister. +func (fd *specialFileFD) EventRegister(e *waiter.Entry, mask waiter.EventMask) { + if fd.mayBlock { + fd.queue.EventRegister(e, mask) + return + } + fd.fileDescription.EventRegister(e, mask) +} + +// EventUnregister implements waiter.Waitable.EventUnregister. +func (fd *specialFileFD) EventUnregister(e *waiter.Entry) { + if fd.mayBlock { + fd.queue.EventUnregister(e) + return + } + fd.fileDescription.EventUnregister(e) +} + // PRead implements vfs.FileDescriptionImpl.PRead. func (fd *specialFileFD) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { - if offset < 0 { + if fd.seekable && offset < 0 { return 0, syserror.EINVAL } if opts.Flags != 0 { @@ -76,10 +138,13 @@ func (fd *specialFileFD) PRead(ctx context.Context, dst usermem.IOSequence, offs // hold here since specialFileFD doesn't client-cache data. Just buffer the // read instead. if d := fd.dentry(); d.fs.opts.interop != InteropModeShared { - d.touchAtime(ctx, fd.vfsfd.Mount()) + d.touchAtime(fd.vfsfd.Mount()) } buf := make([]byte, dst.NumBytes()) n, err := fd.handle.readToBlocksAt(ctx, safemem.BlockSeqOf(safemem.BlockFromSafeSlice(buf)), uint64(offset)) + if err == syserror.EAGAIN { + err = syserror.ErrWouldBlock + } if n == 0 { return 0, err } @@ -91,6 +156,10 @@ func (fd *specialFileFD) PRead(ctx context.Context, dst usermem.IOSequence, offs // Read implements vfs.FileDescriptionImpl.Read. func (fd *specialFileFD) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { + if !fd.seekable { + return fd.PRead(ctx, dst, -1, opts) + } + fd.mu.Lock() n, err := fd.PRead(ctx, dst, fd.off, opts) fd.off += n @@ -100,16 +169,24 @@ func (fd *specialFileFD) Read(ctx context.Context, dst usermem.IOSequence, opts // PWrite implements vfs.FileDescriptionImpl.PWrite. func (fd *specialFileFD) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) { - if offset < 0 { + if fd.seekable && offset < 0 { return 0, syserror.EINVAL } if opts.Flags != 0 { return 0, syserror.EOPNOTSUPP } + if fd.seekable { + limit, err := vfs.CheckLimit(ctx, offset, src.NumBytes()) + if err != nil { + return 0, err + } + src = src.TakeFirst64(limit) + } + // Do a buffered write. See rationale in PRead. if d := fd.dentry(); d.fs.opts.interop != InteropModeShared { - d.touchCMtime(ctx) + d.touchCMtime() } buf := make([]byte, src.NumBytes()) // Don't do partial writes if we get a partial read from src. @@ -117,11 +194,18 @@ func (fd *specialFileFD) PWrite(ctx context.Context, src usermem.IOSequence, off return 0, err } n, err := fd.handle.writeFromBlocksAt(ctx, safemem.BlockSeqOf(safemem.BlockFromSafeSlice(buf)), uint64(offset)) + if err == syserror.EAGAIN { + err = syserror.ErrWouldBlock + } return int64(n), err } // Write implements vfs.FileDescriptionImpl.Write. func (fd *specialFileFD) Write(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) { + if !fd.seekable { + return fd.PWrite(ctx, src, -1, opts) + } + fd.mu.Lock() n, err := fd.PWrite(ctx, src, fd.off, opts) fd.off += n @@ -131,6 +215,9 @@ func (fd *specialFileFD) Write(ctx context.Context, src usermem.IOSequence, opts // Seek implements vfs.FileDescriptionImpl.Seek. func (fd *specialFileFD) Seek(ctx context.Context, offset int64, whence int32) (int64, error) { + if !fd.seekable { + return 0, syserror.ESPIPE + } fd.mu.Lock() defer fd.mu.Unlock() switch whence { diff --git a/pkg/sentry/fsimpl/gofer/symlink.go b/pkg/sentry/fsimpl/gofer/symlink.go index adf43be60..2ec819f86 100644 --- a/pkg/sentry/fsimpl/gofer/symlink.go +++ b/pkg/sentry/fsimpl/gofer/symlink.go @@ -27,7 +27,7 @@ func (d *dentry) isSymlink() bool { // Precondition: d.isSymlink(). func (d *dentry) readlink(ctx context.Context, mnt *vfs.Mount) (string, error) { if d.fs.opts.interop != InteropModeShared { - d.touchAtime(ctx, mnt) + d.touchAtime(mnt) d.dataMu.Lock() if d.haveTarget { target := d.target diff --git a/pkg/sentry/fsimpl/gofer/time.go b/pkg/sentry/fsimpl/gofer/time.go index 7598ec6a8..1d5aa82dc 100644 --- a/pkg/sentry/fsimpl/gofer/time.go +++ b/pkg/sentry/fsimpl/gofer/time.go @@ -18,8 +18,6 @@ import ( "sync/atomic" "gvisor.dev/gvisor/pkg/abi/linux" - "gvisor.dev/gvisor/pkg/context" - ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" "gvisor.dev/gvisor/pkg/sentry/vfs" ) @@ -38,23 +36,15 @@ func statxTimestampFromDentry(ns int64) linux.StatxTimestamp { } } -func nowFromContext(ctx context.Context) (int64, bool) { - if clock := ktime.RealtimeClockFromContext(ctx); clock != nil { - return clock.Now().Nanoseconds(), true - } - return 0, false -} - // Preconditions: fs.interop != InteropModeShared. -func (d *dentry) touchAtime(ctx context.Context, mnt *vfs.Mount) { - if err := mnt.CheckBeginWrite(); err != nil { +func (d *dentry) touchAtime(mnt *vfs.Mount) { + if mnt.Flags.NoATime { return } - now, ok := nowFromContext(ctx) - if !ok { - mnt.EndWrite() + if err := mnt.CheckBeginWrite(); err != nil { return } + now := d.fs.clock.Now().Nanoseconds() d.metadataMu.Lock() atomic.StoreInt64(&d.atime, now) d.metadataMu.Unlock() @@ -63,13 +53,25 @@ func (d *dentry) touchAtime(ctx context.Context, mnt *vfs.Mount) { // Preconditions: fs.interop != InteropModeShared. The caller has successfully // called vfs.Mount.CheckBeginWrite(). -func (d *dentry) touchCMtime(ctx context.Context) { - now, ok := nowFromContext(ctx) - if !ok { - return - } +func (d *dentry) touchCtime() { + now := d.fs.clock.Now().Nanoseconds() + d.metadataMu.Lock() + atomic.StoreInt64(&d.ctime, now) + d.metadataMu.Unlock() +} + +// Preconditions: fs.interop != InteropModeShared. The caller has successfully +// called vfs.Mount.CheckBeginWrite(). +func (d *dentry) touchCMtime() { + now := d.fs.clock.Now().Nanoseconds() d.metadataMu.Lock() atomic.StoreInt64(&d.mtime, now) atomic.StoreInt64(&d.ctime, now) d.metadataMu.Unlock() } + +func (d *dentry) touchCMtimeLocked() { + now := d.fs.clock.Now().Nanoseconds() + atomic.StoreInt64(&d.mtime, now) + atomic.StoreInt64(&d.ctime, now) +} diff --git a/pkg/sentry/fsimpl/host/BUILD b/pkg/sentry/fsimpl/host/BUILD new file mode 100644 index 000000000..ca0fe6d2b --- /dev/null +++ b/pkg/sentry/fsimpl/host/BUILD @@ -0,0 +1,51 @@ +load("//tools:defs.bzl", "go_library") + +licenses(["notice"]) + +go_library( + name = "host", + srcs = [ + "control.go", + "host.go", + "ioctl_unsafe.go", + "mmap.go", + "socket.go", + "socket_iovec.go", + "socket_unsafe.go", + "tty.go", + "util.go", + "util_unsafe.go", + ], + visibility = ["//pkg/sentry:internal"], + deps = [ + "//pkg/abi/linux", + "//pkg/context", + "//pkg/fdnotifier", + "//pkg/fspath", + "//pkg/log", + "//pkg/refs", + "//pkg/safemem", + "//pkg/sentry/arch", + "//pkg/sentry/fs/fsutil", + "//pkg/sentry/fsimpl/kernfs", + "//pkg/sentry/hostfd", + "//pkg/sentry/kernel", + "//pkg/sentry/kernel/auth", + "//pkg/sentry/memmap", + "//pkg/sentry/platform", + "//pkg/sentry/socket/control", + "//pkg/sentry/socket/unix", + "//pkg/sentry/socket/unix/transport", + "//pkg/sentry/unimpl", + "//pkg/sentry/uniqueid", + "//pkg/sentry/vfs", + "//pkg/sync", + "//pkg/syserr", + "//pkg/syserror", + "//pkg/tcpip", + "//pkg/unet", + "//pkg/usermem", + "//pkg/waiter", + "@org_golang_x_sys//unix:go_default_library", + ], +) diff --git a/pkg/sentry/fsimpl/host/control.go b/pkg/sentry/fsimpl/host/control.go new file mode 100644 index 000000000..b9082a20f --- /dev/null +++ b/pkg/sentry/fsimpl/host/control.go @@ -0,0 +1,96 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package host + +import ( + "syscall" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/socket/control" + "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" + "gvisor.dev/gvisor/pkg/sentry/vfs" +) + +type scmRights struct { + fds []int +} + +func newSCMRights(fds []int) control.SCMRightsVFS2 { + return &scmRights{fds} +} + +// Files implements control.SCMRights.Files. +func (c *scmRights) Files(ctx context.Context, max int) (control.RightsFilesVFS2, bool) { + n := max + var trunc bool + if l := len(c.fds); n > l { + n = l + } else if n < l { + trunc = true + } + + rf := control.RightsFilesVFS2(fdsToFiles(ctx, c.fds[:n])) + + // Only consume converted FDs (fdsToFiles may convert fewer than n FDs). + c.fds = c.fds[len(rf):] + return rf, trunc +} + +// Clone implements transport.RightsControlMessage.Clone. +func (c *scmRights) Clone() transport.RightsControlMessage { + // Host rights never need to be cloned. + return nil +} + +// Release implements transport.RightsControlMessage.Release. +func (c *scmRights) Release() { + for _, fd := range c.fds { + syscall.Close(fd) + } + c.fds = nil +} + +// If an error is encountered, only files created before the error will be +// returned. This is what Linux does. +func fdsToFiles(ctx context.Context, fds []int) []*vfs.FileDescription { + files := make([]*vfs.FileDescription, 0, len(fds)) + for _, fd := range fds { + // Get flags. We do it here because they may be modified + // by subsequent functions. + fileFlags, _, errno := syscall.Syscall(syscall.SYS_FCNTL, uintptr(fd), syscall.F_GETFL, 0) + if errno != 0 { + ctx.Warningf("Error retrieving host FD flags: %v", error(errno)) + break + } + + // Create the file backed by hostFD. + file, err := ImportFD(ctx, kernel.KernelFromContext(ctx).HostMount(), fd, false /* isTTY */) + if err != nil { + ctx.Warningf("Error creating file from host FD: %v", err) + break + } + + if err := file.SetStatusFlags(ctx, auth.CredentialsFromContext(ctx), uint32(fileFlags&linux.O_NONBLOCK)); err != nil { + ctx.Warningf("Error setting flags on host FD file: %v", err) + break + } + + files = append(files, file) + } + return files +} diff --git a/pkg/sentry/fsimpl/host/host.go b/pkg/sentry/fsimpl/host/host.go new file mode 100644 index 000000000..18b127521 --- /dev/null +++ b/pkg/sentry/fsimpl/host/host.go @@ -0,0 +1,714 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// Package host provides a filesystem implementation for host files imported as +// file descriptors. +package host + +import ( + "fmt" + "math" + "syscall" + + "golang.org/x/sys/unix" + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/fdnotifier" + "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/log" + "gvisor.dev/gvisor/pkg/refs" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs" + "gvisor.dev/gvisor/pkg/sentry/hostfd" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/memmap" + unixsocket "gvisor.dev/gvisor/pkg/sentry/socket/unix" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" + "gvisor.dev/gvisor/pkg/waiter" +) + +// NewFDOptions contains options to NewFD. +type NewFDOptions struct { + // If IsTTY is true, the file descriptor is a TTY. + IsTTY bool + + // If HaveFlags is true, use Flags for the new file description. Otherwise, + // the new file description will inherit flags from hostFD. + HaveFlags bool + Flags uint32 +} + +// NewFD returns a vfs.FileDescription representing the given host file +// descriptor. mnt must be Kernel.HostMount(). +func NewFD(ctx context.Context, mnt *vfs.Mount, hostFD int, opts *NewFDOptions) (*vfs.FileDescription, error) { + fs, ok := mnt.Filesystem().Impl().(*filesystem) + if !ok { + return nil, fmt.Errorf("can't import host FDs into filesystems of type %T", mnt.Filesystem().Impl()) + } + + // Retrieve metadata. + var s unix.Stat_t + if err := unix.Fstat(hostFD, &s); err != nil { + return nil, err + } + + flags := opts.Flags + if !opts.HaveFlags { + // Get flags for the imported FD. + flagsInt, err := unix.FcntlInt(uintptr(hostFD), syscall.F_GETFL, 0) + if err != nil { + return nil, err + } + flags = uint32(flagsInt) + } + + fileMode := linux.FileMode(s.Mode) + fileType := fileMode.FileType() + + // Determine if hostFD is seekable. If not, this syscall will return ESPIPE + // (see fs/read_write.c:llseek), e.g. for pipes, sockets, and some character + // devices. + _, err := unix.Seek(hostFD, 0, linux.SEEK_CUR) + seekable := err != syserror.ESPIPE + + i := &inode{ + hostFD: hostFD, + ino: fs.NextIno(), + isTTY: opts.IsTTY, + wouldBlock: wouldBlock(uint32(fileType)), + seekable: seekable, + canMap: canMap(uint32(fileType)), + } + i.pf.inode = i + + // Non-seekable files can't be memory mapped, assert this. + if !i.seekable && i.canMap { + panic("files that can return EWOULDBLOCK (sockets, pipes, etc.) cannot be memory mapped") + } + + // If the hostFD would block, we must set it to non-blocking and handle + // blocking behavior in the sentry. + if i.wouldBlock { + if err := syscall.SetNonblock(i.hostFD, true); err != nil { + return nil, err + } + if err := fdnotifier.AddFD(int32(i.hostFD), &i.queue); err != nil { + return nil, err + } + } + + d := &kernfs.Dentry{} + d.Init(i) + + // i.open will take a reference on d. + defer d.DecRef() + + // 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) +} + +// ImportFD sets up and returns a vfs.FileDescription from a donated fd. +func ImportFD(ctx context.Context, mnt *vfs.Mount, hostFD int, isTTY bool) (*vfs.FileDescription, error) { + return NewFD(ctx, mnt, hostFD, &NewFDOptions{ + IsTTY: isTTY, + }) +} + +// filesystemType implements vfs.FilesystemType. +type filesystemType struct{} + +// GetFilesystem implements FilesystemType.GetFilesystem. +func (filesystemType) GetFilesystem(context.Context, *vfs.VirtualFilesystem, *auth.Credentials, string, vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { + panic("host.filesystemType.GetFilesystem should never be called") +} + +// Name implements FilesystemType.Name. +func (filesystemType) Name() string { + return "none" +} + +// NewFilesystem sets up and returns a new hostfs filesystem. +// +// Note that there should only ever be one instance of host.filesystem, +// a global mount for host fds. +func NewFilesystem(vfsObj *vfs.VirtualFilesystem) (*vfs.Filesystem, error) { + devMinor, err := vfsObj.GetAnonBlockDevMinor() + if err != nil { + return nil, err + } + fs := &filesystem{ + devMinor: devMinor, + } + fs.VFSFilesystem().Init(vfsObj, filesystemType{}, fs) + return fs.VFSFilesystem(), nil +} + +// filesystem implements vfs.FilesystemImpl. +type filesystem struct { + kernfs.Filesystem + + devMinor uint32 +} + +func (fs *filesystem) Release() { + fs.VFSFilesystem().VirtualFilesystem().PutAnonBlockDevMinor(fs.devMinor) + fs.Filesystem.Release() +} + +func (fs *filesystem) PrependPath(ctx context.Context, vfsroot, vd vfs.VirtualDentry, b *fspath.Builder) error { + d := vd.Dentry().Impl().(*kernfs.Dentry) + inode := d.Inode().(*inode) + b.PrependComponent(fmt.Sprintf("host:[%d]", inode.ino)) + return vfs.PrependPathSyntheticError{} +} + +// inode implements kernfs.Inode. +type inode struct { + kernfs.InodeNotDirectory + kernfs.InodeNotSymlink + + // When the reference count reaches zero, the host fd is closed. + refs.AtomicRefCount + + // hostFD contains the host fd that this file was originally created from, + // which must be available at time of restore. + // + // This field is initialized at creation time and is immutable. + hostFD int + + // ino is an inode number unique within this filesystem. + // + // This field is initialized at creation time and is immutable. + ino uint64 + + // isTTY is true if this file represents a TTY. + // + // This field is initialized at creation time and is immutable. + isTTY bool + + // seekable is false if the host fd points to a file representing a stream, + // e.g. a socket or a pipe. Such files are not seekable and can return + // EWOULDBLOCK for I/O operations. + // + // This field is initialized at creation time and is immutable. + seekable bool + + // wouldBlock is true if the host FD would return EWOULDBLOCK for + // operations that would block. + // + // This field is initialized at creation time and is immutable. + wouldBlock bool + + // Event queue for blocking operations. + queue waiter.Queue + + // canMap specifies whether we allow the file to be memory mapped. + // + // This field is initialized at creation time and is immutable. + canMap bool + + // mapsMu protects mappings. + mapsMu sync.Mutex + + // If canMap is true, mappings tracks mappings of hostFD into + // memmap.MappingSpaces. + mappings memmap.MappingSet + + // pf implements platform.File for mappings of hostFD. + pf inodePlatformFile +} + +// CheckPermissions implements kernfs.Inode. +func (i *inode) CheckPermissions(ctx context.Context, creds *auth.Credentials, ats vfs.AccessTypes) error { + var s syscall.Stat_t + if err := syscall.Fstat(i.hostFD, &s); err != nil { + return err + } + return vfs.GenericCheckPermissions(creds, ats, linux.FileMode(s.Mode), auth.KUID(s.Uid), auth.KGID(s.Gid)) +} + +// Mode implements kernfs.Inode. +func (i *inode) Mode() linux.FileMode { + var s syscall.Stat_t + if err := syscall.Fstat(i.hostFD, &s); err != nil { + // Retrieving the mode from the host fd using fstat(2) should not fail. + // If the syscall does not succeed, something is fundamentally wrong. + panic(fmt.Sprintf("failed to retrieve mode from host fd %d: %v", i.hostFD, err)) + } + return linux.FileMode(s.Mode) +} + +// Stat implements kernfs.Inode. +func (i *inode) Stat(vfsfs *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) { + if opts.Mask&linux.STATX__RESERVED != 0 { + return linux.Statx{}, syserror.EINVAL + } + if opts.Sync&linux.AT_STATX_SYNC_TYPE == linux.AT_STATX_SYNC_TYPE { + return linux.Statx{}, syserror.EINVAL + } + + fs := vfsfs.Impl().(*filesystem) + + // Limit our host call only to known flags. + mask := opts.Mask & linux.STATX_ALL + var s unix.Statx_t + err := unix.Statx(i.hostFD, "", int(unix.AT_EMPTY_PATH|opts.Sync), int(mask), &s) + if err == syserror.ENOSYS { + // Fallback to fstat(2), if statx(2) is not supported on the host. + // + // TODO(b/151263641): Remove fallback. + return i.fstat(fs) + } + if err != nil { + return linux.Statx{}, err + } + + // Unconditionally fill blksize, attributes, and device numbers, as + // indicated by /include/uapi/linux/stat.h. Inode number is always + // available, since we use our own rather than the host's. + ls := linux.Statx{ + Mask: linux.STATX_INO, + Blksize: s.Blksize, + Attributes: s.Attributes, + Ino: i.ino, + AttributesMask: s.Attributes_mask, + DevMajor: linux.UNNAMED_MAJOR, + DevMinor: fs.devMinor, + } + + // Copy other fields that were returned by the host. RdevMajor/RdevMinor + // are never copied (and therefore left as zero), so as not to expose host + // device numbers. + ls.Mask |= s.Mask & linux.STATX_ALL + if s.Mask&linux.STATX_TYPE != 0 { + ls.Mode |= s.Mode & linux.S_IFMT + } + if s.Mask&linux.STATX_MODE != 0 { + ls.Mode |= s.Mode &^ linux.S_IFMT + } + if s.Mask&linux.STATX_NLINK != 0 { + ls.Nlink = s.Nlink + } + if s.Mask&linux.STATX_UID != 0 { + ls.UID = s.Uid + } + if s.Mask&linux.STATX_GID != 0 { + ls.GID = s.Gid + } + if s.Mask&linux.STATX_ATIME != 0 { + ls.Atime = unixToLinuxStatxTimestamp(s.Atime) + } + if s.Mask&linux.STATX_BTIME != 0 { + ls.Btime = unixToLinuxStatxTimestamp(s.Btime) + } + if s.Mask&linux.STATX_CTIME != 0 { + ls.Ctime = unixToLinuxStatxTimestamp(s.Ctime) + } + if s.Mask&linux.STATX_MTIME != 0 { + ls.Mtime = unixToLinuxStatxTimestamp(s.Mtime) + } + if s.Mask&linux.STATX_SIZE != 0 { + ls.Size = s.Size + } + if s.Mask&linux.STATX_BLOCKS != 0 { + ls.Blocks = s.Blocks + } + + return ls, nil +} + +// fstat is a best-effort fallback for inode.Stat() if the host does not +// support statx(2). +// +// We ignore the mask and sync flags in opts and simply supply +// STATX_BASIC_STATS, as fstat(2) itself does not allow the specification +// of a mask or sync flags. fstat(2) does not provide any metadata +// equivalent to Statx.Attributes, Statx.AttributesMask, or Statx.Btime, so +// those fields remain empty. +func (i *inode) fstat(fs *filesystem) (linux.Statx, error) { + var s unix.Stat_t + if err := unix.Fstat(i.hostFD, &s); err != nil { + return linux.Statx{}, err + } + + // As with inode.Stat(), we always use internal device and inode numbers, + // and never expose the host's represented device numbers. + return linux.Statx{ + Mask: linux.STATX_BASIC_STATS, + Blksize: uint32(s.Blksize), + Nlink: uint32(s.Nlink), + UID: s.Uid, + GID: s.Gid, + Mode: uint16(s.Mode), + Ino: i.ino, + Size: uint64(s.Size), + Blocks: uint64(s.Blocks), + Atime: timespecToStatxTimestamp(s.Atim), + Ctime: timespecToStatxTimestamp(s.Ctim), + Mtime: timespecToStatxTimestamp(s.Mtim), + DevMajor: linux.UNNAMED_MAJOR, + DevMinor: fs.devMinor, + }, nil +} + +// SetStat implements kernfs.Inode. +func (i *inode) SetStat(ctx context.Context, fs *vfs.Filesystem, creds *auth.Credentials, opts vfs.SetStatOptions) error { + s := opts.Stat + + m := s.Mask + if m == 0 { + return nil + } + if m&^(linux.STATX_MODE|linux.STATX_SIZE|linux.STATX_ATIME|linux.STATX_MTIME) != 0 { + return syserror.EPERM + } + var hostStat syscall.Stat_t + if err := syscall.Fstat(i.hostFD, &hostStat); err != nil { + return err + } + if err := vfs.CheckSetStat(ctx, creds, &s, linux.FileMode(hostStat.Mode&linux.PermissionsMask), auth.KUID(hostStat.Uid), auth.KGID(hostStat.Gid)); err != nil { + return err + } + + if m&linux.STATX_MODE != 0 { + if err := syscall.Fchmod(i.hostFD, uint32(s.Mode)); err != nil { + return err + } + } + if m&linux.STATX_SIZE != 0 { + if err := syscall.Ftruncate(i.hostFD, int64(s.Size)); err != nil { + return err + } + oldSize := uint64(hostStat.Size) + if s.Size < oldSize { + oldpgend, _ := usermem.PageRoundUp(oldSize) + newpgend, _ := usermem.PageRoundUp(s.Size) + if oldpgend != newpgend { + i.mapsMu.Lock() + i.mappings.Invalidate(memmap.MappableRange{newpgend, oldpgend}, memmap.InvalidateOpts{ + // Compare Linux's mm/truncate.c:truncate_setsize() => + // truncate_pagecache() => + // mm/memory.c:unmap_mapping_range(evencows=1). + InvalidatePrivate: true, + }) + i.mapsMu.Unlock() + } + } + } + if m&(linux.STATX_ATIME|linux.STATX_MTIME) != 0 { + ts := [2]syscall.Timespec{ + toTimespec(s.Atime, m&linux.STATX_ATIME == 0), + toTimespec(s.Mtime, m&linux.STATX_MTIME == 0), + } + if err := setTimestamps(i.hostFD, &ts); err != nil { + return err + } + } + return nil +} + +// DecRef implements kernfs.Inode. +func (i *inode) DecRef() { + i.AtomicRefCount.DecRefWithDestructor(i.Destroy) +} + +// Destroy implements kernfs.Inode. +func (i *inode) Destroy() { + if i.wouldBlock { + fdnotifier.RemoveFD(int32(i.hostFD)) + } + if err := unix.Close(i.hostFD); err != nil { + log.Warningf("failed to close host fd %d: %v", i.hostFD, err) + } +} + +// Open implements kernfs.Inode. +func (i *inode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.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) +} + +func (i *inode) open(ctx context.Context, d *vfs.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 + } + fileType := s.Mode & linux.FileTypeMask + + // Constrain flags to a subset we can handle. + // TODO(gvisor.dev/issue/1672): implement behavior corresponding to these allowed flags. + flags &= syscall.O_ACCMODE | syscall.O_DIRECT | syscall.O_NONBLOCK | syscall.O_DSYNC | syscall.O_SYNC | syscall.O_APPEND + + if fileType == syscall.S_IFSOCK { + if i.isTTY { + log.Warningf("cannot use host socket fd %d as TTY", i.hostFD) + return nil, syserror.ENOTTY + } + + ep, err := newEndpoint(ctx, i.hostFD, &i.queue) + if err != nil { + return nil, err + } + // Currently, we only allow Unix sockets to be imported. + return unixsocket.NewFileDescription(ep, ep.Type(), flags, mnt, d) + } + + // TODO(gvisor.dev/issue/1672): Whitelist specific file types here, so that + // we don't allow importing arbitrary file types without proper support. + if i.isTTY { + fd := &TTYFileDescription{ + fileDescription: fileDescription{inode: i}, + termios: linux.DefaultSlaveTermios, + } + vfsfd := &fd.vfsfd + if err := vfsfd.Init(fd, flags, mnt, d, &vfs.FileDescriptionOptions{}); err != nil { + return nil, err + } + return vfsfd, nil + } + + fd := &fileDescription{inode: i} + vfsfd := &fd.vfsfd + if err := vfsfd.Init(fd, flags, mnt, d, &vfs.FileDescriptionOptions{}); err != nil { + return nil, err + } + return vfsfd, nil +} + +// fileDescription is embedded by host fd implementations of FileDescriptionImpl. +type fileDescription struct { + vfsfd vfs.FileDescription + vfs.FileDescriptionDefaultImpl + + // inode is vfsfd.Dentry().Impl().(*kernfs.Dentry).Inode().(*inode), but + // cached to reduce indirections and casting. fileDescription does not hold + // a reference on the inode through the inode field (since one is already + // held via the Dentry). + // + // inode is immutable after fileDescription creation. + inode *inode + + // offsetMu protects offset. + offsetMu sync.Mutex + + // offset specifies the current file offset. It is only meaningful when + // inode.seekable is true. + offset int64 +} + +// SetStat implements vfs.FileDescriptionImpl. +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. +func (f *fileDescription) Stat(_ context.Context, opts vfs.StatOptions) (linux.Statx, error) { + return f.inode.Stat(f.vfsfd.Mount().Filesystem(), opts) +} + +// Release implements vfs.FileDescriptionImpl. +func (f *fileDescription) Release() { + // noop +} + +// PRead implements FileDescriptionImpl. +func (f *fileDescription) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { + i := f.inode + if !i.seekable { + return 0, syserror.ESPIPE + } + + return readFromHostFD(ctx, i.hostFD, dst, offset, opts.Flags) +} + +// Read implements FileDescriptionImpl. +func (f *fileDescription) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { + i := f.inode + if !i.seekable { + n, err := readFromHostFD(ctx, i.hostFD, dst, -1, opts.Flags) + if isBlockError(err) { + // If we got any data at all, return it as a "completed" partial read + // rather than retrying until complete. + if n != 0 { + err = nil + } else { + err = syserror.ErrWouldBlock + } + } + return n, err + } + // TODO(gvisor.dev/issue/1672): Cache pages, when forced to do so. + f.offsetMu.Lock() + n, err := readFromHostFD(ctx, i.hostFD, dst, f.offset, opts.Flags) + f.offset += n + f.offsetMu.Unlock() + return n, err +} + +func readFromHostFD(ctx context.Context, hostFD int, dst usermem.IOSequence, offset int64, flags uint32) (int64, error) { + // TODO(gvisor.dev/issue/1672): Support select preadv2 flags. + if flags != 0 { + return 0, syserror.EOPNOTSUPP + } + reader := hostfd.GetReadWriterAt(int32(hostFD), offset, flags) + n, err := dst.CopyOutFrom(ctx, reader) + hostfd.PutReadWriterAt(reader) + return int64(n), err +} + +// PWrite implements FileDescriptionImpl. +func (f *fileDescription) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) { + i := f.inode + if !i.seekable { + return 0, syserror.ESPIPE + } + + return writeToHostFD(ctx, i.hostFD, src, offset, opts.Flags) +} + +// Write implements FileDescriptionImpl. +func (f *fileDescription) Write(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) { + i := f.inode + if !i.seekable { + n, err := writeToHostFD(ctx, i.hostFD, src, -1, opts.Flags) + if isBlockError(err) { + err = syserror.ErrWouldBlock + } + return n, err + } + // TODO(gvisor.dev/issue/1672): Cache pages, when forced to do so. + // TODO(gvisor.dev/issue/1672): Write to end of file and update offset if O_APPEND is set on this file. + f.offsetMu.Lock() + n, err := writeToHostFD(ctx, i.hostFD, src, f.offset, opts.Flags) + f.offset += n + f.offsetMu.Unlock() + return n, err +} + +func writeToHostFD(ctx context.Context, hostFD int, src usermem.IOSequence, offset int64, flags uint32) (int64, error) { + // TODO(gvisor.dev/issue/1672): Support select pwritev2 flags. + if flags != 0 { + return 0, syserror.EOPNOTSUPP + } + writer := hostfd.GetReadWriterAt(int32(hostFD), offset, flags) + n, err := src.CopyInTo(ctx, writer) + hostfd.PutReadWriterAt(writer) + return int64(n), err +} + +// Seek implements FileDescriptionImpl. +// +// Note that we do not support seeking on directories, since we do not even +// allow directory fds to be imported at all. +func (f *fileDescription) Seek(_ context.Context, offset int64, whence int32) (int64, error) { + i := f.inode + if !i.seekable { + return 0, syserror.ESPIPE + } + + f.offsetMu.Lock() + defer f.offsetMu.Unlock() + + switch whence { + case linux.SEEK_SET: + if offset < 0 { + return f.offset, syserror.EINVAL + } + f.offset = offset + + case linux.SEEK_CUR: + // Check for overflow. Note that underflow cannot occur, since f.offset >= 0. + if offset > math.MaxInt64-f.offset { + return f.offset, syserror.EOVERFLOW + } + if f.offset+offset < 0 { + return f.offset, syserror.EINVAL + } + f.offset += offset + + case linux.SEEK_END: + var s syscall.Stat_t + if err := syscall.Fstat(i.hostFD, &s); err != nil { + return f.offset, err + } + size := s.Size + + // Check for overflow. Note that underflow cannot occur, since size >= 0. + if offset > math.MaxInt64-size { + return f.offset, syserror.EOVERFLOW + } + if size+offset < 0 { + return f.offset, syserror.EINVAL + } + f.offset = size + offset + + case linux.SEEK_DATA, linux.SEEK_HOLE: + // Modifying the offset in the host file table should not matter, since + // this is the only place where we use it. + // + // For reading and writing, we always rely on our internal offset. + n, err := unix.Seek(i.hostFD, offset, int(whence)) + if err != nil { + return f.offset, err + } + f.offset = n + + default: + // Invalid whence. + return f.offset, syserror.EINVAL + } + + return f.offset, nil +} + +// Sync implements FileDescriptionImpl. +func (f *fileDescription) Sync(context.Context) error { + // TODO(gvisor.dev/issue/1672): Currently we do not support the SyncData optimization, so we always sync everything. + return unix.Fsync(f.inode.hostFD) +} + +// ConfigureMMap implements FileDescriptionImpl. +func (f *fileDescription) ConfigureMMap(_ context.Context, opts *memmap.MMapOpts) error { + if !f.inode.canMap { + return syserror.ENODEV + } + i := f.inode + i.pf.fileMapperInitOnce.Do(i.pf.fileMapper.Init) + return vfs.GenericConfigureMMap(&f.vfsfd, i, opts) +} + +// EventRegister implements waiter.Waitable.EventRegister. +func (f *fileDescription) EventRegister(e *waiter.Entry, mask waiter.EventMask) { + f.inode.queue.EventRegister(e, mask) + fdnotifier.UpdateFD(int32(f.inode.hostFD)) +} + +// EventUnregister implements waiter.Waitable.EventUnregister. +func (f *fileDescription) EventUnregister(e *waiter.Entry) { + f.inode.queue.EventUnregister(e) + fdnotifier.UpdateFD(int32(f.inode.hostFD)) +} + +// Readiness uses the poll() syscall to check the status of the underlying FD. +func (f *fileDescription) Readiness(mask waiter.EventMask) waiter.EventMask { + return fdnotifier.NonBlockingPoll(int32(f.inode.hostFD), mask) +} diff --git a/pkg/sentry/fsimpl/host/ioctl_unsafe.go b/pkg/sentry/fsimpl/host/ioctl_unsafe.go new file mode 100644 index 000000000..0983bf7d8 --- /dev/null +++ b/pkg/sentry/fsimpl/host/ioctl_unsafe.go @@ -0,0 +1,56 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package host + +import ( + "syscall" + "unsafe" + + "gvisor.dev/gvisor/pkg/abi/linux" +) + +func ioctlGetTermios(fd int) (*linux.Termios, error) { + var t linux.Termios + _, _, errno := syscall.Syscall(syscall.SYS_IOCTL, uintptr(fd), linux.TCGETS, uintptr(unsafe.Pointer(&t))) + if errno != 0 { + return nil, errno + } + return &t, nil +} + +func ioctlSetTermios(fd int, req uint64, t *linux.Termios) error { + _, _, errno := syscall.Syscall(syscall.SYS_IOCTL, uintptr(fd), uintptr(req), uintptr(unsafe.Pointer(t))) + if errno != 0 { + return errno + } + return nil +} + +func ioctlGetWinsize(fd int) (*linux.Winsize, error) { + var w linux.Winsize + _, _, errno := syscall.Syscall(syscall.SYS_IOCTL, uintptr(fd), linux.TIOCGWINSZ, uintptr(unsafe.Pointer(&w))) + if errno != 0 { + return nil, errno + } + return &w, nil +} + +func ioctlSetWinsize(fd int, w *linux.Winsize) error { + _, _, errno := syscall.Syscall(syscall.SYS_IOCTL, uintptr(fd), linux.TIOCSWINSZ, uintptr(unsafe.Pointer(w))) + if errno != 0 { + return errno + } + return nil +} diff --git a/pkg/sentry/fsimpl/host/mmap.go b/pkg/sentry/fsimpl/host/mmap.go new file mode 100644 index 000000000..8545a82f0 --- /dev/null +++ b/pkg/sentry/fsimpl/host/mmap.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 host + +import ( + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/safemem" + "gvisor.dev/gvisor/pkg/sentry/fs/fsutil" + "gvisor.dev/gvisor/pkg/sentry/memmap" + "gvisor.dev/gvisor/pkg/sentry/platform" + "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/pkg/usermem" +) + +// inodePlatformFile implements platform.File. It exists solely because inode +// cannot implement both kernfs.Inode.IncRef and platform.File.IncRef. +// +// inodePlatformFile should only be used if inode.canMap is true. +type inodePlatformFile struct { + *inode + + // fdRefsMu protects fdRefs. + fdRefsMu sync.Mutex + + // fdRefs counts references on platform.File offsets. It is used solely for + // memory accounting. + fdRefs fsutil.FrameRefSet + + // fileMapper caches mappings of the host file represented by this inode. + fileMapper fsutil.HostFileMapper + + // fileMapperInitOnce is used to lazily initialize fileMapper. + fileMapperInitOnce sync.Once +} + +// IncRef implements platform.File.IncRef. +// +// Precondition: i.inode.canMap must be true. +func (i *inodePlatformFile) IncRef(fr platform.FileRange) { + i.fdRefsMu.Lock() + i.fdRefs.IncRefAndAccount(fr) + i.fdRefsMu.Unlock() +} + +// DecRef implements platform.File.DecRef. +// +// Precondition: i.inode.canMap must be true. +func (i *inodePlatformFile) DecRef(fr platform.FileRange) { + i.fdRefsMu.Lock() + i.fdRefs.DecRefAndAccount(fr) + i.fdRefsMu.Unlock() +} + +// MapInternal implements platform.File.MapInternal. +// +// Precondition: i.inode.canMap must be true. +func (i *inodePlatformFile) MapInternal(fr platform.FileRange, at usermem.AccessType) (safemem.BlockSeq, error) { + return i.fileMapper.MapInternal(fr, i.hostFD, at.Write) +} + +// FD implements platform.File.FD. +func (i *inodePlatformFile) FD() int { + return i.hostFD +} + +// AddMapping implements memmap.Mappable.AddMapping. +// +// Precondition: i.inode.canMap must be true. +func (i *inode) AddMapping(ctx context.Context, ms memmap.MappingSpace, ar usermem.AddrRange, offset uint64, writable bool) error { + i.mapsMu.Lock() + mapped := i.mappings.AddMapping(ms, ar, offset, writable) + for _, r := range mapped { + i.pf.fileMapper.IncRefOn(r) + } + i.mapsMu.Unlock() + return nil +} + +// RemoveMapping implements memmap.Mappable.RemoveMapping. +// +// Precondition: i.inode.canMap must be true. +func (i *inode) RemoveMapping(ctx context.Context, ms memmap.MappingSpace, ar usermem.AddrRange, offset uint64, writable bool) { + i.mapsMu.Lock() + unmapped := i.mappings.RemoveMapping(ms, ar, offset, writable) + for _, r := range unmapped { + i.pf.fileMapper.DecRefOn(r) + } + i.mapsMu.Unlock() +} + +// CopyMapping implements memmap.Mappable.CopyMapping. +// +// Precondition: i.inode.canMap must be true. +func (i *inode) CopyMapping(ctx context.Context, ms memmap.MappingSpace, srcAR, dstAR usermem.AddrRange, offset uint64, writable bool) error { + return i.AddMapping(ctx, ms, dstAR, offset, writable) +} + +// Translate implements memmap.Mappable.Translate. +// +// Precondition: i.inode.canMap must be true. +func (i *inode) Translate(ctx context.Context, required, optional memmap.MappableRange, at usermem.AccessType) ([]memmap.Translation, error) { + mr := optional + return []memmap.Translation{ + { + Source: mr, + File: &i.pf, + Offset: mr.Start, + Perms: usermem.AnyAccess, + }, + }, nil +} + +// InvalidateUnsavable implements memmap.Mappable.InvalidateUnsavable. +// +// Precondition: i.inode.canMap must be true. +func (i *inode) InvalidateUnsavable(ctx context.Context) error { + // We expect the same host fd across save/restore, so all translations + // should be valid. + return nil +} diff --git a/pkg/sentry/fsimpl/host/socket.go b/pkg/sentry/fsimpl/host/socket.go new file mode 100644 index 000000000..38f1fbfba --- /dev/null +++ b/pkg/sentry/fsimpl/host/socket.go @@ -0,0 +1,394 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package host + +import ( + "fmt" + "syscall" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/fdnotifier" + "gvisor.dev/gvisor/pkg/log" + "gvisor.dev/gvisor/pkg/refs" + "gvisor.dev/gvisor/pkg/sentry/socket/control" + "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" + "gvisor.dev/gvisor/pkg/sentry/uniqueid" + "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/pkg/syserr" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/tcpip" + "gvisor.dev/gvisor/pkg/unet" + "gvisor.dev/gvisor/pkg/waiter" +) + +// Create a new host-backed endpoint from the given fd and its corresponding +// notification queue. +func newEndpoint(ctx context.Context, hostFD int, queue *waiter.Queue) (transport.Endpoint, error) { + // Set up an external transport.Endpoint using the host fd. + addr := fmt.Sprintf("hostfd:[%d]", hostFD) + e, err := NewConnectedEndpoint(ctx, hostFD, addr, true /* saveable */) + if err != nil { + return nil, err.ToError() + } + ep := transport.NewExternal(ctx, e.stype, uniqueid.GlobalProviderFromContext(ctx), queue, e, e) + return ep, nil +} + +// maxSendBufferSize is the maximum host send buffer size allowed for endpoint. +// +// N.B. 8MB is the default maximum on Linux (2 * sysctl_wmem_max). +const maxSendBufferSize = 8 << 20 + +// ConnectedEndpoint is an implementation of transport.ConnectedEndpoint and +// transport.Receiver. It is backed by a host fd that was imported at sentry +// startup. This fd is shared with a hostfs inode, which retains ownership of +// it. +// +// ConnectedEndpoint is saveable, since we expect that the host will provide +// the same fd upon restore. +// +// As of this writing, we only allow Unix sockets to be imported. +// +// +stateify savable +type ConnectedEndpoint struct { + // ref keeps track of references to a ConnectedEndpoint. + ref refs.AtomicRefCount + + // mu protects fd below. + mu sync.RWMutex `state:"nosave"` + + // fd is the host fd backing this endpoint. + fd int + + // addr is the address at which this endpoint is bound. + addr string + + // sndbuf is the size of the send buffer. + // + // N.B. When this is smaller than the host size, we present it via + // GetSockOpt and message splitting/rejection in SendMsg, but do not + // prevent lots of small messages from filling the real send buffer + // size on the host. + sndbuf int64 `state:"nosave"` + + // stype is the type of Unix socket. + stype linux.SockType +} + +// init performs initialization required for creating new ConnectedEndpoints and +// for restoring them. +func (c *ConnectedEndpoint) init() *syserr.Error { + family, err := syscall.GetsockoptInt(c.fd, syscall.SOL_SOCKET, syscall.SO_DOMAIN) + if err != nil { + return syserr.FromError(err) + } + + if family != syscall.AF_UNIX { + // We only allow Unix sockets. + return syserr.ErrInvalidEndpointState + } + + stype, err := syscall.GetsockoptInt(c.fd, syscall.SOL_SOCKET, syscall.SO_TYPE) + if err != nil { + return syserr.FromError(err) + } + + if err := syscall.SetNonblock(c.fd, true); err != nil { + return syserr.FromError(err) + } + + sndbuf, err := syscall.GetsockoptInt(c.fd, syscall.SOL_SOCKET, syscall.SO_SNDBUF) + if err != nil { + return syserr.FromError(err) + } + if sndbuf > maxSendBufferSize { + log.Warningf("Socket send buffer too large: %d", sndbuf) + return syserr.ErrInvalidEndpointState + } + + c.stype = linux.SockType(stype) + c.sndbuf = int64(sndbuf) + + return nil +} + +// NewConnectedEndpoint creates a new ConnectedEndpoint backed by a host fd +// imported at sentry startup, +// +// The caller is responsible for calling Init(). Additionaly, Release needs to +// be called twice because ConnectedEndpoint is both a transport.Receiver and +// transport.ConnectedEndpoint. +func NewConnectedEndpoint(ctx context.Context, hostFD int, addr string, saveable bool) (*ConnectedEndpoint, *syserr.Error) { + e := ConnectedEndpoint{ + fd: hostFD, + addr: addr, + } + + if err := e.init(); err != nil { + return nil, err + } + + // AtomicRefCounters start off with a single reference. We need two. + e.ref.IncRef() + e.ref.EnableLeakCheck("host.ConnectedEndpoint") + return &e, nil +} + +// Send implements transport.ConnectedEndpoint.Send. +func (c *ConnectedEndpoint) Send(data [][]byte, controlMessages transport.ControlMessages, from tcpip.FullAddress) (int64, bool, *syserr.Error) { + c.mu.RLock() + defer c.mu.RUnlock() + + if !controlMessages.Empty() { + return 0, false, syserr.ErrInvalidEndpointState + } + + // Since stream sockets don't preserve message boundaries, we can write + // only as much of the message as fits in the send buffer. + truncate := c.stype == linux.SOCK_STREAM + + n, totalLen, err := fdWriteVec(c.fd, data, c.sndbuf, truncate) + if n < totalLen && err == nil { + // The host only returns a short write if it would otherwise + // block (and only for stream sockets). + err = syserror.EAGAIN + } + if n > 0 && err != syserror.EAGAIN { + // The caller may need to block to send more data, but + // otherwise there isn't anything that can be done about an + // error with a partial write. + err = nil + } + + // There is no need for the callee to call SendNotify because fdWriteVec + // uses the host's sendmsg(2) and the host kernel's queue. + return n, false, syserr.FromError(err) +} + +// SendNotify implements transport.ConnectedEndpoint.SendNotify. +func (c *ConnectedEndpoint) SendNotify() {} + +// CloseSend implements transport.ConnectedEndpoint.CloseSend. +func (c *ConnectedEndpoint) CloseSend() { + c.mu.Lock() + defer c.mu.Unlock() + + if err := syscall.Shutdown(c.fd, syscall.SHUT_WR); err != nil { + // A well-formed UDS shutdown can't fail. See + // net/unix/af_unix.c:unix_shutdown. + panic(fmt.Sprintf("failed write shutdown on host socket %+v: %v", c, err)) + } +} + +// CloseNotify implements transport.ConnectedEndpoint.CloseNotify. +func (c *ConnectedEndpoint) CloseNotify() {} + +// Writable implements transport.ConnectedEndpoint.Writable. +func (c *ConnectedEndpoint) Writable() bool { + c.mu.RLock() + defer c.mu.RUnlock() + + return fdnotifier.NonBlockingPoll(int32(c.fd), waiter.EventOut)&waiter.EventOut != 0 +} + +// Passcred implements transport.ConnectedEndpoint.Passcred. +func (c *ConnectedEndpoint) Passcred() bool { + // We don't support credential passing for host sockets. + return false +} + +// GetLocalAddress implements transport.ConnectedEndpoint.GetLocalAddress. +func (c *ConnectedEndpoint) GetLocalAddress() (tcpip.FullAddress, *tcpip.Error) { + return tcpip.FullAddress{Addr: tcpip.Address(c.addr)}, nil +} + +// EventUpdate implements transport.ConnectedEndpoint.EventUpdate. +func (c *ConnectedEndpoint) EventUpdate() { + c.mu.RLock() + defer c.mu.RUnlock() + if c.fd != -1 { + fdnotifier.UpdateFD(int32(c.fd)) + } +} + +// Recv implements transport.Receiver.Recv. +func (c *ConnectedEndpoint) Recv(data [][]byte, creds bool, numRights int, peek bool) (int64, int64, transport.ControlMessages, bool, tcpip.FullAddress, bool, *syserr.Error) { + c.mu.RLock() + defer c.mu.RUnlock() + + var cm unet.ControlMessage + if numRights > 0 { + cm.EnableFDs(int(numRights)) + } + + // N.B. Unix sockets don't have a receive buffer, the send buffer + // serves both purposes. + rl, ml, cl, cTrunc, err := fdReadVec(c.fd, data, []byte(cm), peek, c.sndbuf) + if rl > 0 && err != nil { + // We got some data, so all we need to do on error is return + // the data that we got. Short reads are fine, no need to + // block. + err = nil + } + if err != nil { + return 0, 0, transport.ControlMessages{}, false, tcpip.FullAddress{}, false, syserr.FromError(err) + } + + // There is no need for the callee to call RecvNotify because fdReadVec uses + // the host's recvmsg(2) and the host kernel's queue. + + // Trim the control data if we received less than the full amount. + if cl < uint64(len(cm)) { + cm = cm[:cl] + } + + // Avoid extra allocations in the case where there isn't any control data. + if len(cm) == 0 { + return rl, ml, transport.ControlMessages{}, cTrunc, tcpip.FullAddress{Addr: tcpip.Address(c.addr)}, false, nil + } + + fds, err := cm.ExtractFDs() + if err != nil { + return 0, 0, transport.ControlMessages{}, false, tcpip.FullAddress{}, false, syserr.FromError(err) + } + + if len(fds) == 0 { + return rl, ml, transport.ControlMessages{}, cTrunc, tcpip.FullAddress{Addr: tcpip.Address(c.addr)}, false, nil + } + return rl, ml, control.NewVFS2(nil, nil, newSCMRights(fds)), cTrunc, tcpip.FullAddress{Addr: tcpip.Address(c.addr)}, false, nil +} + +// RecvNotify implements transport.Receiver.RecvNotify. +func (c *ConnectedEndpoint) RecvNotify() {} + +// CloseRecv implements transport.Receiver.CloseRecv. +func (c *ConnectedEndpoint) CloseRecv() { + c.mu.Lock() + defer c.mu.Unlock() + + if err := syscall.Shutdown(c.fd, syscall.SHUT_RD); err != nil { + // A well-formed UDS shutdown can't fail. See + // net/unix/af_unix.c:unix_shutdown. + panic(fmt.Sprintf("failed read shutdown on host socket %+v: %v", c, err)) + } +} + +// Readable implements transport.Receiver.Readable. +func (c *ConnectedEndpoint) Readable() bool { + c.mu.RLock() + defer c.mu.RUnlock() + + return fdnotifier.NonBlockingPoll(int32(c.fd), waiter.EventIn)&waiter.EventIn != 0 +} + +// SendQueuedSize implements transport.Receiver.SendQueuedSize. +func (c *ConnectedEndpoint) SendQueuedSize() int64 { + // TODO(gvisor.dev/issue/273): SendQueuedSize isn't supported for host + // sockets because we don't allow the sentry to call ioctl(2). + return -1 +} + +// RecvQueuedSize implements transport.Receiver.RecvQueuedSize. +func (c *ConnectedEndpoint) RecvQueuedSize() int64 { + // TODO(gvisor.dev/issue/273): RecvQueuedSize isn't supported for host + // sockets because we don't allow the sentry to call ioctl(2). + return -1 +} + +// SendMaxQueueSize implements transport.Receiver.SendMaxQueueSize. +func (c *ConnectedEndpoint) SendMaxQueueSize() int64 { + return int64(c.sndbuf) +} + +// RecvMaxQueueSize implements transport.Receiver.RecvMaxQueueSize. +func (c *ConnectedEndpoint) RecvMaxQueueSize() int64 { + // N.B. Unix sockets don't use the receive buffer. We'll claim it is + // the same size as the send buffer. + return int64(c.sndbuf) +} + +func (c *ConnectedEndpoint) destroyLocked() { + c.fd = -1 +} + +// Release implements transport.ConnectedEndpoint.Release and +// transport.Receiver.Release. +func (c *ConnectedEndpoint) Release() { + c.ref.DecRefWithDestructor(func() { + c.mu.Lock() + c.destroyLocked() + c.mu.Unlock() + }) +} + +// CloseUnread implements transport.ConnectedEndpoint.CloseUnread. +func (c *ConnectedEndpoint) CloseUnread() {} + +// SCMConnectedEndpoint represents an endpoint backed by a host fd that was +// passed through a gofer Unix socket. It resembles ConnectedEndpoint, with the +// following differences: +// - SCMConnectedEndpoint is not saveable, because the host cannot guarantee +// the same descriptor number across S/R. +// - SCMConnectedEndpoint holds ownership of its fd and notification queue. +type SCMConnectedEndpoint struct { + ConnectedEndpoint + + queue *waiter.Queue +} + +// Init will do the initialization required without holding other locks. +func (e *SCMConnectedEndpoint) Init() error { + return fdnotifier.AddFD(int32(e.fd), e.queue) +} + +// Release implements transport.ConnectedEndpoint.Release and +// transport.Receiver.Release. +func (e *SCMConnectedEndpoint) Release() { + e.ref.DecRefWithDestructor(func() { + e.mu.Lock() + 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() + }) +} + +// NewSCMEndpoint creates a new SCMConnectedEndpoint backed by a host fd that +// was passed through a Unix socket. +// +// The caller is responsible for calling Init(). Additionaly, Release needs to +// be called twice because ConnectedEndpoint is both a transport.Receiver and +// transport.ConnectedEndpoint. +func NewSCMEndpoint(ctx context.Context, hostFD int, queue *waiter.Queue, addr string) (*SCMConnectedEndpoint, *syserr.Error) { + e := SCMConnectedEndpoint{ + ConnectedEndpoint: ConnectedEndpoint{ + fd: hostFD, + addr: addr, + }, + queue: queue, + } + + if err := e.init(); err != nil { + return nil, err + } + + // AtomicRefCounters start off with a single reference. We need two. + e.ref.IncRef() + e.ref.EnableLeakCheck("host.SCMConnectedEndpoint") + return &e, nil +} diff --git a/pkg/sentry/fsimpl/host/socket_iovec.go b/pkg/sentry/fsimpl/host/socket_iovec.go new file mode 100644 index 000000000..584c247d2 --- /dev/null +++ b/pkg/sentry/fsimpl/host/socket_iovec.go @@ -0,0 +1,113 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package host + +import ( + "syscall" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/syserror" +) + +// maxIovs is the maximum number of iovecs to pass to the host. +var maxIovs = linux.UIO_MAXIOV + +// copyToMulti copies as many bytes from src to dst as possible. +func copyToMulti(dst [][]byte, src []byte) { + for _, d := range dst { + done := copy(d, src) + src = src[done:] + if len(src) == 0 { + break + } + } +} + +// copyFromMulti copies as many bytes from src to dst as possible. +func copyFromMulti(dst []byte, src [][]byte) { + for _, s := range src { + done := copy(dst, s) + dst = dst[done:] + if len(dst) == 0 { + break + } + } +} + +// buildIovec builds an iovec slice from the given []byte slice. +// +// If truncate, truncate bufs > maxlen. Otherwise, immediately return an error. +// +// If length < the total length of bufs, err indicates why, even when returning +// a truncated iovec. +// +// If intermediate != nil, iovecs references intermediate rather than bufs and +// the caller must copy to/from bufs as necessary. +func buildIovec(bufs [][]byte, maxlen int64, truncate bool) (length int64, iovecs []syscall.Iovec, intermediate []byte, err error) { + var iovsRequired int + for _, b := range bufs { + length += int64(len(b)) + if len(b) > 0 { + iovsRequired++ + } + } + + stopLen := length + if length > maxlen { + if truncate { + stopLen = maxlen + err = syserror.EAGAIN + } else { + return 0, nil, nil, syserror.EMSGSIZE + } + } + + if iovsRequired > maxIovs { + // The kernel will reject our call if we pass this many iovs. + // Use a single intermediate buffer instead. + b := make([]byte, stopLen) + + return stopLen, []syscall.Iovec{{ + Base: &b[0], + Len: uint64(stopLen), + }}, b, err + } + + var total int64 + iovecs = make([]syscall.Iovec, 0, iovsRequired) + for i := range bufs { + l := len(bufs[i]) + if l == 0 { + continue + } + + stop := int64(l) + if total+stop > stopLen { + stop = stopLen - total + } + + iovecs = append(iovecs, syscall.Iovec{ + Base: &bufs[i][0], + Len: uint64(stop), + }) + + total += stop + if total >= stopLen { + break + } + } + + return total, iovecs, nil, err +} diff --git a/pkg/sentry/fsimpl/host/socket_unsafe.go b/pkg/sentry/fsimpl/host/socket_unsafe.go new file mode 100644 index 000000000..35ded24bc --- /dev/null +++ b/pkg/sentry/fsimpl/host/socket_unsafe.go @@ -0,0 +1,101 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package host + +import ( + "syscall" + "unsafe" +) + +// fdReadVec receives from fd to bufs. +// +// If the total length of bufs is > maxlen, fdReadVec will do a partial read +// and err will indicate why the message was truncated. +func fdReadVec(fd int, bufs [][]byte, control []byte, peek bool, maxlen int64) (readLen int64, msgLen int64, controlLen uint64, controlTrunc bool, err error) { + flags := uintptr(syscall.MSG_DONTWAIT | syscall.MSG_TRUNC) + if peek { + flags |= syscall.MSG_PEEK + } + + // Always truncate the receive buffer. All socket types will truncate + // received messages. + length, iovecs, intermediate, err := buildIovec(bufs, maxlen, true) + if err != nil && len(iovecs) == 0 { + // No partial write to do, return error immediately. + return 0, 0, 0, false, err + } + + var msg syscall.Msghdr + if len(control) != 0 { + msg.Control = &control[0] + msg.Controllen = uint64(len(control)) + } + + if len(iovecs) != 0 { + msg.Iov = &iovecs[0] + msg.Iovlen = uint64(len(iovecs)) + } + + rawN, _, e := syscall.RawSyscall(syscall.SYS_RECVMSG, uintptr(fd), uintptr(unsafe.Pointer(&msg)), flags) + if e != 0 { + // N.B. prioritize the syscall error over the buildIovec error. + return 0, 0, 0, false, e + } + n := int64(rawN) + + // Copy data back to bufs. + if intermediate != nil { + copyToMulti(bufs, intermediate) + } + + controlTrunc = msg.Flags&syscall.MSG_CTRUNC == syscall.MSG_CTRUNC + + if n > length { + return length, n, msg.Controllen, controlTrunc, err + } + + return n, n, msg.Controllen, controlTrunc, err +} + +// fdWriteVec sends from bufs to fd. +// +// If the total length of bufs is > maxlen && truncate, fdWriteVec will do a +// partial write and err will indicate why the message was truncated. +func fdWriteVec(fd int, bufs [][]byte, maxlen int64, truncate bool) (int64, int64, error) { + length, iovecs, intermediate, err := buildIovec(bufs, maxlen, truncate) + if err != nil && len(iovecs) == 0 { + // No partial write to do, return error immediately. + return 0, length, err + } + + // Copy data to intermediate buf. + if intermediate != nil { + copyFromMulti(intermediate, bufs) + } + + var msg syscall.Msghdr + if len(iovecs) > 0 { + msg.Iov = &iovecs[0] + msg.Iovlen = uint64(len(iovecs)) + } + + n, _, e := syscall.RawSyscall(syscall.SYS_SENDMSG, uintptr(fd), uintptr(unsafe.Pointer(&msg)), syscall.MSG_DONTWAIT|syscall.MSG_NOSIGNAL) + if e != 0 { + // N.B. prioritize the syscall error over the buildIovec error. + return 0, length, e + } + + return int64(n), length, err +} diff --git a/pkg/sentry/fsimpl/host/tty.go b/pkg/sentry/fsimpl/host/tty.go new file mode 100644 index 000000000..68af6e5af --- /dev/null +++ b/pkg/sentry/fsimpl/host/tty.go @@ -0,0 +1,379 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package host + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/unimpl" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +// TTYFileDescription implements vfs.FileDescriptionImpl for a host file +// descriptor that wraps a TTY FD. +type TTYFileDescription struct { + fileDescription + + // mu protects the fields below. + mu sync.Mutex `state:"nosave"` + + // session is the session attached to this TTYFileDescription. + session *kernel.Session + + // fgProcessGroup is the foreground process group that is currently + // connected to this TTY. + fgProcessGroup *kernel.ProcessGroup + + // termios contains the terminal attributes for this TTY. + termios linux.KernelTermios +} + +// InitForegroundProcessGroup sets the foreground process group and session for +// the TTY. This should only be called once, after the foreground process group +// has been created, but before it has started running. +func (t *TTYFileDescription) InitForegroundProcessGroup(pg *kernel.ProcessGroup) { + t.mu.Lock() + defer t.mu.Unlock() + if t.fgProcessGroup != nil { + panic("foreground process group is already set") + } + t.fgProcessGroup = pg + t.session = pg.Session() +} + +// ForegroundProcessGroup returns the foreground process for the TTY. +func (t *TTYFileDescription) ForegroundProcessGroup() *kernel.ProcessGroup { + t.mu.Lock() + defer t.mu.Unlock() + return t.fgProcessGroup +} + +// Release implements fs.FileOperations.Release. +func (t *TTYFileDescription) Release() { + t.mu.Lock() + t.fgProcessGroup = nil + t.mu.Unlock() + + t.fileDescription.Release() +} + +// PRead implements vfs.FileDescriptionImpl. +// +// Reading from a TTY is only allowed for foreground process groups. Background +// process groups will either get EIO or a SIGTTIN. +func (t *TTYFileDescription) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { + t.mu.Lock() + defer t.mu.Unlock() + + // Are we allowed to do the read? + // drivers/tty/n_tty.c:n_tty_read()=>job_control()=>tty_check_change(). + if err := t.checkChange(ctx, linux.SIGTTIN); err != nil { + return 0, err + } + + // Do the read. + return t.fileDescription.PRead(ctx, dst, offset, opts) +} + +// Read implements vfs.FileDescriptionImpl. +// +// Reading from a TTY is only allowed for foreground process groups. Background +// process groups will either get EIO or a SIGTTIN. +func (t *TTYFileDescription) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { + t.mu.Lock() + defer t.mu.Unlock() + + // Are we allowed to do the read? + // drivers/tty/n_tty.c:n_tty_read()=>job_control()=>tty_check_change(). + if err := t.checkChange(ctx, linux.SIGTTIN); err != nil { + return 0, err + } + + // Do the read. + return t.fileDescription.Read(ctx, dst, opts) +} + +// PWrite implements vfs.FileDescriptionImpl. +func (t *TTYFileDescription) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) { + t.mu.Lock() + defer t.mu.Unlock() + + // Check whether TOSTOP is enabled. This corresponds to the check in + // drivers/tty/n_tty.c:n_tty_write(). + if t.termios.LEnabled(linux.TOSTOP) { + if err := t.checkChange(ctx, linux.SIGTTOU); err != nil { + return 0, err + } + } + return t.fileDescription.PWrite(ctx, src, offset, opts) +} + +// Write implements vfs.FileDescriptionImpl. +func (t *TTYFileDescription) Write(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) { + t.mu.Lock() + defer t.mu.Unlock() + + // Check whether TOSTOP is enabled. This corresponds to the check in + // drivers/tty/n_tty.c:n_tty_write(). + if t.termios.LEnabled(linux.TOSTOP) { + if err := t.checkChange(ctx, linux.SIGTTOU); err != nil { + return 0, err + } + } + return t.fileDescription.Write(ctx, src, opts) +} + +// Ioctl implements vfs.FileDescriptionImpl. +func (t *TTYFileDescription) Ioctl(ctx context.Context, io usermem.IO, args arch.SyscallArguments) (uintptr, error) { + // Ignore arg[0]. This is the real FD: + fd := t.inode.hostFD + ioctl := args[1].Uint64() + switch ioctl { + case linux.TCGETS: + termios, err := ioctlGetTermios(fd) + if err != nil { + return 0, err + } + _, err = usermem.CopyObjectOut(ctx, io, args[2].Pointer(), termios, usermem.IOOpts{ + AddressSpaceActive: true, + }) + return 0, err + + case linux.TCSETS, linux.TCSETSW, linux.TCSETSF: + t.mu.Lock() + defer t.mu.Unlock() + + if err := t.checkChange(ctx, linux.SIGTTOU); err != nil { + return 0, err + } + + var termios linux.Termios + if _, err := usermem.CopyObjectIn(ctx, io, args[2].Pointer(), &termios, usermem.IOOpts{ + AddressSpaceActive: true, + }); err != nil { + return 0, err + } + err := ioctlSetTermios(fd, ioctl, &termios) + if err == nil { + t.termios.FromTermios(termios) + } + return 0, err + + case linux.TIOCGPGRP: + // Args: pid_t *argp + // When successful, equivalent to *argp = tcgetpgrp(fd). + // Get the process group ID of the foreground process group on this + // terminal. + + pidns := kernel.PIDNamespaceFromContext(ctx) + if pidns == nil { + return 0, syserror.ENOTTY + } + + t.mu.Lock() + defer t.mu.Unlock() + + // Map the ProcessGroup into a ProcessGroupID in the task's PID namespace. + pgID := pidns.IDOfProcessGroup(t.fgProcessGroup) + _, err := usermem.CopyObjectOut(ctx, io, args[2].Pointer(), &pgID, usermem.IOOpts{ + AddressSpaceActive: true, + }) + return 0, err + + case linux.TIOCSPGRP: + // Args: const pid_t *argp + // Equivalent to tcsetpgrp(fd, *argp). + // Set the foreground process group ID of this terminal. + + task := kernel.TaskFromContext(ctx) + if task == nil { + return 0, syserror.ENOTTY + } + + t.mu.Lock() + defer t.mu.Unlock() + + // Check that we are allowed to set the process group. + if err := t.checkChange(ctx, linux.SIGTTOU); err != nil { + // drivers/tty/tty_io.c:tiocspgrp() converts -EIO from tty_check_change() + // to -ENOTTY. + if err == syserror.EIO { + return 0, syserror.ENOTTY + } + return 0, err + } + + // Check that calling task's process group is in the TTY session. + if task.ThreadGroup().Session() != t.session { + return 0, syserror.ENOTTY + } + + var pgID kernel.ProcessGroupID + if _, err := usermem.CopyObjectIn(ctx, io, args[2].Pointer(), &pgID, usermem.IOOpts{ + AddressSpaceActive: true, + }); err != nil { + return 0, err + } + + // pgID must be non-negative. + if pgID < 0 { + return 0, syserror.EINVAL + } + + // Process group with pgID must exist in this PID namespace. + pidns := task.PIDNamespace() + pg := pidns.ProcessGroupWithID(pgID) + if pg == nil { + return 0, syserror.ESRCH + } + + // Check that new process group is in the TTY session. + if pg.Session() != t.session { + return 0, syserror.EPERM + } + + t.fgProcessGroup = pg + return 0, nil + + case linux.TIOCGWINSZ: + // Args: struct winsize *argp + // Get window size. + winsize, err := ioctlGetWinsize(fd) + if err != nil { + return 0, err + } + _, err = usermem.CopyObjectOut(ctx, io, args[2].Pointer(), winsize, usermem.IOOpts{ + AddressSpaceActive: true, + }) + return 0, err + + case linux.TIOCSWINSZ: + // Args: const struct winsize *argp + // Set window size. + + // Unlike setting the termios, any process group (even background ones) can + // set the winsize. + + var winsize linux.Winsize + if _, err := usermem.CopyObjectIn(ctx, io, args[2].Pointer(), &winsize, usermem.IOOpts{ + AddressSpaceActive: true, + }); err != nil { + return 0, err + } + err := ioctlSetWinsize(fd, &winsize) + return 0, err + + // Unimplemented commands. + case linux.TIOCSETD, + linux.TIOCSBRK, + linux.TIOCCBRK, + linux.TCSBRK, + linux.TCSBRKP, + linux.TIOCSTI, + linux.TIOCCONS, + linux.FIONBIO, + linux.TIOCEXCL, + linux.TIOCNXCL, + linux.TIOCGEXCL, + linux.TIOCNOTTY, + linux.TIOCSCTTY, + linux.TIOCGSID, + linux.TIOCGETD, + linux.TIOCVHANGUP, + linux.TIOCGDEV, + linux.TIOCMGET, + linux.TIOCMSET, + linux.TIOCMBIC, + linux.TIOCMBIS, + linux.TIOCGICOUNT, + linux.TCFLSH, + linux.TIOCSSERIAL, + linux.TIOCGPTPEER: + + unimpl.EmitUnimplementedEvent(ctx) + fallthrough + default: + return 0, syserror.ENOTTY + } +} + +// checkChange checks that the process group is allowed to read, write, or +// change the state of the TTY. +// +// This corresponds to Linux drivers/tty/tty_io.c:tty_check_change(). The logic +// is a bit convoluted, but documented inline. +// +// Preconditions: t.mu must be held. +func (t *TTYFileDescription) checkChange(ctx context.Context, sig linux.Signal) error { + task := kernel.TaskFromContext(ctx) + if task == nil { + // No task? Linux does not have an analog for this case, but + // tty_check_change is more of a blacklist of cases than a + // whitelist, and is surprisingly permissive. Allowing the + // change seems most appropriate. + return nil + } + + tg := task.ThreadGroup() + pg := tg.ProcessGroup() + + // If the session for the task is different than the session for the + // controlling TTY, then the change is allowed. Seems like a bad idea, + // but that's exactly what linux does. + if tg.Session() != t.fgProcessGroup.Session() { + return nil + } + + // If we are the foreground process group, then the change is allowed. + if pg == t.fgProcessGroup { + return nil + } + + // We are not the foreground process group. + + // Is the provided signal blocked or ignored? + if (task.SignalMask()&linux.SignalSetOf(sig) != 0) || tg.SignalHandlers().IsIgnored(sig) { + // If the signal is SIGTTIN, then we are attempting to read + // from the TTY. Don't send the signal and return EIO. + if sig == linux.SIGTTIN { + return syserror.EIO + } + + // Otherwise, we are writing or changing terminal state. This is allowed. + return nil + } + + // If the process group is an orphan, return EIO. + if pg.IsOrphan() { + return syserror.EIO + } + + // Otherwise, send the signal to the process group and return ERESTARTSYS. + // + // Note that Linux also unconditionally sets TIF_SIGPENDING on current, + // but this isn't necessary in gVisor because the rationale given in + // 040b6362d58f "tty: fix leakage of -ERESTARTSYS to userland" doesn't + // apply: the sentry will handle -ERESTARTSYS in + // kernel.runApp.execute() even if the kernel.Task isn't interrupted. + // + // Linux ignores the result of kill_pgrp(). + _ = pg.SendSignal(kernel.SignalInfoPriv(sig)) + return kernel.ERESTARTSYS +} diff --git a/pkg/sentry/fsimpl/host/util.go b/pkg/sentry/fsimpl/host/util.go new file mode 100644 index 000000000..2bc757b1a --- /dev/null +++ b/pkg/sentry/fsimpl/host/util.go @@ -0,0 +1,66 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package host + +import ( + "syscall" + + "golang.org/x/sys/unix" + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/syserror" +) + +func toTimespec(ts linux.StatxTimestamp, omit bool) syscall.Timespec { + if omit { + return syscall.Timespec{ + Sec: 0, + Nsec: unix.UTIME_OMIT, + } + } + return syscall.Timespec{ + Sec: ts.Sec, + Nsec: int64(ts.Nsec), + } +} + +func unixToLinuxStatxTimestamp(ts unix.StatxTimestamp) linux.StatxTimestamp { + return linux.StatxTimestamp{Sec: ts.Sec, Nsec: ts.Nsec} +} + +func timespecToStatxTimestamp(ts unix.Timespec) linux.StatxTimestamp { + return linux.StatxTimestamp{Sec: int64(ts.Sec), Nsec: uint32(ts.Nsec)} +} + +// wouldBlock returns true for file types that can return EWOULDBLOCK +// for blocking operations, e.g. pipes, character devices, and sockets. +func wouldBlock(fileType uint32) bool { + return fileType == syscall.S_IFIFO || fileType == syscall.S_IFCHR || fileType == syscall.S_IFSOCK +} + +// canMap returns true if a file with fileType is allowed to be memory mapped. +// This is ported over from VFS1, but it's probably not the best way for us +// to check if a file can be memory mapped. +func canMap(fileType uint32) bool { + // TODO(gvisor.dev/issue/1672): Also allow "special files" to be mapped (see fs/host:canMap()). + // + // TODO(b/38213152): Some obscure character devices can be mapped. + return fileType == syscall.S_IFREG +} + +// isBlockError checks if an error is EAGAIN or EWOULDBLOCK. +// If so, they can be transformed into syserror.ErrWouldBlock. +func isBlockError(err error) bool { + return err == syserror.EAGAIN || err == syserror.EWOULDBLOCK +} diff --git a/pkg/sentry/syscalls/linux/vfs2/linux64_override_arm64.go b/pkg/sentry/fsimpl/host/util_unsafe.go index 6af5c400f..5136ac844 100644 --- a/pkg/sentry/syscalls/linux/vfs2/linux64_override_arm64.go +++ b/pkg/sentry/fsimpl/host/util_unsafe.go @@ -12,14 +12,23 @@ // See the License for the specific language governing permissions and // limitations under the License. -package vfs2 +package host import ( - "gvisor.dev/gvisor/pkg/sentry/kernel" - "gvisor.dev/gvisor/pkg/sentry/syscalls" + "syscall" + "unsafe" ) -// Override syscall table to add syscalls implementations from this package. -func Override(table map[uintptr]kernel.Syscall) { - table[63] = syscalls.Supported("read", Read) +func setTimestamps(fd int, ts *[2]syscall.Timespec) error { + _, _, errno := syscall.Syscall6( + syscall.SYS_UTIMENSAT, + uintptr(fd), + 0, /* path */ + uintptr(unsafe.Pointer(ts)), + 0, /* flags */ + 0, 0) + if errno != 0 { + return errno + } + return nil } diff --git a/pkg/sentry/fsimpl/kernfs/BUILD b/pkg/sentry/fsimpl/kernfs/BUILD index e73f1f857..ef34cb28a 100644 --- a/pkg/sentry/fsimpl/kernfs/BUILD +++ b/pkg/sentry/fsimpl/kernfs/BUILD @@ -4,6 +4,17 @@ load("//tools/go_generics:defs.bzl", "go_template_instance") licenses(["notice"]) go_template_instance( + name = "fstree", + out = "fstree.go", + package = "kernfs", + prefix = "generic", + template = "//pkg/sentry/vfs/genericfstree:generic_fstree", + types = { + "Dentry": "Dentry", + }, +) + +go_template_instance( name = "slot_list", out = "slot_list.go", package = "kernfs", @@ -21,6 +32,7 @@ go_library( "dynamic_bytes_file.go", "fd_impl_util.go", "filesystem.go", + "fstree.go", "inode_impl_util.go", "kernfs.go", "slot_list.go", @@ -35,6 +47,7 @@ go_library( "//pkg/refs", "//pkg/sentry/kernel/auth", "//pkg/sentry/memmap", + "//pkg/sentry/socket/unix/transport", "//pkg/sentry/vfs", "//pkg/sync", "//pkg/syserror", diff --git a/pkg/sentry/fsimpl/kernfs/dynamic_bytes_file.go b/pkg/sentry/fsimpl/kernfs/dynamic_bytes_file.go index d092ccb2a..1568a9d49 100644 --- a/pkg/sentry/fsimpl/kernfs/dynamic_bytes_file.go +++ b/pkg/sentry/fsimpl/kernfs/dynamic_bytes_file.go @@ -44,16 +44,16 @@ type DynamicBytesFile struct { var _ Inode = (*DynamicBytesFile)(nil) // Init initializes a dynamic bytes file. -func (f *DynamicBytesFile) Init(creds *auth.Credentials, ino uint64, data vfs.DynamicBytesSource, perm linux.FileMode) { +func (f *DynamicBytesFile) Init(creds *auth.Credentials, devMajor, devMinor uint32, ino uint64, data vfs.DynamicBytesSource, perm linux.FileMode) { if perm&^linux.PermissionsMask != 0 { panic(fmt.Sprintf("Only permission mask must be set: %x", perm&linux.PermissionsMask)) } - f.InodeAttrs.Init(creds, ino, linux.ModeRegular|perm) + f.InodeAttrs.Init(creds, devMajor, devMinor, ino, linux.ModeRegular|perm) f.data = data } // Open implements Inode.Open. -func (f *DynamicBytesFile) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { +func (f *DynamicBytesFile) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { fd := &DynamicBytesFD{} if err := fd.Init(rp.Mount(), vfsd, f.data, opts.Flags); err != nil { return nil, err @@ -61,9 +61,10 @@ func (f *DynamicBytesFile) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vf return &fd.vfsfd, nil } -// SetStat implements Inode.SetStat. -func (f *DynamicBytesFile) SetStat(*vfs.Filesystem, vfs.SetStatOptions) error { - // DynamicBytesFiles are immutable. +// SetStat implements Inode.SetStat. By default DynamicBytesFile doesn't allow +// inode attributes to be changed. Override SetStat() making it call +// f.InodeAttrs to allow it. +func (*DynamicBytesFile) SetStat(context.Context, *vfs.Filesystem, *auth.Credentials, vfs.SetStatOptions) error { return syserror.EPERM } @@ -122,7 +123,7 @@ func (fd *DynamicBytesFD) Release() {} // Stat implements vfs.FileDescriptionImpl.Stat. func (fd *DynamicBytesFD) Stat(ctx context.Context, opts vfs.StatOptions) (linux.Statx, error) { fs := fd.vfsfd.VirtualDentry().Mount().Filesystem() - return fd.inode.Stat(fs), nil + return fd.inode.Stat(fs, opts) } // SetStat implements vfs.FileDescriptionImpl.SetStat. diff --git a/pkg/sentry/fsimpl/kernfs/fd_impl_util.go b/pkg/sentry/fsimpl/kernfs/fd_impl_util.go index eda781155..8284e76a7 100644 --- a/pkg/sentry/fsimpl/kernfs/fd_impl_util.go +++ b/pkg/sentry/fsimpl/kernfs/fd_impl_util.go @@ -15,10 +15,14 @@ package kernfs import ( + "math" + "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/memmap" "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" ) @@ -33,24 +37,43 @@ import ( // inode. // // Must be initialize with Init before first use. +// +// Lock ordering: mu => children.mu. type GenericDirectoryFD struct { vfs.FileDescriptionDefaultImpl vfs.DirectoryFileDescriptionDefaultImpl vfsfd vfs.FileDescription children *OrderedChildren - off int64 + + // mu protects the fields below. + mu sync.Mutex + + // off is the current directory offset. Protected by "mu". + off int64 +} + +// NewGenericDirectoryFD creates a new GenericDirectoryFD and returns its +// dentry. +func NewGenericDirectoryFD(m *vfs.Mount, d *vfs.Dentry, children *OrderedChildren, opts *vfs.OpenOptions) (*GenericDirectoryFD, error) { + fd := &GenericDirectoryFD{} + if err := fd.Init(children, opts); err != nil { + return nil, err + } + if err := fd.vfsfd.Init(fd, opts.Flags, m, d, &vfs.FileDescriptionOptions{}); err != nil { + return nil, err + } + return fd, nil } -// Init initializes a GenericDirectoryFD. -func (fd *GenericDirectoryFD) Init(m *vfs.Mount, d *vfs.Dentry, children *OrderedChildren, opts *vfs.OpenOptions) error { +// Init initializes a GenericDirectoryFD. Use it when overriding +// GenericDirectoryFD. Caller must call fd.VFSFileDescription.Init() with the +// correct implementation. +func (fd *GenericDirectoryFD) Init(children *OrderedChildren, opts *vfs.OpenOptions) error { if vfs.AccessTypesForOpenFlags(opts)&vfs.MayWrite != 0 { // Can't open directories for writing. return syserror.EISDIR } - if err := fd.vfsfd.Init(fd, opts.Flags, m, d, &vfs.FileDescriptionOptions{}); err != nil { - return err - } fd.children = children return nil } @@ -100,40 +123,44 @@ func (fd *GenericDirectoryFD) inode() Inode { // IterDirents implements vfs.FileDecriptionImpl.IterDirents. IterDirents holds // o.mu when calling cb. func (fd *GenericDirectoryFD) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback) error { - vfsFS := fd.filesystem() - fs := vfsFS.Impl().(*Filesystem) - vfsd := fd.vfsfd.VirtualDentry().Dentry() - - fs.mu.Lock() - defer fs.mu.Unlock() + fd.mu.Lock() + defer fd.mu.Unlock() + opts := vfs.StatOptions{Mask: linux.STATX_INO} // Handle ".". if fd.off == 0 { - stat := fd.inode().Stat(vfsFS) + stat, err := fd.inode().Stat(fd.filesystem(), opts) + if err != nil { + return err + } dirent := vfs.Dirent{ Name: ".", Type: linux.DT_DIR, Ino: stat.Ino, NextOff: 1, } - if !cb.Handle(dirent) { - return nil + if err := cb.Handle(dirent); err != nil { + return err } fd.off++ } // Handle "..". if fd.off == 1 { - parentInode := vfsd.ParentOrSelf().Impl().(*Dentry).inode - stat := parentInode.Stat(vfsFS) + vfsd := fd.vfsfd.VirtualDentry().Dentry() + parentInode := genericParentOrSelf(vfsd.Impl().(*Dentry)).inode + stat, err := parentInode.Stat(fd.filesystem(), opts) + if err != nil { + return err + } dirent := vfs.Dirent{ Name: "..", Type: linux.FileMode(stat.Mode).DirentType(), Ino: stat.Ino, NextOff: 2, } - if !cb.Handle(dirent) { - return nil + if err := cb.Handle(dirent); err != nil { + return err } fd.off++ } @@ -146,15 +173,18 @@ func (fd *GenericDirectoryFD) IterDirents(ctx context.Context, cb vfs.IterDirent childIdx := fd.off - 2 for it := fd.children.nthLocked(childIdx); it != nil; it = it.Next() { inode := it.Dentry.Impl().(*Dentry).inode - stat := inode.Stat(vfsFS) + stat, err := inode.Stat(fd.filesystem(), opts) + if err != nil { + return err + } dirent := vfs.Dirent{ Name: it.Name, Type: linux.FileMode(stat.Mode).DirentType(), Ino: stat.Ino, NextOff: fd.off + 1, } - if !cb.Handle(dirent) { - return nil + if err := cb.Handle(dirent); err != nil { + return err } fd.off++ } @@ -167,15 +197,18 @@ func (fd *GenericDirectoryFD) IterDirents(ctx context.Context, cb vfs.IterDirent // Seek implements vfs.FileDecriptionImpl.Seek. func (fd *GenericDirectoryFD) Seek(ctx context.Context, offset int64, whence int32) (int64, error) { - fs := fd.filesystem().Impl().(*Filesystem) - fs.mu.Lock() - defer fs.mu.Unlock() + fd.mu.Lock() + defer fd.mu.Unlock() switch whence { case linux.SEEK_SET: // Use offset as given. case linux.SEEK_CUR: offset += fd.off + case linux.SEEK_END: + // TODO(gvisor.dev/issue/1193): This can prevent new files from showing up + // if they are added after SEEK_END. + offset = math.MaxInt64 default: return 0, syserror.EINVAL } @@ -190,12 +223,12 @@ func (fd *GenericDirectoryFD) Seek(ctx context.Context, offset int64, whence int func (fd *GenericDirectoryFD) Stat(ctx context.Context, opts vfs.StatOptions) (linux.Statx, error) { fs := fd.filesystem() inode := fd.inode() - return inode.Stat(fs), nil + return inode.Stat(fs, opts) } // SetStat implements vfs.FileDescriptionImpl.SetStat. func (fd *GenericDirectoryFD) SetStat(ctx context.Context, opts vfs.SetStatOptions) error { - fs := fd.filesystem() + creds := auth.CredentialsFromContext(ctx) inode := fd.vfsfd.VirtualDentry().Dentry().Impl().(*Dentry).inode - return inode.SetStat(fs, opts) + return inode.SetStat(ctx, fd.filesystem(), creds, opts) } diff --git a/pkg/sentry/fsimpl/kernfs/filesystem.go b/pkg/sentry/fsimpl/kernfs/filesystem.go index ee98eb66a..4a12ae245 100644 --- a/pkg/sentry/fsimpl/kernfs/filesystem.go +++ b/pkg/sentry/fsimpl/kernfs/filesystem.go @@ -12,16 +12,18 @@ // See the License for the specific language governing permissions and // limitations under the License. -// This file implements vfs.FilesystemImpl for kernfs. - package kernfs +// This file implements vfs.FilesystemImpl for kernfs. + import ( "fmt" "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserror" ) @@ -54,36 +56,48 @@ afterSymlink: return vfsd, nil } if name == ".." { - nextVFSD, err := rp.ResolveParent(vfsd) - if err != nil { + if isRoot, err := rp.CheckRoot(vfsd); err != nil { + return nil, err + } else if isRoot || d.parent == nil { + rp.Advance() + return vfsd, nil + } + if err := rp.CheckMount(&d.parent.vfsd); err != nil { return nil, err } rp.Advance() - return nextVFSD, nil + return &d.parent.vfsd, nil + } + if len(name) > linux.NAME_MAX { + return nil, syserror.ENAMETOOLONG } d.dirMu.Lock() - nextVFSD, err := rp.ResolveChild(vfsd, name) + next, err := fs.revalidateChildLocked(ctx, rp.VirtualFilesystem(), d, name, d.children[name]) + d.dirMu.Unlock() if err != nil { - d.dirMu.Unlock() return nil, err } - next, err := fs.revalidateChildLocked(ctx, rp.VirtualFilesystem(), d, name, nextVFSD) - d.dirMu.Unlock() - if err != nil { + if err := rp.CheckMount(&next.vfsd); err != nil { return nil, err } // Resolve any symlink at current path component. if rp.ShouldFollowSymlink() && next.isSymlink() { - // TODO: VFS2 needs something extra for /proc/[pid]/fd/ "magic symlinks". - target, err := next.inode.Readlink(ctx) + targetVD, targetPathname, err := next.inode.Getlink(ctx, rp.Mount()) if err != nil { return nil, err } - if err := rp.HandleSymlink(target); err != nil { - return nil, err + if targetVD.Ok() { + err := rp.HandleJump(targetVD) + targetVD.DecRef() + if err != nil { + return nil, err + } + } else { + if err := rp.HandleSymlink(targetPathname); err != nil { + return nil, err + } } goto afterSymlink - } rp.Advance() return &next.vfsd, nil @@ -97,17 +111,17 @@ afterSymlink: // parent.dirMu must be locked. parent.isDir(). name is not "." or "..". // // Postconditions: Caller must call fs.processDeferredDecRefs*. -func (fs *Filesystem) revalidateChildLocked(ctx context.Context, vfsObj *vfs.VirtualFilesystem, parent *Dentry, name string, childVFSD *vfs.Dentry) (*Dentry, error) { - if childVFSD != nil { +func (fs *Filesystem) revalidateChildLocked(ctx context.Context, vfsObj *vfs.VirtualFilesystem, parent *Dentry, name string, child *Dentry) (*Dentry, error) { + if child != nil { // Cached dentry exists, revalidate. - child := childVFSD.Impl().(*Dentry) if !child.inode.Valid(ctx) { - vfsObj.ForceDeleteDentry(childVFSD) - fs.deferDecRef(childVFSD) // Reference from Lookup. - childVFSD = nil + delete(parent.children, name) + vfsObj.InvalidateDentry(&child.vfsd) + fs.deferDecRef(&child.vfsd) // Reference from Lookup. + child = nil } } - if childVFSD == nil { + if child == nil { // Dentry isn't cached; it either doesn't exist or failed // revalidation. Attempt to resolve it via Lookup. // @@ -115,15 +129,15 @@ func (fs *Filesystem) revalidateChildLocked(ctx context.Context, vfsObj *vfs.Vir // *(kernfs.)Dentry, not *vfs.Dentry, since (kernfs.)Filesystem assumes // that all dentries in the filesystem are (kernfs.)Dentry and performs // vfs.DentryImpl casts accordingly. - var err error - childVFSD, err = parent.inode.Lookup(ctx, name) + childVFSD, err := parent.inode.Lookup(ctx, name) if err != nil { return nil, err } // Reference on childVFSD dropped by a corresponding Valid. - parent.insertChildLocked(name, childVFSD) + child = childVFSD.Impl().(*Dentry) + parent.insertChildLocked(name, child) } - return childVFSD.Impl().(*Dentry), nil + return child, nil } // walkExistingLocked resolves rp to an existing file. @@ -189,14 +203,14 @@ func checkCreateLocked(ctx context.Context, rp *vfs.ResolvingPath, parentVFSD *v if pc == "." || pc == ".." { return "", syserror.EEXIST } - childVFSD, err := rp.ResolveChild(parentVFSD, pc) - if err != nil { - return "", err + if len(pc) > linux.NAME_MAX { + return "", syserror.ENAMETOOLONG } - if childVFSD != nil { + // FIXME(gvisor.dev/issue/1193): Data race due to not holding dirMu. + if _, ok := parentVFSD.Impl().(*Dentry).children[pc]; ok { return "", syserror.EEXIST } - if parentVFSD.IsDisowned() { + if parentVFSD.IsDead() { return "", syserror.ENOENT } return pc, nil @@ -206,14 +220,14 @@ func checkCreateLocked(ctx context.Context, rp *vfs.ResolvingPath, parentVFSD *v // // Preconditions: Filesystem.mu must be locked for at least reading. func checkDeleteLocked(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry) error { - parentVFSD := vfsd.Parent() - if parentVFSD == nil { + parent := vfsd.Impl().(*Dentry).parent + if parent == nil { return syserror.EBUSY } - if parentVFSD.IsDisowned() { + if parent.vfsd.IsDead() { return syserror.ENOENT } - if err := parentVFSD.Impl().(*Dentry).inode.CheckPermissions(ctx, rp.Credentials(), vfs.MayWrite|vfs.MayExec); err != nil { + if err := parent.inode.CheckPermissions(ctx, rp.Credentials(), vfs.MayWrite|vfs.MayExec); err != nil { return err } return nil @@ -229,6 +243,19 @@ func (fs *Filesystem) Sync(ctx context.Context) error { return nil } +// AccessAt implements vfs.Filesystem.Impl.AccessAt. +func (fs *Filesystem) AccessAt(ctx context.Context, rp *vfs.ResolvingPath, creds *auth.Credentials, ats vfs.AccessTypes) error { + fs.mu.RLock() + defer fs.processDeferredDecRefs() + defer fs.mu.RUnlock() + + _, inode, err := fs.walkExistingLocked(ctx, rp) + if err != nil { + return err + } + return inode.CheckPermissions(ctx, creds, ats) +} + // GetDentryAt implements vfs.FilesystemImpl.GetDentryAt. func (fs *Filesystem) GetDentryAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.GetDentryOptions) (*vfs.Dentry, error) { fs.mu.RLock() @@ -294,11 +321,11 @@ func (fs *Filesystem) LinkAt(ctx context.Context, rp *vfs.ResolvingPath, vd vfs. return syserror.EPERM } - child, err := parentInode.NewLink(ctx, pc, d.inode) + childVFSD, err := parentInode.NewLink(ctx, pc, d.inode) if err != nil { return err } - parentVFSD.Impl().(*Dentry).InsertChild(pc, child) + parentVFSD.Impl().(*Dentry).InsertChild(pc, childVFSD.Impl().(*Dentry)) return nil } @@ -322,11 +349,11 @@ func (fs *Filesystem) MkdirAt(ctx context.Context, rp *vfs.ResolvingPath, opts v return err } defer rp.Mount().EndWrite() - child, err := parentInode.NewDir(ctx, pc, opts) + childVFSD, err := parentInode.NewDir(ctx, pc, opts) if err != nil { return err } - parentVFSD.Impl().(*Dentry).InsertChild(pc, child) + parentVFSD.Impl().(*Dentry).InsertChild(pc, childVFSD.Impl().(*Dentry)) return nil } @@ -350,11 +377,11 @@ func (fs *Filesystem) MknodAt(ctx context.Context, rp *vfs.ResolvingPath, opts v return err } defer rp.Mount().EndWrite() - new, err := parentInode.NewNode(ctx, pc, opts) + newVFSD, err := parentInode.NewNode(ctx, pc, opts) if err != nil { return err } - parentVFSD.Impl().(*Dentry).InsertChild(pc, new) + parentVFSD.Impl().(*Dentry).InsertChild(pc, newVFSD.Impl().(*Dentry)) return nil } @@ -364,7 +391,7 @@ func (fs *Filesystem) OpenAt(ctx context.Context, rp *vfs.ResolvingPath, opts vf // O_NOFOLLOW have no effect here (they're handled by VFS by setting // appropriate bits in rp), but are returned by // FileDescriptionImpl.StatusFlags(). - opts.Flags &= linux.O_ACCMODE | linux.O_CREAT | linux.O_EXCL | linux.O_TRUNC | linux.O_DIRECTORY | linux.O_NOFOLLOW + opts.Flags &= linux.O_ACCMODE | linux.O_CREAT | linux.O_EXCL | linux.O_TRUNC | linux.O_DIRECTORY | linux.O_NOFOLLOW | linux.O_NONBLOCK ats := vfs.AccessTypesForOpenFlags(&opts) // Do not create new file. @@ -379,7 +406,7 @@ func (fs *Filesystem) OpenAt(ctx context.Context, rp *vfs.ResolvingPath, opts vf if err := inode.CheckPermissions(ctx, rp.Credentials(), ats); err != nil { return nil, err } - return inode.Open(rp, vfsd, opts) + return inode.Open(ctx, rp, vfsd, opts) } // May create new file. @@ -398,7 +425,7 @@ func (fs *Filesystem) OpenAt(ctx context.Context, rp *vfs.ResolvingPath, opts vf if err := inode.CheckPermissions(ctx, rp.Credentials(), ats); err != nil { return nil, err } - return inode.Open(rp, vfsd, opts) + return inode.Open(ctx, rp, vfsd, opts) } afterTrailingSymlink: parentVFSD, parentInode, err := fs.walkParentDirLocked(ctx, rp) @@ -418,12 +445,12 @@ afterTrailingSymlink: if pc == "." || pc == ".." { return nil, syserror.EISDIR } - // Determine whether or not we need to create a file. - childVFSD, err := rp.ResolveChild(parentVFSD, pc) - if err != nil { - return nil, err + if len(pc) > linux.NAME_MAX { + return nil, syserror.ENAMETOOLONG } - if childVFSD == nil { + // Determine whether or not we need to create a file. + childVFSD, err := fs.stepExistingLocked(ctx, rp, parentVFSD) + if err == syserror.ENOENT { // Already checked for searchability above; now check for writability. if err := parentInode.CheckPermissions(ctx, rp.Credentials(), vfs.MayWrite); err != nil { return nil, err @@ -433,37 +460,46 @@ afterTrailingSymlink: } defer rp.Mount().EndWrite() // Create and open the child. - child, err := parentInode.NewFile(ctx, pc, opts) + childVFSD, err = parentInode.NewFile(ctx, pc, opts) if err != nil { return nil, err } + child := childVFSD.Impl().(*Dentry) parentVFSD.Impl().(*Dentry).InsertChild(pc, child) - return child.Impl().(*Dentry).inode.Open(rp, child, opts) + return child.inode.Open(ctx, rp, childVFSD, opts) + } + if err != nil { + return nil, err } // Open existing file or follow symlink. if mustCreate { return nil, syserror.EEXIST } - childDentry := childVFSD.Impl().(*Dentry) - childInode := childDentry.inode - if rp.ShouldFollowSymlink() { - if childDentry.isSymlink() { - target, err := childInode.Readlink(ctx) + child := childVFSD.Impl().(*Dentry) + if rp.ShouldFollowSymlink() && child.isSymlink() { + targetVD, targetPathname, err := child.inode.Getlink(ctx, rp.Mount()) + if err != nil { + return nil, err + } + if targetVD.Ok() { + err := rp.HandleJump(targetVD) + targetVD.DecRef() if err != nil { return nil, err } - if err := rp.HandleSymlink(target); err != nil { + } else { + if err := rp.HandleSymlink(targetPathname); err != nil { return nil, err } - // rp.Final() may no longer be true since we now need to resolve the - // symlink target. - goto afterTrailingSymlink } + // rp.Final() may no longer be true since we now need to resolve the + // symlink target. + goto afterTrailingSymlink } - if err := childInode.CheckPermissions(ctx, rp.Credentials(), ats); err != nil { + if err := child.inode.CheckPermissions(ctx, rp.Credentials(), ats); err != nil { return nil, err } - return childInode.Open(rp, childVFSD, opts) + return child.inode.Open(ctx, rp, &child.vfsd, opts) } // ReadlinkAt implements vfs.FilesystemImpl.ReadlinkAt. @@ -490,15 +526,16 @@ func (fs *Filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa noReplace := opts.Flags&linux.RENAME_NOREPLACE != 0 fs.mu.Lock() - defer fs.mu.Lock() + defer fs.processDeferredDecRefsLocked() + defer fs.mu.Unlock() // Resolve the destination directory first to verify that it's on this // Mount. dstDirVFSD, dstDirInode, err := fs.walkParentDirLocked(ctx, rp) - fs.processDeferredDecRefsLocked() if err != nil { return err } + dstDir := dstDirVFSD.Impl().(*Dentry) mnt := rp.Mount() if mnt != oldParentVD.Mount() { return syserror.EXDEV @@ -511,9 +548,8 @@ func (fs *Filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa srcDirVFSD := oldParentVD.Dentry() srcDir := srcDirVFSD.Impl().(*Dentry) srcDir.dirMu.Lock() - src, err := fs.revalidateChildLocked(ctx, rp.VirtualFilesystem(), srcDir, oldName, srcDirVFSD.Child(oldName)) + src, err := fs.revalidateChildLocked(ctx, rp.VirtualFilesystem(), srcDir, oldName, srcDir.children[oldName]) srcDir.dirMu.Unlock() - fs.processDeferredDecRefsLocked() if err != nil { return err } @@ -525,7 +561,7 @@ func (fs *Filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa } // Can we create the dst dentry? - var dstVFSD *vfs.Dentry + var dst *Dentry pc, err := checkCreateLocked(ctx, rp, dstDirVFSD, dstDirInode) switch err { case nil: @@ -535,37 +571,51 @@ func (fs *Filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa // Won't overwrite existing node since RENAME_NOREPLACE was requested. return syserror.EEXIST } - dstVFSD, err = rp.ResolveChild(dstDirVFSD, pc) - if err != nil { + dst = dstDir.children[pc] + if dst == nil { panic(fmt.Sprintf("Child %q for parent Dentry %+v disappeared inside atomic section?", pc, dstDirVFSD)) } default: return err } + var dstVFSD *vfs.Dentry + if dst != nil { + dstVFSD = &dst.vfsd + } mntns := vfs.MountNamespaceFromContext(ctx) + defer mntns.DecRef() virtfs := rp.VirtualFilesystem() - srcDirDentry := srcDirVFSD.Impl().(*Dentry) - dstDirDentry := dstDirVFSD.Impl().(*Dentry) - // We can't deadlock here due to lock ordering because we're protected from // concurrent renames by fs.mu held for writing. - srcDirDentry.dirMu.Lock() - defer srcDirDentry.dirMu.Unlock() - dstDirDentry.dirMu.Lock() - defer dstDirDentry.dirMu.Unlock() + srcDir.dirMu.Lock() + defer srcDir.dirMu.Unlock() + if srcDir != dstDir { + dstDir.dirMu.Lock() + defer dstDir.dirMu.Unlock() + } if err := virtfs.PrepareRenameDentry(mntns, srcVFSD, dstVFSD); err != nil { return err } - srcDirInode := srcDirDentry.inode - replaced, err := srcDirInode.Rename(ctx, srcVFSD.Name(), pc, srcVFSD, dstDirVFSD) + replaced, err := srcDir.inode.Rename(ctx, src.name, pc, srcVFSD, dstDirVFSD) if err != nil { virtfs.AbortRenameDentry(srcVFSD, dstVFSD) return err } - virtfs.CommitRenameReplaceDentry(srcVFSD, dstDirVFSD, pc, replaced) + delete(srcDir.children, src.name) + if srcDir != dstDir { + fs.deferDecRef(srcDirVFSD) + dstDir.IncRef() + } + src.parent = dstDir + src.name = pc + if dstDir.children == nil { + dstDir.children = make(map[string]*Dentry) + } + dstDir.children[pc] = src + virtfs.CommitRenameReplaceDentry(srcVFSD, replaced) return nil } @@ -585,17 +635,21 @@ func (fs *Filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error if err := checkDeleteLocked(ctx, rp, vfsd); err != nil { return err } - if !vfsd.Impl().(*Dentry).isDir() { + d := vfsd.Impl().(*Dentry) + if !d.isDir() { return syserror.ENOTDIR } if inode.HasChildren() { return syserror.ENOTEMPTY } virtfs := rp.VirtualFilesystem() - parentDentry := vfsd.Parent().Impl().(*Dentry) + parentDentry := d.parent parentDentry.dirMu.Lock() defer parentDentry.dirMu.Unlock() - if err := virtfs.PrepareDeleteDentry(vfs.MountNamespaceFromContext(ctx), vfsd); err != nil { + + mntns := vfs.MountNamespaceFromContext(ctx) + defer mntns.DecRef() + if err := virtfs.PrepareDeleteDentry(mntns, vfsd); err != nil { return err } if err := parentDentry.inode.RmDir(ctx, rp.Component(), vfsd); err != nil { @@ -618,7 +672,7 @@ func (fs *Filesystem) SetStatAt(ctx context.Context, rp *vfs.ResolvingPath, opts if opts.Stat.Mask == 0 { return nil } - return inode.SetStat(fs.VFSFilesystem(), opts) + return inode.SetStat(ctx, fs.VFSFilesystem(), rp.Credentials(), opts) } // StatAt implements vfs.FilesystemImpl.StatAt. @@ -630,7 +684,7 @@ func (fs *Filesystem) StatAt(ctx context.Context, rp *vfs.ResolvingPath, opts vf if err != nil { return linux.Statx{}, err } - return inode.Stat(fs.VFSFilesystem()), nil + return inode.Stat(fs.VFSFilesystem(), opts) } // StatFSAt implements vfs.FilesystemImpl.StatFSAt. @@ -642,7 +696,7 @@ func (fs *Filesystem) StatFSAt(ctx context.Context, rp *vfs.ResolvingPath) (linu if err != nil { return linux.Statfs{}, err } - // TODO: actually implement statfs + // TODO(gvisor.dev/issue/1193): actually implement statfs. return linux.Statfs{}, syserror.ENOSYS } @@ -666,11 +720,11 @@ func (fs *Filesystem) SymlinkAt(ctx context.Context, rp *vfs.ResolvingPath, targ return err } defer rp.Mount().EndWrite() - child, err := parentInode.NewSymlink(ctx, pc, target) + childVFSD, err := parentInode.NewSymlink(ctx, pc, target) if err != nil { return err } - parentVFSD.Impl().(*Dentry).InsertChild(pc, child) + parentVFSD.Impl().(*Dentry).InsertChild(pc, childVFSD.Impl().(*Dentry)) return nil } @@ -690,14 +744,17 @@ func (fs *Filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error if err := checkDeleteLocked(ctx, rp, vfsd); err != nil { return err } - if vfsd.Impl().(*Dentry).isDir() { + d := vfsd.Impl().(*Dentry) + if d.isDir() { return syserror.EISDIR } virtfs := rp.VirtualFilesystem() - parentDentry := vfsd.Parent().Impl().(*Dentry) + parentDentry := d.parent parentDentry.dirMu.Lock() defer parentDentry.dirMu.Unlock() - if err := virtfs.PrepareDeleteDentry(vfs.MountNamespaceFromContext(ctx), vfsd); err != nil { + mntns := vfs.MountNamespaceFromContext(ctx) + defer mntns.DecRef() + if err := virtfs.PrepareDeleteDentry(mntns, vfsd); err != nil { return err } if err := parentDentry.inode.Unlink(ctx, rp.Component(), vfsd); err != nil { @@ -708,8 +765,23 @@ func (fs *Filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error return nil } +// BoundEndpointAt implements FilesystemImpl.BoundEndpointAt. +func (fs *Filesystem) BoundEndpointAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.BoundEndpointOptions) (transport.BoundEndpoint, error) { + fs.mu.RLock() + _, inode, err := fs.walkExistingLocked(ctx, rp) + fs.mu.RUnlock() + fs.processDeferredDecRefs() + if err != nil { + return nil, err + } + if err := inode.CheckPermissions(ctx, rp.Credentials(), vfs.MayWrite); err != nil { + return nil, err + } + return nil, syserror.ECONNREFUSED +} + // ListxattrAt implements vfs.FilesystemImpl.ListxattrAt. -func (fs *Filesystem) ListxattrAt(ctx context.Context, rp *vfs.ResolvingPath) ([]string, error) { +func (fs *Filesystem) ListxattrAt(ctx context.Context, rp *vfs.ResolvingPath, size uint64) ([]string, error) { fs.mu.RLock() _, _, err := fs.walkExistingLocked(ctx, rp) fs.mu.RUnlock() @@ -722,7 +794,7 @@ func (fs *Filesystem) ListxattrAt(ctx context.Context, rp *vfs.ResolvingPath) ([ } // GetxattrAt implements vfs.FilesystemImpl.GetxattrAt. -func (fs *Filesystem) GetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) (string, error) { +func (fs *Filesystem) GetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.GetxattrOptions) (string, error) { fs.mu.RLock() _, _, err := fs.walkExistingLocked(ctx, rp) fs.mu.RUnlock() @@ -764,5 +836,5 @@ func (fs *Filesystem) RemovexattrAt(ctx context.Context, rp *vfs.ResolvingPath, func (fs *Filesystem) PrependPath(ctx context.Context, vfsroot, vd vfs.VirtualDentry, b *fspath.Builder) error { fs.mu.RLock() defer fs.mu.RUnlock() - return vfs.GenericPrependPath(vfsroot, vd, b) + return genericPrependPath(vfsroot, vd.Mount(), vd.Dentry().Impl().(*Dentry), b) } diff --git a/pkg/sentry/fsimpl/kernfs/inode_impl_util.go b/pkg/sentry/fsimpl/kernfs/inode_impl_util.go index 099d70a16..982daa2e6 100644 --- a/pkg/sentry/fsimpl/kernfs/inode_impl_util.go +++ b/pkg/sentry/fsimpl/kernfs/inode_impl_util.go @@ -36,20 +36,20 @@ type InodeNoopRefCount struct { } // IncRef implements Inode.IncRef. -func (n *InodeNoopRefCount) IncRef() { +func (InodeNoopRefCount) IncRef() { } // DecRef implements Inode.DecRef. -func (n *InodeNoopRefCount) DecRef() { +func (InodeNoopRefCount) DecRef() { } // TryIncRef implements Inode.TryIncRef. -func (n *InodeNoopRefCount) TryIncRef() bool { +func (InodeNoopRefCount) TryIncRef() bool { return true } // Destroy implements Inode.Destroy. -func (n *InodeNoopRefCount) Destroy() { +func (InodeNoopRefCount) Destroy() { } // InodeDirectoryNoNewChildren partially implements the Inode interface. @@ -58,27 +58,27 @@ func (n *InodeNoopRefCount) Destroy() { type InodeDirectoryNoNewChildren struct{} // NewFile implements Inode.NewFile. -func (*InodeDirectoryNoNewChildren) NewFile(context.Context, string, vfs.OpenOptions) (*vfs.Dentry, error) { +func (InodeDirectoryNoNewChildren) NewFile(context.Context, string, vfs.OpenOptions) (*vfs.Dentry, error) { return nil, syserror.EPERM } // NewDir implements Inode.NewDir. -func (*InodeDirectoryNoNewChildren) NewDir(context.Context, string, vfs.MkdirOptions) (*vfs.Dentry, error) { +func (InodeDirectoryNoNewChildren) NewDir(context.Context, string, vfs.MkdirOptions) (*vfs.Dentry, error) { return nil, syserror.EPERM } // NewLink implements Inode.NewLink. -func (*InodeDirectoryNoNewChildren) NewLink(context.Context, string, Inode) (*vfs.Dentry, error) { +func (InodeDirectoryNoNewChildren) NewLink(context.Context, string, Inode) (*vfs.Dentry, error) { return nil, syserror.EPERM } // NewSymlink implements Inode.NewSymlink. -func (*InodeDirectoryNoNewChildren) NewSymlink(context.Context, string, string) (*vfs.Dentry, error) { +func (InodeDirectoryNoNewChildren) NewSymlink(context.Context, string, string) (*vfs.Dentry, error) { return nil, syserror.EPERM } // NewNode implements Inode.NewNode. -func (*InodeDirectoryNoNewChildren) NewNode(context.Context, string, vfs.MknodOptions) (*vfs.Dentry, error) { +func (InodeDirectoryNoNewChildren) NewNode(context.Context, string, vfs.MknodOptions) (*vfs.Dentry, error) { return nil, syserror.EPERM } @@ -90,62 +90,62 @@ type InodeNotDirectory struct { } // HasChildren implements Inode.HasChildren. -func (*InodeNotDirectory) HasChildren() bool { +func (InodeNotDirectory) HasChildren() bool { return false } // NewFile implements Inode.NewFile. -func (*InodeNotDirectory) NewFile(context.Context, string, vfs.OpenOptions) (*vfs.Dentry, error) { +func (InodeNotDirectory) NewFile(context.Context, string, vfs.OpenOptions) (*vfs.Dentry, error) { panic("NewFile called on non-directory inode") } // NewDir implements Inode.NewDir. -func (*InodeNotDirectory) NewDir(context.Context, string, vfs.MkdirOptions) (*vfs.Dentry, error) { +func (InodeNotDirectory) NewDir(context.Context, string, vfs.MkdirOptions) (*vfs.Dentry, error) { panic("NewDir called on non-directory inode") } // NewLink implements Inode.NewLinkink. -func (*InodeNotDirectory) NewLink(context.Context, string, Inode) (*vfs.Dentry, error) { +func (InodeNotDirectory) NewLink(context.Context, string, Inode) (*vfs.Dentry, error) { panic("NewLink called on non-directory inode") } // NewSymlink implements Inode.NewSymlink. -func (*InodeNotDirectory) NewSymlink(context.Context, string, string) (*vfs.Dentry, error) { +func (InodeNotDirectory) NewSymlink(context.Context, string, string) (*vfs.Dentry, error) { panic("NewSymlink called on non-directory inode") } // NewNode implements Inode.NewNode. -func (*InodeNotDirectory) NewNode(context.Context, string, vfs.MknodOptions) (*vfs.Dentry, error) { +func (InodeNotDirectory) NewNode(context.Context, string, vfs.MknodOptions) (*vfs.Dentry, error) { panic("NewNode called on non-directory inode") } // Unlink implements Inode.Unlink. -func (*InodeNotDirectory) Unlink(context.Context, string, *vfs.Dentry) error { +func (InodeNotDirectory) Unlink(context.Context, string, *vfs.Dentry) error { panic("Unlink called on non-directory inode") } // RmDir implements Inode.RmDir. -func (*InodeNotDirectory) RmDir(context.Context, string, *vfs.Dentry) error { +func (InodeNotDirectory) RmDir(context.Context, string, *vfs.Dentry) error { panic("RmDir called on non-directory inode") } // Rename implements Inode.Rename. -func (*InodeNotDirectory) Rename(context.Context, string, string, *vfs.Dentry, *vfs.Dentry) (*vfs.Dentry, error) { +func (InodeNotDirectory) Rename(context.Context, string, string, *vfs.Dentry, *vfs.Dentry) (*vfs.Dentry, error) { panic("Rename called on non-directory inode") } // Lookup implements Inode.Lookup. -func (*InodeNotDirectory) Lookup(ctx context.Context, name string) (*vfs.Dentry, error) { +func (InodeNotDirectory) Lookup(ctx context.Context, name string) (*vfs.Dentry, error) { panic("Lookup called on non-directory inode") } // IterDirents implements Inode.IterDirents. -func (*InodeNotDirectory) IterDirents(ctx context.Context, callback vfs.IterDirentsCallback, offset, relOffset int64) (newOffset int64, err error) { +func (InodeNotDirectory) IterDirents(ctx context.Context, callback vfs.IterDirentsCallback, offset, relOffset int64) (newOffset int64, err error) { panic("IterDirents called on non-directory inode") } // Valid implements Inode.Valid. -func (*InodeNotDirectory) Valid(context.Context) bool { +func (InodeNotDirectory) Valid(context.Context) bool { return true } @@ -157,17 +157,17 @@ func (*InodeNotDirectory) Valid(context.Context) bool { type InodeNoDynamicLookup struct{} // Lookup implements Inode.Lookup. -func (*InodeNoDynamicLookup) Lookup(ctx context.Context, name string) (*vfs.Dentry, error) { +func (InodeNoDynamicLookup) Lookup(ctx context.Context, name string) (*vfs.Dentry, error) { return nil, syserror.ENOENT } // IterDirents implements Inode.IterDirents. -func (*InodeNoDynamicLookup) IterDirents(ctx context.Context, callback vfs.IterDirentsCallback, offset, relOffset int64) (int64, error) { +func (InodeNoDynamicLookup) IterDirents(ctx context.Context, callback vfs.IterDirentsCallback, offset, relOffset int64) (int64, error) { return offset, nil } // Valid implements Inode.Valid. -func (*InodeNoDynamicLookup) Valid(ctx context.Context) bool { +func (InodeNoDynamicLookup) Valid(ctx context.Context) bool { return true } @@ -177,25 +177,32 @@ func (*InodeNoDynamicLookup) Valid(ctx context.Context) bool { type InodeNotSymlink struct{} // Readlink implements Inode.Readlink. -func (*InodeNotSymlink) Readlink(context.Context) (string, error) { +func (InodeNotSymlink) Readlink(context.Context) (string, error) { return "", syserror.EINVAL } +// Getlink implements Inode.Getlink. +func (InodeNotSymlink) Getlink(context.Context, *vfs.Mount) (vfs.VirtualDentry, string, error) { + return vfs.VirtualDentry{}, "", syserror.EINVAL +} + // InodeAttrs partially implements the Inode interface, specifically the // inodeMetadata sub interface. InodeAttrs provides functionality related to // inode attributes. // // Must be initialized by Init prior to first use. type InodeAttrs struct { - ino uint64 - mode uint32 - uid uint32 - gid uint32 - nlink uint32 + devMajor uint32 + devMinor uint32 + ino uint64 + mode uint32 + uid uint32 + gid uint32 + nlink uint32 } // Init initializes this InodeAttrs. -func (a *InodeAttrs) Init(creds *auth.Credentials, ino uint64, mode linux.FileMode) { +func (a *InodeAttrs) Init(creds *auth.Credentials, devMajor, devMinor uint32, ino uint64, mode linux.FileMode) { if mode.FileType() == 0 { panic(fmt.Sprintf("No file type specified in 'mode' for InodeAttrs.Init(): mode=0%o", mode)) } @@ -204,6 +211,8 @@ func (a *InodeAttrs) Init(creds *auth.Credentials, ino uint64, mode linux.FileMo if mode.FileType() == linux.ModeDirectory { nlink = 2 } + a.devMajor = devMajor + a.devMinor = devMinor atomic.StoreUint64(&a.ino, ino) atomic.StoreUint32(&a.mode, uint32(mode)) atomic.StoreUint32(&a.uid, uint32(creds.EffectiveKUID)) @@ -211,6 +220,21 @@ func (a *InodeAttrs) Init(creds *auth.Credentials, ino uint64, mode linux.FileMo atomic.StoreUint32(&a.nlink, nlink) } +// DevMajor returns the device major number. +func (a *InodeAttrs) DevMajor() uint32 { + return a.devMajor +} + +// DevMinor returns the device minor number. +func (a *InodeAttrs) DevMinor() uint32 { + return a.devMinor +} + +// Ino returns the inode id. +func (a *InodeAttrs) Ino() uint64 { + return atomic.LoadUint64(&a.ino) +} + // Mode implements Inode.Mode. func (a *InodeAttrs) Mode() linux.FileMode { return linux.FileMode(atomic.LoadUint32(&a.mode)) @@ -219,22 +243,34 @@ func (a *InodeAttrs) Mode() linux.FileMode { // Stat partially implements Inode.Stat. Note that this function doesn't provide // all the stat fields, and the embedder should consider extending the result // with filesystem-specific fields. -func (a *InodeAttrs) Stat(*vfs.Filesystem) linux.Statx { +func (a *InodeAttrs) Stat(*vfs.Filesystem, vfs.StatOptions) (linux.Statx, error) { var stat linux.Statx stat.Mask = linux.STATX_TYPE | linux.STATX_MODE | linux.STATX_UID | linux.STATX_GID | linux.STATX_INO | linux.STATX_NLINK + stat.DevMajor = a.devMajor + stat.DevMinor = a.devMinor stat.Ino = atomic.LoadUint64(&a.ino) stat.Mode = uint16(a.Mode()) stat.UID = atomic.LoadUint32(&a.uid) stat.GID = atomic.LoadUint32(&a.gid) stat.Nlink = atomic.LoadUint32(&a.nlink) - // TODO: Implement other stat fields like timestamps. + // TODO(gvisor.dev/issue/1193): Implement other stat fields like timestamps. - return stat + return stat, nil } // SetStat implements Inode.SetStat. -func (a *InodeAttrs) SetStat(_ *vfs.Filesystem, opts vfs.SetStatOptions) error { +func (a *InodeAttrs) SetStat(ctx context.Context, fs *vfs.Filesystem, creds *auth.Credentials, opts vfs.SetStatOptions) error { + if opts.Stat.Mask == 0 { + return nil + } + if opts.Stat.Mask&^(linux.STATX_MODE|linux.STATX_UID|linux.STATX_GID) != 0 { + return syserror.EPERM + } + if err := vfs.CheckSetStat(ctx, creds, &opts.Stat, a.Mode(), auth.KUID(atomic.LoadUint32(&a.uid)), auth.KGID(atomic.LoadUint32(&a.gid))); err != nil { + return err + } + stat := opts.Stat if stat.Mask&linux.STATX_MODE != 0 { for { @@ -256,19 +292,17 @@ func (a *InodeAttrs) SetStat(_ *vfs.Filesystem, opts vfs.SetStatOptions) error { // Note that not all fields are modifiable. For example, the file type and // inode numbers are immutable after node creation. - // TODO: Implement other stat fields like timestamps. + // TODO(gvisor.dev/issue/1193): Implement other stat fields like timestamps. return nil } // CheckPermissions implements Inode.CheckPermissions. func (a *InodeAttrs) CheckPermissions(_ context.Context, creds *auth.Credentials, ats vfs.AccessTypes) error { - mode := a.Mode() return vfs.GenericCheckPermissions( creds, ats, - mode.FileType() == linux.ModeDirectory, - uint16(mode), + a.Mode(), auth.KUID(atomic.LoadUint32(&a.uid)), auth.KGID(atomic.LoadUint32(&a.gid)), ) @@ -346,8 +380,8 @@ func (o *OrderedChildren) Destroy() { // cache. Populate returns the number of directories inserted, which the caller // may use to update the link count for the parent directory. // -// Precondition: d.Impl() must be a kernfs Dentry. d must represent a directory -// inode. children must not contain any conflicting entries already in o. +// Precondition: d must represent a directory inode. children must not contain +// any conflicting entries already in o. func (o *OrderedChildren) Populate(d *Dentry, children map[string]*Dentry) uint32 { var links uint32 for name, child := range children { @@ -357,7 +391,7 @@ func (o *OrderedChildren) Populate(d *Dentry, children map[string]*Dentry) uint3 if err := o.Insert(name, child.VFSDentry()); err != nil { panic(fmt.Sprintf("Collision when attempting to insert child %q (%+v) into %+v", name, child, d)) } - d.InsertChild(name, child.VFSDentry()) + d.InsertChild(name, child) } return links } @@ -507,7 +541,7 @@ type InodeSymlink struct { } // Open implements Inode.Open. -func (InodeSymlink) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { +func (InodeSymlink) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { return nil, syserror.ELOOP } @@ -526,9 +560,9 @@ type StaticDirectory struct { var _ Inode = (*StaticDirectory)(nil) // NewStaticDir creates a new static directory and returns its dentry. -func NewStaticDir(creds *auth.Credentials, ino uint64, perm linux.FileMode, children map[string]*Dentry) *Dentry { +func NewStaticDir(creds *auth.Credentials, devMajor, devMinor uint32, ino uint64, perm linux.FileMode, children map[string]*Dentry) *Dentry { inode := &StaticDirectory{} - inode.Init(creds, ino, perm) + inode.Init(creds, devMajor, devMinor, ino, perm) dentry := &Dentry{} dentry.Init(inode) @@ -541,16 +575,31 @@ func NewStaticDir(creds *auth.Credentials, ino uint64, perm linux.FileMode, chil } // Init initializes StaticDirectory. -func (s *StaticDirectory) Init(creds *auth.Credentials, ino uint64, perm linux.FileMode) { +func (s *StaticDirectory) Init(creds *auth.Credentials, devMajor, devMinor uint32, ino uint64, perm linux.FileMode) { if perm&^linux.PermissionsMask != 0 { panic(fmt.Sprintf("Only permission mask must be set: %x", perm&linux.PermissionsMask)) } - s.InodeAttrs.Init(creds, ino, linux.ModeDirectory|perm) + s.InodeAttrs.Init(creds, devMajor, devMinor, ino, linux.ModeDirectory|perm) } // Open implements kernfs.Inode. -func (s *StaticDirectory) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { - fd := &GenericDirectoryFD{} - fd.Init(rp.Mount(), vfsd, &s.OrderedChildren, &opts) +func (s *StaticDirectory) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd, err := NewGenericDirectoryFD(rp.Mount(), vfsd, &s.OrderedChildren, &opts) + if err != nil { + return nil, err + } return fd.VFSFileDescription(), nil } + +// SetStat implements Inode.SetStat not allowing inode attributes to be changed. +func (*StaticDirectory) SetStat(context.Context, *vfs.Filesystem, *auth.Credentials, vfs.SetStatOptions) error { + return syserror.EPERM +} + +// AlwaysValid partially implements kernfs.inodeDynamicLookup. +type AlwaysValid struct{} + +// Valid implements kernfs.inodeDynamicLookup. +func (*AlwaysValid) Valid(context.Context) bool { + return true +} diff --git a/pkg/sentry/fsimpl/kernfs/kernfs.go b/pkg/sentry/fsimpl/kernfs/kernfs.go index c74fa999b..bbee8ccda 100644 --- a/pkg/sentry/fsimpl/kernfs/kernfs.go +++ b/pkg/sentry/fsimpl/kernfs/kernfs.go @@ -63,9 +63,6 @@ import ( "gvisor.dev/gvisor/pkg/sync" ) -// FilesystemType implements vfs.FilesystemType. -type FilesystemType struct{} - // Filesystem mostly implements vfs.FilesystemImpl for a generic in-memory // filesystem. Concrete implementations are expected to embed this in their own // Filesystem type. @@ -135,13 +132,6 @@ func (fs *Filesystem) processDeferredDecRefsLocked() { fs.droppedDentriesMu.Unlock() } -// Init initializes a kernfs filesystem. This should be called from during -// vfs.FilesystemType.NewFilesystem for the concrete filesystem embedding -// kernfs. -func (fs *Filesystem) Init(vfsObj *vfs.VirtualFilesystem) { - fs.vfsfs.Init(vfsObj, fs) -} - // VFSFilesystem returns the generic vfs filesystem object. func (fs *Filesystem) VFSFilesystem() *vfs.Filesystem { return &fs.vfsfs @@ -171,19 +161,22 @@ const ( // // Must be initialized by Init prior to first use. type Dentry struct { - refs.AtomicRefCount - - vfsd vfs.Dentry - inode Inode + vfsd vfs.Dentry - refs uint64 + refs.AtomicRefCount // flags caches useful information about the dentry from the inode. See the // dflags* consts above. Must be accessed by atomic ops. flags uint32 - // dirMu protects vfsd.children for directory dentries. - dirMu sync.Mutex + parent *Dentry + name string + + // dirMu protects children and the names of child Dentries. + dirMu sync.Mutex + children map[string]*Dentry + + inode Inode } // Init initializes this dentry. @@ -227,18 +220,30 @@ func (d *Dentry) DecRef() { func (d *Dentry) destroy() { d.inode.DecRef() // IncRef from Init. d.inode = nil - if parent := d.vfsd.Parent(); parent != nil { - parent.DecRef() // IncRef from Dentry.InsertChild. + if d.parent != nil { + d.parent.DecRef() // IncRef from Dentry.InsertChild. } } +// InotifyWithParent implements vfs.DentryImpl.InotifyWithParent. +// +// TODO(gvisor.dev/issue/1479): Implement inotify. +func (d *Dentry) InotifyWithParent(events uint32, cookie uint32, et vfs.EventType) {} + +// Watches implements vfs.DentryImpl.Watches. +// +// TODO(gvisor.dev/issue/1479): Implement inotify. +func (d *Dentry) Watches() *vfs.Watches { + return nil +} + // 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 -// it's own isn't sufficient to insert a child into a directory. InsertChild +// its own isn't sufficient to insert a child into a directory. InsertChild // updates the link count on d if required. // // Precondition: d must represent a directory inode. -func (d *Dentry) InsertChild(name string, child *vfs.Dentry) { +func (d *Dentry) InsertChild(name string, child *Dentry) { d.dirMu.Lock() d.insertChildLocked(name, child) d.dirMu.Unlock() @@ -248,13 +253,22 @@ func (d *Dentry) InsertChild(name string, child *vfs.Dentry) { // preconditions. // // Precondition: d.dirMu must be locked. -func (d *Dentry) insertChildLocked(name string, child *vfs.Dentry) { +func (d *Dentry) insertChildLocked(name string, child *Dentry) { if !d.isDir() { panic(fmt.Sprintf("InsertChild called on non-directory Dentry: %+v.", d)) } - vfsDentry := d.VFSDentry() - vfsDentry.IncRef() // DecRef in child's Dentry.destroy. - vfsDentry.InsertChild(child, name) + d.IncRef() // DecRef in child's Dentry.destroy. + child.parent = d + child.name = name + if d.children == nil { + d.children = make(map[string]*Dentry) + } + d.children[name] = child +} + +// Inode returns the dentry's inode. +func (d *Dentry) Inode() Inode { + return d.inode } // The Inode interface maps filesystem-level operations that operate on paths to @@ -302,8 +316,9 @@ type Inode interface { // this inode. The returned file description should hold a reference on the // inode for its lifetime. // - // Precondition: !rp.Done(). vfsd.Impl() must be a kernfs Dentry. - Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) + // 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) } type inodeRefs interface { @@ -320,7 +335,7 @@ type inodeMetadata interface { // CheckPermissions checks that creds may access this inode for the // requested access type, per the the rules of // fs/namei.c:generic_permission(). - CheckPermissions(ctx context.Context, creds *auth.Credentials, atx vfs.AccessTypes) error + CheckPermissions(ctx context.Context, creds *auth.Credentials, ats vfs.AccessTypes) error // Mode returns the (struct stat)::st_mode value for this inode. This is // separated from Stat for performance. @@ -328,11 +343,13 @@ type inodeMetadata interface { // Stat returns the metadata for this inode. This corresponds to // vfs.FilesystemImpl.StatAt. - Stat(fs *vfs.Filesystem) linux.Statx + Stat(fs *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) // SetStat updates the metadata for this inode. This corresponds to - // vfs.FilesystemImpl.SetStatAt. - SetStat(fs *vfs.Filesystem, opts vfs.SetStatOptions) error + // vfs.FilesystemImpl.SetStatAt. Implementations are responsible for checking + // if the operation can be performed (see vfs.CheckSetStat() for common + // checks). + SetStat(ctx context.Context, fs *vfs.Filesystem, creds *auth.Credentials, opts vfs.SetStatOptions) error } // Precondition: All methods in this interface may only be called on directory @@ -416,7 +433,21 @@ type inodeDynamicLookup interface { } type inodeSymlink interface { - // Readlink resolves the target of a symbolic link. If an inode is not a + // Readlink returns the target of a symbolic link. If an inode is not a // symlink, the implementation should return EINVAL. Readlink(ctx context.Context) (string, error) + + // Getlink returns the target of a symbolic link, as used by path + // resolution: + // + // - If the inode is a "magic link" (a link whose target is most accurately + // represented as a VirtualDentry), Getlink returns (ok VirtualDentry, "", + // nil). A reference is taken on the returned VirtualDentry. + // + // - If the inode is an ordinary symlink, Getlink returns (zero-value + // VirtualDentry, symlink target, nil). + // + // - If the inode is not a symlink, Getlink returns (zero-value + // VirtualDentry, "", EINVAL). + Getlink(ctx context.Context, mnt *vfs.Mount) (vfs.VirtualDentry, string, error) } diff --git a/pkg/sentry/fsimpl/kernfs/kernfs_test.go b/pkg/sentry/fsimpl/kernfs/kernfs_test.go index 96a16e654..412cf6ac9 100644 --- a/pkg/sentry/fsimpl/kernfs/kernfs_test.go +++ b/pkg/sentry/fsimpl/kernfs/kernfs_test.go @@ -45,7 +45,10 @@ type RootDentryFn func(*auth.Credentials, *filesystem) *kernfs.Dentry func newTestSystem(t *testing.T, rootFn RootDentryFn) *testutil.System { ctx := contexttest.Context(t) creds := auth.CredentialsFromContext(ctx) - v := vfs.New() + v := &vfs.VirtualFilesystem{} + if err := v.Init(); err != nil { + t.Fatalf("VFS init: %v", err) + } v.MustRegisterFilesystemType("testfs", &fsType{rootFn: rootFn}, &vfs.RegisterFilesystemTypeOptions{ AllowUserMount: true, }) @@ -72,7 +75,7 @@ type file struct { func (fs *filesystem) newFile(creds *auth.Credentials, content string) *kernfs.Dentry { f := &file{} f.content = content - f.DynamicBytesFile.Init(creds, fs.NextIno(), f, 0777) + f.DynamicBytesFile.Init(creds, 0 /* devMajor */, 0 /* devMinor */, fs.NextIno(), f, 0777) d := &kernfs.Dentry{} d.Init(f) @@ -88,7 +91,7 @@ type attrs struct { kernfs.InodeAttrs } -func (a *attrs) SetStat(fs *vfs.Filesystem, opt vfs.SetStatOptions) error { +func (*attrs) SetStat(context.Context, *vfs.Filesystem, *auth.Credentials, vfs.SetStatOptions) error { return syserror.EPERM } @@ -104,7 +107,7 @@ type readonlyDir struct { func (fs *filesystem) newReadonlyDir(creds *auth.Credentials, mode linux.FileMode, contents map[string]*kernfs.Dentry) *kernfs.Dentry { dir := &readonlyDir{} - dir.attrs.Init(creds, fs.NextIno(), linux.ModeDirectory|mode) + dir.attrs.Init(creds, 0 /* devMajor */, 0 /* devMinor */, fs.NextIno(), linux.ModeDirectory|mode) dir.OrderedChildren.Init(kernfs.OrderedChildrenOptions{}) dir.dentry.Init(dir) @@ -113,9 +116,9 @@ func (fs *filesystem) newReadonlyDir(creds *auth.Credentials, mode linux.FileMod return &dir.dentry } -func (d *readonlyDir) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { - fd := &kernfs.GenericDirectoryFD{} - if err := fd.Init(rp.Mount(), vfsd, &d.OrderedChildren, &opts); err != nil { +func (d *readonlyDir) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), vfsd, &d.OrderedChildren, &opts) + if err != nil { return nil, err } return fd.VFSFileDescription(), nil @@ -134,7 +137,7 @@ type dir struct { func (fs *filesystem) newDir(creds *auth.Credentials, mode linux.FileMode, contents map[string]*kernfs.Dentry) *kernfs.Dentry { dir := &dir{} dir.fs = fs - dir.attrs.Init(creds, fs.NextIno(), linux.ModeDirectory|mode) + dir.attrs.Init(creds, 0 /* devMajor */, 0 /* devMinor */, fs.NextIno(), linux.ModeDirectory|mode) dir.OrderedChildren.Init(kernfs.OrderedChildrenOptions{Writable: true}) dir.dentry.Init(dir) @@ -143,9 +146,11 @@ func (fs *filesystem) newDir(creds *auth.Credentials, mode linux.FileMode, conte return &dir.dentry } -func (d *dir) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { - fd := &kernfs.GenericDirectoryFD{} - fd.Init(rp.Mount(), vfsd, &d.OrderedChildren, &opts) +func (d *dir) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), vfsd, &d.OrderedChildren, &opts) + if err != nil { + return nil, err + } return fd.VFSFileDescription(), nil } @@ -184,9 +189,13 @@ func (*dir) NewNode(context.Context, string, vfs.MknodOptions) (*vfs.Dentry, err return nil, syserror.EPERM } -func (fst *fsType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opt vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { +func (fsType) Name() string { + return "kernfs" +} + +func (fst fsType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opt vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { fs := &filesystem{} - fs.Init(vfsObj) + fs.VFSFilesystem().Init(vfsObj, &fst, fs) root := fst.rootFn(creds, fs) return fs.VFSFilesystem(), root.VFSDentry(), nil } diff --git a/pkg/sentry/fsimpl/kernfs/symlink.go b/pkg/sentry/fsimpl/kernfs/symlink.go index 0ee7eb9b7..2ab3f53fd 100644 --- a/pkg/sentry/fsimpl/kernfs/symlink.go +++ b/pkg/sentry/fsimpl/kernfs/symlink.go @@ -18,6 +18,8 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" ) // StaticSymlink provides an Inode implementation for symlinks that point to @@ -33,9 +35,9 @@ type StaticSymlink struct { var _ Inode = (*StaticSymlink)(nil) // NewStaticSymlink creates a new symlink file pointing to 'target'. -func NewStaticSymlink(creds *auth.Credentials, ino uint64, target string) *Dentry { +func NewStaticSymlink(creds *auth.Credentials, devMajor, devMinor uint32, ino uint64, target string) *Dentry { inode := &StaticSymlink{} - inode.Init(creds, ino, target) + inode.Init(creds, devMajor, devMinor, ino, target) d := &Dentry{} d.Init(inode) @@ -43,12 +45,22 @@ func NewStaticSymlink(creds *auth.Credentials, ino uint64, target string) *Dentr } // Init initializes the instance. -func (s *StaticSymlink) Init(creds *auth.Credentials, ino uint64, target string) { +func (s *StaticSymlink) Init(creds *auth.Credentials, devMajor uint32, devMinor uint32, ino uint64, target string) { s.target = target - s.InodeAttrs.Init(creds, ino, linux.ModeSymlink|0777) + s.InodeAttrs.Init(creds, devMajor, devMinor, ino, linux.ModeSymlink|0777) } // Readlink implements Inode. func (s *StaticSymlink) Readlink(_ context.Context) (string, error) { return s.target, nil } + +// Getlink implements Inode.Getlink. +func (s *StaticSymlink) Getlink(context.Context, *vfs.Mount) (vfs.VirtualDentry, string, error) { + return vfs.VirtualDentry{}, s.target, nil +} + +// SetStat implements Inode.SetStat not allowing inode attributes to be changed. +func (*StaticSymlink) SetStat(context.Context, *vfs.Filesystem, *auth.Credentials, vfs.SetStatOptions) error { + return syserror.EPERM +} diff --git a/pkg/sentry/fsimpl/pipefs/BUILD b/pkg/sentry/fsimpl/pipefs/BUILD new file mode 100644 index 000000000..5950a2d59 --- /dev/null +++ b/pkg/sentry/fsimpl/pipefs/BUILD @@ -0,0 +1,21 @@ +load("//tools:defs.bzl", "go_library") + +licenses(["notice"]) + +go_library( + name = "pipefs", + srcs = ["pipefs.go"], + visibility = ["//pkg/sentry:internal"], + deps = [ + "//pkg/abi/linux", + "//pkg/context", + "//pkg/fspath", + "//pkg/sentry/fsimpl/kernfs", + "//pkg/sentry/kernel/auth", + "//pkg/sentry/kernel/pipe", + "//pkg/sentry/kernel/time", + "//pkg/sentry/vfs", + "//pkg/syserror", + "//pkg/usermem", + ], +) diff --git a/pkg/sentry/fsimpl/pipefs/pipefs.go b/pkg/sentry/fsimpl/pipefs/pipefs.go new file mode 100644 index 000000000..cab771211 --- /dev/null +++ b/pkg/sentry/fsimpl/pipefs/pipefs.go @@ -0,0 +1,164 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// Package pipefs provides the filesystem implementation backing +// Kernel.PipeMount. +package pipefs + +import ( + "fmt" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/kernel/pipe" + ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +type filesystemType struct{} + +// Name implements vfs.FilesystemType.Name. +func (filesystemType) Name() string { + return "pipefs" +} + +// GetFilesystem implements vfs.FilesystemType.GetFilesystem. +func (filesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { + panic("pipefs.filesystemType.GetFilesystem should never be called") +} + +type filesystem struct { + kernfs.Filesystem + + devMinor uint32 +} + +// NewFilesystem sets up and returns a new vfs.Filesystem implemented by pipefs. +func NewFilesystem(vfsObj *vfs.VirtualFilesystem) (*vfs.Filesystem, error) { + devMinor, err := vfsObj.GetAnonBlockDevMinor() + if err != nil { + return nil, err + } + fs := &filesystem{ + devMinor: devMinor, + } + fs.Filesystem.VFSFilesystem().Init(vfsObj, filesystemType{}, fs) + return fs.Filesystem.VFSFilesystem(), nil +} + +// Release implements vfs.FilesystemImpl.Release. +func (fs *filesystem) Release() { + fs.Filesystem.VFSFilesystem().VirtualFilesystem().PutAnonBlockDevMinor(fs.devMinor) + fs.Filesystem.Release() +} + +// PrependPath implements vfs.FilesystemImpl.PrependPath. +func (fs *filesystem) PrependPath(ctx context.Context, vfsroot, vd vfs.VirtualDentry, b *fspath.Builder) error { + inode := vd.Dentry().Impl().(*kernfs.Dentry).Inode().(*inode) + b.PrependComponent(fmt.Sprintf("pipe:[%d]", inode.ino)) + return vfs.PrependPathSyntheticError{} +} + +// inode implements kernfs.Inode. +type inode struct { + kernfs.InodeNotDirectory + kernfs.InodeNotSymlink + kernfs.InodeNoopRefCount + + pipe *pipe.VFSPipe + + ino uint64 + uid auth.KUID + gid auth.KGID + // We use the creation timestamp for all of atime, mtime, and ctime. + ctime ktime.Time +} + +func newInode(ctx context.Context, fs *filesystem) *inode { + creds := auth.CredentialsFromContext(ctx) + return &inode{ + pipe: pipe.NewVFSPipe(false /* isNamed */, pipe.DefaultPipeSize, usermem.PageSize), + ino: fs.Filesystem.NextIno(), + uid: creds.EffectiveKUID, + gid: creds.EffectiveKGID, + ctime: ktime.NowFromContext(ctx), + } +} + +const pipeMode = 0600 | linux.S_IFIFO + +// CheckPermissions implements kernfs.Inode.CheckPermissions. +func (i *inode) CheckPermissions(ctx context.Context, creds *auth.Credentials, ats vfs.AccessTypes) error { + return vfs.GenericCheckPermissions(creds, ats, pipeMode, i.uid, i.gid) +} + +// Mode implements kernfs.Inode.Mode. +func (i *inode) Mode() linux.FileMode { + return pipeMode +} + +// Stat implements kernfs.Inode.Stat. +func (i *inode) Stat(vfsfs *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) { + ts := linux.NsecToStatxTimestamp(i.ctime.Nanoseconds()) + return linux.Statx{ + Mask: linux.STATX_TYPE | linux.STATX_MODE | linux.STATX_NLINK | linux.STATX_UID | linux.STATX_GID | linux.STATX_ATIME | linux.STATX_MTIME | linux.STATX_CTIME | linux.STATX_INO | linux.STATX_SIZE | linux.STATX_BLOCKS, + Blksize: usermem.PageSize, + Nlink: 1, + UID: uint32(i.uid), + GID: uint32(i.gid), + Mode: pipeMode, + Ino: i.ino, + Size: 0, + Blocks: 0, + Atime: ts, + Ctime: ts, + Mtime: ts, + DevMajor: linux.UNNAMED_MAJOR, + DevMinor: vfsfs.Impl().(*filesystem).devMinor, + }, nil +} + +// SetStat implements kernfs.Inode.SetStat. +func (i *inode) SetStat(ctx context.Context, vfsfs *vfs.Filesystem, creds *auth.Credentials, opts vfs.SetStatOptions) error { + if opts.Stat.Mask == 0 { + return nil + } + return syserror.EPERM +} + +// TODO(gvisor.dev/issue/1193): kernfs does not provide a way to implement +// statfs, from which we should indicate PIPEFS_MAGIC. + +// Open implements kernfs.Inode.Open. +func (i *inode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + return i.pipe.Open(ctx, rp.Mount(), vfsd, opts.Flags) +} + +// NewConnectedPipeFDs returns a pair of FileDescriptions representing the read +// and write ends of a newly-created pipe, as for pipe(2) and pipe2(2). +// +// Preconditions: mnt.Filesystem() must have been returned by NewFilesystem(). +func NewConnectedPipeFDs(ctx context.Context, mnt *vfs.Mount, flags uint32) (*vfs.FileDescription, *vfs.FileDescription) { + fs := mnt.Filesystem().Impl().(*filesystem) + inode := newInode(ctx, fs) + var d kernfs.Dentry + d.Init(inode) + defer d.DecRef() + return inode.pipe.ReaderWriterPair(mnt, d.VFSDentry(), flags) +} diff --git a/pkg/sentry/fsimpl/proc/BUILD b/pkg/sentry/fsimpl/proc/BUILD index 12aac2e6a..17c1342b5 100644 --- a/pkg/sentry/fsimpl/proc/BUILD +++ b/pkg/sentry/fsimpl/proc/BUILD @@ -8,18 +8,21 @@ go_library( "filesystem.go", "subtasks.go", "task.go", + "task_fds.go", "task_files.go", + "task_net.go", "tasks.go", "tasks_files.go", - "tasks_net.go", "tasks_sys.go", ], + visibility = ["//pkg/sentry:internal"], deps = [ "//pkg/abi/linux", "//pkg/context", "//pkg/log", + "//pkg/refs", "//pkg/safemem", - "//pkg/sentry/fs", + "//pkg/sentry/fsbridge", "//pkg/sentry/fsimpl/kernfs", "//pkg/sentry/inet", "//pkg/sentry/kernel", @@ -52,6 +55,7 @@ go_test( "//pkg/fspath", "//pkg/sentry/contexttest", "//pkg/sentry/fsimpl/testutil", + "//pkg/sentry/fsimpl/tmpfs", "//pkg/sentry/inet", "//pkg/sentry/kernel", "//pkg/sentry/kernel/auth", diff --git a/pkg/sentry/fsimpl/proc/filesystem.go b/pkg/sentry/fsimpl/proc/filesystem.go index 11477b6a9..609210253 100644 --- a/pkg/sentry/fsimpl/proc/filesystem.go +++ b/pkg/sentry/fsimpl/proc/filesystem.go @@ -26,15 +26,29 @@ import ( "gvisor.dev/gvisor/pkg/sentry/vfs" ) -// procFSType is the factory class for procfs. +// Name is the default filesystem name. +const Name = "proc" + +// FilesystemType is the factory class for procfs. // // +stateify savable -type procFSType struct{} +type FilesystemType struct{} + +var _ vfs.FilesystemType = (*FilesystemType)(nil) + +// Name implements vfs.FilesystemType.Name. +func (FilesystemType) Name() string { + return Name +} + +type filesystem struct { + kernfs.Filesystem -var _ vfs.FilesystemType = (*procFSType)(nil) + devMinor uint32 +} -// GetFilesystem implements vfs.FilesystemType. -func (ft *procFSType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { +// GetFilesystem implements vfs.FilesystemType.GetFilesystem. +func (ft FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { k := kernel.KernelFromContext(ctx) if k == nil { return nil, nil, fmt.Errorf("procfs requires a kernel") @@ -43,30 +57,42 @@ func (ft *procFSType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFile if pidns == nil { return nil, nil, fmt.Errorf("procfs requires a PID namespace") } + devMinor, err := vfsObj.GetAnonBlockDevMinor() + if err != nil { + return nil, nil, err + } + procfs := &filesystem{ + devMinor: devMinor, + } + procfs.VFSFilesystem().Init(vfsObj, &ft, procfs) - procfs := &kernfs.Filesystem{} - procfs.VFSFilesystem().Init(vfsObj, procfs) - - var data *InternalData + var cgroups map[string]string if opts.InternalData != nil { - data = opts.InternalData.(*InternalData) + data := opts.InternalData.(*InternalData) + cgroups = data.Cgroups } - _, dentry := newTasksInode(procfs, k, pidns, data.Cgroups) + _, dentry := procfs.newTasksInode(k, pidns, cgroups) return procfs.VFSFilesystem(), dentry.VFSDentry(), nil } +// Release implements vfs.FilesystemImpl.Release. +func (fs *filesystem) Release() { + fs.Filesystem.VFSFilesystem().VirtualFilesystem().PutAnonBlockDevMinor(fs.devMinor) + fs.Filesystem.Release() +} + // dynamicInode is an overfitted interface for common Inodes with // dynamicByteSource types used in procfs. type dynamicInode interface { kernfs.Inode vfs.DynamicBytesSource - Init(creds *auth.Credentials, ino uint64, data vfs.DynamicBytesSource, perm linux.FileMode) + Init(creds *auth.Credentials, devMajor, devMinor uint32, ino uint64, data vfs.DynamicBytesSource, perm linux.FileMode) } -func newDentry(creds *auth.Credentials, ino uint64, perm linux.FileMode, inode dynamicInode) *kernfs.Dentry { - inode.Init(creds, ino, inode, perm) +func (fs *filesystem) newDentry(creds *auth.Credentials, ino uint64, perm linux.FileMode, inode dynamicInode) *kernfs.Dentry { + inode.Init(creds, linux.UNNAMED_MAJOR, fs.devMinor, ino, inode, perm) d := &kernfs.Dentry{} d.Init(inode) diff --git a/pkg/sentry/fsimpl/proc/subtasks.go b/pkg/sentry/fsimpl/proc/subtasks.go index 102af0e93..36a911db4 100644 --- a/pkg/sentry/fsimpl/proc/subtasks.go +++ b/pkg/sentry/fsimpl/proc/subtasks.go @@ -22,6 +22,7 @@ import ( "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs" "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserror" ) @@ -34,24 +35,25 @@ type subtasksInode struct { kernfs.InodeDirectoryNoNewChildren kernfs.InodeAttrs kernfs.OrderedChildren + kernfs.AlwaysValid + fs *filesystem task *kernel.Task pidns *kernel.PIDNamespace - inoGen InoGenerator cgroupControllers map[string]string } var _ kernfs.Inode = (*subtasksInode)(nil) -func newSubtasks(task *kernel.Task, pidns *kernel.PIDNamespace, inoGen InoGenerator, cgroupControllers map[string]string) *kernfs.Dentry { +func (fs *filesystem) newSubtasks(task *kernel.Task, pidns *kernel.PIDNamespace, cgroupControllers map[string]string) *kernfs.Dentry { subInode := &subtasksInode{ + fs: fs, task: task, pidns: pidns, - inoGen: inoGen, cgroupControllers: cgroupControllers, } // Note: credentials are overridden by taskOwnedInode. - subInode.InodeAttrs.Init(task.Credentials(), inoGen.NextIno(), linux.ModeDirectory|0555) + subInode.InodeAttrs.Init(task.Credentials(), linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), linux.ModeDirectory|0555) subInode.OrderedChildren.Init(kernfs.OrderedChildrenOptions{}) inode := &taskOwnedInode{Inode: subInode, owner: task} @@ -61,11 +63,6 @@ func newSubtasks(task *kernel.Task, pidns *kernel.PIDNamespace, inoGen InoGenera return dentry } -// Valid implements kernfs.inodeDynamicLookup. -func (i *subtasksInode) Valid(ctx context.Context) bool { - return true -} - // Lookup implements kernfs.inodeDynamicLookup. func (i *subtasksInode) Lookup(ctx context.Context, name string) (*vfs.Dentry, error) { tid, err := strconv.ParseUint(name, 10, 32) @@ -81,7 +78,7 @@ func (i *subtasksInode) Lookup(ctx context.Context, name string) (*vfs.Dentry, e return nil, syserror.ENOENT } - subTaskDentry := newTaskInode(i.inoGen, subTask, i.pidns, false, i.cgroupControllers) + subTaskDentry := i.fs.newTaskInode(subTask, i.pidns, false, i.cgroupControllers) return subTaskDentry.VFSDentry(), nil } @@ -91,6 +88,9 @@ func (i *subtasksInode) IterDirents(ctx context.Context, cb vfs.IterDirentsCallb if len(tasks) == 0 { return offset, syserror.ENOENT } + if relOffset >= int64(len(tasks)) { + return offset, nil + } tids := make([]int, 0, len(tasks)) for _, tid := range tasks { @@ -102,27 +102,79 @@ func (i *subtasksInode) IterDirents(ctx context.Context, cb vfs.IterDirentsCallb dirent := vfs.Dirent{ Name: strconv.FormatUint(uint64(tid), 10), Type: linux.DT_DIR, - Ino: i.inoGen.NextIno(), + Ino: i.fs.NextIno(), NextOff: offset + 1, } - if !cb.Handle(dirent) { - return offset, nil + if err := cb.Handle(dirent); err != nil { + return offset, err } offset++ } return offset, nil } +type subtasksFD struct { + kernfs.GenericDirectoryFD + + task *kernel.Task +} + +func (fd *subtasksFD) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback) error { + if fd.task.ExitState() >= kernel.TaskExitZombie { + return syserror.ENOENT + } + return fd.GenericDirectoryFD.IterDirents(ctx, cb) +} + +// Seek implements vfs.FileDecriptionImpl.Seek. +func (fd *subtasksFD) Seek(ctx context.Context, offset int64, whence int32) (int64, error) { + if fd.task.ExitState() >= kernel.TaskExitZombie { + return 0, syserror.ENOENT + } + return fd.GenericDirectoryFD.Seek(ctx, offset, whence) +} + +// Stat implements vfs.FileDescriptionImpl.Stat. +func (fd *subtasksFD) Stat(ctx context.Context, opts vfs.StatOptions) (linux.Statx, error) { + if fd.task.ExitState() >= kernel.TaskExitZombie { + return linux.Statx{}, syserror.ENOENT + } + return fd.GenericDirectoryFD.Stat(ctx, opts) +} + +// SetStat implements vfs.FileDescriptionImpl.SetStat. +func (fd *subtasksFD) SetStat(ctx context.Context, opts vfs.SetStatOptions) error { + if fd.task.ExitState() >= kernel.TaskExitZombie { + return syserror.ENOENT + } + return fd.GenericDirectoryFD.SetStat(ctx, opts) +} + // Open implements kernfs.Inode. -func (i *subtasksInode) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { - fd := &kernfs.GenericDirectoryFD{} - fd.Init(rp.Mount(), vfsd, &i.OrderedChildren, &opts) +func (i *subtasksInode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd := &subtasksFD{task: i.task} + if err := fd.Init(&i.OrderedChildren, &opts); err != nil { + return nil, err + } + if err := fd.VFSFileDescription().Init(fd, opts.Flags, rp.Mount(), vfsd, &vfs.FileDescriptionOptions{}); err != nil { + return nil, err + } return fd.VFSFileDescription(), nil } // Stat implements kernfs.Inode. -func (i *subtasksInode) Stat(vsfs *vfs.Filesystem) linux.Statx { - stat := i.InodeAttrs.Stat(vsfs) - stat.Nlink += uint32(i.task.ThreadGroup().Count()) - return stat +func (i *subtasksInode) Stat(vsfs *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) { + stat, err := i.InodeAttrs.Stat(vsfs, opts) + if err != nil { + return linux.Statx{}, err + } + if opts.Mask&linux.STATX_NLINK != 0 { + stat.Nlink += uint32(i.task.ThreadGroup().Count()) + } + return stat, nil +} + +// SetStat implements Inode.SetStat not allowing inode attributes to be changed. +func (*subtasksInode) SetStat(context.Context, *vfs.Filesystem, *auth.Credentials, vfs.SetStatOptions) error { + return syserror.EPERM } diff --git a/pkg/sentry/fsimpl/proc/task.go b/pkg/sentry/fsimpl/proc/task.go index 2d814668a..482055db1 100644 --- a/pkg/sentry/fsimpl/proc/task.go +++ b/pkg/sentry/fsimpl/proc/task.go @@ -43,41 +43,45 @@ type taskInode struct { var _ kernfs.Inode = (*taskInode)(nil) -func newTaskInode(inoGen InoGenerator, task *kernel.Task, pidns *kernel.PIDNamespace, isThreadGroup bool, cgroupControllers map[string]string) *kernfs.Dentry { +func (fs *filesystem) newTaskInode(task *kernel.Task, pidns *kernel.PIDNamespace, isThreadGroup bool, cgroupControllers map[string]string) *kernfs.Dentry { + // TODO(gvisor.dev/issue/164): Fail with ESRCH if task exited. contents := map[string]*kernfs.Dentry{ - "auxv": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &auxvData{task: task}), - "cmdline": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &cmdlineData{task: task, arg: cmdlineDataArg}), - "comm": newComm(task, inoGen.NextIno(), 0444), - "environ": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &cmdlineData{task: task, arg: environDataArg}), - //"exe": newExe(t, msrc), - //"fd": newFdDir(t, msrc), - //"fdinfo": newFdInfoDir(t, msrc), - "gid_map": newTaskOwnedFile(task, inoGen.NextIno(), 0644, &idMapData{task: task, gids: true}), - "io": newTaskOwnedFile(task, inoGen.NextIno(), 0400, newIO(task, isThreadGroup)), - "maps": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &mapsData{task: task}), - //"mountinfo": seqfile.NewSeqFileInode(t, &mountInfoFile{t: t}, msrc), - //"mounts": seqfile.NewSeqFileInode(t, &mountsFile{t: t}, msrc), - "ns": newTaskOwnedDir(task, inoGen.NextIno(), 0511, map[string]*kernfs.Dentry{ - "net": newNamespaceSymlink(task, inoGen.NextIno(), "net"), - "pid": newNamespaceSymlink(task, inoGen.NextIno(), "pid"), - "user": newNamespaceSymlink(task, inoGen.NextIno(), "user"), + "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), + "environ": fs.newTaskOwnedFile(task, fs.NextIno(), 0444, &cmdlineData{task: task, arg: environDataArg}), + "exe": fs.newExeSymlink(task, fs.NextIno()), + "fd": fs.newFDDirInode(task), + "fdinfo": fs.newFDInfoDirInode(task), + "gid_map": fs.newTaskOwnedFile(task, fs.NextIno(), 0644, &idMapData{task: task, gids: true}), + "io": fs.newTaskOwnedFile(task, fs.NextIno(), 0400, newIO(task, isThreadGroup)), + "maps": fs.newTaskOwnedFile(task, fs.NextIno(), 0444, &mapsData{task: task}), + "mountinfo": fs.newTaskOwnedFile(task, fs.NextIno(), 0444, &mountInfoData{task: task}), + "mounts": fs.newTaskOwnedFile(task, fs.NextIno(), 0444, &mountsData{task: task}), + "net": fs.newTaskNetDir(task), + "ns": fs.newTaskOwnedDir(task, fs.NextIno(), 0511, map[string]*kernfs.Dentry{ + "net": fs.newNamespaceSymlink(task, fs.NextIno(), "net"), + "pid": fs.newNamespaceSymlink(task, fs.NextIno(), "pid"), + "user": fs.newNamespaceSymlink(task, fs.NextIno(), "user"), }), - "smaps": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &smapsData{task: task}), - "stat": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &taskStatData{task: task, pidns: pidns, tgstats: isThreadGroup}), - "statm": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &statmData{task: task}), - "status": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &statusData{task: task, pidns: pidns}), - "uid_map": newTaskOwnedFile(task, inoGen.NextIno(), 0644, &idMapData{task: task, gids: false}), + "oom_score": fs.newTaskOwnedFile(task, fs.NextIno(), 0444, newStaticFile("0\n")), + "oom_score_adj": fs.newTaskOwnedFile(task, fs.NextIno(), 0644, &oomScoreAdj{task: task}), + "smaps": fs.newTaskOwnedFile(task, fs.NextIno(), 0444, &smapsData{task: task}), + "stat": fs.newTaskOwnedFile(task, fs.NextIno(), 0444, &taskStatData{task: task, pidns: pidns, tgstats: isThreadGroup}), + "statm": fs.newTaskOwnedFile(task, fs.NextIno(), 0444, &statmData{task: task}), + "status": fs.newTaskOwnedFile(task, fs.NextIno(), 0444, &statusData{task: task, pidns: pidns}), + "uid_map": fs.newTaskOwnedFile(task, fs.NextIno(), 0644, &idMapData{task: task, gids: false}), } if isThreadGroup { - contents["task"] = newSubtasks(task, pidns, inoGen, cgroupControllers) + contents["task"] = fs.newSubtasks(task, pidns, cgroupControllers) } if len(cgroupControllers) > 0 { - contents["cgroup"] = newTaskOwnedFile(task, inoGen.NextIno(), 0444, newCgroupData(cgroupControllers)) + contents["cgroup"] = fs.newTaskOwnedFile(task, fs.NextIno(), 0444, newCgroupData(cgroupControllers)) } taskInode := &taskInode{task: task} // Note: credentials are overridden by taskOwnedInode. - taskInode.InodeAttrs.Init(task.Credentials(), inoGen.NextIno(), linux.ModeDirectory|0555) + taskInode.InodeAttrs.Init(task.Credentials(), linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), linux.ModeDirectory|0555) inode := &taskOwnedInode{Inode: taskInode, owner: task} dentry := &kernfs.Dentry{} @@ -98,19 +102,17 @@ func (i *taskInode) Valid(ctx context.Context) bool { } // Open implements kernfs.Inode. -func (i *taskInode) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { - fd := &kernfs.GenericDirectoryFD{} - fd.Init(rp.Mount(), vfsd, &i.OrderedChildren, &opts) +func (i *taskInode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), vfsd, &i.OrderedChildren, &opts) + if err != nil { + return nil, err + } return fd.VFSFileDescription(), nil } -// SetStat implements kernfs.Inode. -func (i *taskInode) SetStat(_ *vfs.Filesystem, opts vfs.SetStatOptions) error { - stat := opts.Stat - if stat.Mask&linux.STATX_MODE != 0 { - return syserror.EPERM - } - return nil +// SetStat implements Inode.SetStat not allowing inode attributes to be changed. +func (*taskInode) SetStat(context.Context, *vfs.Filesystem, *auth.Credentials, vfs.SetStatOptions) error { + return syserror.EPERM } // taskOwnedInode implements kernfs.Inode and overrides inode owner with task @@ -124,9 +126,9 @@ type taskOwnedInode struct { var _ kernfs.Inode = (*taskOwnedInode)(nil) -func newTaskOwnedFile(task *kernel.Task, ino uint64, perm linux.FileMode, inode dynamicInode) *kernfs.Dentry { +func (fs *filesystem) newTaskOwnedFile(task *kernel.Task, ino uint64, perm linux.FileMode, inode dynamicInode) *kernfs.Dentry { // Note: credentials are overridden by taskOwnedInode. - inode.Init(task.Credentials(), ino, inode, perm) + inode.Init(task.Credentials(), linux.UNNAMED_MAJOR, fs.devMinor, ino, inode, perm) taskInode := &taskOwnedInode{Inode: inode, owner: task} d := &kernfs.Dentry{} @@ -134,11 +136,11 @@ func newTaskOwnedFile(task *kernel.Task, ino uint64, perm linux.FileMode, inode return d } -func newTaskOwnedDir(task *kernel.Task, ino uint64, perm linux.FileMode, children map[string]*kernfs.Dentry) *kernfs.Dentry { +func (fs *filesystem) newTaskOwnedDir(task *kernel.Task, ino uint64, perm linux.FileMode, children map[string]*kernfs.Dentry) *kernfs.Dentry { dir := &kernfs.StaticDirectory{} // Note: credentials are overridden by taskOwnedInode. - dir.Init(task.Credentials(), ino, perm) + dir.Init(task.Credentials(), linux.UNNAMED_MAJOR, fs.devMinor, ino, perm) inode := &taskOwnedInode{Inode: dir, owner: task} d := &kernfs.Dentry{} @@ -152,26 +154,28 @@ func newTaskOwnedDir(task *kernel.Task, ino uint64, perm linux.FileMode, childre } // Stat implements kernfs.Inode. -func (i *taskOwnedInode) Stat(fs *vfs.Filesystem) linux.Statx { - stat := i.Inode.Stat(fs) - uid, gid := i.getOwner(linux.FileMode(stat.Mode)) - stat.UID = uint32(uid) - stat.GID = uint32(gid) - return stat +func (i *taskOwnedInode) Stat(fs *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) { + stat, err := i.Inode.Stat(fs, opts) + if err != nil { + return linux.Statx{}, err + } + if opts.Mask&(linux.STATX_UID|linux.STATX_GID) != 0 { + uid, gid := i.getOwner(linux.FileMode(stat.Mode)) + if opts.Mask&linux.STATX_UID != 0 { + stat.UID = uint32(uid) + } + if opts.Mask&linux.STATX_GID != 0 { + stat.GID = uint32(gid) + } + } + return stat, nil } // CheckPermissions implements kernfs.Inode. func (i *taskOwnedInode) CheckPermissions(_ context.Context, creds *auth.Credentials, ats vfs.AccessTypes) error { mode := i.Mode() uid, gid := i.getOwner(mode) - return vfs.GenericCheckPermissions( - creds, - ats, - mode.FileType() == linux.ModeDirectory, - uint16(mode), - uid, - gid, - ) + return vfs.GenericCheckPermissions(creds, ats, mode, uid, gid) } func (i *taskOwnedInode) getOwner(mode linux.FileMode) (auth.KUID, auth.KGID) { @@ -213,28 +217,12 @@ func newIO(t *kernel.Task, isThreadGroup bool) *ioData { return &ioData{ioUsage: t} } -func newNamespaceSymlink(task *kernel.Task, ino uint64, ns string) *kernfs.Dentry { - // Namespace symlinks should contain the namespace name and the inode number - // for the namespace instance, so for example user:[123456]. We currently fake - // the inode number by sticking the symlink inode in its place. - target := fmt.Sprintf("%s:[%d]", ns, ino) - - inode := &kernfs.StaticSymlink{} - // Note: credentials are overridden by taskOwnedInode. - inode.Init(task.Credentials(), ino, target) - - taskInode := &taskOwnedInode{Inode: inode, owner: task} - d := &kernfs.Dentry{} - d.Init(taskInode) - return d -} - // newCgroupData creates inode that shows cgroup information. // From man 7 cgroups: "For each cgroup hierarchy of which the process is a // member, there is one entry containing three colon-separated fields: // hierarchy-ID:controller-list:cgroup-path" func newCgroupData(controllers map[string]string) dynamicInode { - buf := bytes.Buffer{} + var buf bytes.Buffer // The hierarchy ids must be positive integers (for cgroup v1), but the // exact number does not matter, so long as they are unique. We can diff --git a/pkg/sentry/fsimpl/proc/task_fds.go b/pkg/sentry/fsimpl/proc/task_fds.go new file mode 100644 index 000000000..44ccc9e4a --- /dev/null +++ b/pkg/sentry/fsimpl/proc/task_fds.go @@ -0,0 +1,306 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package proc + +import ( + "bytes" + "fmt" + "sort" + "strconv" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/refs" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" +) + +func getTaskFD(t *kernel.Task, fd int32) (*vfs.FileDescription, kernel.FDFlags) { + var ( + file *vfs.FileDescription + flags kernel.FDFlags + ) + t.WithMuLocked(func(t *kernel.Task) { + if fdt := t.FDTable(); fdt != nil { + file, flags = fdt.GetVFS2(fd) + } + }) + return file, flags +} + +func taskFDExists(t *kernel.Task, fd int32) bool { + file, _ := getTaskFD(t, fd) + if file == nil { + return false + } + file.DecRef() + return true +} + +type fdDir struct { + fs *filesystem + task *kernel.Task + + // When produceSymlinks is set, dirents produces for the FDs are reported + // as symlink. Otherwise, they are reported as regular files. + produceSymlink bool +} + +// IterDirents implements kernfs.inodeDynamicLookup. +func (i *fdDir) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback, absOffset, relOffset int64) (int64, error) { + var fds []int32 + i.task.WithMuLocked(func(t *kernel.Task) { + if fdTable := t.FDTable(); fdTable != nil { + fds = fdTable.GetFDs() + } + }) + + offset := absOffset + relOffset + typ := uint8(linux.DT_REG) + if i.produceSymlink { + typ = linux.DT_LNK + } + + // Find the appropriate starting point. + idx := sort.Search(len(fds), func(i int) bool { return fds[i] >= int32(relOffset) }) + if idx >= len(fds) { + return offset, nil + } + for _, fd := range fds[idx:] { + dirent := vfs.Dirent{ + Name: strconv.FormatUint(uint64(fd), 10), + Type: typ, + Ino: i.fs.NextIno(), + NextOff: offset + 1, + } + if err := cb.Handle(dirent); err != nil { + return offset, err + } + offset++ + } + return offset, nil +} + +// fdDirInode represents the inode for /proc/[pid]/fd directory. +// +// +stateify savable +type fdDirInode struct { + kernfs.InodeNotSymlink + kernfs.InodeDirectoryNoNewChildren + kernfs.InodeAttrs + kernfs.OrderedChildren + kernfs.AlwaysValid + fdDir +} + +var _ kernfs.Inode = (*fdDirInode)(nil) + +func (fs *filesystem) newFDDirInode(task *kernel.Task) *kernfs.Dentry { + inode := &fdDirInode{ + fdDir: fdDir{ + fs: fs, + task: task, + produceSymlink: true, + }, + } + inode.InodeAttrs.Init(task.Credentials(), linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), linux.ModeDirectory|0555) + + dentry := &kernfs.Dentry{} + dentry.Init(inode) + inode.OrderedChildren.Init(kernfs.OrderedChildrenOptions{}) + + return dentry +} + +// Lookup implements kernfs.inodeDynamicLookup. +func (i *fdDirInode) Lookup(ctx context.Context, name string) (*vfs.Dentry, error) { + fdInt, err := strconv.ParseInt(name, 10, 32) + if err != nil { + return nil, syserror.ENOENT + } + fd := int32(fdInt) + if !taskFDExists(i.task, fd) { + return nil, syserror.ENOENT + } + taskDentry := i.fs.newFDSymlink(i.task, fd, i.fs.NextIno()) + return taskDentry.VFSDentry(), nil +} + +// Open implements kernfs.Inode. +func (i *fdDirInode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), vfsd, &i.OrderedChildren, &opts) + if err != nil { + return nil, err + } + return fd.VFSFileDescription(), nil +} + +// CheckPermissions implements kernfs.Inode. +// +// This is to match Linux, which uses a special permission handler to guarantee +// that a process can still access /proc/self/fd after it has executed +// setuid. See fs/proc/fd.c:proc_fd_permission. +func (i *fdDirInode) CheckPermissions(ctx context.Context, creds *auth.Credentials, ats vfs.AccessTypes) error { + err := i.InodeAttrs.CheckPermissions(ctx, creds, ats) + if err == nil { + // Access granted, no extra check needed. + return nil + } + if t := kernel.TaskFromContext(ctx); t != nil { + // Allow access if the task trying to access it is in the thread group + // corresponding to this directory. + if i.task.ThreadGroup() == t.ThreadGroup() { + // Access granted (overridden). + return nil + } + } + return err +} + +// fdSymlink is an symlink for the /proc/[pid]/fd/[fd] file. +// +// +stateify savable +type fdSymlink struct { + kernfs.InodeAttrs + kernfs.InodeNoopRefCount + kernfs.InodeSymlink + + task *kernel.Task + fd int32 +} + +var _ kernfs.Inode = (*fdSymlink)(nil) + +func (fs *filesystem) newFDSymlink(task *kernel.Task, fd int32, ino uint64) *kernfs.Dentry { + inode := &fdSymlink{ + task: task, + fd: fd, + } + inode.Init(task.Credentials(), linux.UNNAMED_MAJOR, fs.devMinor, ino, linux.ModeSymlink|0777) + + d := &kernfs.Dentry{} + d.Init(inode) + return d +} + +func (s *fdSymlink) Readlink(ctx context.Context) (string, error) { + file, _ := getTaskFD(s.task, s.fd) + if file == nil { + return "", syserror.ENOENT + } + defer file.DecRef() + root := vfs.RootFromContext(ctx) + defer root.DecRef() + return s.task.Kernel().VFS().PathnameWithDeleted(ctx, root, file.VirtualDentry()) +} + +func (s *fdSymlink) Getlink(ctx context.Context, mnt *vfs.Mount) (vfs.VirtualDentry, string, error) { + file, _ := getTaskFD(s.task, s.fd) + if file == nil { + return vfs.VirtualDentry{}, "", syserror.ENOENT + } + defer file.DecRef() + vd := file.VirtualDentry() + vd.IncRef() + return vd, "", nil +} + +// fdInfoDirInode represents the inode for /proc/[pid]/fdinfo directory. +// +// +stateify savable +type fdInfoDirInode struct { + kernfs.InodeNotSymlink + kernfs.InodeDirectoryNoNewChildren + kernfs.InodeAttrs + kernfs.OrderedChildren + kernfs.AlwaysValid + fdDir +} + +var _ kernfs.Inode = (*fdInfoDirInode)(nil) + +func (fs *filesystem) newFDInfoDirInode(task *kernel.Task) *kernfs.Dentry { + inode := &fdInfoDirInode{ + fdDir: fdDir{ + fs: fs, + task: task, + }, + } + inode.InodeAttrs.Init(task.Credentials(), linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), linux.ModeDirectory|0555) + + dentry := &kernfs.Dentry{} + dentry.Init(inode) + inode.OrderedChildren.Init(kernfs.OrderedChildrenOptions{}) + + return dentry +} + +// Lookup implements kernfs.inodeDynamicLookup. +func (i *fdInfoDirInode) Lookup(ctx context.Context, name string) (*vfs.Dentry, error) { + fdInt, err := strconv.ParseInt(name, 10, 32) + if err != nil { + return nil, syserror.ENOENT + } + fd := int32(fdInt) + if !taskFDExists(i.task, fd) { + return nil, syserror.ENOENT + } + data := &fdInfoData{ + task: i.task, + fd: fd, + } + dentry := i.fs.newTaskOwnedFile(i.task, i.fs.NextIno(), 0444, data) + return dentry.VFSDentry(), nil +} + +// Open implements kernfs.Inode. +func (i *fdInfoDirInode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), vfsd, &i.OrderedChildren, &opts) + if err != nil { + return nil, err + } + return fd.VFSFileDescription(), nil +} + +// fdInfoData implements vfs.DynamicBytesSource for /proc/[pid]/fdinfo/[fd]. +// +// +stateify savable +type fdInfoData struct { + kernfs.DynamicBytesFile + refs.AtomicRefCount + + task *kernel.Task + fd int32 +} + +var _ dynamicInode = (*fdInfoData)(nil) + +// Generate implements vfs.DynamicBytesSource.Generate. +func (d *fdInfoData) Generate(ctx context.Context, buf *bytes.Buffer) error { + file, descriptorFlags := getTaskFD(d.task, d.fd) + if file == nil { + return syserror.ENOENT + } + defer file.DecRef() + // TODO(b/121266871): Include pos, locks, and other data. For now we only + // have flags. + // See https://www.kernel.org/doc/Documentation/filesystems/proc.txt + flags := uint(file.StatusFlags()) | descriptorFlags.ToLinuxFileFlags() + fmt.Fprintf(buf, "flags:\t0%o\n", flags) + return nil +} diff --git a/pkg/sentry/fsimpl/proc/task_files.go b/pkg/sentry/fsimpl/proc/task_files.go index efd3b3453..2f297e48a 100644 --- a/pkg/sentry/fsimpl/proc/task_files.go +++ b/pkg/sentry/fsimpl/proc/task_files.go @@ -22,6 +22,7 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/safemem" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" "gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" @@ -63,6 +64,16 @@ func getMMIncRef(task *kernel.Task) (*mm.MemoryManager, error) { return m, nil } +func checkTaskState(t *kernel.Task) error { + switch t.ExitState() { + case kernel.TaskExitZombie: + return syserror.EACCES + case kernel.TaskExitDead: + return syserror.ESRCH + } + return nil +} + type bufferWriter struct { buf *bytes.Buffer } @@ -100,17 +111,18 @@ func (d *auxvData) Generate(ctx context.Context, buf *bytes.Buffer) error { } defer m.DecUsers(ctx) - // Space for buffer with AT_NULL (0) terminator at the end. auxv := m.Auxv() + // Space for buffer with AT_NULL (0) terminator at the end. buf.Grow((len(auxv) + 1) * 16) for _, e := range auxv { - var tmp [8]byte - usermem.ByteOrder.PutUint64(tmp[:], e.Key) - buf.Write(tmp[:]) - - usermem.ByteOrder.PutUint64(tmp[:], uint64(e.Value)) + var tmp [16]byte + usermem.ByteOrder.PutUint64(tmp[:8], e.Key) + usermem.ByteOrder.PutUint64(tmp[8:], uint64(e.Value)) buf.Write(tmp[:]) } + var atNull [16]byte + buf.Write(atNull[:]) + return nil } @@ -229,9 +241,9 @@ type commInode struct { task *kernel.Task } -func newComm(task *kernel.Task, ino uint64, perm linux.FileMode) *kernfs.Dentry { +func (fs *filesystem) newComm(task *kernel.Task, ino uint64, perm linux.FileMode) *kernfs.Dentry { inode := &commInode{task: task} - inode.DynamicBytesFile.Init(task.Credentials(), ino, &commData{task: task}, perm) + inode.DynamicBytesFile.Init(task.Credentials(), linux.UNNAMED_MAJOR, fs.devMinor, ino, &commData{task: task}, perm) d := &kernfs.Dentry{} d.Init(inode) @@ -496,7 +508,7 @@ func (s *statusData) Generate(ctx context.Context, buf *bytes.Buffer) error { return nil } -// ioUsage is the /proc/<pid>/io and /proc/<pid>/task/<tid>/io data provider. +// ioUsage is the /proc/[pid]/io and /proc/[pid]/task/[tid]/io data provider. type ioUsage interface { // IOUsage returns the io usage data. IOUsage() *usage.IO @@ -525,3 +537,296 @@ func (i *ioData) Generate(ctx context.Context, buf *bytes.Buffer) error { fmt.Fprintf(buf, "cancelled_write_bytes: %d\n", io.BytesWriteCancelled) return nil } + +// oomScoreAdj is a stub of the /proc/<pid>/oom_score_adj file. +// +// +stateify savable +type oomScoreAdj struct { + kernfs.DynamicBytesFile + + task *kernel.Task +} + +var _ vfs.WritableDynamicBytesSource = (*oomScoreAdj)(nil) + +// Generate implements vfs.DynamicBytesSource.Generate. +func (o *oomScoreAdj) Generate(ctx context.Context, buf *bytes.Buffer) error { + if o.task.ExitState() == kernel.TaskExitDead { + return syserror.ESRCH + } + fmt.Fprintf(buf, "%d\n", o.task.OOMScoreAdj()) + return nil +} + +// Write implements vfs.WritableDynamicBytesSource.Write. +func (o *oomScoreAdj) Write(ctx context.Context, src usermem.IOSequence, offset int64) (int64, error) { + if src.NumBytes() == 0 { + return 0, nil + } + + // Limit input size so as not to impact performance if input size is large. + src = src.TakeFirst(usermem.PageSize - 1) + + var v int32 + n, err := usermem.CopyInt32StringInVec(ctx, src.IO, src.Addrs, &v, src.Opts) + if err != nil { + return 0, err + } + + if o.task.ExitState() == kernel.TaskExitDead { + return 0, syserror.ESRCH + } + if err := o.task.SetOOMScoreAdj(v); err != nil { + return 0, err + } + + return n, nil +} + +// exeSymlink is an symlink for the /proc/[pid]/exe file. +// +// +stateify savable +type exeSymlink struct { + kernfs.InodeAttrs + kernfs.InodeNoopRefCount + kernfs.InodeSymlink + + task *kernel.Task +} + +var _ kernfs.Inode = (*exeSymlink)(nil) + +func (fs *filesystem) newExeSymlink(task *kernel.Task, ino uint64) *kernfs.Dentry { + inode := &exeSymlink{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. +func (s *exeSymlink) Readlink(ctx context.Context) (string, error) { + if !kernel.ContextCanTrace(ctx, s.task, false) { + return "", syserror.EACCES + } + + // Pull out the executable for /proc/[pid]/exe. + exec, err := s.executable() + if err != nil { + return "", err + } + defer exec.DecRef() + + return exec.PathnameWithDeleted(ctx), nil +} + +// Getlink implements kernfs.Inode.Getlink. +func (s *exeSymlink) Getlink(ctx context.Context, _ *vfs.Mount) (vfs.VirtualDentry, string, error) { + if !kernel.ContextCanTrace(ctx, s.task, false) { + return vfs.VirtualDentry{}, "", syserror.EACCES + } + + exec, err := s.executable() + if err != nil { + return vfs.VirtualDentry{}, "", err + } + defer exec.DecRef() + + vd := exec.(*fsbridge.VFSFile).FileDescription().VirtualDentry() + vd.IncRef() + return vd, "", nil +} + +func (s *exeSymlink) executable() (file fsbridge.File, err error) { + if err := checkTaskState(s.task); err != nil { + return nil, err + } + + s.task.WithMuLocked(func(t *kernel.Task) { + mm := t.MemoryManager() + if mm == nil { + err = syserror.EACCES + return + } + + // The MemoryManager may be destroyed, in which case + // MemoryManager.destroy will simply set the executable to nil + // (with locks held). + file = mm.Executable() + if file == nil { + err = syserror.ESRCH + } + }) + return +} + +// mountInfoData is used to implement /proc/[pid]/mountinfo. +// +// +stateify savable +type mountInfoData struct { + kernfs.DynamicBytesFile + + task *kernel.Task +} + +var _ dynamicInode = (*mountInfoData)(nil) + +// Generate implements vfs.DynamicBytesSource.Generate. +func (i *mountInfoData) Generate(ctx context.Context, buf *bytes.Buffer) error { + var fsctx *kernel.FSContext + i.task.WithMuLocked(func(t *kernel.Task) { + fsctx = t.FSContext() + }) + if fsctx == nil { + // The task has been destroyed. Nothing to show here. + return nil + } + rootDir := fsctx.RootDirectoryVFS2() + if !rootDir.Ok() { + // Root has been destroyed. Don't try to read mounts. + return nil + } + defer rootDir.DecRef() + i.task.Kernel().VFS().GenerateProcMountInfo(ctx, rootDir, buf) + return nil +} + +// mountsData is used to implement /proc/[pid]/mounts. +// +// +stateify savable +type mountsData struct { + kernfs.DynamicBytesFile + + task *kernel.Task +} + +var _ dynamicInode = (*mountsData)(nil) + +// Generate implements vfs.DynamicBytesSource.Generate. +func (i *mountsData) Generate(ctx context.Context, buf *bytes.Buffer) error { + var fsctx *kernel.FSContext + i.task.WithMuLocked(func(t *kernel.Task) { + fsctx = t.FSContext() + }) + if fsctx == nil { + // The task has been destroyed. Nothing to show here. + return nil + } + rootDir := fsctx.RootDirectoryVFS2() + if !rootDir.Ok() { + // Root has been destroyed. Don't try to read mounts. + return nil + } + defer rootDir.DecRef() + i.task.Kernel().VFS().GenerateProcMounts(ctx, rootDir, buf) + return nil +} + +type namespaceSymlink struct { + kernfs.StaticSymlink + + task *kernel.Task +} + +func (fs *filesystem) newNamespaceSymlink(task *kernel.Task, ino uint64, ns string) *kernfs.Dentry { + // Namespace symlinks should contain the namespace name and the inode number + // for the namespace instance, so for example user:[123456]. We currently fake + // the inode number by sticking the symlink inode in its place. + target := fmt.Sprintf("%s:[%d]", ns, ino) + + inode := &namespaceSymlink{task: task} + // Note: credentials are overridden by taskOwnedInode. + inode.Init(task.Credentials(), linux.UNNAMED_MAJOR, fs.devMinor, ino, target) + + taskInode := &taskOwnedInode{Inode: inode, owner: task} + d := &kernfs.Dentry{} + d.Init(taskInode) + return d +} + +// Readlink implements Inode. +func (s *namespaceSymlink) Readlink(ctx context.Context) (string, error) { + if err := checkTaskState(s.task); err != nil { + return "", err + } + return s.StaticSymlink.Readlink(ctx) +} + +// Getlink implements Inode.Getlink. +func (s *namespaceSymlink) Getlink(ctx context.Context, mnt *vfs.Mount) (vfs.VirtualDentry, string, error) { + if err := checkTaskState(s.task); err != nil { + return vfs.VirtualDentry{}, "", err + } + + // Create a synthetic inode to represent the namespace. + dentry := &kernfs.Dentry{} + dentry.Init(&namespaceInode{}) + vd := vfs.MakeVirtualDentry(mnt, dentry.VFSDentry()) + vd.IncRef() + dentry.DecRef() + return vd, "", nil +} + +// namespaceInode is a synthetic inode created to represent a namespace in +// /proc/[pid]/ns/*. +type namespaceInode struct { + kernfs.InodeAttrs + kernfs.InodeNoopRefCount + kernfs.InodeNotDirectory + kernfs.InodeNotSymlink +} + +var _ kernfs.Inode = (*namespaceInode)(nil) + +// Init initializes a namespace inode. +func (i *namespaceInode) Init(creds *auth.Credentials, devMajor, devMinor uint32, ino uint64, perm linux.FileMode) { + if perm&^linux.PermissionsMask != 0 { + panic(fmt.Sprintf("Only permission mask must be set: %x", perm&linux.PermissionsMask)) + } + i.InodeAttrs.Init(creds, 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) { + fd := &namespaceFD{inode: i} + i.IncRef() + if err := fd.vfsfd.Init(fd, opts.Flags, rp.Mount(), vfsd, &vfs.FileDescriptionOptions{}); err != nil { + return nil, err + } + return &fd.vfsfd, nil +} + +// namespace FD is a synthetic file that represents a namespace in +// /proc/[pid]/ns/*. +type namespaceFD struct { + vfs.FileDescriptionDefaultImpl + + vfsfd vfs.FileDescription + inode *namespaceInode +} + +var _ vfs.FileDescriptionImpl = (*namespaceFD)(nil) + +// Stat implements FileDescriptionImpl. +func (fd *namespaceFD) Stat(ctx context.Context, opts vfs.StatOptions) (linux.Statx, error) { + vfs := fd.vfsfd.VirtualDentry().Mount().Filesystem() + return fd.inode.Stat(vfs, opts) +} + +// SetStat implements FileDescriptionImpl. +func (fd *namespaceFD) SetStat(ctx context.Context, opts vfs.SetStatOptions) error { + vfs := fd.vfsfd.VirtualDentry().Mount().Filesystem() + creds := auth.CredentialsFromContext(ctx) + return fd.inode.SetStat(ctx, vfs, creds, opts) +} + +// Release implements FileDescriptionImpl. +func (fd *namespaceFD) Release() { + fd.inode.DecRef() +} + +// OnClose implements FileDescriptionImpl. +func (*namespaceFD) OnClose(context.Context) error { + return nil +} diff --git a/pkg/sentry/fsimpl/proc/tasks_net.go b/pkg/sentry/fsimpl/proc/task_net.go index 608fec017..6bde27376 100644 --- a/pkg/sentry/fsimpl/proc/tasks_net.go +++ b/pkg/sentry/fsimpl/proc/task_net.go @@ -24,7 +24,6 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/log" - "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs" "gvisor.dev/gvisor/pkg/sentry/inet" "gvisor.dev/gvisor/pkg/sentry/kernel" @@ -32,14 +31,19 @@ import ( "gvisor.dev/gvisor/pkg/sentry/socket" "gvisor.dev/gvisor/pkg/sentry/socket/unix" "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" + "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/tcpip/header" "gvisor.dev/gvisor/pkg/usermem" ) -func newNetDir(root *auth.Credentials, inoGen InoGenerator, k *kernel.Kernel) *kernfs.Dentry { +func (fs *filesystem) newTaskNetDir(task *kernel.Task) *kernfs.Dentry { + k := task.Kernel() + pidns := task.PIDNamespace() + root := auth.NewRootCredentials(pidns.UserNamespace()) + var contents map[string]*kernfs.Dentry - if stack := k.NetworkStack(); stack != nil { + if stack := task.NetworkNamespace().Stack(); stack != nil { const ( arp = "IP address HW type Flags HW address Mask Device\n" netlink = "sk Eth Pid Groups Rmem Wmem Dump Locks Drops Inode\n" @@ -50,38 +54,40 @@ func newNetDir(root *auth.Credentials, inoGen InoGenerator, k *kernel.Kernel) *k ) psched := fmt.Sprintf("%08x %08x %08x %08x\n", uint64(time.Microsecond/time.Nanosecond), 64, 1000000, uint64(time.Second/time.Nanosecond)) + // TODO(gvisor.dev/issue/1833): Make sure file contents reflect the task + // network namespace. contents = map[string]*kernfs.Dentry{ - "dev": newDentry(root, inoGen.NextIno(), 0444, &netDevData{stack: stack}), - "snmp": newDentry(root, inoGen.NextIno(), 0444, &netSnmpData{stack: stack}), + "dev": fs.newDentry(root, fs.NextIno(), 0444, &netDevData{stack: stack}), + "snmp": fs.newDentry(root, fs.NextIno(), 0444, &netSnmpData{stack: stack}), // The following files are simple stubs until they are implemented in // netstack, if the file contains a header the stub is just the header // otherwise it is an empty file. - "arp": newDentry(root, inoGen.NextIno(), 0444, newStaticFile(arp)), - "netlink": newDentry(root, inoGen.NextIno(), 0444, newStaticFile(netlink)), - "netstat": newDentry(root, inoGen.NextIno(), 0444, &netStatData{}), - "packet": newDentry(root, inoGen.NextIno(), 0444, newStaticFile(packet)), - "protocols": newDentry(root, inoGen.NextIno(), 0444, newStaticFile(protocols)), + "arp": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile(arp)), + "netlink": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile(netlink)), + "netstat": fs.newDentry(root, fs.NextIno(), 0444, &netStatData{}), + "packet": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile(packet)), + "protocols": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile(protocols)), // Linux sets psched values to: nsec per usec, psched tick in ns, 1000000, // high res timer ticks per sec (ClockGetres returns 1ns resolution). - "psched": newDentry(root, inoGen.NextIno(), 0444, newStaticFile(psched)), - "ptype": newDentry(root, inoGen.NextIno(), 0444, newStaticFile(ptype)), - "route": newDentry(root, inoGen.NextIno(), 0444, &netRouteData{stack: stack}), - "tcp": newDentry(root, inoGen.NextIno(), 0444, &netTCPData{kernel: k}), - "udp": newDentry(root, inoGen.NextIno(), 0444, &netUDPData{kernel: k}), - "unix": newDentry(root, inoGen.NextIno(), 0444, &netUnixData{kernel: k}), + "psched": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile(psched)), + "ptype": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile(ptype)), + "route": fs.newDentry(root, fs.NextIno(), 0444, &netRouteData{stack: stack}), + "tcp": fs.newDentry(root, fs.NextIno(), 0444, &netTCPData{kernel: k}), + "udp": fs.newDentry(root, fs.NextIno(), 0444, &netUDPData{kernel: k}), + "unix": fs.newDentry(root, fs.NextIno(), 0444, &netUnixData{kernel: k}), } if stack.SupportsIPv6() { - contents["if_inet6"] = newDentry(root, inoGen.NextIno(), 0444, &ifinet6{stack: stack}) - contents["ipv6_route"] = newDentry(root, inoGen.NextIno(), 0444, newStaticFile("")) - contents["tcp6"] = newDentry(root, inoGen.NextIno(), 0444, &netTCP6Data{kernel: k}) - contents["udp6"] = newDentry(root, inoGen.NextIno(), 0444, newStaticFile(upd6)) + contents["if_inet6"] = fs.newDentry(root, fs.NextIno(), 0444, &ifinet6{stack: stack}) + contents["ipv6_route"] = fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("")) + contents["tcp6"] = fs.newDentry(root, fs.NextIno(), 0444, &netTCP6Data{kernel: k}) + contents["udp6"] = fs.newDentry(root, fs.NextIno(), 0444, newStaticFile(upd6)) } } - return kernfs.NewStaticDir(root, inoGen.NextIno(), 0555, contents) + return fs.newTaskOwnedDir(task, fs.NextIno(), 0555, contents) } // ifinet6 implements vfs.DynamicBytesSource for /proc/net/if_inet6. @@ -200,22 +206,21 @@ var _ dynamicInode = (*netUnixData)(nil) func (n *netUnixData) Generate(ctx context.Context, buf *bytes.Buffer) error { buf.WriteString("Num RefCount Protocol Flags Type St Inode Path\n") for _, se := range n.kernel.ListSockets() { - s := se.Sock.Get() - if s == nil { - log.Debugf("Couldn't resolve weakref %v in socket table, racing with destruction?", se.Sock) + s := se.SockVFS2 + if !s.TryIncRef() { + log.Debugf("Couldn't get reference on %v in socket table, racing with destruction?", s) continue } - sfile := s.(*fs.File) - if family, _, _ := sfile.FileOperations.(socket.Socket).Type(); family != linux.AF_UNIX { + if family, _, _ := s.Impl().(socket.SocketVFS2).Type(); family != linux.AF_UNIX { s.DecRef() // Not a unix socket. continue } - sops := sfile.FileOperations.(*unix.SocketOperations) + sops := s.Impl().(*unix.SocketVFS2) addr, err := sops.Endpoint().GetLocalAddress() if err != nil { - log.Warningf("Failed to retrieve socket name from %+v: %v", sfile, err) + log.Warningf("Failed to retrieve socket name from %+v: %v", s, err) addr.Addr = "<unknown>" } @@ -228,6 +233,15 @@ func (n *netUnixData) Generate(ctx context.Context, buf *bytes.Buffer) error { } } + // Get inode number. + var ino uint64 + stat, statErr := s.Stat(ctx, vfs.StatOptions{Mask: linux.STATX_INO}) + if statErr != nil || stat.Mask&linux.STATX_INO == 0 { + log.Warningf("Failed to retrieve ino for socket file: %v", statErr) + } else { + ino = stat.Ino + } + // In the socket entry below, the value for the 'Num' field requires // some consideration. Linux prints the address to the struct // unix_sock representing a socket in the kernel, but may redact the @@ -246,14 +260,14 @@ func (n *netUnixData) Generate(ctx context.Context, buf *bytes.Buffer) error { // the definition of this struct changes over time. // // For now, we always redact this pointer. - fmt.Fprintf(buf, "%#016p: %08X %08X %08X %04X %02X %5d", + fmt.Fprintf(buf, "%#016p: %08X %08X %08X %04X %02X %8d", (*unix.SocketOperations)(nil), // Num, pointer to kernel socket struct. - sfile.ReadRefs()-1, // RefCount, don't count our own ref. + s.Refs()-1, // RefCount, don't count our own ref. 0, // Protocol, always 0 for UDS. sockFlags, // Flags. sops.Endpoint().Type(), // Type. sops.State(), // State. - sfile.InodeID(), // Inode. + ino, // Inode. ) // Path @@ -335,15 +349,14 @@ func commonGenerateTCP(ctx context.Context, buf *bytes.Buffer, k *kernel.Kernel, t := kernel.TaskFromContext(ctx) for _, se := range k.ListSockets() { - s := se.Sock.Get() - if s == nil { - log.Debugf("Couldn't resolve weakref with ID %v in socket table, racing with destruction?", se.ID) + s := se.SockVFS2 + if !s.TryIncRef() { + log.Debugf("Couldn't get reference on %v in socket table, racing with destruction?", s) continue } - sfile := s.(*fs.File) - sops, ok := sfile.FileOperations.(socket.Socket) + sops, ok := s.Impl().(socket.SocketVFS2) if !ok { - panic(fmt.Sprintf("Found non-socket file in socket table: %+v", sfile)) + panic(fmt.Sprintf("Found non-socket file in socket table: %+v", s)) } if fa, stype, _ := sops.Type(); !(family == fa && stype == linux.SOCK_STREAM) { s.DecRef() @@ -392,14 +405,15 @@ func commonGenerateTCP(ctx context.Context, buf *bytes.Buffer, k *kernel.Kernel, // Unimplemented. fmt.Fprintf(buf, "%08X ", 0) + stat, statErr := s.Stat(ctx, vfs.StatOptions{Mask: linux.STATX_UID | linux.STATX_INO}) + // Field: uid. - uattr, err := sfile.Dirent.Inode.UnstableAttr(ctx) - if err != nil { - log.Warningf("Failed to retrieve unstable attr for socket file: %v", err) + if statErr != nil || stat.Mask&linux.STATX_UID == 0 { + log.Warningf("Failed to retrieve uid for socket file: %v", statErr) fmt.Fprintf(buf, "%5d ", 0) } else { creds := auth.CredentialsFromContext(ctx) - fmt.Fprintf(buf, "%5d ", uint32(uattr.Owner.UID.In(creds.UserNamespace).OrOverflow())) + fmt.Fprintf(buf, "%5d ", uint32(auth.KUID(stat.UID).In(creds.UserNamespace).OrOverflow())) } // Field: timeout; number of unanswered 0-window probes. @@ -407,11 +421,16 @@ func commonGenerateTCP(ctx context.Context, buf *bytes.Buffer, k *kernel.Kernel, fmt.Fprintf(buf, "%8d ", 0) // Field: inode. - fmt.Fprintf(buf, "%8d ", sfile.InodeID()) + if statErr != nil || stat.Mask&linux.STATX_INO == 0 { + log.Warningf("Failed to retrieve inode for socket file: %v", statErr) + fmt.Fprintf(buf, "%8d ", 0) + } else { + fmt.Fprintf(buf, "%8d ", stat.Ino) + } // Field: refcount. Don't count the ref we obtain while deferencing // the weakref to this socket. - fmt.Fprintf(buf, "%d ", sfile.ReadRefs()-1) + fmt.Fprintf(buf, "%d ", s.Refs()-1) // Field: Socket struct address. Redacted due to the same reason as // the 'Num' field in /proc/net/unix, see netUnix.ReadSeqFileData. @@ -492,16 +511,17 @@ func (d *netUDPData) Generate(ctx context.Context, buf *bytes.Buffer) error { // degrade gracefully and retrieve what we can. t := kernel.TaskFromContext(ctx) + buf.WriteString(" sl local_address rem_address st tx_queue rx_queue tr tm->when retrnsmt uid timeout inode ref pointer drops \n") + for _, se := range d.kernel.ListSockets() { - s := se.Sock.Get() - if s == nil { - log.Debugf("Couldn't resolve weakref with ID %v in socket table, racing with destruction?", se.ID) + s := se.SockVFS2 + if !s.TryIncRef() { + log.Debugf("Couldn't get reference on %v in socket table, racing with destruction?", s) continue } - sfile := s.(*fs.File) - sops, ok := sfile.FileOperations.(socket.Socket) + sops, ok := s.Impl().(socket.SocketVFS2) if !ok { - panic(fmt.Sprintf("Found non-socket file in socket table: %+v", sfile)) + panic(fmt.Sprintf("Found non-socket file in socket table: %+v", s)) } if family, stype, _ := sops.Type(); family != linux.AF_INET || stype != linux.SOCK_DGRAM { s.DecRef() @@ -545,25 +565,31 @@ func (d *netUDPData) Generate(ctx context.Context, buf *bytes.Buffer) error { // Field: retrnsmt. Always 0 for UDP. fmt.Fprintf(buf, "%08X ", 0) + stat, statErr := s.Stat(ctx, vfs.StatOptions{Mask: linux.STATX_UID | linux.STATX_INO}) + // Field: uid. - uattr, err := sfile.Dirent.Inode.UnstableAttr(ctx) - if err != nil { - log.Warningf("Failed to retrieve unstable attr for socket file: %v", err) + if statErr != nil || stat.Mask&linux.STATX_UID == 0 { + log.Warningf("Failed to retrieve uid for socket file: %v", statErr) fmt.Fprintf(buf, "%5d ", 0) } else { creds := auth.CredentialsFromContext(ctx) - fmt.Fprintf(buf, "%5d ", uint32(uattr.Owner.UID.In(creds.UserNamespace).OrOverflow())) + fmt.Fprintf(buf, "%5d ", uint32(auth.KUID(stat.UID).In(creds.UserNamespace).OrOverflow())) } // Field: timeout. Always 0 for UDP. fmt.Fprintf(buf, "%8d ", 0) // Field: inode. - fmt.Fprintf(buf, "%8d ", sfile.InodeID()) + if statErr != nil || stat.Mask&linux.STATX_INO == 0 { + log.Warningf("Failed to retrieve inode for socket file: %v", statErr) + fmt.Fprintf(buf, "%8d ", 0) + } else { + fmt.Fprintf(buf, "%8d ", stat.Ino) + } // Field: ref; reference count on the socket inode. Don't count the ref // we obtain while deferencing the weakref to this socket. - fmt.Fprintf(buf, "%d ", sfile.ReadRefs()-1) + fmt.Fprintf(buf, "%d ", s.Refs()-1) // Field: Socket struct address. Redacted due to the same reason as // the 'Num' field in /proc/net/unix, see netUnix.ReadSeqFileData. @@ -664,9 +690,9 @@ func (d *netSnmpData) Generate(ctx context.Context, buf *bytes.Buffer) error { if line.prefix == "Tcp" { tcp := stat.(*inet.StatSNMPTCP) // "Tcp" needs special processing because MaxConn is signed. RFC 2012. - fmt.Sprintf("%s: %s %d %s\n", line.prefix, sprintSlice(tcp[:3]), int64(tcp[3]), sprintSlice(tcp[4:])) + fmt.Fprintf(buf, "%s: %s %d %s\n", line.prefix, sprintSlice(tcp[:3]), int64(tcp[3]), sprintSlice(tcp[4:])) } else { - fmt.Sprintf("%s: %s\n", line.prefix, sprintSlice(toSlice(stat))) + fmt.Fprintf(buf, "%s: %s\n", line.prefix, sprintSlice(toSlice(stat))) } } return nil diff --git a/pkg/sentry/fsimpl/proc/tasks.go b/pkg/sentry/fsimpl/proc/tasks.go index ebe21630c..b51d43954 100644 --- a/pkg/sentry/fsimpl/proc/tasks.go +++ b/pkg/sentry/fsimpl/proc/tasks.go @@ -33,11 +33,6 @@ const ( threadSelfName = "thread-self" ) -// InoGenerator generates unique inode numbers for a given filesystem. -type InoGenerator interface { - NextIno() uint64 -} - // tasksInode represents the inode for /proc/ directory. // // +stateify savable @@ -46,9 +41,10 @@ type tasksInode struct { kernfs.InodeDirectoryNoNewChildren kernfs.InodeAttrs kernfs.OrderedChildren + kernfs.AlwaysValid - inoGen InoGenerator - pidns *kernel.PIDNamespace + fs *filesystem + pidns *kernel.PIDNamespace // '/proc/self' and '/proc/thread-self' have custom directory offsets in // Linux. So handle them outside of OrderedChildren. @@ -63,29 +59,29 @@ type tasksInode struct { var _ kernfs.Inode = (*tasksInode)(nil) -func newTasksInode(inoGen InoGenerator, k *kernel.Kernel, pidns *kernel.PIDNamespace, cgroupControllers map[string]string) (*tasksInode, *kernfs.Dentry) { +func (fs *filesystem) newTasksInode(k *kernel.Kernel, pidns *kernel.PIDNamespace, cgroupControllers map[string]string) (*tasksInode, *kernfs.Dentry) { root := auth.NewRootCredentials(pidns.UserNamespace()) contents := map[string]*kernfs.Dentry{ - "cpuinfo": newDentry(root, inoGen.NextIno(), 0444, newStaticFile(cpuInfoData(k))), - //"filesystems": newDentry(root, inoGen.NextIno(), 0444, &filesystemsData{}), - "loadavg": newDentry(root, inoGen.NextIno(), 0444, &loadavgData{}), - "sys": newSysDir(root, inoGen, k), - "meminfo": newDentry(root, inoGen.NextIno(), 0444, &meminfoData{}), - "mounts": kernfs.NewStaticSymlink(root, inoGen.NextIno(), "self/mounts"), - "net": newNetDir(root, inoGen, k), - "stat": newDentry(root, inoGen.NextIno(), 0444, &statData{}), - "uptime": newDentry(root, inoGen.NextIno(), 0444, &uptimeData{}), - "version": newDentry(root, inoGen.NextIno(), 0444, &versionData{}), + "cpuinfo": fs.newDentry(root, fs.NextIno(), 0444, newStaticFileSetStat(cpuInfoData(k))), + "filesystems": fs.newDentry(root, fs.NextIno(), 0444, &filesystemsData{}), + "loadavg": fs.newDentry(root, fs.NextIno(), 0444, &loadavgData{}), + "sys": fs.newSysDir(root, k), + "meminfo": fs.newDentry(root, fs.NextIno(), 0444, &meminfoData{}), + "mounts": kernfs.NewStaticSymlink(root, linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), "self/mounts"), + "net": kernfs.NewStaticSymlink(root, linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), "self/net"), + "stat": fs.newDentry(root, fs.NextIno(), 0444, &statData{}), + "uptime": fs.newDentry(root, fs.NextIno(), 0444, &uptimeData{}), + "version": fs.newDentry(root, fs.NextIno(), 0444, &versionData{}), } inode := &tasksInode{ pidns: pidns, - inoGen: inoGen, - selfSymlink: newSelfSymlink(root, inoGen.NextIno(), 0444, pidns).VFSDentry(), - threadSelfSymlink: newThreadSelfSymlink(root, inoGen.NextIno(), 0444, pidns).VFSDentry(), + fs: fs, + selfSymlink: fs.newSelfSymlink(root, fs.NextIno(), pidns).VFSDentry(), + threadSelfSymlink: fs.newThreadSelfSymlink(root, fs.NextIno(), pidns).VFSDentry(), cgroupControllers: cgroupControllers, } - inode.InodeAttrs.Init(root, inoGen.NextIno(), linux.ModeDirectory|0555) + inode.InodeAttrs.Init(root, linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), linux.ModeDirectory|0555) dentry := &kernfs.Dentry{} dentry.Init(inode) @@ -117,15 +113,10 @@ func (i *tasksInode) Lookup(ctx context.Context, name string) (*vfs.Dentry, erro return nil, syserror.ENOENT } - taskDentry := newTaskInode(i.inoGen, task, i.pidns, true, i.cgroupControllers) + taskDentry := i.fs.newTaskInode(task, i.pidns, true, i.cgroupControllers) return taskDentry.VFSDentry(), nil } -// Valid implements kernfs.inodeDynamicLookup. -func (i *tasksInode) Valid(ctx context.Context) bool { - return true -} - // IterDirents implements kernfs.inodeDynamicLookup. func (i *tasksInode) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback, offset, _ int64) (int64, error) { // fs/proc/internal.h: #define FIRST_PROCESS_ENTRY 256 @@ -148,11 +139,11 @@ func (i *tasksInode) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback dirent := vfs.Dirent{ Name: selfName, Type: linux.DT_LNK, - Ino: i.inoGen.NextIno(), + Ino: i.fs.NextIno(), NextOff: offset + 1, } - if !cb.Handle(dirent) { - return offset, nil + if err := cb.Handle(dirent); err != nil { + return offset, err } offset++ } @@ -160,11 +151,11 @@ func (i *tasksInode) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback dirent := vfs.Dirent{ Name: threadSelfName, Type: linux.DT_LNK, - Ino: i.inoGen.NextIno(), + Ino: i.fs.NextIno(), NextOff: offset + 1, } - if !cb.Handle(dirent) { - return offset, nil + if err := cb.Handle(dirent); err != nil { + return offset, err } offset++ } @@ -193,11 +184,11 @@ func (i *tasksInode) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback dirent := vfs.Dirent{ Name: strconv.FormatUint(uint64(tid), 10), Type: linux.DT_DIR, - Ino: i.inoGen.NextIno(), + Ino: i.fs.NextIno(), NextOff: FIRST_PROCESS_ENTRY + 2 + int64(tid) + 1, } - if !cb.Handle(dirent) { - return offset, nil + if err := cb.Handle(dirent); err != nil { + return offset, err } offset++ } @@ -205,23 +196,44 @@ func (i *tasksInode) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback } // Open implements kernfs.Inode. -func (i *tasksInode) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { - fd := &kernfs.GenericDirectoryFD{} - fd.Init(rp.Mount(), vfsd, &i.OrderedChildren, &opts) +func (i *tasksInode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), vfsd, &i.OrderedChildren, &opts) + if err != nil { + return nil, err + } return fd.VFSFileDescription(), nil } -func (i *tasksInode) Stat(vsfs *vfs.Filesystem) linux.Statx { - stat := i.InodeAttrs.Stat(vsfs) +func (i *tasksInode) Stat(vsfs *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) { + stat, err := i.InodeAttrs.Stat(vsfs, opts) + if err != nil { + return linux.Statx{}, err + } - // Add dynamic children to link count. - for _, tg := range i.pidns.ThreadGroups() { - if leader := tg.Leader(); leader != nil { - stat.Nlink++ + if opts.Mask&linux.STATX_NLINK != 0 { + // Add dynamic children to link count. + for _, tg := range i.pidns.ThreadGroups() { + if leader := tg.Leader(); leader != nil { + stat.Nlink++ + } } } - return stat + return stat, nil +} + +// staticFileSetStat implements a special static file that allows inode +// attributes to be set. This is to support /proc files that are readonly, but +// allow attributes to be set. +type staticFileSetStat struct { + dynamicBytesFileSetAttr + vfs.StaticData +} + +var _ dynamicInode = (*staticFileSetStat)(nil) + +func newStaticFileSetStat(data string) *staticFileSetStat { + return &staticFileSetStat{StaticData: vfs.StaticData{Data: data}} } func cpuInfoData(k *kernel.Kernel) string { diff --git a/pkg/sentry/fsimpl/proc/tasks_files.go b/pkg/sentry/fsimpl/proc/tasks_files.go index 434998910..7d8983aa5 100644 --- a/pkg/sentry/fsimpl/proc/tasks_files.go +++ b/pkg/sentry/fsimpl/proc/tasks_files.go @@ -26,6 +26,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/kernel/time" "gvisor.dev/gvisor/pkg/sentry/usage" + "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" ) @@ -40,9 +41,9 @@ type selfSymlink struct { var _ kernfs.Inode = (*selfSymlink)(nil) -func newSelfSymlink(creds *auth.Credentials, ino uint64, perm linux.FileMode, pidns *kernel.PIDNamespace) *kernfs.Dentry { +func (fs *filesystem) newSelfSymlink(creds *auth.Credentials, ino uint64, pidns *kernel.PIDNamespace) *kernfs.Dentry { inode := &selfSymlink{pidns: pidns} - inode.Init(creds, ino, linux.ModeSymlink|perm) + inode.Init(creds, linux.UNNAMED_MAJOR, fs.devMinor, ino, linux.ModeSymlink|0777) d := &kernfs.Dentry{} d.Init(inode) @@ -62,6 +63,16 @@ func (s *selfSymlink) Readlink(ctx context.Context) (string, error) { return strconv.FormatUint(uint64(tgid), 10), nil } +func (s *selfSymlink) Getlink(ctx context.Context, _ *vfs.Mount) (vfs.VirtualDentry, string, error) { + target, err := s.Readlink(ctx) + return vfs.VirtualDentry{}, target, err +} + +// SetStat implements Inode.SetStat not allowing inode attributes to be changed. +func (*selfSymlink) SetStat(context.Context, *vfs.Filesystem, *auth.Credentials, vfs.SetStatOptions) error { + return syserror.EPERM +} + type threadSelfSymlink struct { kernfs.InodeAttrs kernfs.InodeNoopRefCount @@ -72,9 +83,9 @@ type threadSelfSymlink struct { var _ kernfs.Inode = (*threadSelfSymlink)(nil) -func newThreadSelfSymlink(creds *auth.Credentials, ino uint64, perm linux.FileMode, pidns *kernel.PIDNamespace) *kernfs.Dentry { +func (fs *filesystem) newThreadSelfSymlink(creds *auth.Credentials, ino uint64, pidns *kernel.PIDNamespace) *kernfs.Dentry { inode := &threadSelfSymlink{pidns: pidns} - inode.Init(creds, ino, linux.ModeSymlink|perm) + inode.Init(creds, linux.UNNAMED_MAJOR, fs.devMinor, ino, linux.ModeSymlink|0777) d := &kernfs.Dentry{} d.Init(inode) @@ -95,6 +106,28 @@ func (s *threadSelfSymlink) Readlink(ctx context.Context) (string, error) { return fmt.Sprintf("%d/task/%d", tgid, tid), nil } +func (s *threadSelfSymlink) Getlink(ctx context.Context, _ *vfs.Mount) (vfs.VirtualDentry, string, error) { + target, err := s.Readlink(ctx) + return vfs.VirtualDentry{}, target, err +} + +// SetStat implements Inode.SetStat not allowing inode attributes to be changed. +func (*threadSelfSymlink) SetStat(context.Context, *vfs.Filesystem, *auth.Credentials, vfs.SetStatOptions) error { + return syserror.EPERM +} + +// dynamicBytesFileSetAttr implements a special file that allows inode +// attributes to be set. This is to support /proc files that are readonly, but +// allow attributes to be set. +type dynamicBytesFileSetAttr struct { + kernfs.DynamicBytesFile +} + +// SetStat implements Inode.SetStat. +func (d *dynamicBytesFileSetAttr) SetStat(ctx context.Context, fs *vfs.Filesystem, creds *auth.Credentials, opts vfs.SetStatOptions) error { + return d.DynamicBytesFile.InodeAttrs.SetStat(ctx, fs, creds, opts) +} + // cpuStats contains the breakdown of CPU time for /proc/stat. type cpuStats struct { // user is time spent in userspace tasks with non-positive niceness. @@ -137,22 +170,20 @@ func (c cpuStats) String() string { // // +stateify savable type statData struct { - kernfs.DynamicBytesFile - - // k is the owning Kernel. - k *kernel.Kernel + dynamicBytesFileSetAttr } var _ dynamicInode = (*statData)(nil) // Generate implements vfs.DynamicBytesSource.Generate. -func (s *statData) Generate(ctx context.Context, buf *bytes.Buffer) error { +func (*statData) Generate(ctx context.Context, buf *bytes.Buffer) error { // TODO(b/37226836): We currently export only zero CPU stats. We could // at least provide some aggregate stats. var cpu cpuStats fmt.Fprintf(buf, "cpu %s\n", cpu) - for c, max := uint(0), s.k.ApplicationCores(); c < max; c++ { + k := kernel.KernelFromContext(ctx) + for c, max := uint(0), k.ApplicationCores(); c < max; c++ { fmt.Fprintf(buf, "cpu%d %s\n", c, cpu) } @@ -176,7 +207,7 @@ func (s *statData) Generate(ctx context.Context, buf *bytes.Buffer) error { fmt.Fprintf(buf, "ctxt 0\n") // CLOCK_REALTIME timestamp from boot, in seconds. - fmt.Fprintf(buf, "btime %d\n", s.k.Timekeeper().BootTime().Seconds()) + fmt.Fprintf(buf, "btime %d\n", k.Timekeeper().BootTime().Seconds()) // Total number of clones. // TODO(b/37226836): Count this. @@ -203,13 +234,13 @@ func (s *statData) Generate(ctx context.Context, buf *bytes.Buffer) error { // // +stateify savable type loadavgData struct { - kernfs.DynamicBytesFile + dynamicBytesFileSetAttr } var _ dynamicInode = (*loadavgData)(nil) // Generate implements vfs.DynamicBytesSource.Generate. -func (d *loadavgData) Generate(ctx context.Context, buf *bytes.Buffer) error { +func (*loadavgData) Generate(ctx context.Context, buf *bytes.Buffer) error { // TODO(b/62345059): Include real data in fields. // Column 1-3: CPU and IO utilization of the last 1, 5, and 10 minute periods. // Column 4-5: currently running processes and the total number of processes. @@ -222,17 +253,15 @@ func (d *loadavgData) Generate(ctx context.Context, buf *bytes.Buffer) error { // // +stateify savable type meminfoData struct { - kernfs.DynamicBytesFile - - // k is the owning Kernel. - k *kernel.Kernel + dynamicBytesFileSetAttr } var _ dynamicInode = (*meminfoData)(nil) // Generate implements vfs.DynamicBytesSource.Generate. -func (d *meminfoData) Generate(ctx context.Context, buf *bytes.Buffer) error { - mf := d.k.MemoryFile() +func (*meminfoData) Generate(ctx context.Context, buf *bytes.Buffer) error { + k := kernel.KernelFromContext(ctx) + mf := k.MemoryFile() mf.UpdateUsage() snapshot, totalUsage := usage.MemoryAccounting.Copy() totalSize := usage.TotalMemory(mf.TotalSize(), totalUsage) @@ -243,12 +272,16 @@ func (d *meminfoData) Generate(ctx context.Context, buf *bytes.Buffer) error { inactiveFile := file - activeFile fmt.Fprintf(buf, "MemTotal: %8d kB\n", totalSize/1024) - memFree := (totalSize - totalUsage) / 1024 + memFree := totalSize - totalUsage + if memFree > totalSize { + // Underflow. + memFree = 0 + } // We use MemFree as MemAvailable because we don't swap. // TODO(rahat): When reclaim is implemented the value of MemAvailable // should change. - fmt.Fprintf(buf, "MemFree: %8d kB\n", memFree) - fmt.Fprintf(buf, "MemAvailable: %8d kB\n", memFree) + fmt.Fprintf(buf, "MemFree: %8d kB\n", memFree/1024) + fmt.Fprintf(buf, "MemAvailable: %8d kB\n", memFree/1024) fmt.Fprintf(buf, "Buffers: 0 kB\n") // memory usage by block devices fmt.Fprintf(buf, "Cached: %8d kB\n", (file+snapshot.Tmpfs)/1024) // Emulate a system with no swap, which disables inactivation of anon pages. @@ -275,7 +308,7 @@ func (d *meminfoData) Generate(ctx context.Context, buf *bytes.Buffer) error { // // +stateify savable type uptimeData struct { - kernfs.DynamicBytesFile + dynamicBytesFileSetAttr } var _ dynamicInode = (*uptimeData)(nil) @@ -294,17 +327,15 @@ func (*uptimeData) Generate(ctx context.Context, buf *bytes.Buffer) error { // // +stateify savable type versionData struct { - kernfs.DynamicBytesFile - - // k is the owning Kernel. - k *kernel.Kernel + dynamicBytesFileSetAttr } var _ dynamicInode = (*versionData)(nil) // Generate implements vfs.DynamicBytesSource.Generate. -func (v *versionData) Generate(ctx context.Context, buf *bytes.Buffer) error { - init := v.k.GlobalInit() +func (*versionData) Generate(ctx context.Context, buf *bytes.Buffer) error { + k := kernel.KernelFromContext(ctx) + init := k.GlobalInit() if init == nil { // Attempted to read before the init Task is created. This can // only occur during startup, which should never need to read @@ -335,3 +366,19 @@ func (v *versionData) Generate(ctx context.Context, buf *bytes.Buffer) error { fmt.Fprintf(buf, "%s version %s %s\n", ver.Sysname, ver.Release, ver.Version) return nil } + +// filesystemsData backs /proc/filesystems. +// +// +stateify savable +type filesystemsData struct { + kernfs.DynamicBytesFile +} + +var _ dynamicInode = (*filesystemsData)(nil) + +// Generate implements vfs.DynamicBytesSource.Generate. +func (d *filesystemsData) Generate(ctx context.Context, buf *bytes.Buffer) error { + k := kernel.KernelFromContext(ctx) + k.VFS().GenerateProcFilesystems(buf) + return nil +} diff --git a/pkg/sentry/fsimpl/proc/tasks_sys.go b/pkg/sentry/fsimpl/proc/tasks_sys.go index c7ce74883..6dac2afa4 100644 --- a/pkg/sentry/fsimpl/proc/tasks_sys.go +++ b/pkg/sentry/fsimpl/proc/tasks_sys.go @@ -30,89 +30,89 @@ import ( ) // newSysDir returns the dentry corresponding to /proc/sys directory. -func newSysDir(root *auth.Credentials, inoGen InoGenerator, k *kernel.Kernel) *kernfs.Dentry { - return kernfs.NewStaticDir(root, inoGen.NextIno(), 0555, map[string]*kernfs.Dentry{ - "kernel": kernfs.NewStaticDir(root, inoGen.NextIno(), 0555, map[string]*kernfs.Dentry{ - "hostname": newDentry(root, inoGen.NextIno(), 0444, &hostnameData{}), - "shmall": newDentry(root, inoGen.NextIno(), 0444, shmData(linux.SHMALL)), - "shmmax": newDentry(root, inoGen.NextIno(), 0444, shmData(linux.SHMMAX)), - "shmmni": newDentry(root, inoGen.NextIno(), 0444, shmData(linux.SHMMNI)), +func (fs *filesystem) newSysDir(root *auth.Credentials, k *kernel.Kernel) *kernfs.Dentry { + return kernfs.NewStaticDir(root, linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), 0555, map[string]*kernfs.Dentry{ + "kernel": kernfs.NewStaticDir(root, linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), 0555, map[string]*kernfs.Dentry{ + "hostname": fs.newDentry(root, fs.NextIno(), 0444, &hostnameData{}), + "shmall": fs.newDentry(root, fs.NextIno(), 0444, shmData(linux.SHMALL)), + "shmmax": fs.newDentry(root, fs.NextIno(), 0444, shmData(linux.SHMMAX)), + "shmmni": fs.newDentry(root, fs.NextIno(), 0444, shmData(linux.SHMMNI)), }), - "vm": kernfs.NewStaticDir(root, inoGen.NextIno(), 0555, map[string]*kernfs.Dentry{ - "mmap_min_addr": newDentry(root, inoGen.NextIno(), 0444, &mmapMinAddrData{}), - "overcommit_memory": newDentry(root, inoGen.NextIno(), 0444, newStaticFile("0\n")), + "vm": kernfs.NewStaticDir(root, linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), 0555, map[string]*kernfs.Dentry{ + "mmap_min_addr": fs.newDentry(root, fs.NextIno(), 0444, &mmapMinAddrData{k: k}), + "overcommit_memory": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("0\n")), }), - "net": newSysNetDir(root, inoGen, k), + "net": fs.newSysNetDir(root, k), }) } // newSysNetDir returns the dentry corresponding to /proc/sys/net directory. -func newSysNetDir(root *auth.Credentials, inoGen InoGenerator, k *kernel.Kernel) *kernfs.Dentry { +func (fs *filesystem) newSysNetDir(root *auth.Credentials, k *kernel.Kernel) *kernfs.Dentry { var contents map[string]*kernfs.Dentry - if stack := k.NetworkStack(); stack != nil { + // TODO(gvisor.dev/issue/1833): Support for using the network stack in the + // network namespace of the calling process. + if stack := k.RootNetworkNamespace().Stack(); stack != nil { contents = map[string]*kernfs.Dentry{ - "ipv4": kernfs.NewStaticDir(root, inoGen.NextIno(), 0555, map[string]*kernfs.Dentry{ - "tcp_sack": newDentry(root, inoGen.NextIno(), 0644, &tcpSackData{stack: stack}), + "ipv4": kernfs.NewStaticDir(root, linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), 0555, map[string]*kernfs.Dentry{ + "tcp_sack": fs.newDentry(root, fs.NextIno(), 0644, &tcpSackData{stack: stack}), // 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 actual netstack behavior or any empty file, all // of these files will have mode 0444 (read-only for all users). - "ip_local_port_range": newDentry(root, inoGen.NextIno(), 0444, newStaticFile("16000 65535")), - "ip_local_reserved_ports": newDentry(root, inoGen.NextIno(), 0444, newStaticFile("")), - "ipfrag_time": newDentry(root, inoGen.NextIno(), 0444, newStaticFile("30")), - "ip_nonlocal_bind": newDentry(root, inoGen.NextIno(), 0444, newStaticFile("0")), - "ip_no_pmtu_disc": newDentry(root, inoGen.NextIno(), 0444, newStaticFile("1")), + "ip_local_port_range": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("16000 65535")), + "ip_local_reserved_ports": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("")), + "ipfrag_time": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("30")), + "ip_nonlocal_bind": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("0")), + "ip_no_pmtu_disc": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("1")), // tcp_allowed_congestion_control tell the user what they are able to // do as an unprivledged process so we leave it empty. - "tcp_allowed_congestion_control": newDentry(root, inoGen.NextIno(), 0444, newStaticFile("")), - "tcp_available_congestion_control": newDentry(root, inoGen.NextIno(), 0444, newStaticFile("reno")), - "tcp_congestion_control": newDentry(root, inoGen.NextIno(), 0444, newStaticFile("reno")), + "tcp_allowed_congestion_control": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("")), + "tcp_available_congestion_control": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("reno")), + "tcp_congestion_control": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("reno")), // Many of the following stub files are features netstack doesn't // support. The unsupported features return "0" to indicate they are // disabled. - "tcp_base_mss": newDentry(root, inoGen.NextIno(), 0444, newStaticFile("1280")), - "tcp_dsack": newDentry(root, inoGen.NextIno(), 0444, newStaticFile("0")), - "tcp_early_retrans": newDentry(root, inoGen.NextIno(), 0444, newStaticFile("0")), - "tcp_fack": newDentry(root, inoGen.NextIno(), 0444, newStaticFile("0")), - "tcp_fastopen": newDentry(root, inoGen.NextIno(), 0444, newStaticFile("0")), - "tcp_fastopen_key": newDentry(root, inoGen.NextIno(), 0444, newStaticFile("")), - "tcp_invalid_ratelimit": newDentry(root, inoGen.NextIno(), 0444, newStaticFile("0")), - "tcp_keepalive_intvl": newDentry(root, inoGen.NextIno(), 0444, newStaticFile("0")), - "tcp_keepalive_probes": newDentry(root, inoGen.NextIno(), 0444, newStaticFile("0")), - "tcp_keepalive_time": newDentry(root, inoGen.NextIno(), 0444, newStaticFile("7200")), - "tcp_mtu_probing": newDentry(root, inoGen.NextIno(), 0444, newStaticFile("0")), - "tcp_no_metrics_save": newDentry(root, inoGen.NextIno(), 0444, newStaticFile("1")), - "tcp_probe_interval": newDentry(root, inoGen.NextIno(), 0444, newStaticFile("0")), - "tcp_probe_threshold": newDentry(root, inoGen.NextIno(), 0444, newStaticFile("0")), - "tcp_retries1": newDentry(root, inoGen.NextIno(), 0444, newStaticFile("3")), - "tcp_retries2": newDentry(root, inoGen.NextIno(), 0444, newStaticFile("15")), - "tcp_rfc1337": newDentry(root, inoGen.NextIno(), 0444, newStaticFile("1")), - "tcp_slow_start_after_idle": newDentry(root, inoGen.NextIno(), 0444, newStaticFile("1")), - "tcp_synack_retries": newDentry(root, inoGen.NextIno(), 0444, newStaticFile("5")), - "tcp_syn_retries": newDentry(root, inoGen.NextIno(), 0444, newStaticFile("3")), - "tcp_timestamps": newDentry(root, inoGen.NextIno(), 0444, newStaticFile("1")), + "tcp_base_mss": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("1280")), + "tcp_dsack": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("0")), + "tcp_early_retrans": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("0")), + "tcp_fack": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("0")), + "tcp_fastopen": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("0")), + "tcp_fastopen_key": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("")), + "tcp_invalid_ratelimit": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("0")), + "tcp_keepalive_intvl": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("0")), + "tcp_keepalive_probes": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("0")), + "tcp_keepalive_time": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("7200")), + "tcp_mtu_probing": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("0")), + "tcp_no_metrics_save": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("1")), + "tcp_probe_interval": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("0")), + "tcp_probe_threshold": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("0")), + "tcp_retries1": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("3")), + "tcp_retries2": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("15")), + "tcp_rfc1337": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("1")), + "tcp_slow_start_after_idle": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("1")), + "tcp_synack_retries": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("5")), + "tcp_syn_retries": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("3")), + "tcp_timestamps": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("1")), }), - "core": kernfs.NewStaticDir(root, inoGen.NextIno(), 0555, map[string]*kernfs.Dentry{ - "default_qdisc": newDentry(root, inoGen.NextIno(), 0444, newStaticFile("pfifo_fast")), - "message_burst": newDentry(root, inoGen.NextIno(), 0444, newStaticFile("10")), - "message_cost": newDentry(root, inoGen.NextIno(), 0444, newStaticFile("5")), - "optmem_max": newDentry(root, inoGen.NextIno(), 0444, newStaticFile("0")), - "rmem_default": newDentry(root, inoGen.NextIno(), 0444, newStaticFile("212992")), - "rmem_max": newDentry(root, inoGen.NextIno(), 0444, newStaticFile("212992")), - "somaxconn": newDentry(root, inoGen.NextIno(), 0444, newStaticFile("128")), - "wmem_default": newDentry(root, inoGen.NextIno(), 0444, newStaticFile("212992")), - "wmem_max": newDentry(root, inoGen.NextIno(), 0444, newStaticFile("212992")), + "core": kernfs.NewStaticDir(root, linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), 0555, map[string]*kernfs.Dentry{ + "default_qdisc": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("pfifo_fast")), + "message_burst": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("10")), + "message_cost": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("5")), + "optmem_max": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("0")), + "rmem_default": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("212992")), + "rmem_max": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("212992")), + "somaxconn": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("128")), + "wmem_default": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("212992")), + "wmem_max": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("212992")), }), } } - return kernfs.NewStaticDir(root, inoGen.NextIno(), 0555, map[string]*kernfs.Dentry{ - "net": kernfs.NewStaticDir(root, inoGen.NextIno(), 0555, contents), - }) + return kernfs.NewStaticDir(root, linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), 0555, contents) } // mmapMinAddrData implements vfs.DynamicBytesSource for diff --git a/pkg/sentry/fsimpl/proc/tasks_test.go b/pkg/sentry/fsimpl/proc/tasks_test.go index 6fc3524db..19abb5034 100644 --- a/pkg/sentry/fsimpl/proc/tasks_test.go +++ b/pkg/sentry/fsimpl/proc/tasks_test.go @@ -25,6 +25,7 @@ import ( "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/fspath" "gvisor.dev/gvisor/pkg/sentry/fsimpl/testutil" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/tmpfs" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/vfs" @@ -47,10 +48,11 @@ var ( var ( tasksStaticFiles = map[string]testutil.DirentType{ "cpuinfo": linux.DT_REG, + "filesystems": linux.DT_REG, "loadavg": linux.DT_REG, "meminfo": linux.DT_REG, "mounts": linux.DT_LNK, - "net": linux.DT_DIR, + "net": linux.DT_LNK, "self": linux.DT_LNK, "stat": linux.DT_REG, "sys": linux.DT_DIR, @@ -63,21 +65,29 @@ var ( "thread-self": threadSelfLink.NextOff, } taskStaticFiles = map[string]testutil.DirentType{ - "auxv": linux.DT_REG, - "cgroup": linux.DT_REG, - "cmdline": linux.DT_REG, - "comm": linux.DT_REG, - "environ": linux.DT_REG, - "gid_map": linux.DT_REG, - "io": linux.DT_REG, - "maps": linux.DT_REG, - "ns": linux.DT_DIR, - "smaps": linux.DT_REG, - "stat": linux.DT_REG, - "statm": linux.DT_REG, - "status": linux.DT_REG, - "task": linux.DT_DIR, - "uid_map": linux.DT_REG, + "auxv": linux.DT_REG, + "cgroup": linux.DT_REG, + "cmdline": linux.DT_REG, + "comm": linux.DT_REG, + "environ": linux.DT_REG, + "exe": linux.DT_LNK, + "fd": linux.DT_DIR, + "fdinfo": linux.DT_DIR, + "gid_map": linux.DT_REG, + "io": linux.DT_REG, + "maps": linux.DT_REG, + "mountinfo": linux.DT_REG, + "mounts": linux.DT_REG, + "net": linux.DT_DIR, + "ns": linux.DT_DIR, + "oom_score": linux.DT_REG, + "oom_score_adj": linux.DT_REG, + "smaps": linux.DT_REG, + "stat": linux.DT_REG, + "statm": linux.DT_REG, + "status": linux.DT_REG, + "task": linux.DT_DIR, + "uid_map": linux.DT_REG, } ) @@ -90,30 +100,49 @@ func setup(t *testing.T) *testutil.System { ctx := k.SupervisorContext() creds := auth.CredentialsFromContext(ctx) - vfsObj := vfs.New() - vfsObj.MustRegisterFilesystemType("procfs", &procFSType{}, &vfs.RegisterFilesystemTypeOptions{ + k.VFS().MustRegisterFilesystemType(Name, &FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ AllowUserMount: true, }) - fsOpts := vfs.GetFilesystemOptions{ - InternalData: &InternalData{ - Cgroups: map[string]string{ - "cpuset": "/foo/cpuset", - "memory": "/foo/memory", + + mntns, err := k.VFS().NewMountNamespace(ctx, creds, "", tmpfs.Name, &vfs.GetFilesystemOptions{}) + if err != nil { + t.Fatalf("NewMountNamespace(): %v", err) + } + pop := &vfs.PathOperation{ + Root: mntns.Root(), + Start: mntns.Root(), + Path: fspath.Parse("/proc"), + } + if err := k.VFS().MkdirAt(ctx, creds, pop, &vfs.MkdirOptions{Mode: 0777}); err != nil { + t.Fatalf("MkDir(/proc): %v", err) + } + + pop = &vfs.PathOperation{ + Root: mntns.Root(), + Start: mntns.Root(), + Path: fspath.Parse("/proc"), + } + mntOpts := &vfs.MountOptions{ + GetFilesystemOptions: vfs.GetFilesystemOptions{ + InternalData: &InternalData{ + Cgroups: map[string]string{ + "cpuset": "/foo/cpuset", + "memory": "/foo/memory", + }, }, }, } - mntns, err := vfsObj.NewMountNamespace(ctx, creds, "", "procfs", &fsOpts) - if err != nil { - t.Fatalf("NewMountNamespace(): %v", err) + if err := k.VFS().MountAt(ctx, creds, "", pop, Name, mntOpts); err != nil { + t.Fatalf("MountAt(/proc): %v", err) } - return testutil.NewSystem(ctx, t, vfsObj, mntns) + return testutil.NewSystem(ctx, t, k.VFS(), mntns) } func TestTasksEmpty(t *testing.T) { s := setup(t) defer s.Destroy() - collector := s.ListDirents(s.PathOpAtRoot("/")) + collector := s.ListDirents(s.PathOpAtRoot("/proc")) s.AssertAllDirentTypes(collector, tasksStaticFiles) s.AssertDirentOffsets(collector, tasksStaticFilesNextOffs) } @@ -131,7 +160,7 @@ func TestTasks(t *testing.T) { var tasks []*kernel.Task for i := 0; i < 5; i++ { tc := k.NewThreadGroup(nil, k.RootPIDNamespace(), kernel.NewSignalHandlers(), linux.SIGCHLD, k.GlobalInit().Limits()) - task, err := testutil.CreateTask(s.Ctx, fmt.Sprintf("name-%d", i), tc) + task, err := testutil.CreateTask(s.Ctx, fmt.Sprintf("name-%d", i), tc, s.MntNs, s.Root, s.Root) if err != nil { t.Fatalf("CreateTask(): %v", err) } @@ -139,7 +168,7 @@ func TestTasks(t *testing.T) { expectedDirents[fmt.Sprintf("%d", i+1)] = linux.DT_DIR } - collector := s.ListDirents(s.PathOpAtRoot("/")) + collector := s.ListDirents(s.PathOpAtRoot("/proc")) s.AssertAllDirentTypes(collector, expectedDirents) s.AssertDirentOffsets(collector, tasksStaticFilesNextOffs) @@ -179,7 +208,7 @@ func TestTasks(t *testing.T) { } // Test lookup. - for _, path := range []string{"/1", "/2"} { + for _, path := range []string{"/proc/1", "/proc/2"} { fd, err := s.VFS.OpenAt( s.Ctx, s.Creds, @@ -189,6 +218,7 @@ func TestTasks(t *testing.T) { if err != nil { t.Fatalf("vfsfs.OpenAt(%q) failed: %v", path, err) } + defer fd.DecRef() buf := make([]byte, 1) bufIOSeq := usermem.BytesIOSequence(buf) if _, err := fd.Read(s.Ctx, bufIOSeq, vfs.ReadOptions{}); err != syserror.EISDIR { @@ -199,10 +229,10 @@ func TestTasks(t *testing.T) { if _, err := s.VFS.OpenAt( s.Ctx, s.Creds, - s.PathOpAtRoot("/9999"), + s.PathOpAtRoot("/proc/9999"), &vfs.OpenOptions{}, ); err != syserror.ENOENT { - t.Fatalf("wrong error from vfsfs.OpenAt(/9999): %v", err) + t.Fatalf("wrong error from vfsfs.OpenAt(/proc/9999): %v", err) } } @@ -213,7 +243,7 @@ func TestTasksOffset(t *testing.T) { k := kernel.KernelFromContext(s.Ctx) for i := 0; i < 3; i++ { tc := k.NewThreadGroup(nil, k.RootPIDNamespace(), kernel.NewSignalHandlers(), linux.SIGCHLD, k.GlobalInit().Limits()) - if _, err := testutil.CreateTask(s.Ctx, fmt.Sprintf("name-%d", i), tc); err != nil { + if _, err := testutil.CreateTask(s.Ctx, fmt.Sprintf("name-%d", i), tc, s.MntNs, s.Root, s.Root); err != nil { t.Fatalf("CreateTask(): %v", err) } } @@ -300,12 +330,13 @@ func TestTasksOffset(t *testing.T) { fd, err := s.VFS.OpenAt( s.Ctx, s.Creds, - s.PathOpAtRoot("/"), + s.PathOpAtRoot("/proc"), &vfs.OpenOptions{}, ) if err != nil { t.Fatalf("vfsfs.OpenAt(/) failed: %v", err) } + defer fd.DecRef() if _, err := fd.Seek(s.Ctx, tc.offset, linux.SEEK_SET); err != nil { t.Fatalf("Seek(%d, SEEK_SET): %v", tc.offset, err) } @@ -337,12 +368,12 @@ func TestTask(t *testing.T) { k := kernel.KernelFromContext(s.Ctx) tc := k.NewThreadGroup(nil, k.RootPIDNamespace(), kernel.NewSignalHandlers(), linux.SIGCHLD, k.GlobalInit().Limits()) - _, err := testutil.CreateTask(s.Ctx, "name", tc) + _, err := testutil.CreateTask(s.Ctx, "name", tc, s.MntNs, s.Root, s.Root) if err != nil { t.Fatalf("CreateTask(): %v", err) } - collector := s.ListDirents(s.PathOpAtRoot("/1")) + collector := s.ListDirents(s.PathOpAtRoot("/proc/1")) s.AssertAllDirentTypes(collector, taskStaticFiles) } @@ -352,7 +383,7 @@ func TestProcSelf(t *testing.T) { k := kernel.KernelFromContext(s.Ctx) tc := k.NewThreadGroup(nil, k.RootPIDNamespace(), kernel.NewSignalHandlers(), linux.SIGCHLD, k.GlobalInit().Limits()) - task, err := testutil.CreateTask(s.Ctx, "name", tc) + task, err := testutil.CreateTask(s.Ctx, "name", tc, s.MntNs, s.Root, s.Root) if err != nil { t.Fatalf("CreateTask(): %v", err) } @@ -360,14 +391,14 @@ func TestProcSelf(t *testing.T) { collector := s.WithTemporaryContext(task).ListDirents(&vfs.PathOperation{ Root: s.Root, Start: s.Root, - Path: fspath.Parse("/self/"), + Path: fspath.Parse("/proc/self/"), FollowFinalSymlink: true, }) s.AssertAllDirentTypes(collector, taskStaticFiles) } func iterateDir(ctx context.Context, t *testing.T, s *testutil.System, fd *vfs.FileDescription) { - t.Logf("Iterating: /proc%s", fd.MappedName(ctx)) + t.Logf("Iterating: %s", fd.MappedName(ctx)) var collector testutil.DirentCollector if err := fd.IterDirents(ctx, &collector); err != nil { @@ -384,35 +415,36 @@ func iterateDir(ctx context.Context, t *testing.T, s *testutil.System, fd *vfs.F if d.Name == "." || d.Name == ".." { continue } - childPath := path.Join(fd.MappedName(ctx), d.Name) + absPath := path.Join(fd.MappedName(ctx), d.Name) if d.Type == linux.DT_LNK { link, err := s.VFS.ReadlinkAt( ctx, auth.CredentialsFromContext(ctx), - &vfs.PathOperation{Root: s.Root, Start: s.Root, Path: fspath.Parse(childPath)}, + &vfs.PathOperation{Root: s.Root, Start: s.Root, Path: fspath.Parse(absPath)}, ) if err != nil { - t.Errorf("vfsfs.ReadlinkAt(%v) failed: %v", childPath, err) + t.Errorf("vfsfs.ReadlinkAt(%v) failed: %v", absPath, err) } else { - t.Logf("Skipping symlink: /proc%s => %s", childPath, link) + t.Logf("Skipping symlink: %s => %s", absPath, link) } continue } - t.Logf("Opening: /proc%s", childPath) + t.Logf("Opening: %s", absPath) child, err := s.VFS.OpenAt( ctx, auth.CredentialsFromContext(ctx), - &vfs.PathOperation{Root: s.Root, Start: s.Root, Path: fspath.Parse(childPath)}, + &vfs.PathOperation{Root: s.Root, Start: s.Root, Path: fspath.Parse(absPath)}, &vfs.OpenOptions{}, ) if err != nil { - t.Errorf("vfsfs.OpenAt(%v) failed: %v", childPath, err) + t.Errorf("vfsfs.OpenAt(%v) failed: %v", absPath, err) continue } + defer child.DecRef() stat, err := child.Stat(ctx, vfs.StatOptions{}) if err != nil { - t.Errorf("Stat(%v) failed: %v", childPath, err) + t.Errorf("Stat(%v) failed: %v", absPath, err) } if got := linux.FileMode(stat.Mode).DirentType(); got != d.Type { t.Errorf("wrong file mode, stat: %v, dirent: %v", got, d.Type) @@ -430,13 +462,31 @@ func TestTree(t *testing.T) { defer s.Destroy() k := kernel.KernelFromContext(s.Ctx) + + pop := &vfs.PathOperation{ + Root: s.Root, + Start: s.Root, + Path: fspath.Parse("test-file"), + } + opts := &vfs.OpenOptions{ + Flags: linux.O_RDONLY | linux.O_CREAT, + Mode: 0777, + } + file, err := s.VFS.OpenAt(s.Ctx, s.Creds, pop, opts) + if err != nil { + t.Fatalf("failed to create test file: %v", err) + } + defer file.DecRef() + var tasks []*kernel.Task for i := 0; i < 5; i++ { tc := k.NewThreadGroup(nil, k.RootPIDNamespace(), kernel.NewSignalHandlers(), linux.SIGCHLD, k.GlobalInit().Limits()) - task, err := testutil.CreateTask(s.Ctx, fmt.Sprintf("name-%d", i), tc) + task, err := testutil.CreateTask(s.Ctx, fmt.Sprintf("name-%d", i), tc, s.MntNs, s.Root, s.Root) if err != nil { t.Fatalf("CreateTask(): %v", err) } + // Add file to populate /proc/[pid]/fd and fdinfo directories. + task.FDTable().NewFDVFS2(task, 0, file, kernel.FDFlags{}) tasks = append(tasks, task) } @@ -444,11 +494,12 @@ func TestTree(t *testing.T) { fd, err := s.VFS.OpenAt( ctx, auth.CredentialsFromContext(s.Ctx), - &vfs.PathOperation{Root: s.Root, Start: s.Root, Path: fspath.Parse("/")}, + &vfs.PathOperation{Root: s.Root, Start: s.Root, Path: fspath.Parse("/proc")}, &vfs.OpenOptions{}, ) if err != nil { - t.Fatalf("vfsfs.OpenAt(/) failed: %v", err) + t.Fatalf("vfsfs.OpenAt(/proc) failed: %v", err) } iterateDir(ctx, t, s, fd) + fd.DecRef() } diff --git a/pkg/sentry/fsimpl/signalfd/BUILD b/pkg/sentry/fsimpl/signalfd/BUILD new file mode 100644 index 000000000..067c1657f --- /dev/null +++ b/pkg/sentry/fsimpl/signalfd/BUILD @@ -0,0 +1,20 @@ +load("//tools:defs.bzl", "go_library") + +package(licenses = ["notice"]) + +go_library( + name = "signalfd", + srcs = ["signalfd.go"], + visibility = ["//pkg/sentry:internal"], + deps = [ + "//pkg/abi/linux", + "//pkg/binary", + "//pkg/context", + "//pkg/sentry/kernel", + "//pkg/sentry/vfs", + "//pkg/sync", + "//pkg/syserror", + "//pkg/usermem", + "//pkg/waiter", + ], +) diff --git a/pkg/sentry/fsimpl/signalfd/signalfd.go b/pkg/sentry/fsimpl/signalfd/signalfd.go new file mode 100644 index 000000000..d29ef3f83 --- /dev/null +++ b/pkg/sentry/fsimpl/signalfd/signalfd.go @@ -0,0 +1,135 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package 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" + "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" + "gvisor.dev/gvisor/pkg/waiter" +) + +// SignalFileDescription implements FileDescriptionImpl for signal fds. +type SignalFileDescription struct { + vfsfd vfs.FileDescription + vfs.FileDescriptionDefaultImpl + vfs.DentryMetadataFileDescriptionImpl + + // target is the original signal target task. + // + // The semantics here are a bit broken. Linux will always use current + // for all reads, regardless of where the signalfd originated. We can't + // do exactly that because we need to plumb the context through + // EventRegister in order to support proper blocking behavior. This + // will undoubtedly become very complicated quickly. + target *kernel.Task + + // mu protects mask. + mu sync.Mutex + + // mask is the signal mask. Protected by mu. + mask linux.SignalSet +} + +var _ vfs.FileDescriptionImpl = (*SignalFileDescription)(nil) + +// New creates a new signal fd. +func New(vfsObj *vfs.VirtualFilesystem, target *kernel.Task, mask linux.SignalSet, flags uint32) (*vfs.FileDescription, error) { + vd := vfsObj.NewAnonVirtualDentry("[signalfd]") + defer vd.DecRef() + sfd := &SignalFileDescription{ + target: target, + mask: mask, + } + if err := sfd.vfsfd.Init(sfd, flags, vd.Mount(), vd.Dentry(), &vfs.FileDescriptionOptions{ + UseDentryMetadata: true, + DenyPRead: true, + DenyPWrite: true, + }); err != nil { + return nil, err + } + return &sfd.vfsfd, nil +} + +// Mask returns the signal mask. +func (sfd *SignalFileDescription) Mask() linux.SignalSet { + sfd.mu.Lock() + defer sfd.mu.Unlock() + return sfd.mask +} + +// SetMask sets the signal mask. +func (sfd *SignalFileDescription) SetMask(mask linux.SignalSet) { + sfd.mu.Lock() + defer sfd.mu.Unlock() + sfd.mask = mask +} + +// Read implements 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) + if err != nil { + // There must be no signal available. + return 0, syserror.ErrWouldBlock + } + + // Copy out the signal info using the specified format. + var buf [128]byte + binary.Marshal(buf[:0], usermem.ByteOrder, &linux.SignalfdSiginfo{ + Signo: uint32(info.Signo), + Errno: info.Errno, + Code: info.Code, + PID: uint32(info.Pid()), + UID: uint32(info.Uid()), + Status: info.Status(), + Overrun: uint32(info.Overrun()), + Addr: info.Addr(), + }) + n, err := dst.CopyOut(ctx, buf[:]) + return int64(n), err +} + +// Readiness implements waiter.Waitable.Readiness. +func (sfd *SignalFileDescription) Readiness(mask waiter.EventMask) waiter.EventMask { + sfd.mu.Lock() + defer sfd.mu.Unlock() + if mask&waiter.EventIn != 0 && sfd.target.PendingSignals()&sfd.mask != 0 { + return waiter.EventIn // Pending signals. + } + return 0 +} + +// EventRegister implements waiter.Waitable.EventRegister. +func (sfd *SignalFileDescription) EventRegister(entry *waiter.Entry, _ waiter.EventMask) { + sfd.mu.Lock() + defer sfd.mu.Unlock() + // Register for the signal set; ignore the passed events. + sfd.target.SignalRegister(entry, waiter.EventMask(sfd.mask)) +} + +// EventUnregister implements waiter.Waitable.EventUnregister. +func (sfd *SignalFileDescription) EventUnregister(entry *waiter.Entry) { + // Unregister the original entry. + sfd.target.SignalUnregister(entry) +} + +// Release implements FileDescriptionImpl.Release() +func (sfd *SignalFileDescription) Release() {} diff --git a/pkg/sentry/fsimpl/sockfs/BUILD b/pkg/sentry/fsimpl/sockfs/BUILD new file mode 100644 index 000000000..9453277b8 --- /dev/null +++ b/pkg/sentry/fsimpl/sockfs/BUILD @@ -0,0 +1,18 @@ +load("//tools:defs.bzl", "go_library") + +licenses(["notice"]) + +go_library( + name = "sockfs", + srcs = ["sockfs.go"], + visibility = ["//pkg/sentry:internal"], + deps = [ + "//pkg/abi/linux", + "//pkg/context", + "//pkg/fspath", + "//pkg/sentry/fsimpl/kernfs", + "//pkg/sentry/kernel/auth", + "//pkg/sentry/vfs", + "//pkg/syserror", + ], +) diff --git a/pkg/sentry/fsimpl/sockfs/sockfs.go b/pkg/sentry/fsimpl/sockfs/sockfs.go new file mode 100644 index 000000000..ee0828a15 --- /dev/null +++ b/pkg/sentry/fsimpl/sockfs/sockfs.go @@ -0,0 +1,109 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// Package sockfs provides a filesystem implementation for anonymous sockets. +package sockfs + +import ( + "fmt" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" +) + +// filesystemType implements vfs.FilesystemType. +type filesystemType struct{} + +// GetFilesystem implements FilesystemType.GetFilesystem. +func (fsType filesystemType) GetFilesystem(_ context.Context, vfsObj *vfs.VirtualFilesystem, _ *auth.Credentials, _ string, _ vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { + panic("sockfs.filesystemType.GetFilesystem should never be called") +} + +// Name implements FilesystemType.Name. +// +// Note that registering sockfs is unnecessary, except for the fact that it +// will not show up under /proc/filesystems as a result. This is a very minor +// discrepancy from Linux. +func (filesystemType) Name() string { + return "sockfs" +} + +type filesystem struct { + kernfs.Filesystem + + devMinor uint32 +} + +// NewFilesystem sets up and returns a new sockfs filesystem. +// +// Note that there should only ever be one instance of sockfs.Filesystem, +// backing a global socket mount. +func NewFilesystem(vfsObj *vfs.VirtualFilesystem) (*vfs.Filesystem, error) { + devMinor, err := vfsObj.GetAnonBlockDevMinor() + if err != nil { + return nil, err + } + fs := &filesystem{ + devMinor: devMinor, + } + fs.Filesystem.VFSFilesystem().Init(vfsObj, filesystemType{}, fs) + return fs.Filesystem.VFSFilesystem(), nil +} + +// Release implements vfs.FilesystemImpl.Release. +func (fs *filesystem) Release() { + fs.Filesystem.VFSFilesystem().VirtualFilesystem().PutAnonBlockDevMinor(fs.devMinor) + fs.Filesystem.Release() +} + +// PrependPath implements vfs.FilesystemImpl.PrependPath. +func (fs *filesystem) PrependPath(ctx context.Context, vfsroot, vd vfs.VirtualDentry, b *fspath.Builder) error { + inode := vd.Dentry().Impl().(*kernfs.Dentry).Inode().(*inode) + b.PrependComponent(fmt.Sprintf("socket:[%d]", inode.InodeAttrs.Ino())) + return vfs.PrependPathSyntheticError{} +} + +// inode implements kernfs.Inode. +type inode struct { + kernfs.InodeNotDirectory + kernfs.InodeNotSymlink + kernfs.InodeAttrs + kernfs.InodeNoopRefCount +} + +// Open implements kernfs.Inode.Open. +func (i *inode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + return nil, syserror.ENXIO +} + +// NewDentry constructs and returns a sockfs dentry. +// +// Preconditions: mnt.Filesystem() must have been returned by NewFilesystem(). +func NewDentry(creds *auth.Credentials, mnt *vfs.Mount) *vfs.Dentry { + fs := mnt.Filesystem().Impl().(*filesystem) + + // File mode matches net/socket.c:sock_alloc. + filemode := linux.FileMode(linux.S_IFSOCK | 0600) + i := &inode{} + i.InodeAttrs.Init(creds, linux.UNNAMED_MAJOR, fs.devMinor, fs.Filesystem.NextIno(), filemode) + + d := &kernfs.Dentry{} + d.Init(i) + return d.VFSDentry() +} diff --git a/pkg/sentry/fsimpl/sys/BUILD b/pkg/sentry/fsimpl/sys/BUILD index 66c0d8bc8..a741e2bb6 100644 --- a/pkg/sentry/fsimpl/sys/BUILD +++ b/pkg/sentry/fsimpl/sys/BUILD @@ -7,6 +7,7 @@ go_library( srcs = [ "sys.go", ], + visibility = ["//pkg/sentry:internal"], deps = [ "//pkg/abi/linux", "//pkg/context", diff --git a/pkg/sentry/fsimpl/sys/sys.go b/pkg/sentry/fsimpl/sys/sys.go index d693fceae..0af373604 100644 --- a/pkg/sentry/fsimpl/sys/sys.go +++ b/pkg/sentry/fsimpl/sys/sys.go @@ -28,18 +28,35 @@ import ( "gvisor.dev/gvisor/pkg/syserror" ) +// Name is the default filesystem name. +const Name = "sysfs" + // FilesystemType implements vfs.FilesystemType. type FilesystemType struct{} // filesystem implements vfs.FilesystemImpl. type filesystem struct { kernfs.Filesystem + + devMinor uint32 +} + +// Name implements vfs.FilesystemType.Name. +func (FilesystemType) Name() string { + return Name } // GetFilesystem implements vfs.FilesystemType.GetFilesystem. -func (FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { - fs := &filesystem{} - fs.Filesystem.Init(vfsObj) +func (fsType FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { + devMinor, err := vfsObj.GetAnonBlockDevMinor() + if err != nil { + return nil, nil, err + } + + fs := &filesystem{ + devMinor: devMinor, + } + fs.VFSFilesystem().Init(vfsObj, &fsType, fs) k := kernel.KernelFromContext(ctx) maxCPUCores := k.ApplicationCores() defaultSysDirMode := linux.FileMode(0755) @@ -69,6 +86,12 @@ func (FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFile return fs.VFSFilesystem(), root.VFSDentry(), nil } +// Release implements vfs.FilesystemImpl.Release. +func (fs *filesystem) Release() { + fs.Filesystem.VFSFilesystem().VirtualFilesystem().PutAnonBlockDevMinor(fs.devMinor) + fs.Filesystem.Release() +} + // dir implements kernfs.Inode. type dir struct { kernfs.InodeAttrs @@ -82,7 +105,7 @@ type dir struct { func (fs *filesystem) newDir(creds *auth.Credentials, mode linux.FileMode, contents map[string]*kernfs.Dentry) *kernfs.Dentry { d := &dir{} - d.InodeAttrs.Init(creds, fs.NextIno(), linux.ModeDirectory|0755) + d.InodeAttrs.Init(creds, linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), linux.ModeDirectory|0755) d.OrderedChildren.Init(kernfs.OrderedChildrenOptions{}) d.dentry.Init(d) @@ -91,15 +114,17 @@ func (fs *filesystem) newDir(creds *auth.Credentials, mode linux.FileMode, conte return &d.dentry } -// SetStat implements kernfs.Inode.SetStat. -func (d *dir) SetStat(fs *vfs.Filesystem, opts vfs.SetStatOptions) error { +// SetStat implements Inode.SetStat not allowing inode attributes to be changed. +func (*dir) SetStat(context.Context, *vfs.Filesystem, *auth.Credentials, vfs.SetStatOptions) error { return syserror.EPERM } // Open implements kernfs.Inode.Open. -func (d *dir) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { - fd := &kernfs.GenericDirectoryFD{} - fd.Init(rp.Mount(), vfsd, &d.OrderedChildren, &opts) +func (d *dir) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), vfsd, &d.OrderedChildren, &opts) + if err != nil { + return nil, err + } return fd.VFSFileDescription(), nil } @@ -117,7 +142,7 @@ func (c *cpuFile) Generate(ctx context.Context, buf *bytes.Buffer) error { func (fs *filesystem) newCPUFile(creds *auth.Credentials, maxCores uint, mode linux.FileMode) *kernfs.Dentry { c := &cpuFile{maxCores: maxCores} - c.DynamicBytesFile.Init(creds, fs.NextIno(), c, mode) + c.DynamicBytesFile.Init(creds, linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), c, mode) d := &kernfs.Dentry{} d.Init(c) return d diff --git a/pkg/sentry/fsimpl/sys/sys_test.go b/pkg/sentry/fsimpl/sys/sys_test.go index 8b1cf0bd0..4b3602d47 100644 --- a/pkg/sentry/fsimpl/sys/sys_test.go +++ b/pkg/sentry/fsimpl/sys/sys_test.go @@ -34,16 +34,15 @@ func newTestSystem(t *testing.T) *testutil.System { } ctx := k.SupervisorContext() creds := auth.CredentialsFromContext(ctx) - v := vfs.New() - v.MustRegisterFilesystemType("sysfs", sys.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ + k.VFS().MustRegisterFilesystemType(sys.Name, sys.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ AllowUserMount: true, }) - mns, err := v.NewMountNamespace(ctx, creds, "", "sysfs", &vfs.GetFilesystemOptions{}) + mns, err := k.VFS().NewMountNamespace(ctx, creds, "", sys.Name, &vfs.GetFilesystemOptions{}) if err != nil { t.Fatalf("Failed to create new mount namespace: %v", err) } - return testutil.NewSystem(ctx, t, v, mns) + return testutil.NewSystem(ctx, t, k.VFS(), mns) } func TestReadCPUFile(t *testing.T) { diff --git a/pkg/sentry/fsimpl/testutil/BUILD b/pkg/sentry/fsimpl/testutil/BUILD index efd5974c4..0e4053a46 100644 --- a/pkg/sentry/fsimpl/testutil/BUILD +++ b/pkg/sentry/fsimpl/testutil/BUILD @@ -16,12 +16,14 @@ go_library( "//pkg/cpuid", "//pkg/fspath", "//pkg/memutil", - "//pkg/sentry/fs", + "//pkg/sentry/fsbridge", + "//pkg/sentry/fsimpl/tmpfs", "//pkg/sentry/kernel", "//pkg/sentry/kernel/auth", "//pkg/sentry/kernel/sched", "//pkg/sentry/limits", "//pkg/sentry/loader", + "//pkg/sentry/mm", "//pkg/sentry/pgalloc", "//pkg/sentry/platform", "//pkg/sentry/platform/kvm", diff --git a/pkg/sentry/fsimpl/testutil/kernel.go b/pkg/sentry/fsimpl/testutil/kernel.go index 89f8c4915..c16a36cdb 100644 --- a/pkg/sentry/fsimpl/testutil/kernel.go +++ b/pkg/sentry/fsimpl/testutil/kernel.go @@ -23,16 +23,20 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/cpuid" + "gvisor.dev/gvisor/pkg/fspath" "gvisor.dev/gvisor/pkg/memutil" - "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/tmpfs" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/kernel/sched" "gvisor.dev/gvisor/pkg/sentry/limits" "gvisor.dev/gvisor/pkg/sentry/loader" + "gvisor.dev/gvisor/pkg/sentry/mm" "gvisor.dev/gvisor/pkg/sentry/pgalloc" "gvisor.dev/gvisor/pkg/sentry/platform" "gvisor.dev/gvisor/pkg/sentry/time" + "gvisor.dev/gvisor/pkg/sentry/vfs" // Platforms are plugable. _ "gvisor.dev/gvisor/pkg/sentry/platform/kvm" @@ -99,40 +103,67 @@ func Boot() (*kernel.Kernel, error) { return nil, fmt.Errorf("initializing kernel: %v", err) } - ctx := k.SupervisorContext() + kernel.VFS2Enabled = true - // Create mount namespace without root as it's the minimum required to create - // the global thread group. - mntns, err := fs.NewMountNamespace(ctx, nil) - if err != nil { - return nil, err + if err := k.VFS().Init(); err != nil { + return nil, fmt.Errorf("VFS init: %v", err) } + k.VFS().MustRegisterFilesystemType(tmpfs.Name, &tmpfs.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ + AllowUserMount: true, + AllowUserList: true, + }) + ls, err := limits.NewLinuxLimitSet() if err != nil { return nil, err } - tg := k.NewThreadGroup(mntns, k.RootPIDNamespace(), kernel.NewSignalHandlers(), linux.SIGCHLD, ls) + tg := k.NewThreadGroup(nil, k.RootPIDNamespace(), kernel.NewSignalHandlers(), linux.SIGCHLD, ls) k.TestOnly_SetGlobalInit(tg) return k, nil } // CreateTask creates a new bare bones task for tests. -func CreateTask(ctx context.Context, name string, tc *kernel.ThreadGroup) (*kernel.Task, error) { +func CreateTask(ctx context.Context, name string, tc *kernel.ThreadGroup, mntns *vfs.MountNamespace, root, cwd vfs.VirtualDentry) (*kernel.Task, error) { k := kernel.KernelFromContext(ctx) + exe, err := newFakeExecutable(ctx, k.VFS(), auth.CredentialsFromContext(ctx), root) + if err != nil { + return nil, err + } + m := mm.NewMemoryManager(k, k, k.SleepForAddressSpaceActivation) + m.SetExecutable(fsbridge.NewVFSFile(exe)) + config := &kernel.TaskConfig{ Kernel: k, ThreadGroup: tc, - TaskContext: &kernel.TaskContext{Name: name}, + TaskContext: &kernel.TaskContext{Name: name, MemoryManager: m}, Credentials: auth.CredentialsFromContext(ctx), + NetworkNamespace: k.RootNetworkNamespace(), AllowedCPUMask: sched.NewFullCPUSet(k.ApplicationCores()), UTSNamespace: kernel.UTSNamespaceFromContext(ctx), IPCNamespace: kernel.IPCNamespaceFromContext(ctx), AbstractSocketNamespace: kernel.NewAbstractSocketNamespace(), + MountNamespaceVFS2: mntns, + FSContext: kernel.NewFSContextVFS2(root, cwd, 0022), + FDTable: k.NewFDTable(), } return k.TaskSet().NewTask(config) } +func newFakeExecutable(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, root vfs.VirtualDentry) (*vfs.FileDescription, error) { + const name = "executable" + pop := &vfs.PathOperation{ + Root: root, + Start: root, + Path: fspath.Parse(name), + } + opts := &vfs.OpenOptions{ + Flags: linux.O_RDONLY | linux.O_CREAT, + Mode: 0777, + } + return vfsObj.OpenAt(ctx, creds, pop, opts) +} + func createMemoryFile() (*pgalloc.MemoryFile, error) { const memfileName = "test-memory" memfd, err := memutil.CreateMemFD(memfileName, 0) diff --git a/pkg/sentry/fsimpl/testutil/testutil.go b/pkg/sentry/fsimpl/testutil/testutil.go index 69fd84ddd..0556af877 100644 --- a/pkg/sentry/fsimpl/testutil/testutil.go +++ b/pkg/sentry/fsimpl/testutil/testutil.go @@ -41,12 +41,12 @@ type System struct { Creds *auth.Credentials VFS *vfs.VirtualFilesystem Root vfs.VirtualDentry - mns *vfs.MountNamespace + MntNs *vfs.MountNamespace } // NewSystem constructs a System. // -// Precondition: Caller must hold a reference on mns, whose ownership +// Precondition: Caller must hold a reference on MntNs, whose ownership // is transferred to the new System. func NewSystem(ctx context.Context, t *testing.T, v *vfs.VirtualFilesystem, mns *vfs.MountNamespace) *System { s := &System{ @@ -54,7 +54,7 @@ func NewSystem(ctx context.Context, t *testing.T, v *vfs.VirtualFilesystem, mns Ctx: ctx, Creds: auth.CredentialsFromContext(ctx), VFS: v, - mns: mns, + MntNs: mns, Root: mns.Root(), } return s @@ -75,7 +75,7 @@ func (s *System) WithSubtest(t *testing.T) *System { Ctx: s.Ctx, Creds: s.Creds, VFS: s.VFS, - mns: s.mns, + MntNs: s.MntNs, Root: s.Root, } } @@ -90,7 +90,7 @@ func (s *System) WithTemporaryContext(ctx context.Context) *System { Ctx: ctx, Creds: s.Creds, VFS: s.VFS, - mns: s.mns, + MntNs: s.MntNs, Root: s.Root, } } @@ -98,7 +98,7 @@ func (s *System) WithTemporaryContext(ctx context.Context) *System { // Destroy release resources associated with a test system. func (s *System) Destroy() { s.Root.DecRef() - s.mns.DecRef() // Reference on mns passed to NewSystem. + s.MntNs.DecRef() // Reference on MntNs passed to NewSystem. } // ReadToEnd reads the contents of fd until EOF to a string. @@ -162,6 +162,9 @@ func (s *System) ListDirents(pop *vfs.PathOperation) *DirentCollector { // exactly the specified set of expected entries. AssertAllDirentTypes respects // collector.skipDots, and implicitly checks for "." and ".." accordingly. func (s *System) AssertAllDirentTypes(collector *DirentCollector, expected map[string]DirentType) { + if expected == nil { + expected = make(map[string]DirentType) + } // Also implicitly check for "." and "..", if enabled. if !collector.skipDots { expected["."] = linux.DT_DIR @@ -226,7 +229,7 @@ func (d *DirentCollector) SkipDotsChecks(value bool) { } // Handle implements vfs.IterDirentsCallback.Handle. -func (d *DirentCollector) Handle(dirent vfs.Dirent) bool { +func (d *DirentCollector) Handle(dirent vfs.Dirent) error { d.mu.Lock() if d.dirents == nil { d.dirents = make(map[string]*vfs.Dirent) @@ -234,7 +237,7 @@ func (d *DirentCollector) Handle(dirent vfs.Dirent) bool { d.order = append(d.order, &dirent) d.dirents[dirent.Name] = &dirent d.mu.Unlock() - return true + return nil } // Count returns the number of dirents currently in the collector. diff --git a/pkg/sentry/fsimpl/timerfd/BUILD b/pkg/sentry/fsimpl/timerfd/BUILD new file mode 100644 index 000000000..fbb02a271 --- /dev/null +++ b/pkg/sentry/fsimpl/timerfd/BUILD @@ -0,0 +1,17 @@ +load("//tools:defs.bzl", "go_library") + +licenses(["notice"]) + +go_library( + name = "timerfd", + srcs = ["timerfd.go"], + visibility = ["//pkg/sentry:internal"], + deps = [ + "//pkg/context", + "//pkg/sentry/kernel/time", + "//pkg/sentry/vfs", + "//pkg/syserror", + "//pkg/usermem", + "//pkg/waiter", + ], +) diff --git a/pkg/sentry/fsimpl/timerfd/timerfd.go b/pkg/sentry/fsimpl/timerfd/timerfd.go new file mode 100644 index 000000000..60c92d626 --- /dev/null +++ b/pkg/sentry/fsimpl/timerfd/timerfd.go @@ -0,0 +1,143 @@ +// 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 timerfd implements timer fds. +package timerfd + +import ( + "sync/atomic" + + "gvisor.dev/gvisor/pkg/context" + ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" + "gvisor.dev/gvisor/pkg/waiter" +) + +// TimerFileDescription implements FileDescriptionImpl for timer fds. It also +// implements ktime.TimerListener. +type TimerFileDescription struct { + vfsfd vfs.FileDescription + vfs.FileDescriptionDefaultImpl + vfs.DentryMetadataFileDescriptionImpl + + events waiter.Queue + timer *ktime.Timer + + // val is the number of timer expirations since the last successful + // call to PRead, or SetTime. val must be accessed using atomic memory + // operations. + val uint64 +} + +var _ vfs.FileDescriptionImpl = (*TimerFileDescription)(nil) +var _ ktime.TimerListener = (*TimerFileDescription)(nil) + +// New returns a new timer fd. +func New(vfsObj *vfs.VirtualFilesystem, clock ktime.Clock, flags uint32) (*vfs.FileDescription, error) { + vd := vfsObj.NewAnonVirtualDentry("[timerfd]") + defer vd.DecRef() + tfd := &TimerFileDescription{} + tfd.timer = ktime.NewTimer(clock, tfd) + if err := tfd.vfsfd.Init(tfd, flags, vd.Mount(), vd.Dentry(), &vfs.FileDescriptionOptions{ + UseDentryMetadata: true, + DenyPRead: true, + DenyPWrite: true, + }); err != nil { + return nil, err + } + return &tfd.vfsfd, nil +} + +// Read implements FileDescriptionImpl.Read. +func (tfd *TimerFileDescription) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { + const sizeofUint64 = 8 + if dst.NumBytes() < sizeofUint64 { + return 0, syserror.EINVAL + } + if val := atomic.SwapUint64(&tfd.val, 0); val != 0 { + var buf [sizeofUint64]byte + usermem.ByteOrder.PutUint64(buf[:], val) + if _, err := dst.CopyOut(ctx, buf[:]); err != nil { + // Linux does not undo consuming the number of + // expirations even if writing to userspace fails. + return 0, err + } + return sizeofUint64, nil + } + return 0, syserror.ErrWouldBlock +} + +// Clock returns the timer fd's Clock. +func (tfd *TimerFileDescription) Clock() ktime.Clock { + return tfd.timer.Clock() +} + +// GetTime returns the associated Timer's setting and the time at which it was +// observed. +func (tfd *TimerFileDescription) GetTime() (ktime.Time, ktime.Setting) { + return tfd.timer.Get() +} + +// SetTime atomically changes the associated Timer's setting, resets the number +// of expirations to 0, and returns the previous setting and the time at which +// it was observed. +func (tfd *TimerFileDescription) SetTime(s ktime.Setting) (ktime.Time, ktime.Setting) { + return tfd.timer.SwapAnd(s, func() { atomic.StoreUint64(&tfd.val, 0) }) +} + +// Readiness implements waiter.Waitable.Readiness. +func (tfd *TimerFileDescription) Readiness(mask waiter.EventMask) waiter.EventMask { + var ready waiter.EventMask + if atomic.LoadUint64(&tfd.val) != 0 { + ready |= waiter.EventIn + } + return ready +} + +// EventRegister implements waiter.Waitable.EventRegister. +func (tfd *TimerFileDescription) EventRegister(e *waiter.Entry, mask waiter.EventMask) { + tfd.events.EventRegister(e, mask) +} + +// EventUnregister implements waiter.Waitable.EventUnregister. +func (tfd *TimerFileDescription) EventUnregister(e *waiter.Entry) { + tfd.events.EventUnregister(e) +} + +// PauseTimer pauses the associated Timer. +func (tfd *TimerFileDescription) PauseTimer() { + tfd.timer.Pause() +} + +// ResumeTimer resumes the associated Timer. +func (tfd *TimerFileDescription) ResumeTimer() { + tfd.timer.Resume() +} + +// Release implements FileDescriptionImpl.Release() +func (tfd *TimerFileDescription) Release() { + tfd.timer.Destroy() +} + +// Notify implements ktime.TimerListener.Notify. +func (tfd *TimerFileDescription) Notify(exp uint64, setting ktime.Setting) (ktime.Setting, bool) { + atomic.AddUint64(&tfd.val, exp) + tfd.events.Notify(waiter.EventIn) + return ktime.Setting{}, false +} + +// Destroy implements ktime.TimerListener.Destroy. +func (tfd *TimerFileDescription) Destroy() {} diff --git a/pkg/sentry/fsimpl/tmpfs/BUILD b/pkg/sentry/fsimpl/tmpfs/BUILD index 57abd5583..062321cbc 100644 --- a/pkg/sentry/fsimpl/tmpfs/BUILD +++ b/pkg/sentry/fsimpl/tmpfs/BUILD @@ -15,6 +15,17 @@ go_template_instance( }, ) +go_template_instance( + name = "fstree", + out = "fstree.go", + package = "tmpfs", + prefix = "generic", + template = "//pkg/sentry/vfs/genericfstree:generic_fstree", + types = { + "Dentry": "dentry", + }, +) + go_library( name = "tmpfs", srcs = [ @@ -22,8 +33,10 @@ go_library( "device_file.go", "directory.go", "filesystem.go", + "fstree.go", "named_pipe.go", "regular_file.go", + "socket_file.go", "symlink.go", "tmpfs.go", ], @@ -39,16 +52,18 @@ go_library( "//pkg/sentry/fs", "//pkg/sentry/fs/fsutil", "//pkg/sentry/fs/lock", - "//pkg/sentry/kernel", "//pkg/sentry/kernel/auth", "//pkg/sentry/kernel/pipe", "//pkg/sentry/kernel/time", "//pkg/sentry/memmap", "//pkg/sentry/pgalloc", "//pkg/sentry/platform", + "//pkg/sentry/socket/unix/transport", + "//pkg/sentry/uniqueid", "//pkg/sentry/usage", "//pkg/sentry/vfs", "//pkg/sentry/vfs/lock", + "//pkg/sentry/vfs/memxattr", "//pkg/sync", "//pkg/syserror", "//pkg/usermem", @@ -81,6 +96,7 @@ go_test( "pipe_test.go", "regular_file_test.go", "stat_test.go", + "tmpfs_test.go", ], library = ":tmpfs", deps = [ @@ -90,7 +106,6 @@ go_test( "//pkg/sentry/contexttest", "//pkg/sentry/fs/lock", "//pkg/sentry/kernel/auth", - "//pkg/sentry/kernel/contexttest", "//pkg/sentry/vfs", "//pkg/syserror", "//pkg/usermem", diff --git a/pkg/sentry/fsimpl/tmpfs/benchmark_test.go b/pkg/sentry/fsimpl/tmpfs/benchmark_test.go index 9fce5e4b4..2fb5c4d84 100644 --- a/pkg/sentry/fsimpl/tmpfs/benchmark_test.go +++ b/pkg/sentry/fsimpl/tmpfs/benchmark_test.go @@ -168,14 +168,17 @@ func BenchmarkVFS1TmpfsStat(b *testing.B) { } } -func BenchmarkVFS2MemfsStat(b *testing.B) { +func BenchmarkVFS2TmpfsStat(b *testing.B) { for _, depth := range depths { b.Run(fmt.Sprintf("%d", depth), func(b *testing.B) { ctx := contexttest.Context(b) creds := auth.CredentialsFromContext(ctx) // Create VFS. - vfsObj := vfs.New() + vfsObj := vfs.VirtualFilesystem{} + if err := vfsObj.Init(); err != nil { + b.Fatalf("VFS init: %v", err) + } vfsObj.MustRegisterFilesystemType("tmpfs", tmpfs.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ AllowUserMount: true, }) @@ -359,14 +362,17 @@ func BenchmarkVFS1TmpfsMountStat(b *testing.B) { } } -func BenchmarkVFS2MemfsMountStat(b *testing.B) { +func BenchmarkVFS2TmpfsMountStat(b *testing.B) { for _, depth := range depths { b.Run(fmt.Sprintf("%d", depth), func(b *testing.B) { ctx := contexttest.Context(b) creds := auth.CredentialsFromContext(ctx) // Create VFS. - vfsObj := vfs.New() + vfsObj := vfs.VirtualFilesystem{} + if err := vfsObj.Init(); err != nil { + b.Fatalf("VFS init: %v", err) + } vfsObj.MustRegisterFilesystemType("tmpfs", tmpfs.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ AllowUserMount: true, }) @@ -432,13 +438,6 @@ func BenchmarkVFS2MemfsMountStat(b *testing.B) { filePathBuilder.WriteByte('/') } - // Verify that we didn't create any directories under the mount - // point (i.e. they were all created on the submount). - firstDirName := fmt.Sprintf("%d", depth) - if child := mountPoint.Dentry().Child(firstDirName); child != nil { - b.Fatalf("created directory %q under root mount, not submount", firstDirName) - } - // Create the file that will be stat'd. fd, err := vfsObj.OpenAt(ctx, creds, &vfs.PathOperation{ Root: root, diff --git a/pkg/sentry/fsimpl/tmpfs/device_file.go b/pkg/sentry/fsimpl/tmpfs/device_file.go index 84b181b90..ac54d420d 100644 --- a/pkg/sentry/fsimpl/tmpfs/device_file.go +++ b/pkg/sentry/fsimpl/tmpfs/device_file.go @@ -15,6 +15,8 @@ package tmpfs import ( + "fmt" + "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/vfs" @@ -27,13 +29,21 @@ type deviceFile struct { minor uint32 } -func (fs *filesystem) newDeviceFile(creds *auth.Credentials, mode linux.FileMode, kind vfs.DeviceKind, major, minor uint32) *inode { +func (fs *filesystem) newDeviceFile(kuid auth.KUID, kgid auth.KGID, mode linux.FileMode, kind vfs.DeviceKind, major, minor uint32) *inode { file := &deviceFile{ kind: kind, major: major, minor: minor, } - file.inode.init(file, fs, creds, mode) + switch kind { + case vfs.BlockDevice: + mode |= linux.S_IFBLK + case vfs.CharDevice: + mode |= linux.S_IFCHR + default: + panic(fmt.Sprintf("invalid DeviceKind: %v", kind)) + } + file.inode.init(file, fs, kuid, kgid, mode) file.inode.nlink = 1 // from parent directory return &file.inode } diff --git a/pkg/sentry/fsimpl/tmpfs/directory.go b/pkg/sentry/fsimpl/tmpfs/directory.go index dc0d27cf9..913b8a6c5 100644 --- a/pkg/sentry/fsimpl/tmpfs/directory.go +++ b/pkg/sentry/fsimpl/tmpfs/directory.go @@ -15,40 +15,78 @@ package tmpfs import ( + "sync/atomic" + "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/syserror" ) type directory struct { - inode inode + // Since directories can't be hard-linked, each directory can only be + // associated with a single dentry, which we can store in the directory + // struct. + dentry dentry + inode inode + + // childMap maps the names of the directory's children to their dentries. + // childMap is protected by filesystem.mu. + childMap map[string]*dentry - // childList is a list containing (1) child Dentries and (2) fake Dentries + // numChildren is len(childMap), but accessed using atomic memory + // operations to avoid locking in inode.statTo(). + numChildren int64 + + // childList is a list containing (1) child dentries and (2) fake dentries // (with inode == nil) that represent the iteration position of // directoryFDs. childList is used to support directoryFD.IterDirents() - // efficiently. childList is protected by filesystem.mu. + // efficiently. childList is protected by iterMu. + iterMu sync.Mutex childList dentryList } -func (fs *filesystem) newDirectory(creds *auth.Credentials, mode linux.FileMode) *inode { +func (fs *filesystem) newDirectory(kuid auth.KUID, kgid auth.KGID, mode linux.FileMode) *directory { dir := &directory{} - dir.inode.init(dir, fs, creds, mode) + dir.inode.init(dir, fs, kuid, kgid, linux.S_IFDIR|mode) dir.inode.nlink = 2 // from "." and parent directory or ".." for root - return &dir.inode + dir.dentry.inode = &dir.inode + dir.dentry.vfsd.Init(&dir.dentry) + return dir } -func (i *inode) isDir() bool { - _, ok := i.impl.(*directory) - return ok +// Preconditions: filesystem.mu must be locked for writing. dir must not +// already contain a child with the given name. +func (dir *directory) insertChildLocked(child *dentry, name string) { + child.parent = &dir.dentry + child.name = name + if dir.childMap == nil { + dir.childMap = make(map[string]*dentry) + } + dir.childMap[name] = child + atomic.AddInt64(&dir.numChildren, 1) + dir.iterMu.Lock() + dir.childList.PushBack(child) + dir.iterMu.Unlock() +} + +// Preconditions: filesystem.mu must be locked for writing. +func (dir *directory) removeChildLocked(child *dentry) { + delete(dir.childMap, child.name) + atomic.AddInt64(&dir.numChildren, -1) + dir.iterMu.Lock() + dir.childList.Remove(child) + dir.iterMu.Unlock() + child.unlinked = true } type directoryFD struct { fileDescription vfs.DirectoryFileDescriptionDefaultImpl - // Protected by filesystem.mu. + // Protected by directory.iterMu. iter *dentry off int64 } @@ -56,11 +94,10 @@ type directoryFD struct { // Release implements vfs.FileDescriptionImpl.Release. func (fd *directoryFD) Release() { if fd.iter != nil { - fs := fd.filesystem() dir := fd.inode().impl.(*directory) - fs.mu.Lock() + dir.iterMu.Lock() dir.childList.Remove(fd.iter) - fs.mu.Unlock() + dir.iterMu.Unlock() fd.iter = nil } } @@ -68,36 +105,42 @@ func (fd *directoryFD) Release() { // IterDirents implements vfs.FileDescriptionImpl.IterDirents. func (fd *directoryFD) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback) error { fs := fd.filesystem() - vfsd := fd.vfsfd.VirtualDentry().Dentry() + dir := fd.inode().impl.(*directory) - fs.mu.Lock() - defer fs.mu.Unlock() + // fs.mu is required to read d.parent and dentry.name. + fs.mu.RLock() + defer fs.mu.RUnlock() + dir.iterMu.Lock() + defer dir.iterMu.Unlock() + + fd.dentry().InotifyWithParent(linux.IN_ACCESS, 0, vfs.PathEvent) + fd.inode().touchAtime(fd.vfsfd.Mount()) if fd.off == 0 { - if !cb.Handle(vfs.Dirent{ + if err := cb.Handle(vfs.Dirent{ Name: ".", Type: linux.DT_DIR, - Ino: vfsd.Impl().(*dentry).inode.ino, + Ino: dir.inode.ino, NextOff: 1, - }) { - return nil + }); err != nil { + return err } fd.off++ } + if fd.off == 1 { - parentInode := vfsd.ParentOrSelf().Impl().(*dentry).inode - if !cb.Handle(vfs.Dirent{ + parentInode := genericParentOrSelf(&dir.dentry).inode + if err := cb.Handle(vfs.Dirent{ Name: "..", Type: parentInode.direntType(), Ino: parentInode.ino, NextOff: 2, - }) { - return nil + }); err != nil { + return err } fd.off++ } - dir := vfsd.Impl().(*dentry).inode.impl.(*directory) var child *dentry if fd.iter == nil { // Start iteration at the beginning of dir. @@ -111,14 +154,14 @@ func (fd *directoryFD) IterDirents(ctx context.Context, cb vfs.IterDirentsCallba for child != nil { // Skip other directoryFD iterators. if child.inode != nil { - if !cb.Handle(vfs.Dirent{ - Name: child.vfsd.Name(), + if err := cb.Handle(vfs.Dirent{ + Name: child.name, Type: child.inode.direntType(), Ino: child.inode.ino, NextOff: fd.off + 1, - }) { + }); err != nil { dir.childList.InsertBefore(child, fd.iter) - return nil + return err } fd.off++ } @@ -130,9 +173,9 @@ func (fd *directoryFD) IterDirents(ctx context.Context, cb vfs.IterDirentsCallba // Seek implements vfs.FileDescriptionImpl.Seek. func (fd *directoryFD) Seek(ctx context.Context, offset int64, whence int32) (int64, error) { - fs := fd.filesystem() - fs.mu.Lock() - defer fs.mu.Unlock() + dir := fd.inode().impl.(*directory) + dir.iterMu.Lock() + defer dir.iterMu.Unlock() switch whence { case linux.SEEK_SET: @@ -160,8 +203,6 @@ func (fd *directoryFD) Seek(ctx context.Context, offset int64, whence int32) (in remChildren = offset - 2 } - dir := fd.inode().impl.(*directory) - // Ensure that fd.iter exists and is not linked into dir.childList. if fd.iter == nil { fd.iter = &dentry{} diff --git a/pkg/sentry/fsimpl/tmpfs/filesystem.go b/pkg/sentry/fsimpl/tmpfs/filesystem.go index 8785452b6..e801680e8 100644 --- a/pkg/sentry/fsimpl/tmpfs/filesystem.go +++ b/pkg/sentry/fsimpl/tmpfs/filesystem.go @@ -21,8 +21,11 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" ) // Sync implements vfs.FilesystemImpl.Sync. @@ -38,33 +41,52 @@ func (fs *filesystem) Sync(ctx context.Context) error { // // Preconditions: filesystem.mu must be locked. !rp.Done(). func stepLocked(rp *vfs.ResolvingPath, d *dentry) (*dentry, error) { - if !d.inode.isDir() { + dir, ok := d.inode.impl.(*directory) + if !ok { return nil, syserror.ENOTDIR } - if err := d.inode.checkPermissions(rp.Credentials(), vfs.MayExec, true); err != nil { + if err := d.inode.checkPermissions(rp.Credentials(), vfs.MayExec); err != nil { return nil, err } afterSymlink: - nextVFSD, err := rp.ResolveComponent(&d.vfsd) - if err != nil { - return nil, err + name := rp.Component() + if name == "." { + rp.Advance() + return d, nil + } + if name == ".." { + if isRoot, err := rp.CheckRoot(&d.vfsd); err != nil { + return nil, err + } else if isRoot || d.parent == nil { + rp.Advance() + return d, nil + } + if err := rp.CheckMount(&d.parent.vfsd); err != nil { + return nil, err + } + rp.Advance() + return d.parent, nil + } + if len(name) > linux.NAME_MAX { + return nil, syserror.ENAMETOOLONG } - if nextVFSD == nil { - // Since the Dentry tree is the sole source of truth for tmpfs, if it's - // not in the Dentry tree, it doesn't exist. + child, ok := dir.childMap[name] + if !ok { return nil, syserror.ENOENT } - next := nextVFSD.Impl().(*dentry) - if symlink, ok := next.inode.impl.(*symlink); ok && rp.ShouldFollowSymlink() { - // TODO(gvisor.dev/issues/1197): Symlink traversals updates - // access time. + if err := rp.CheckMount(&child.vfsd); err != nil { + return nil, err + } + if symlink, ok := child.inode.impl.(*symlink); ok && rp.ShouldFollowSymlink() { + // Symlink traversal updates access time. + atomic.StoreInt64(&d.inode.atime, d.inode.fs.clock.Now().Nanoseconds()) if err := rp.HandleSymlink(symlink.target); err != nil { return nil, err } goto afterSymlink // don't check the current directory again } rp.Advance() - return next, nil + return child, nil } // walkParentDirLocked resolves all but the last path component of rp to an @@ -76,7 +98,7 @@ afterSymlink: // fs/namei.c:path_parentat(). // // Preconditions: filesystem.mu must be locked. !rp.Done(). -func walkParentDirLocked(rp *vfs.ResolvingPath, d *dentry) (*dentry, error) { +func walkParentDirLocked(rp *vfs.ResolvingPath, d *dentry) (*directory, error) { for !rp.Final() { next, err := stepLocked(rp, d) if err != nil { @@ -84,10 +106,11 @@ func walkParentDirLocked(rp *vfs.ResolvingPath, d *dentry) (*dentry, error) { } d = next } - if !d.inode.isDir() { + dir, ok := d.inode.impl.(*directory) + if !ok { return nil, syserror.ENOTDIR } - return d, nil + return dir, nil } // resolveLocked resolves rp to an existing file. @@ -118,33 +141,32 @@ func resolveLocked(rp *vfs.ResolvingPath) (*dentry, error) { // // Preconditions: !rp.Done(). For the final path component in rp, // !rp.ShouldFollowSymlink(). -func (fs *filesystem) doCreateAt(rp *vfs.ResolvingPath, dir bool, create func(parent *dentry, name string) error) error { +func (fs *filesystem) doCreateAt(rp *vfs.ResolvingPath, dir bool, create func(parentDir *directory, name string) error) error { fs.mu.Lock() defer fs.mu.Unlock() - parent, err := walkParentDirLocked(rp, rp.Start().Impl().(*dentry)) + parentDir, err := walkParentDirLocked(rp, rp.Start().Impl().(*dentry)) if err != nil { return err } - if err := parent.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true /* isDir */); err != nil { + if err := parentDir.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec); err != nil { return err } name := rp.Component() if name == "." || name == ".." { return syserror.EEXIST } - // Call parent.vfsd.Child() instead of stepLocked() or rp.ResolveChild(), - // because if the child exists we want to return EEXIST immediately instead - // of attempting symlink/mount traversal. - if parent.vfsd.Child(name) != nil { + if len(name) > linux.NAME_MAX { + return syserror.ENAMETOOLONG + } + if _, ok := parentDir.childMap[name]; ok { return syserror.EEXIST } if !dir && rp.MustBeDir() { return syserror.ENOENT } - // In memfs, the only way to cause a dentry to be disowned is by removing - // it from the filesystem, so this check is equivalent to checking if - // parent has been removed. - if parent.vfsd.IsDisowned() { + // tmpfs never calls VFS.InvalidateDentry(), so parentDir.dentry can only + // be dead if it was deleted. + if parentDir.dentry.vfsd.IsDead() { return syserror.ENOENT } mnt := rp.Mount() @@ -152,7 +174,28 @@ func (fs *filesystem) doCreateAt(rp *vfs.ResolvingPath, dir bool, create func(pa return err } defer mnt.EndWrite() - return create(parent, name) + if err := create(parentDir, name); err != nil { + return err + } + + ev := linux.IN_CREATE + if dir { + ev |= linux.IN_ISDIR + } + parentDir.inode.watches.Notify(name, uint32(ev), 0, vfs.InodeEvent) + parentDir.inode.touchCMtime() + return nil +} + +// AccessAt implements vfs.Filesystem.Impl.AccessAt. +func (fs *filesystem) AccessAt(ctx context.Context, rp *vfs.ResolvingPath, creds *auth.Credentials, ats vfs.AccessTypes) error { + fs.mu.RLock() + defer fs.mu.RUnlock() + d, err := resolveLocked(rp) + if err != nil { + return err + } + return d.inode.checkPermissions(creds, ats) } // GetDentryAt implements vfs.FilesystemImpl.GetDentryAt. @@ -167,7 +210,7 @@ func (fs *filesystem) GetDentryAt(ctx context.Context, rp *vfs.ResolvingPath, op if !d.inode.isDir() { return nil, syserror.ENOTDIR } - if err := d.inode.checkPermissions(rp.Credentials(), vfs.MayExec, true /* isDir */); err != nil { + if err := d.inode.checkPermissions(rp.Credentials(), vfs.MayExec); err != nil { return nil, err } } @@ -179,17 +222,17 @@ func (fs *filesystem) GetDentryAt(ctx context.Context, rp *vfs.ResolvingPath, op func (fs *filesystem) GetParentDentryAt(ctx context.Context, rp *vfs.ResolvingPath) (*vfs.Dentry, error) { fs.mu.RLock() defer fs.mu.RUnlock() - d, err := walkParentDirLocked(rp, rp.Start().Impl().(*dentry)) + dir, err := walkParentDirLocked(rp, rp.Start().Impl().(*dentry)) if err != nil { return nil, err } - d.IncRef() - return &d.vfsd, nil + dir.dentry.IncRef() + return &dir.dentry.vfsd, nil } // LinkAt implements vfs.FilesystemImpl.LinkAt. func (fs *filesystem) LinkAt(ctx context.Context, rp *vfs.ResolvingPath, vd vfs.VirtualDentry) error { - return fs.doCreateAt(rp, false /* dir */, func(parent *dentry, name string) error { + return fs.doCreateAt(rp, false /* dir */, func(parentDir *directory, name string) error { if rp.Mount() != vd.Mount() { return syserror.EXDEV } @@ -204,49 +247,47 @@ func (fs *filesystem) LinkAt(ctx context.Context, rp *vfs.ResolvingPath, vd vfs. return syserror.EMLINK } d.inode.incLinksLocked() - child := fs.newDentry(d.inode) - parent.vfsd.InsertChild(&child.vfsd, name) - parent.inode.impl.(*directory).childList.PushBack(child) + d.inode.watches.Notify("", linux.IN_ATTRIB, 0, vfs.InodeEvent) + parentDir.insertChildLocked(fs.newDentry(d.inode), name) return nil }) } // MkdirAt implements vfs.FilesystemImpl.MkdirAt. func (fs *filesystem) MkdirAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.MkdirOptions) error { - return fs.doCreateAt(rp, true /* dir */, func(parent *dentry, name string) error { - if parent.inode.nlink == maxLinks { + return fs.doCreateAt(rp, true /* dir */, func(parentDir *directory, name string) error { + creds := rp.Credentials() + if parentDir.inode.nlink == maxLinks { return syserror.EMLINK } - parent.inode.incLinksLocked() // from child's ".." - child := fs.newDentry(fs.newDirectory(rp.Credentials(), opts.Mode)) - parent.vfsd.InsertChild(&child.vfsd, name) - parent.inode.impl.(*directory).childList.PushBack(child) + parentDir.inode.incLinksLocked() // from child's ".." + childDir := fs.newDirectory(creds.EffectiveKUID, creds.EffectiveKGID, opts.Mode) + parentDir.insertChildLocked(&childDir.dentry, name) return nil }) } // MknodAt implements vfs.FilesystemImpl.MknodAt. func (fs *filesystem) MknodAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.MknodOptions) error { - return fs.doCreateAt(rp, false /* dir */, func(parent *dentry, name string) error { + return fs.doCreateAt(rp, false /* dir */, func(parentDir *directory, name string) error { + creds := rp.Credentials() var childInode *inode switch opts.Mode.FileType() { case 0, linux.S_IFREG: - childInode = fs.newRegularFile(rp.Credentials(), opts.Mode) + childInode = fs.newRegularFile(creds.EffectiveKUID, creds.EffectiveKGID, opts.Mode) case linux.S_IFIFO: - childInode = fs.newNamedPipe(rp.Credentials(), opts.Mode) + childInode = fs.newNamedPipe(creds.EffectiveKUID, creds.EffectiveKGID, opts.Mode) case linux.S_IFBLK: - childInode = fs.newDeviceFile(rp.Credentials(), opts.Mode, vfs.BlockDevice, opts.DevMajor, opts.DevMinor) + childInode = fs.newDeviceFile(creds.EffectiveKUID, creds.EffectiveKGID, opts.Mode, vfs.BlockDevice, opts.DevMajor, opts.DevMinor) case linux.S_IFCHR: - childInode = fs.newDeviceFile(rp.Credentials(), opts.Mode, vfs.CharDevice, opts.DevMajor, opts.DevMinor) + childInode = fs.newDeviceFile(creds.EffectiveKUID, creds.EffectiveKGID, opts.Mode, vfs.CharDevice, opts.DevMajor, opts.DevMinor) case linux.S_IFSOCK: - // Not yet supported. - return syserror.EPERM + childInode = fs.newSocketFile(creds.EffectiveKUID, creds.EffectiveKGID, opts.Mode, opts.Endpoint) default: return syserror.EINVAL } child := fs.newDentry(childInode) - parent.vfsd.InsertChild(&child.vfsd, name) - parent.inode.impl.(*directory).childList.PushBack(child) + parentDir.insertChildLocked(child, name) return nil }) } @@ -285,12 +326,12 @@ func (fs *filesystem) OpenAt(ctx context.Context, rp *vfs.ResolvingPath, opts vf return start.open(ctx, rp, &opts, false /* afterCreate */) } afterTrailingSymlink: - parent, err := walkParentDirLocked(rp, start) + parentDir, err := walkParentDirLocked(rp, start) if err != nil { return nil, err } // Check for search permission in the parent directory. - if err := parent.inode.checkPermissions(rp.Credentials(), vfs.MayExec, true); err != nil { + if err := parentDir.inode.checkPermissions(rp.Credentials(), vfs.MayExec); err != nil { return nil, err } // Reject attempts to open directories with O_CREAT. @@ -301,11 +342,14 @@ afterTrailingSymlink: if name == "." || name == ".." { return nil, syserror.EISDIR } + if len(name) > linux.NAME_MAX { + return nil, syserror.ENAMETOOLONG + } // Determine whether or not we need to create a file. - child, err := stepLocked(rp, parent) - if err == syserror.ENOENT { + child, ok := parentDir.childMap[name] + if !ok { // Already checked for searchability above; now check for writability. - if err := parent.inode.checkPermissions(rp.Credentials(), vfs.MayWrite, true); err != nil { + if err := parentDir.inode.checkPermissions(rp.Credentials(), vfs.MayWrite); err != nil { return nil, err } if err := rp.Mount().CheckBeginWrite(); err != nil { @@ -313,17 +357,29 @@ afterTrailingSymlink: } defer rp.Mount().EndWrite() // Create and open the child. - child := fs.newDentry(fs.newRegularFile(rp.Credentials(), opts.Mode)) - parent.vfsd.InsertChild(&child.vfsd, name) - parent.inode.impl.(*directory).childList.PushBack(child) - return child.open(ctx, rp, &opts, true) + creds := rp.Credentials() + child := fs.newDentry(fs.newRegularFile(creds.EffectiveKUID, creds.EffectiveKGID, opts.Mode)) + parentDir.insertChildLocked(child, name) + fd, err := child.open(ctx, rp, &opts, true) + if err != nil { + return nil, err + } + parentDir.inode.watches.Notify(name, linux.IN_CREATE, 0, vfs.PathEvent) + parentDir.inode.touchCMtime() + return fd, nil } - if err != nil { + // Is the file mounted over? + if err := rp.CheckMount(&child.vfsd); err != nil { return nil, err } // Do we need to resolve a trailing symlink? - if !rp.Done() { - start = parent + if symlink, ok := child.inode.impl.(*symlink); ok && rp.ShouldFollowSymlink() { + // Symlink traversal updates access time. + atomic.StoreInt64(&child.inode.atime, child.inode.fs.clock.Now().Nanoseconds()) + if err := rp.HandleSymlink(symlink.target); err != nil { + return nil, err + } + start = &parentDir.dentry goto afterTrailingSymlink } // Open existing file. @@ -336,7 +392,7 @@ afterTrailingSymlink: func (d *dentry) open(ctx context.Context, rp *vfs.ResolvingPath, opts *vfs.OpenOptions, afterCreate bool) (*vfs.FileDescription, error) { ats := vfs.AccessTypesForOpenFlags(opts) if !afterCreate { - if err := d.inode.checkPermissions(rp.Credentials(), ats, d.inode.isDir()); err != nil { + if err := d.inode.checkPermissions(rp.Credentials(), ats); err != nil { return nil, err } } @@ -347,10 +403,9 @@ func (d *dentry) open(ctx context.Context, rp *vfs.ResolvingPath, opts *vfs.Open return nil, err } if opts.Flags&linux.O_TRUNC != 0 { - impl.mu.Lock() - impl.data.Truncate(0, impl.memFile) - atomic.StoreUint64(&impl.size, 0) - impl.mu.Unlock() + if _, err := impl.truncate(0); err != nil { + return nil, err + } } return &fd.vfsfd, nil case *directory: @@ -367,9 +422,11 @@ func (d *dentry) open(ctx context.Context, rp *vfs.ResolvingPath, opts *vfs.Open // Can't open symlinks without O_PATH (which is unimplemented). return nil, syserror.ELOOP case *namedPipe: - return newNamedPipeFD(ctx, impl, rp, &d.vfsd, opts.Flags) + return impl.pipe.Open(ctx, rp.Mount(), &d.vfsd, opts.Flags) case *deviceFile: return rp.VirtualFilesystem().OpenDeviceSpecialFile(ctx, rp.Mount(), &d.vfsd, impl.kind, impl.major, impl.minor, opts) + case *socketFile: + return nil, syserror.ENXIO default: panic(fmt.Sprintf("unknown inode type: %T", d.inode.impl)) } @@ -387,6 +444,7 @@ func (fs *filesystem) ReadlinkAt(ctx context.Context, rp *vfs.ResolvingPath) (st if !ok { return "", syserror.EINVAL } + symlink.inode.touchAtime(rp.Mount()) return symlink.target, nil } @@ -400,7 +458,7 @@ func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa // Resolve newParent first to verify that it's on this Mount. fs.mu.Lock() defer fs.mu.Unlock() - newParent, err := walkParentDirLocked(rp, rp.Start().Impl().(*dentry)) + newParentDir, err := walkParentDirLocked(rp, rp.Start().Impl().(*dentry)) if err != nil { return err } @@ -417,25 +475,24 @@ func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa } defer mnt.EndWrite() - oldParent := oldParentVD.Dentry().Impl().(*dentry) - if err := oldParent.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true /* isDir */); err != nil { + oldParentDir := oldParentVD.Dentry().Impl().(*dentry).inode.impl.(*directory) + if err := oldParentDir.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec); err != nil { return err } - // Call vfs.Dentry.Child() instead of stepLocked() or rp.ResolveChild(), - // because if the existing child is a symlink or mount point then we want - // to rename over it rather than follow it. - renamedVFSD := oldParent.vfsd.Child(oldName) - if renamedVFSD == nil { + renamed, ok := oldParentDir.childMap[oldName] + if !ok { return syserror.ENOENT } - renamed := renamedVFSD.Impl().(*dentry) + // Note that we don't need to call rp.CheckMount(), since if renamed is a + // mount point then we want to rename the mount point, not anything in the + // mounted filesystem. if renamed.inode.isDir() { - if renamed == newParent || renamedVFSD.IsAncestorOf(&newParent.vfsd) { + if renamed == &newParentDir.dentry || genericIsAncestorDentry(renamed, &newParentDir.dentry) { return syserror.EINVAL } - if oldParent != newParent { + if oldParentDir != newParentDir { // Writability is needed to change renamed's "..". - if err := renamed.inode.checkPermissions(rp.Credentials(), vfs.MayWrite, true /* isDir */); err != nil { + if err := renamed.inode.checkPermissions(rp.Credentials(), vfs.MayWrite); err != nil { return err } } @@ -445,18 +502,17 @@ func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa } } - if err := newParent.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true /* isDir */); err != nil { + if err := newParentDir.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec); err != nil { return err } - replacedVFSD := newParent.vfsd.Child(newName) - var replaced *dentry - if replacedVFSD != nil { - replaced = replacedVFSD.Impl().(*dentry) - if replaced.inode.isDir() { + replaced, ok := newParentDir.childMap[newName] + if ok { + replacedDir, ok := replaced.inode.impl.(*directory) + if ok { if !renamed.inode.isDir() { return syserror.EISDIR } - if replaced.vfsd.HasChildren() { + if len(replacedDir.childMap) != 0 { return syserror.ENOTEMPTY } } else { @@ -468,11 +524,13 @@ func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa } } } else { - if renamed.inode.isDir() && newParent.inode.nlink == maxLinks { + if renamed.inode.isDir() && newParentDir.inode.nlink == maxLinks { return syserror.EMLINK } } - if newParent.vfsd.IsDisowned() { + // tmpfs never calls VFS.InvalidateDentry(), so newParentDir.dentry can + // only be dead if it was deleted. + if newParentDir.dentry.vfsd.IsDead() { return syserror.ENOENT } @@ -480,31 +538,40 @@ func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa // simplicity, under the assumption that applications are not intentionally // doing noop renames expecting them to succeed where non-noop renames // would fail. - if renamedVFSD == replacedVFSD { + if renamed == replaced { return nil } vfsObj := rp.VirtualFilesystem() - oldParentDir := oldParent.inode.impl.(*directory) - newParentDir := newParent.inode.impl.(*directory) - if err := vfsObj.PrepareRenameDentry(vfs.MountNamespaceFromContext(ctx), renamedVFSD, replacedVFSD); err != nil { + mntns := vfs.MountNamespaceFromContext(ctx) + defer mntns.DecRef() + var replacedVFSD *vfs.Dentry + if replaced != nil { + replacedVFSD = &replaced.vfsd + } + if err := vfsObj.PrepareRenameDentry(mntns, &renamed.vfsd, replacedVFSD); err != nil { return err } if replaced != nil { - newParentDir.childList.Remove(replaced) + newParentDir.removeChildLocked(replaced) if replaced.inode.isDir() { - newParent.inode.decLinksLocked() // from replaced's ".." + newParentDir.inode.decLinksLocked() // from replaced's ".." } replaced.inode.decLinksLocked() } - oldParentDir.childList.Remove(renamed) - newParentDir.childList.PushBack(renamed) - if renamed.inode.isDir() { - oldParent.inode.decLinksLocked() - newParent.inode.incLinksLocked() + oldParentDir.removeChildLocked(renamed) + newParentDir.insertChildLocked(renamed, newName) + vfsObj.CommitRenameReplaceDentry(&renamed.vfsd, replacedVFSD) + oldParentDir.inode.touchCMtime() + if oldParentDir != newParentDir { + if renamed.inode.isDir() { + oldParentDir.inode.decLinksLocked() + newParentDir.inode.incLinksLocked() + } + newParentDir.inode.touchCMtime() } - // TODO(gvisor.dev/issues/1197): Update timestamps and parent directory - // sizes. - vfsObj.CommitRenameReplaceDentry(renamedVFSD, &newParent.vfsd, newName, replacedVFSD) + renamed.inode.touchCtime() + + vfs.InotifyRename(ctx, &renamed.inode.watches, &oldParentDir.inode.watches, &newParentDir.inode.watches, oldName, newName, renamed.inode.isDir()) return nil } @@ -512,11 +579,11 @@ func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa func (fs *filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error { fs.mu.Lock() defer fs.mu.Unlock() - parent, err := walkParentDirLocked(rp, rp.Start().Impl().(*dentry)) + parentDir, err := walkParentDirLocked(rp, rp.Start().Impl().(*dentry)) if err != nil { return err } - if err := parent.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true /* isDir */); err != nil { + if err := parentDir.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec); err != nil { return err } name := rp.Component() @@ -526,15 +593,15 @@ func (fs *filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error if name == ".." { return syserror.ENOTEMPTY } - childVFSD := parent.vfsd.Child(name) - if childVFSD == nil { + child, ok := parentDir.childMap[name] + if !ok { return syserror.ENOENT } - child := childVFSD.Impl().(*dentry) - if !child.inode.isDir() { + childDir, ok := child.inode.impl.(*directory) + if !ok { return syserror.ENOTDIR } - if childVFSD.HasChildren() { + if len(childDir.childMap) != 0 { return syserror.ENOTEMPTY } mnt := rp.Mount() @@ -543,13 +610,19 @@ func (fs *filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error } defer mnt.EndWrite() vfsObj := rp.VirtualFilesystem() - if err := vfsObj.PrepareDeleteDentry(vfs.MountNamespaceFromContext(ctx), childVFSD); err != nil { + mntns := vfs.MountNamespaceFromContext(ctx) + defer mntns.DecRef() + if err := vfsObj.PrepareDeleteDentry(mntns, &child.vfsd); err != nil { return err } - parent.inode.impl.(*directory).childList.Remove(child) - parent.inode.decLinksLocked() // from child's ".." + parentDir.removeChildLocked(child) + parentDir.inode.watches.Notify(name, linux.IN_DELETE|linux.IN_ISDIR, 0, vfs.InodeEvent) + // Remove links for child, child/., and child/.. child.inode.decLinksLocked() - vfsObj.CommitDeleteDentry(childVFSD) + child.inode.decLinksLocked() + parentDir.inode.decLinksLocked() + vfsObj.CommitDeleteDentry(&child.vfsd) + parentDir.inode.touchCMtime() return nil } @@ -561,7 +634,14 @@ func (fs *filesystem) SetStatAt(ctx context.Context, rp *vfs.ResolvingPath, opts if err != nil { return err } - return d.inode.setStat(opts.Stat) + if err := d.inode.setStat(ctx, rp.Credentials(), &opts.Stat); err != nil { + return err + } + + if ev := vfs.InotifyEventFromStatMask(opts.Stat.Mask); ev != 0 { + d.InotifyWithParent(ev, 0, vfs.InodeEvent) + } + return nil } // StatAt implements vfs.FilesystemImpl.StatAt. @@ -581,20 +661,27 @@ func (fs *filesystem) StatAt(ctx context.Context, rp *vfs.ResolvingPath, opts vf func (fs *filesystem) StatFSAt(ctx context.Context, rp *vfs.ResolvingPath) (linux.Statfs, error) { fs.mu.RLock() defer fs.mu.RUnlock() - _, err := resolveLocked(rp) - if err != nil { + if _, err := resolveLocked(rp); err != nil { return linux.Statfs{}, err } - // TODO(gvisor.dev/issues/1197): Actually implement statfs. - return linux.Statfs{}, syserror.ENOSYS + statfs := linux.Statfs{ + Type: linux.TMPFS_MAGIC, + BlockSize: usermem.PageSize, + FragmentSize: usermem.PageSize, + NameLength: linux.NAME_MAX, + // TODO(b/29637826): Allow configuring a tmpfs size and enforce it. + Blocks: 0, + BlocksFree: 0, + } + return statfs, nil } // SymlinkAt implements vfs.FilesystemImpl.SymlinkAt. func (fs *filesystem) SymlinkAt(ctx context.Context, rp *vfs.ResolvingPath, target string) error { - return fs.doCreateAt(rp, false /* dir */, func(parent *dentry, name string) error { - child := fs.newDentry(fs.newSymlink(rp.Credentials(), target)) - parent.vfsd.InsertChild(&child.vfsd, name) - parent.inode.impl.(*directory).childList.PushBack(child) + return fs.doCreateAt(rp, false /* dir */, func(parentDir *directory, name string) error { + creds := rp.Credentials() + child := fs.newDentry(fs.newSymlink(creds.EffectiveKUID, creds.EffectiveKGID, 0777, target)) + parentDir.insertChildLocked(child, name) return nil }) } @@ -603,22 +690,21 @@ func (fs *filesystem) SymlinkAt(ctx context.Context, rp *vfs.ResolvingPath, targ func (fs *filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error { fs.mu.Lock() defer fs.mu.Unlock() - parent, err := walkParentDirLocked(rp, rp.Start().Impl().(*dentry)) + parentDir, err := walkParentDirLocked(rp, rp.Start().Impl().(*dentry)) if err != nil { return err } - if err := parent.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true /* isDir */); err != nil { + if err := parentDir.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec); err != nil { return err } name := rp.Component() if name == "." || name == ".." { return syserror.EISDIR } - childVFSD := parent.vfsd.Child(name) - if childVFSD == nil { + child, ok := parentDir.childMap[name] + if !ok { return syserror.ENOENT } - child := childVFSD.Impl().(*dentry) if child.inode.isDir() { return syserror.EISDIR } @@ -631,66 +717,119 @@ func (fs *filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error } defer mnt.EndWrite() vfsObj := rp.VirtualFilesystem() - if err := vfsObj.PrepareDeleteDentry(vfs.MountNamespaceFromContext(ctx), childVFSD); err != nil { + mntns := vfs.MountNamespaceFromContext(ctx) + defer mntns.DecRef() + if err := vfsObj.PrepareDeleteDentry(mntns, &child.vfsd); err != nil { return err } - parent.inode.impl.(*directory).childList.Remove(child) + + // Generate inotify events. Note that this must take place before the link + // count of the child is decremented, or else the watches may be dropped + // before these events are added. + vfs.InotifyRemoveChild(&child.inode.watches, &parentDir.inode.watches, name) + + parentDir.removeChildLocked(child) child.inode.decLinksLocked() - vfsObj.CommitDeleteDentry(childVFSD) + vfsObj.CommitDeleteDentry(&child.vfsd) + parentDir.inode.touchCMtime() return nil } +// BoundEndpointAt implements FilesystemImpl.BoundEndpointAt. +func (fs *filesystem) BoundEndpointAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.BoundEndpointOptions) (transport.BoundEndpoint, error) { + fs.mu.RLock() + defer fs.mu.RUnlock() + d, err := resolveLocked(rp) + if err != nil { + return nil, err + } + if err := d.inode.checkPermissions(rp.Credentials(), vfs.MayWrite); err != nil { + return nil, err + } + switch impl := d.inode.impl.(type) { + case *socketFile: + return impl.ep, nil + default: + return nil, syserror.ECONNREFUSED + } +} + // ListxattrAt implements vfs.FilesystemImpl.ListxattrAt. -func (fs *filesystem) ListxattrAt(ctx context.Context, rp *vfs.ResolvingPath) ([]string, error) { +func (fs *filesystem) ListxattrAt(ctx context.Context, rp *vfs.ResolvingPath, size uint64) ([]string, error) { fs.mu.RLock() defer fs.mu.RUnlock() - _, err := resolveLocked(rp) + d, err := resolveLocked(rp) if err != nil { return nil, err } - // TODO(b/127675828): support extended attributes - return nil, syserror.ENOTSUP + return d.inode.listxattr(size) } // GetxattrAt implements vfs.FilesystemImpl.GetxattrAt. -func (fs *filesystem) GetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) (string, error) { +func (fs *filesystem) GetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.GetxattrOptions) (string, error) { fs.mu.RLock() defer fs.mu.RUnlock() - _, err := resolveLocked(rp) + d, err := resolveLocked(rp) if err != nil { return "", err } - // TODO(b/127675828): support extended attributes - return "", syserror.ENOTSUP + return d.inode.getxattr(rp.Credentials(), &opts) } // SetxattrAt implements vfs.FilesystemImpl.SetxattrAt. func (fs *filesystem) SetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetxattrOptions) error { fs.mu.RLock() defer fs.mu.RUnlock() - _, err := resolveLocked(rp) + d, err := resolveLocked(rp) if err != nil { return err } - // TODO(b/127675828): support extended attributes - return syserror.ENOTSUP + if err := d.inode.setxattr(rp.Credentials(), &opts); err != nil { + return err + } + + d.InotifyWithParent(linux.IN_ATTRIB, 0, vfs.InodeEvent) + return nil } // RemovexattrAt implements vfs.FilesystemImpl.RemovexattrAt. func (fs *filesystem) RemovexattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) error { fs.mu.RLock() defer fs.mu.RUnlock() - _, err := resolveLocked(rp) + d, err := resolveLocked(rp) if err != nil { return err } - // TODO(b/127675828): support extended attributes - return syserror.ENOTSUP + if err := d.inode.removexattr(rp.Credentials(), name); err != nil { + return err + } + + d.InotifyWithParent(linux.IN_ATTRIB, 0, vfs.InodeEvent) + return nil } // PrependPath implements vfs.FilesystemImpl.PrependPath. func (fs *filesystem) PrependPath(ctx context.Context, vfsroot, vd vfs.VirtualDentry, b *fspath.Builder) error { fs.mu.RLock() defer fs.mu.RUnlock() - return vfs.GenericPrependPath(vfsroot, vd, b) + mnt := vd.Mount() + d := vd.Dentry().Impl().(*dentry) + for { + if mnt == vfsroot.Mount() && &d.vfsd == vfsroot.Dentry() { + return vfs.PrependPathAtVFSRootError{} + } + if &d.vfsd == mnt.Root() { + return nil + } + if d.parent == nil { + if d.name != "" { + // This must be an anonymous memfd file. + b.PrependComponent("/" + d.name) + return vfs.PrependPathSyntheticError{} + } + return vfs.PrependPathAtNonMountRootError{} + } + b.PrependComponent(d.name) + d = d.parent + } } diff --git a/pkg/sentry/fsimpl/tmpfs/named_pipe.go b/pkg/sentry/fsimpl/tmpfs/named_pipe.go index 0c57fdca3..739350cf0 100644 --- a/pkg/sentry/fsimpl/tmpfs/named_pipe.go +++ b/pkg/sentry/fsimpl/tmpfs/named_pipe.go @@ -16,10 +16,8 @@ package tmpfs import ( "gvisor.dev/gvisor/pkg/abi/linux" - "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/kernel/pipe" - "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/usermem" ) @@ -32,28 +30,9 @@ type namedPipe struct { // Preconditions: // * fs.mu must be locked. // * rp.Mount().CheckBeginWrite() has been called successfully. -func (fs *filesystem) newNamedPipe(creds *auth.Credentials, mode linux.FileMode) *inode { - file := &namedPipe{pipe: pipe.NewVFSPipe(pipe.DefaultPipeSize, usermem.PageSize)} - file.inode.init(file, fs, creds, mode) +func (fs *filesystem) newNamedPipe(kuid auth.KUID, kgid auth.KGID, mode linux.FileMode) *inode { + file := &namedPipe{pipe: pipe.NewVFSPipe(true /* isNamed */, pipe.DefaultPipeSize, usermem.PageSize)} + file.inode.init(file, fs, kuid, kgid, linux.S_IFIFO|mode) file.inode.nlink = 1 // Only the parent has a link. return &file.inode } - -// namedPipeFD implements vfs.FileDescriptionImpl. Methods are implemented -// entirely via struct embedding. -type namedPipeFD struct { - fileDescription - - *pipe.VFSPipeFD -} - -func newNamedPipeFD(ctx context.Context, np *namedPipe, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, flags uint32) (*vfs.FileDescription, error) { - var err error - var fd namedPipeFD - fd.VFSPipeFD, err = np.pipe.NewVFSPipeFD(ctx, vfsd, &fd.vfsfd, flags) - if err != nil { - return nil, err - } - fd.vfsfd.Init(&fd, flags, rp.Mount(), vfsd, &vfs.FileDescriptionOptions{}) - return &fd.vfsfd, nil -} diff --git a/pkg/sentry/fsimpl/tmpfs/pipe_test.go b/pkg/sentry/fsimpl/tmpfs/pipe_test.go index 5ee7f2a72..1614f2c39 100644 --- a/pkg/sentry/fsimpl/tmpfs/pipe_test.go +++ b/pkg/sentry/fsimpl/tmpfs/pipe_test.go @@ -151,7 +151,10 @@ func setup(t *testing.T) (context.Context, *auth.Credentials, *vfs.VirtualFilesy creds := auth.CredentialsFromContext(ctx) // Create VFS. - vfsObj := vfs.New() + vfsObj := &vfs.VirtualFilesystem{} + if err := vfsObj.Init(); err != nil { + t.Fatalf("VFS init: %v", err) + } vfsObj.MustRegisterFilesystemType("tmpfs", FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ AllowUserMount: true, }) diff --git a/pkg/sentry/fsimpl/tmpfs/regular_file.go b/pkg/sentry/fsimpl/tmpfs/regular_file.go index dab346a41..4f2ae04d2 100644 --- a/pkg/sentry/fsimpl/tmpfs/regular_file.go +++ b/pkg/sentry/fsimpl/tmpfs/regular_file.go @@ -15,6 +15,7 @@ package tmpfs import ( + "fmt" "io" "math" "sync/atomic" @@ -22,6 +23,7 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/safemem" + "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fs/fsutil" "gvisor.dev/gvisor/pkg/sentry/fs/lock" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" @@ -34,71 +36,231 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) +// regularFile is a regular (=S_IFREG) tmpfs file. type regularFile struct { inode inode // memFile is a platform.File used to allocate pages to this regularFile. memFile *pgalloc.MemoryFile - // mu protects the fields below. - mu sync.RWMutex + // mapsMu protects mappings. + mapsMu sync.Mutex `state:"nosave"` + + // mappings tracks mappings of the file into memmap.MappingSpaces. + // + // Protected by mapsMu. + mappings memmap.MappingSet + + // writableMappingPages tracks how many pages of virtual memory are mapped + // as potentially writable from this file. If a page has multiple mappings, + // each mapping is counted separately. + // + // This counter is susceptible to overflow as we can potentially count + // mappings from many VMAs. We count pages rather than bytes to slightly + // mitigate this. + // + // Protected by mapsMu. + writableMappingPages uint64 + + // dataMu protects the fields below. + dataMu sync.RWMutex // data maps offsets into the file to offsets into memFile that store // the file's data. + // + // Protected by dataMu. data fsutil.FileRangeSet - // size is the size of data, but accessed using atomic memory - // operations to avoid locking in inode.stat(). - size uint64 - // seals represents file seals on this inode. + // + // Protected by dataMu. seals uint32 + + // size is the size of data. + // + // Protected by both dataMu and inode.mu; reading it requires holding + // either mutex, while writing requires holding both AND using atomics. + // Readers that do not require consistency (like Stat) may read the + // value atomically without holding either lock. + size uint64 } -func (fs *filesystem) newRegularFile(creds *auth.Credentials, mode linux.FileMode) *inode { +func (fs *filesystem) newRegularFile(kuid auth.KUID, kgid auth.KGID, mode linux.FileMode) *inode { file := ®ularFile{ memFile: fs.memFile, + seals: linux.F_SEAL_SEAL, } - file.inode.init(file, fs, creds, mode) + file.inode.init(file, fs, kuid, kgid, linux.S_IFREG|mode) file.inode.nlink = 1 // from parent directory return &file.inode } // truncate grows or shrinks the file to the given size. It returns true if the // file size was updated. -func (rf *regularFile) truncate(size uint64) (bool, error) { - rf.mu.Lock() - defer rf.mu.Unlock() +func (rf *regularFile) truncate(newSize uint64) (bool, error) { + rf.inode.mu.Lock() + defer rf.inode.mu.Unlock() + return rf.truncateLocked(newSize) +} - if size == rf.size { +// Preconditions: rf.inode.mu must be held. +func (rf *regularFile) truncateLocked(newSize uint64) (bool, error) { + oldSize := rf.size + if newSize == oldSize { // Nothing to do. return false, nil } - if size > rf.size { - // Growing the file. + // Need to hold inode.mu and dataMu while modifying size. + rf.dataMu.Lock() + if newSize > oldSize { + // Can we grow the file? if rf.seals&linux.F_SEAL_GROW != 0 { - // Seal does not allow growth. + rf.dataMu.Unlock() return false, syserror.EPERM } - rf.size = size + // We only need to update the file size. + atomic.StoreUint64(&rf.size, newSize) + rf.dataMu.Unlock() return true, nil } - // Shrinking the file + // We are shrinking the file. First check if this is allowed. if rf.seals&linux.F_SEAL_SHRINK != 0 { - // Seal does not allow shrink. + rf.dataMu.Unlock() return false, syserror.EPERM } - // TODO(gvisor.dev/issues/1197): Invalidate mappings once we have - // mappings. + // Update the file size. + atomic.StoreUint64(&rf.size, newSize) + rf.dataMu.Unlock() + + // Invalidate past translations of truncated pages. + oldpgend := fs.OffsetPageEnd(int64(oldSize)) + newpgend := fs.OffsetPageEnd(int64(newSize)) + if newpgend < oldpgend { + rf.mapsMu.Lock() + rf.mappings.Invalidate(memmap.MappableRange{newpgend, oldpgend}, memmap.InvalidateOpts{ + // Compare Linux's mm/shmem.c:shmem_setattr() => + // mm/memory.c:unmap_mapping_range(evencows=1). + InvalidatePrivate: true, + }) + rf.mapsMu.Unlock() + } - rf.data.Truncate(size, rf.memFile) - rf.size = size + // We are now guaranteed that there are no translations of truncated pages, + // and can remove them. + rf.dataMu.Lock() + rf.data.Truncate(newSize, rf.memFile) + rf.dataMu.Unlock() return true, nil } +// AddMapping implements memmap.Mappable.AddMapping. +func (rf *regularFile) AddMapping(ctx context.Context, ms memmap.MappingSpace, ar usermem.AddrRange, offset uint64, writable bool) error { + rf.mapsMu.Lock() + defer rf.mapsMu.Unlock() + rf.dataMu.RLock() + defer rf.dataMu.RUnlock() + + // Reject writable mapping if F_SEAL_WRITE is set. + if rf.seals&linux.F_SEAL_WRITE != 0 && writable { + return syserror.EPERM + } + + rf.mappings.AddMapping(ms, ar, offset, writable) + if writable { + pagesBefore := rf.writableMappingPages + + // ar is guaranteed to be page aligned per memmap.Mappable. + rf.writableMappingPages += uint64(ar.Length() / usermem.PageSize) + + if rf.writableMappingPages < pagesBefore { + panic(fmt.Sprintf("Overflow while mapping potentially writable pages pointing to a tmpfs file. Before %v, after %v", pagesBefore, rf.writableMappingPages)) + } + } + + return nil +} + +// RemoveMapping implements memmap.Mappable.RemoveMapping. +func (rf *regularFile) RemoveMapping(ctx context.Context, ms memmap.MappingSpace, ar usermem.AddrRange, offset uint64, writable bool) { + rf.mapsMu.Lock() + defer rf.mapsMu.Unlock() + + rf.mappings.RemoveMapping(ms, ar, offset, writable) + + if writable { + pagesBefore := rf.writableMappingPages + + // ar is guaranteed to be page aligned per memmap.Mappable. + rf.writableMappingPages -= uint64(ar.Length() / usermem.PageSize) + + if rf.writableMappingPages > pagesBefore { + panic(fmt.Sprintf("Underflow while unmapping potentially writable pages pointing to a tmpfs file. Before %v, after %v", pagesBefore, rf.writableMappingPages)) + } + } +} + +// CopyMapping implements memmap.Mappable.CopyMapping. +func (rf *regularFile) CopyMapping(ctx context.Context, ms memmap.MappingSpace, srcAR, dstAR usermem.AddrRange, offset uint64, writable bool) error { + return rf.AddMapping(ctx, ms, dstAR, offset, writable) +} + +// Translate implements memmap.Mappable.Translate. +func (rf *regularFile) Translate(ctx context.Context, required, optional memmap.MappableRange, at usermem.AccessType) ([]memmap.Translation, error) { + rf.dataMu.Lock() + defer rf.dataMu.Unlock() + + // Constrain translations to f.attr.Size (rounded up) to prevent + // translation to pages that may be concurrently truncated. + pgend := fs.OffsetPageEnd(int64(rf.size)) + var beyondEOF bool + if required.End > pgend { + if required.Start >= pgend { + return nil, &memmap.BusError{io.EOF} + } + beyondEOF = true + required.End = pgend + } + if optional.End > pgend { + optional.End = pgend + } + + cerr := rf.data.Fill(ctx, required, optional, rf.memFile, usage.Tmpfs, func(_ context.Context, dsts safemem.BlockSeq, _ uint64) (uint64, error) { + // Newly-allocated pages are zeroed, so we don't need to do anything. + return dsts.NumBytes(), nil + }) + + var ts []memmap.Translation + var translatedEnd uint64 + for seg := rf.data.FindSegment(required.Start); seg.Ok() && seg.Start() < required.End; seg, _ = seg.NextNonEmpty() { + segMR := seg.Range().Intersect(optional) + ts = append(ts, memmap.Translation{ + Source: segMR, + File: rf.memFile, + Offset: seg.FileRangeOf(segMR).Start, + Perms: usermem.AnyAccess, + }) + translatedEnd = segMR.End + } + + // Don't return the error returned by f.data.Fill if it occurred outside of + // required. + if translatedEnd < required.End && cerr != nil { + return ts, &memmap.BusError{cerr} + } + if beyondEOF { + return ts, &memmap.BusError{io.EOF} + } + return ts, nil +} + +// InvalidateUnsavable implements memmap.Mappable.InvalidateUnsavable. +func (*regularFile) InvalidateUnsavable(context.Context) error { + return nil +} + type regularFileFD struct { fileDescription @@ -125,7 +287,8 @@ func (fd *regularFileFD) PRead(ctx context.Context, dst usermem.IOSequence, offs rw := getRegularFileReadWriter(f, offset) n, err := dst.CopyOutFrom(ctx, rw) putRegularFileReadWriter(rw) - return int64(n), err + fd.inode().touchAtime(fd.vfsfd.Mount()) + return n, err } // Read implements vfs.FileDescriptionImpl.Read. @@ -147,13 +310,23 @@ func (fd *regularFileFD) PWrite(ctx context.Context, src usermem.IOSequence, off return 0, nil } f := fd.inode().impl.(*regularFile) - end := offset + srclen - if end < offset { + if end := offset + srclen; end < offset { // Overflow. - return 0, syserror.EFBIG + return 0, syserror.EINVAL + } + + var err error + srclen, err = vfs.CheckLimit(ctx, offset, srclen) + if err != nil { + return 0, err } + src = src.TakeFirst64(srclen) + + f.inode.mu.Lock() rw := getRegularFileReadWriter(f, offset) n, err := src.CopyInTo(ctx, rw) + fd.inode().touchCMtimeLocked() + f.inode.mu.Unlock() putRegularFileReadWriter(rw) return n, err } @@ -215,6 +388,12 @@ func (fd *regularFileFD) UnlockPOSIX(ctx context.Context, uid lock.UniqueID, rng return nil } +// ConfigureMMap implements vfs.FileDescriptionImpl.ConfigureMMap. +func (fd *regularFileFD) ConfigureMMap(ctx context.Context, opts *memmap.MMapOpts) error { + file := fd.inode().impl.(*regularFile) + return vfs.GenericConfigureMMap(&fd.vfsfd, file, opts) +} + // regularFileReadWriter implements safemem.Reader and Safemem.Writer. type regularFileReadWriter struct { file *regularFile @@ -244,14 +423,15 @@ func putRegularFileReadWriter(rw *regularFileReadWriter) { // ReadToBlocks implements safemem.Reader.ReadToBlocks. func (rw *regularFileReadWriter) ReadToBlocks(dsts safemem.BlockSeq) (uint64, error) { - rw.file.mu.RLock() + rw.file.dataMu.RLock() + defer rw.file.dataMu.RUnlock() + size := rw.file.size // Compute the range to read (limited by file size and overflow-checked). - if rw.off >= rw.file.size { - rw.file.mu.RUnlock() + if rw.off >= size { return 0, io.EOF } - end := rw.file.size + end := size if rend := rw.off + dsts.NumBytes(); rend > rw.off && rend < end { end = rend } @@ -265,7 +445,6 @@ func (rw *regularFileReadWriter) ReadToBlocks(dsts safemem.BlockSeq) (uint64, er // Get internal mappings. ims, err := rw.file.memFile.MapInternal(seg.FileRangeOf(seg.Range().Intersect(mr)), usermem.Read) if err != nil { - rw.file.mu.RUnlock() return done, err } @@ -275,7 +454,6 @@ func (rw *regularFileReadWriter) ReadToBlocks(dsts safemem.BlockSeq) (uint64, er rw.off += uint64(n) dsts = dsts.DropFirst64(n) if err != nil { - rw.file.mu.RUnlock() return done, err } @@ -291,7 +469,6 @@ func (rw *regularFileReadWriter) ReadToBlocks(dsts safemem.BlockSeq) (uint64, er rw.off += uint64(n) dsts = dsts.DropFirst64(n) if err != nil { - rw.file.mu.RUnlock() return done, err } @@ -299,13 +476,16 @@ func (rw *regularFileReadWriter) ReadToBlocks(dsts safemem.BlockSeq) (uint64, er seg, gap = gap.NextSegment(), fsutil.FileRangeGapIterator{} } } - rw.file.mu.RUnlock() return done, nil } // WriteFromBlocks implements safemem.Writer.WriteFromBlocks. +// +// Preconditions: inode.mu must be held. func (rw *regularFileReadWriter) WriteFromBlocks(srcs safemem.BlockSeq) (uint64, error) { - rw.file.mu.Lock() + // Hold dataMu so we can modify size. + rw.file.dataMu.Lock() + defer rw.file.dataMu.Unlock() // Compute the range to write (overflow-checked). end := rw.off + srcs.NumBytes() @@ -316,7 +496,6 @@ func (rw *regularFileReadWriter) WriteFromBlocks(srcs safemem.BlockSeq) (uint64, // Check if seals prevent either file growth or all writes. switch { case rw.file.seals&linux.F_SEAL_WRITE != 0: // Write sealed - rw.file.mu.Unlock() return 0, syserror.EPERM case end > rw.file.size && rw.file.seals&linux.F_SEAL_GROW != 0: // Grow sealed // When growth is sealed, Linux effectively allows writes which would @@ -338,7 +517,6 @@ func (rw *regularFileReadWriter) WriteFromBlocks(srcs safemem.BlockSeq) (uint64, } if end <= rw.off { // Truncation would result in no data being written. - rw.file.mu.Unlock() return 0, syserror.EPERM } } @@ -395,9 +573,49 @@ exitLoop: // If the write ends beyond the file's previous size, it causes the // file to grow. if rw.off > rw.file.size { - atomic.StoreUint64(&rw.file.size, rw.off) + rw.file.size = rw.off } - rw.file.mu.Unlock() return done, retErr } + +// GetSeals returns the current set of seals on a memfd inode. +func GetSeals(fd *vfs.FileDescription) (uint32, error) { + f, ok := fd.Impl().(*regularFileFD) + if !ok { + return 0, syserror.EINVAL + } + rf := f.inode().impl.(*regularFile) + rf.dataMu.RLock() + defer rf.dataMu.RUnlock() + return rf.seals, nil +} + +// AddSeals adds new file seals to a memfd inode. +func AddSeals(fd *vfs.FileDescription, val uint32) error { + f, ok := fd.Impl().(*regularFileFD) + if !ok { + return syserror.EINVAL + } + rf := f.inode().impl.(*regularFile) + rf.mapsMu.Lock() + defer rf.mapsMu.Unlock() + rf.dataMu.RLock() + defer rf.dataMu.RUnlock() + + if rf.seals&linux.F_SEAL_SEAL != 0 { + // Seal applied which prevents addition of any new seals. + return syserror.EPERM + } + + // F_SEAL_WRITE can only be added if there are no active writable maps. + if rf.seals&linux.F_SEAL_WRITE == 0 && val&linux.F_SEAL_WRITE != 0 { + if rf.writableMappingPages > 0 { + return syserror.EBUSY + } + } + + // Seals can only be added, never removed. + rf.seals |= val + return nil +} diff --git a/pkg/sentry/fsimpl/tmpfs/regular_file_test.go b/pkg/sentry/fsimpl/tmpfs/regular_file_test.go index e9f71e334..64e1c40ad 100644 --- a/pkg/sentry/fsimpl/tmpfs/regular_file_test.go +++ b/pkg/sentry/fsimpl/tmpfs/regular_file_test.go @@ -18,148 +18,16 @@ import ( "bytes" "fmt" "io" - "sync/atomic" "testing" "gvisor.dev/gvisor/pkg/abi/linux" - "gvisor.dev/gvisor/pkg/context" - "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/sentry/contexttest" "gvisor.dev/gvisor/pkg/sentry/fs/lock" - "gvisor.dev/gvisor/pkg/sentry/kernel/auth" - "gvisor.dev/gvisor/pkg/sentry/kernel/contexttest" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" ) -// nextFileID is used to generate unique file names. -var nextFileID int64 - -// newTmpfsRoot creates a new tmpfs mount, and returns the root. If the error -// is not nil, then cleanup should be called when the root is no longer needed. -func newTmpfsRoot(ctx context.Context) (*vfs.VirtualFilesystem, vfs.VirtualDentry, func(), error) { - creds := auth.CredentialsFromContext(ctx) - - vfsObj := vfs.New() - vfsObj.MustRegisterFilesystemType("tmpfs", FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ - AllowUserMount: true, - }) - mntns, err := vfsObj.NewMountNamespace(ctx, creds, "", "tmpfs", &vfs.GetFilesystemOptions{}) - if err != nil { - return nil, vfs.VirtualDentry{}, nil, fmt.Errorf("failed to create tmpfs root mount: %v", err) - } - root := mntns.Root() - return vfsObj, root, func() { - root.DecRef() - mntns.DecRef() - }, nil -} - -// newFileFD creates a new file in a new tmpfs mount, and returns the FD. If -// the returned err is not nil, then cleanup should be called when the FD is no -// longer needed. -func newFileFD(ctx context.Context, mode linux.FileMode) (*vfs.FileDescription, func(), error) { - creds := auth.CredentialsFromContext(ctx) - vfsObj, root, cleanup, err := newTmpfsRoot(ctx) - if err != nil { - return nil, nil, err - } - - filename := fmt.Sprintf("tmpfs-test-file-%d", atomic.AddInt64(&nextFileID, 1)) - - // Create the file that will be write/read. - fd, err := vfsObj.OpenAt(ctx, creds, &vfs.PathOperation{ - Root: root, - Start: root, - Path: fspath.Parse(filename), - }, &vfs.OpenOptions{ - Flags: linux.O_RDWR | linux.O_CREAT | linux.O_EXCL, - Mode: linux.ModeRegular | mode, - }) - if err != nil { - cleanup() - return nil, nil, fmt.Errorf("failed to create file %q: %v", filename, err) - } - - return fd, cleanup, nil -} - -// newDirFD is like newFileFD, but for directories. -func newDirFD(ctx context.Context, mode linux.FileMode) (*vfs.FileDescription, func(), error) { - creds := auth.CredentialsFromContext(ctx) - vfsObj, root, cleanup, err := newTmpfsRoot(ctx) - if err != nil { - return nil, nil, err - } - - dirname := fmt.Sprintf("tmpfs-test-dir-%d", atomic.AddInt64(&nextFileID, 1)) - - // Create the dir. - if err := vfsObj.MkdirAt(ctx, creds, &vfs.PathOperation{ - Root: root, - Start: root, - Path: fspath.Parse(dirname), - }, &vfs.MkdirOptions{ - Mode: linux.ModeDirectory | mode, - }); err != nil { - cleanup() - return nil, nil, fmt.Errorf("failed to create directory %q: %v", dirname, err) - } - - // Open the dir and return it. - fd, err := vfsObj.OpenAt(ctx, creds, &vfs.PathOperation{ - Root: root, - Start: root, - Path: fspath.Parse(dirname), - }, &vfs.OpenOptions{ - Flags: linux.O_RDONLY | linux.O_DIRECTORY, - }) - if err != nil { - cleanup() - return nil, nil, fmt.Errorf("failed to open directory %q: %v", dirname, err) - } - - return fd, cleanup, nil -} - -// newPipeFD is like newFileFD, but for pipes. -func newPipeFD(ctx context.Context, mode linux.FileMode) (*vfs.FileDescription, func(), error) { - creds := auth.CredentialsFromContext(ctx) - vfsObj, root, cleanup, err := newTmpfsRoot(ctx) - if err != nil { - return nil, nil, err - } - - pipename := fmt.Sprintf("tmpfs-test-pipe-%d", atomic.AddInt64(&nextFileID, 1)) - - // Create the pipe. - if err := vfsObj.MknodAt(ctx, creds, &vfs.PathOperation{ - Root: root, - Start: root, - Path: fspath.Parse(pipename), - }, &vfs.MknodOptions{ - Mode: linux.ModeNamedPipe | mode, - }); err != nil { - cleanup() - return nil, nil, fmt.Errorf("failed to create pipe %q: %v", pipename, err) - } - - // Open the pipe and return it. - fd, err := vfsObj.OpenAt(ctx, creds, &vfs.PathOperation{ - Root: root, - Start: root, - Path: fspath.Parse(pipename), - }, &vfs.OpenOptions{ - Flags: linux.O_RDWR, - }) - if err != nil { - cleanup() - return nil, nil, fmt.Errorf("failed to open pipe %q: %v", pipename, err) - } - - return fd, cleanup, nil -} - // Test that we can write some data to a file and read it back.` func TestSimpleWriteRead(t *testing.T) { ctx := contexttest.Context(t) diff --git a/pkg/sentry/fsimpl/gofer/pagemath.go b/pkg/sentry/fsimpl/tmpfs/socket_file.go index 847cb0784..3ed650474 100644 --- a/pkg/sentry/fsimpl/gofer/pagemath.go +++ b/pkg/sentry/fsimpl/tmpfs/socket_file.go @@ -12,20 +12,23 @@ // See the License for the specific language governing permissions and // limitations under the License. -package gofer +package tmpfs import ( - "gvisor.dev/gvisor/pkg/usermem" + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" ) -// This are equivalent to usermem.Addr.RoundDown/Up, but without the -// potentially truncating conversion to usermem.Addr. This is necessary because -// there is no way to define generic "PageRoundDown/Up" functions in Go. - -func pageRoundDown(x uint64) uint64 { - return x &^ (usermem.PageSize - 1) +// socketFile is a socket (=S_IFSOCK) tmpfs file. +type socketFile struct { + inode inode + ep transport.BoundEndpoint } -func pageRoundUp(x uint64) uint64 { - return pageRoundDown(x + usermem.PageSize - 1) +func (fs *filesystem) newSocketFile(kuid auth.KUID, kgid auth.KGID, mode linux.FileMode, ep transport.BoundEndpoint) *inode { + file := &socketFile{ep: ep} + file.inode.init(file, fs, kuid, kgid, mode) + file.inode.nlink = 1 // from parent directory + return &file.inode } diff --git a/pkg/sentry/fsimpl/tmpfs/stat_test.go b/pkg/sentry/fsimpl/tmpfs/stat_test.go index ebe035dee..f7ee4aab2 100644 --- a/pkg/sentry/fsimpl/tmpfs/stat_test.go +++ b/pkg/sentry/fsimpl/tmpfs/stat_test.go @@ -19,8 +19,8 @@ import ( "testing" "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/contexttest" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" - "gvisor.dev/gvisor/pkg/sentry/kernel/contexttest" "gvisor.dev/gvisor/pkg/sentry/vfs" ) @@ -29,7 +29,6 @@ func TestStatAfterCreate(t *testing.T) { mode := linux.FileMode(0644) // Run with different file types. - // TODO(gvisor.dev/issues/1197): Also test symlinks and sockets. for _, typ := range []string{"file", "dir", "pipe"} { t.Run(fmt.Sprintf("type=%q", typ), func(t *testing.T) { var ( @@ -71,9 +70,15 @@ func TestStatAfterCreate(t *testing.T) { t.Errorf("got btime %d, want 0", got.Btime.ToNsec()) } - // Size should be 0. - if got.Size != 0 { - t.Errorf("got size %d, want 0", got.Size) + // Size should be 0 (except for directories, which make up a size + // of 20 per entry, including the "." and ".." entries present in + // otherwise-empty directories). + wantSize := uint64(0) + if typ == "dir" { + wantSize = 40 + } + if got.Size != wantSize { + t.Errorf("got size %d, want %d", got.Size, wantSize) } // Nlink should be 1 for files, 2 for dirs. @@ -140,7 +145,7 @@ func TestSetStatAtime(t *testing.T) { Mask: 0, Atime: linux.NsecToStatxTimestamp(100), }}); err != nil { - t.Errorf("SetStat atime without mask failed: %v") + t.Errorf("SetStat atime without mask failed: %v", err) } // Atime should be unchanged. if gotStat, err := fd.Stat(ctx, allStatOptions); err != nil { @@ -155,7 +160,7 @@ func TestSetStatAtime(t *testing.T) { Atime: linux.NsecToStatxTimestamp(100), } if err := fd.SetStat(ctx, vfs.SetStatOptions{Stat: setStat}); err != nil { - t.Errorf("SetStat atime with mask failed: %v") + t.Errorf("SetStat atime with mask failed: %v", err) } if gotStat, err := fd.Stat(ctx, allStatOptions); err != nil { t.Errorf("Stat got error: %v", err) @@ -169,7 +174,6 @@ func TestSetStat(t *testing.T) { mode := linux.FileMode(0644) // Run with different file types. - // TODO(gvisor.dev/issues/1197): Also test symlinks and sockets. for _, typ := range []string{"file", "dir", "pipe"} { t.Run(fmt.Sprintf("type=%q", typ), func(t *testing.T) { var ( @@ -205,7 +209,7 @@ func TestSetStat(t *testing.T) { Mask: 0, Atime: linux.NsecToStatxTimestamp(100), }}); err != nil { - t.Errorf("SetStat atime without mask failed: %v") + t.Errorf("SetStat atime without mask failed: %v", err) } // Atime should be unchanged. if gotStat, err := fd.Stat(ctx, allStatOptions); err != nil { @@ -220,7 +224,7 @@ func TestSetStat(t *testing.T) { Atime: linux.NsecToStatxTimestamp(100), } if err := fd.SetStat(ctx, vfs.SetStatOptions{Stat: setStat}); err != nil { - t.Errorf("SetStat atime with mask failed: %v") + t.Errorf("SetStat atime with mask failed: %v", err) } if gotStat, err := fd.Stat(ctx, allStatOptions); err != nil { t.Errorf("Stat got error: %v", err) diff --git a/pkg/sentry/fsimpl/tmpfs/symlink.go b/pkg/sentry/fsimpl/tmpfs/symlink.go index 5246aca84..b0de5fabe 100644 --- a/pkg/sentry/fsimpl/tmpfs/symlink.go +++ b/pkg/sentry/fsimpl/tmpfs/symlink.go @@ -15,6 +15,7 @@ package tmpfs import ( + "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" ) @@ -23,11 +24,11 @@ type symlink struct { target string // immutable } -func (fs *filesystem) newSymlink(creds *auth.Credentials, target string) *inode { +func (fs *filesystem) newSymlink(kuid auth.KUID, kgid auth.KGID, mode linux.FileMode, target string) *inode { link := &symlink{ target: target, } - link.inode.init(link, fs, creds, 0777) + link.inode.init(link, fs, kuid, kgid, linux.S_IFLNK|mode) link.inode.nlink = 1 // from parent directory return &link.inode } diff --git a/pkg/sentry/fsimpl/tmpfs/tmpfs.go b/pkg/sentry/fsimpl/tmpfs/tmpfs.go index 2108d0f4d..7ce1b86c7 100644 --- a/pkg/sentry/fsimpl/tmpfs/tmpfs.go +++ b/pkg/sentry/fsimpl/tmpfs/tmpfs.go @@ -12,20 +12,26 @@ // See the License for the specific language governing permissions and // limitations under the License. -// Package tmpfs provides a filesystem implementation that behaves like tmpfs: -// the Dentry tree is the sole source of truth for the state of the filesystem. +// Package tmpfs provides an in-memory filesystem whose contents are +// application-mutable, consistent with Linux's tmpfs. // // Lock order: // // filesystem.mu -// regularFileFD.offMu -// regularFile.mu // inode.mu +// regularFileFD.offMu +// *** "memmap.Mappable locks" below this point +// regularFile.mapsMu +// *** "memmap.Mappable locks taken by Translate" below this point +// regularFile.dataMu +// directory.iterMu package tmpfs import ( "fmt" "math" + "strconv" + "strings" "sync/atomic" "gvisor.dev/gvisor/pkg/abi/linux" @@ -36,10 +42,15 @@ import ( "gvisor.dev/gvisor/pkg/sentry/pgalloc" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/sentry/vfs/lock" + "gvisor.dev/gvisor/pkg/sentry/vfs/memxattr" "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" ) +// Name is the default filesystem name. +const Name = "tmpfs" + // FilesystemType implements vfs.FilesystemType. type FilesystemType struct{} @@ -53,50 +64,176 @@ type filesystem struct { // clock is a realtime clock used to set timestamps in file operations. clock time.Clock + // devMinor is the filesystem's minor device number. devMinor is immutable. + devMinor uint32 + // mu serializes changes to the Dentry tree. mu sync.RWMutex nextInoMinusOne uint64 // accessed using atomic memory operations } +// Name implements vfs.FilesystemType.Name. +func (FilesystemType) Name() string { + return Name +} + +// FilesystemOpts is used to pass configuration data to tmpfs. +type FilesystemOpts struct { + // RootFileType is the FileType of the filesystem root. Valid values + // are: S_IFDIR, S_IFREG, and S_IFLNK. Defaults to S_IFDIR. + RootFileType uint16 + + // RootSymlinkTarget is the target of the root symlink. Only valid if + // RootFileType == S_IFLNK. + RootSymlinkTarget string + + // FilesystemType allows setting a different FilesystemType for this + // tmpfs filesystem. This allows tmpfs to "impersonate" other + // filesystems, like ramdiskfs and cgroupfs. + FilesystemType vfs.FilesystemType +} + // GetFilesystem implements vfs.FilesystemType.GetFilesystem. -func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { +func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, _ string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { memFileProvider := pgalloc.MemoryFileProviderFromContext(ctx) if memFileProvider == nil { panic("MemoryFileProviderFromContext returned nil") } + + rootFileType := uint16(linux.S_IFDIR) + newFSType := vfs.FilesystemType(&fstype) + tmpfsOpts, ok := opts.InternalData.(FilesystemOpts) + if ok { + if tmpfsOpts.RootFileType != 0 { + rootFileType = tmpfsOpts.RootFileType + } + if tmpfsOpts.FilesystemType != nil { + newFSType = tmpfsOpts.FilesystemType + } + } + + mopts := vfs.GenericParseMountOptions(opts.Data) + rootMode := linux.FileMode(0777) + if rootFileType == linux.S_IFDIR { + rootMode = 01777 + } + modeStr, ok := mopts["mode"] + if ok { + delete(mopts, "mode") + mode, err := strconv.ParseUint(modeStr, 8, 32) + if err != nil { + ctx.Warningf("tmpfs.FilesystemType.GetFilesystem: invalid mode: %q", modeStr) + return nil, nil, syserror.EINVAL + } + rootMode = linux.FileMode(mode & 07777) + } + rootKUID := creds.EffectiveKUID + uidStr, ok := mopts["uid"] + if ok { + delete(mopts, "uid") + uid, err := strconv.ParseUint(uidStr, 10, 32) + if err != nil { + ctx.Warningf("tmpfs.FilesystemType.GetFilesystem: invalid uid: %q", uidStr) + return nil, nil, syserror.EINVAL + } + kuid := creds.UserNamespace.MapToKUID(auth.UID(uid)) + if !kuid.Ok() { + ctx.Warningf("tmpfs.FilesystemType.GetFilesystem: unmapped uid: %d", uid) + return nil, nil, syserror.EINVAL + } + rootKUID = kuid + } + rootKGID := creds.EffectiveKGID + gidStr, ok := mopts["gid"] + if ok { + delete(mopts, "gid") + gid, err := strconv.ParseUint(gidStr, 10, 32) + if err != nil { + ctx.Warningf("tmpfs.FilesystemType.GetFilesystem: invalid gid: %q", gidStr) + return nil, nil, syserror.EINVAL + } + kgid := creds.UserNamespace.MapToKGID(auth.GID(gid)) + if !kgid.Ok() { + ctx.Warningf("tmpfs.FilesystemType.GetFilesystem: unmapped gid: %d", gid) + return nil, nil, syserror.EINVAL + } + rootKGID = kgid + } + if len(mopts) != 0 { + ctx.Warningf("tmpfs.FilesystemType.GetFilesystem: unknown options: %v", mopts) + return nil, nil, syserror.EINVAL + } + + devMinor, err := vfsObj.GetAnonBlockDevMinor() + if err != nil { + return nil, nil, err + } clock := time.RealtimeClockFromContext(ctx) fs := filesystem{ - memFile: memFileProvider.MemoryFile(), - clock: clock, + memFile: memFileProvider.MemoryFile(), + clock: clock, + devMinor: devMinor, + } + fs.vfsfs.Init(vfsObj, newFSType, &fs) + + var root *dentry + switch rootFileType { + case linux.S_IFREG: + root = fs.newDentry(fs.newRegularFile(rootKUID, rootKGID, rootMode)) + case linux.S_IFLNK: + root = fs.newDentry(fs.newSymlink(rootKUID, rootKGID, rootMode, tmpfsOpts.RootSymlinkTarget)) + case linux.S_IFDIR: + root = &fs.newDirectory(rootKUID, rootKGID, rootMode).dentry + default: + fs.vfsfs.DecRef() + return nil, nil, fmt.Errorf("invalid tmpfs root file type: %#o", rootFileType) } - fs.vfsfs.Init(vfsObj, &fs) - root := fs.newDentry(fs.newDirectory(creds, 01777)) return &fs.vfsfs, &root.vfsd, nil } +// NewFilesystem returns a new tmpfs filesystem. +func NewFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials) (*vfs.Filesystem, *vfs.Dentry, error) { + return FilesystemType{}.GetFilesystem(ctx, vfsObj, creds, "", vfs.GetFilesystemOptions{}) +} + // Release implements vfs.FilesystemImpl.Release. func (fs *filesystem) Release() { + fs.vfsfs.VirtualFilesystem().PutAnonBlockDevMinor(fs.devMinor) } // dentry implements vfs.DentryImpl. type dentry struct { vfsd vfs.Dentry + // parent is this dentry's parent directory. Each referenced dentry holds a + // reference on parent.dentry. If this dentry is a filesystem root, parent + // is nil. parent is protected by filesystem.mu. + parent *dentry + + // name is the name of this dentry in its parent. If this dentry is a + // filesystem root, name is the empty string. name is protected by + // filesystem.mu. + name string + + // unlinked indicates whether this dentry has been unlinked from its parent. + // It is only set to true on an unlink operation, and never set from true to + // false. unlinked is protected by filesystem.mu. + unlinked bool + + // dentryEntry (ugh) links dentries into their parent directory.childList. + dentryEntry + // inode is the inode represented by this dentry. Multiple Dentries may // share a single non-directory inode (with hard links). inode is // immutable. - inode *inode - + // // tmpfs doesn't count references on dentries; because the dentry tree is // the sole source of truth, it is by definition always consistent with the // state of the filesystem. However, it does count references on inodes, // because inode resources are released when all references are dropped. - // (tmpfs doesn't really have resources to release, but we implement - // reference counting because tmpfs regular files will.) - - // dentryEntry (ugh) links dentries into their parent directory.childList. - dentryEntry + // dentry therefore forwards reference counting directly to inode. + inode *inode } func (fs *filesystem) newDentry(inode *inode) *dentry { @@ -122,25 +259,48 @@ func (d *dentry) DecRef() { d.inode.decRef() } +// InotifyWithParent implements vfs.DentryImpl.InotifyWithParent. +func (d *dentry) InotifyWithParent(events uint32, cookie uint32, et vfs.EventType) { + if d.inode.isDir() { + events |= linux.IN_ISDIR + } + + // The ordering below is important, Linux always notifies the parent first. + if d.parent != nil { + // Note that d.parent or d.name may be stale if there is a concurrent + // rename operation. Inotify does not provide consistency guarantees. + d.parent.inode.watches.NotifyWithExclusions(d.name, events, cookie, et, d.unlinked) + } + d.inode.watches.Notify("", events, cookie, et) +} + +// Watches implements vfs.DentryImpl.Watches. +func (d *dentry) Watches() *vfs.Watches { + return &d.inode.watches +} + // inode represents a filesystem object. type inode struct { - // clock is a realtime clock used to set timestamps in file operations. - clock time.Clock + // fs is the owning filesystem. fs is immutable. + fs *filesystem // refs is a reference count. refs is accessed using atomic memory // operations. // - // A reference is held on all inodes that are reachable in the filesystem - // tree. For non-directories (which may have multiple hard links), this - // means that a reference is dropped when nlink reaches 0. For directories, - // nlink never reaches 0 due to the "." entry; instead, - // filesystem.RmdirAt() drops the reference. + // A reference is held on all inodes as long as they are reachable in the + // filesystem tree, i.e. nlink is nonzero. This reference is dropped when + // nlink reaches 0. refs int64 + // xattrs implements extended attributes. + // + // TODO(b/148380782): Support xattrs other than user.* + xattrs memxattr.SimpleExtendedAttributes + // Inode metadata. Writing multiple fields atomically requires holding // mu, othewise atomic operations can be used. mu sync.Mutex - mode uint32 // excluding file type bits, which are based on impl + mode uint32 // file type and mode nlink uint32 // protected by filesystem.mu instead of inode.mu uid uint32 // auth.KUID, but stored as raw uint32 for sync/atomic gid uint32 // auth.KGID, but ... @@ -151,31 +311,34 @@ type inode struct { ctime int64 // nanoseconds mtime int64 // nanoseconds - // Only meaningful for device special files. - rdevMajor uint32 - rdevMinor uint32 - // Advisory file locks, which lock at the inode level. locks lock.FileLocks + // Inotify watches for this inode. + watches vfs.Watches + impl interface{} // immutable } const maxLinks = math.MaxUint32 -func (i *inode) init(impl interface{}, fs *filesystem, creds *auth.Credentials, mode linux.FileMode) { - i.clock = fs.clock +func (i *inode) init(impl interface{}, fs *filesystem, kuid auth.KUID, kgid auth.KGID, mode linux.FileMode) { + if mode.FileType() == 0 { + panic("file type is required in FileMode") + } + i.fs = fs i.refs = 1 i.mode = uint32(mode) - i.uid = uint32(creds.EffectiveKUID) - i.gid = uint32(creds.EffectiveKGID) + i.uid = uint32(kuid) + i.gid = uint32(kgid) i.ino = atomic.AddUint64(&fs.nextInoMinusOne, 1) // Tmpfs creation sets atime, ctime, and mtime to current time. - now := i.clock.Now().Nanoseconds() + now := fs.clock.Now().Nanoseconds() i.atime = now i.ctime = now i.mtime = now // i.nlink initialized by caller + i.watches = vfs.Watches{} i.impl = impl } @@ -188,19 +351,22 @@ func (i *inode) incLinksLocked() { panic("tmpfs.inode.incLinksLocked() called with no existing links") } if i.nlink == maxLinks { - panic("memfs.inode.incLinksLocked() called with maximum link count") + panic("tmpfs.inode.incLinksLocked() called with maximum link count") } atomic.AddUint32(&i.nlink, 1) } -// decLinksLocked decrements i's link count. +// decLinksLocked decrements i's link count. If the link count reaches 0, we +// remove a reference on i as well. // // Preconditions: filesystem.mu must be locked for writing. i.nlink != 0. func (i *inode) decLinksLocked() { if i.nlink == 0 { panic("tmpfs.inode.decLinksLocked() called with no existing links") } - atomic.AddUint32(&i.nlink, ^uint32(0)) + if atomic.AddUint32(&i.nlink, ^uint32(0)) == 0 { + i.decRef() + } } func (i *inode) incRef() { @@ -223,20 +389,21 @@ func (i *inode) tryIncRef() bool { func (i *inode) decRef() { if refs := atomic.AddInt64(&i.refs, -1); refs == 0 { - // This is unnecessary; it's mostly to simulate what tmpfs would do. + i.watches.HandleDeletion() if regFile, ok := i.impl.(*regularFile); ok { - regFile.mu.Lock() + // Release memory used by regFile to store data. Since regFile is + // no longer usable, we don't need to grab any locks or update any + // metadata. regFile.data.DropAll(regFile.memFile) - atomic.StoreUint64(®File.size, 0) - regFile.mu.Unlock() } } else if refs < 0 { panic("tmpfs.inode.decRef() called without holding a reference") } } -func (i *inode) checkPermissions(creds *auth.Credentials, ats vfs.AccessTypes, isDir bool) error { - return vfs.GenericCheckPermissions(creds, ats, isDir, uint16(atomic.LoadUint32(&i.mode)), auth.KUID(atomic.LoadUint32(&i.uid)), auth.KGID(atomic.LoadUint32(&i.gid))) +func (i *inode) checkPermissions(creds *auth.Credentials, ats vfs.AccessTypes) error { + mode := linux.FileMode(atomic.LoadUint32(&i.mode)) + return vfs.GenericCheckPermissions(creds, ats, mode, auth.KUID(atomic.LoadUint32(&i.uid)), auth.KGID(atomic.LoadUint32(&i.gid))) } // Go won't inline this function, and returning linux.Statx (which is quite @@ -247,44 +414,38 @@ func (i *inode) checkPermissions(creds *auth.Credentials, ats vfs.AccessTypes, i // a concurrent modification), so we do not require holding inode.mu. func (i *inode) statTo(stat *linux.Statx) { stat.Mask = linux.STATX_TYPE | linux.STATX_MODE | linux.STATX_NLINK | - linux.STATX_UID | linux.STATX_GID | linux.STATX_INO | linux.STATX_ATIME | - linux.STATX_BTIME | linux.STATX_CTIME | linux.STATX_MTIME - stat.Blksize = 1 // usermem.PageSize in tmpfs + linux.STATX_UID | linux.STATX_GID | linux.STATX_INO | linux.STATX_SIZE | + linux.STATX_BLOCKS | linux.STATX_ATIME | linux.STATX_CTIME | + linux.STATX_MTIME + stat.Blksize = usermem.PageSize stat.Nlink = atomic.LoadUint32(&i.nlink) stat.UID = atomic.LoadUint32(&i.uid) stat.GID = atomic.LoadUint32(&i.gid) stat.Mode = uint16(atomic.LoadUint32(&i.mode)) stat.Ino = i.ino - // Linux's tmpfs has no concept of btime, so zero-value is returned. stat.Atime = linux.NsecToStatxTimestamp(i.atime) stat.Ctime = linux.NsecToStatxTimestamp(i.ctime) stat.Mtime = linux.NsecToStatxTimestamp(i.mtime) - // TODO(gvisor.dev/issues/1197): Device number. + stat.DevMajor = linux.UNNAMED_MAJOR + stat.DevMinor = i.fs.devMinor switch impl := i.impl.(type) { case *regularFile: - stat.Mode |= linux.S_IFREG stat.Mask |= linux.STATX_SIZE | linux.STATX_BLOCKS stat.Size = uint64(atomic.LoadUint64(&impl.size)) - // In tmpfs, this will be FileRangeSet.Span() / 512 (but also cached in - // a uint64 accessed using atomic memory operations to avoid taking - // locks). + // TODO(jamieliu): This should be impl.data.Span() / 512, but this is + // too expensive to compute here. Cache it in regularFile. stat.Blocks = allocatedBlocksForSize(stat.Size) case *directory: - stat.Mode |= linux.S_IFDIR + // "20" is mm/shmem.c:BOGO_DIRENT_SIZE. + stat.Size = 20 * (2 + uint64(atomic.LoadInt64(&impl.numChildren))) + // stat.Blocks is 0. case *symlink: - stat.Mode |= linux.S_IFLNK - stat.Mask |= linux.STATX_SIZE | linux.STATX_BLOCKS stat.Size = uint64(len(impl.target)) - stat.Blocks = allocatedBlocksForSize(stat.Size) - case *namedPipe: - stat.Mode |= linux.S_IFIFO + // stat.Blocks is 0. + case *namedPipe, *socketFile: + // stat.Size and stat.Blocks are 0. case *deviceFile: - switch impl.kind { - case vfs.BlockDevice: - stat.Mode |= linux.S_IFBLK - case vfs.CharDevice: - stat.Mode |= linux.S_IFCHR - } + // stat.Size and stat.Blocks are 0. stat.RdevMajor = impl.major stat.RdevMinor = impl.minor default: @@ -292,18 +453,27 @@ func (i *inode) statTo(stat *linux.Statx) { } } -func (i *inode) setStat(stat linux.Statx) error { +func (i *inode) setStat(ctx context.Context, creds *auth.Credentials, stat *linux.Statx) error { if stat.Mask == 0 { return nil } + if stat.Mask&^(linux.STATX_MODE|linux.STATX_UID|linux.STATX_GID|linux.STATX_ATIME|linux.STATX_MTIME|linux.STATX_CTIME|linux.STATX_SIZE) != 0 { + return syserror.EPERM + } + mode := linux.FileMode(atomic.LoadUint32(&i.mode)) + if err := vfs.CheckSetStat(ctx, creds, stat, mode, auth.KUID(atomic.LoadUint32(&i.uid)), auth.KGID(atomic.LoadUint32(&i.gid))); err != nil { + return err + } i.mu.Lock() + defer i.mu.Unlock() var ( needsMtimeBump bool needsCtimeBump bool ) mask := stat.Mask if mask&linux.STATX_MODE != 0 { - atomic.StoreUint32(&i.mode, uint32(stat.Mode)) + ft := atomic.LoadUint32(&i.mode) & linux.S_IFMT + atomic.StoreUint32(&i.mode, ft|uint32(stat.Mode&^linux.S_IFMT)) needsCtimeBump = true } if mask&linux.STATX_UID != 0 { @@ -317,7 +487,7 @@ func (i *inode) setStat(stat linux.Statx) error { if mask&linux.STATX_SIZE != 0 { switch impl := i.impl.(type) { case *regularFile: - updated, err := impl.truncate(stat.Size) + updated, err := impl.truncateLocked(stat.Size) if err != nil { return err } @@ -331,29 +501,41 @@ func (i *inode) setStat(stat linux.Statx) error { return syserror.EINVAL } } + now := i.fs.clock.Now().Nanoseconds() if mask&linux.STATX_ATIME != 0 { - atomic.StoreInt64(&i.atime, stat.Atime.ToNsecCapped()) + if stat.Atime.Nsec == linux.UTIME_NOW { + atomic.StoreInt64(&i.atime, now) + } else { + atomic.StoreInt64(&i.atime, stat.Atime.ToNsecCapped()) + } needsCtimeBump = true } if mask&linux.STATX_MTIME != 0 { - atomic.StoreInt64(&i.mtime, stat.Mtime.ToNsecCapped()) + if stat.Mtime.Nsec == linux.UTIME_NOW { + atomic.StoreInt64(&i.mtime, now) + } else { + atomic.StoreInt64(&i.mtime, stat.Mtime.ToNsecCapped()) + } needsCtimeBump = true // Ignore the mtime bump, since we just set it ourselves. needsMtimeBump = false } if mask&linux.STATX_CTIME != 0 { - atomic.StoreInt64(&i.ctime, stat.Ctime.ToNsecCapped()) + if stat.Ctime.Nsec == linux.UTIME_NOW { + atomic.StoreInt64(&i.ctime, now) + } else { + atomic.StoreInt64(&i.ctime, stat.Ctime.ToNsecCapped()) + } // Ignore the ctime bump, since we just set it ourselves. needsCtimeBump = false } - now := i.clock.Now().Nanoseconds() if needsMtimeBump { atomic.StoreInt64(&i.mtime, now) } if needsCtimeBump { atomic.StoreInt64(&i.ctime, now) } - i.mu.Unlock() + return nil } @@ -412,6 +594,8 @@ func (i *inode) direntType() uint8 { return linux.DT_DIR case *symlink: return linux.DT_LNK + case *socketFile: + return linux.DT_SOCK case *deviceFile: switch impl.kind { case vfs.BlockDevice: @@ -426,6 +610,99 @@ func (i *inode) direntType() uint8 { } } +func (i *inode) isDir() bool { + return linux.FileMode(i.mode).FileType() == linux.S_IFDIR +} + +func (i *inode) touchAtime(mnt *vfs.Mount) { + if mnt.Flags.NoATime { + return + } + if err := mnt.CheckBeginWrite(); err != nil { + return + } + now := i.fs.clock.Now().Nanoseconds() + i.mu.Lock() + atomic.StoreInt64(&i.atime, now) + i.mu.Unlock() + mnt.EndWrite() +} + +// Preconditions: The caller has called vfs.Mount.CheckBeginWrite(). +func (i *inode) touchCtime() { + now := i.fs.clock.Now().Nanoseconds() + i.mu.Lock() + atomic.StoreInt64(&i.ctime, now) + i.mu.Unlock() +} + +// Preconditions: The caller has called vfs.Mount.CheckBeginWrite(). +func (i *inode) touchCMtime() { + now := i.fs.clock.Now().Nanoseconds() + i.mu.Lock() + atomic.StoreInt64(&i.mtime, now) + atomic.StoreInt64(&i.ctime, now) + i.mu.Unlock() +} + +// Preconditions: The caller has called vfs.Mount.CheckBeginWrite() and holds +// inode.mu. +func (i *inode) touchCMtimeLocked() { + now := i.fs.clock.Now().Nanoseconds() + atomic.StoreInt64(&i.mtime, now) + atomic.StoreInt64(&i.ctime, now) +} + +func (i *inode) listxattr(size uint64) ([]string, error) { + return i.xattrs.Listxattr(size) +} + +func (i *inode) getxattr(creds *auth.Credentials, opts *vfs.GetxattrOptions) (string, error) { + if err := i.checkPermissions(creds, vfs.MayRead); err != nil { + return "", err + } + if !strings.HasPrefix(opts.Name, linux.XATTR_USER_PREFIX) { + return "", syserror.EOPNOTSUPP + } + if !i.userXattrSupported() { + return "", syserror.ENODATA + } + return i.xattrs.Getxattr(opts) +} + +func (i *inode) setxattr(creds *auth.Credentials, opts *vfs.SetxattrOptions) error { + if err := i.checkPermissions(creds, vfs.MayWrite); err != nil { + return err + } + if !strings.HasPrefix(opts.Name, linux.XATTR_USER_PREFIX) { + return syserror.EOPNOTSUPP + } + if !i.userXattrSupported() { + return syserror.EPERM + } + return i.xattrs.Setxattr(opts) +} + +func (i *inode) removexattr(creds *auth.Credentials, name string) error { + if err := i.checkPermissions(creds, vfs.MayWrite); err != nil { + return err + } + if !strings.HasPrefix(name, linux.XATTR_USER_PREFIX) { + return syserror.EOPNOTSUPP + } + if !i.userXattrSupported() { + return syserror.EPERM + } + return i.xattrs.Removexattr(name) +} + +// Extended attributes in the user.* namespace are only supported for regular +// files and directories. +func (i *inode) userXattrSupported() bool { + filetype := linux.S_IFMT & atomic.LoadUint32(&i.mode) + return filetype == linux.S_IFREG || filetype == linux.S_IFDIR +} + // fileDescription is embedded by tmpfs implementations of // vfs.FileDescriptionImpl. type fileDescription struct { @@ -437,8 +714,12 @@ func (fd *fileDescription) filesystem() *filesystem { return fd.vfsfd.Mount().Filesystem().Impl().(*filesystem) } +func (fd *fileDescription) dentry() *dentry { + return fd.vfsfd.Dentry().Impl().(*dentry) +} + func (fd *fileDescription) inode() *inode { - return fd.vfsfd.Dentry().Impl().(*dentry).inode + return fd.dentry().inode } // Stat implements vfs.FileDescriptionImpl.Stat. @@ -450,5 +731,78 @@ func (fd *fileDescription) Stat(ctx context.Context, opts vfs.StatOptions) (linu // SetStat implements vfs.FileDescriptionImpl.SetStat. func (fd *fileDescription) SetStat(ctx context.Context, opts vfs.SetStatOptions) error { - return fd.inode().setStat(opts.Stat) + creds := auth.CredentialsFromContext(ctx) + d := fd.dentry() + if err := d.inode.setStat(ctx, creds, &opts.Stat); err != nil { + return err + } + + if ev := vfs.InotifyEventFromStatMask(opts.Stat.Mask); ev != 0 { + d.InotifyWithParent(ev, 0, vfs.InodeEvent) + } + return nil +} + +// Listxattr implements vfs.FileDescriptionImpl.Listxattr. +func (fd *fileDescription) Listxattr(ctx context.Context, size uint64) ([]string, error) { + return fd.inode().listxattr(size) +} + +// Getxattr implements vfs.FileDescriptionImpl.Getxattr. +func (fd *fileDescription) Getxattr(ctx context.Context, opts vfs.GetxattrOptions) (string, error) { + return fd.inode().getxattr(auth.CredentialsFromContext(ctx), &opts) +} + +// Setxattr implements vfs.FileDescriptionImpl.Setxattr. +func (fd *fileDescription) Setxattr(ctx context.Context, opts vfs.SetxattrOptions) error { + d := fd.dentry() + if err := d.inode.setxattr(auth.CredentialsFromContext(ctx), &opts); err != nil { + return err + } + + // Generate inotify events. + d.InotifyWithParent(linux.IN_ATTRIB, 0, vfs.InodeEvent) + return nil +} + +// Removexattr implements vfs.FileDescriptionImpl.Removexattr. +func (fd *fileDescription) Removexattr(ctx context.Context, name string) error { + d := fd.dentry() + if err := d.inode.removexattr(auth.CredentialsFromContext(ctx), name); err != nil { + return err + } + + // Generate inotify events. + d.InotifyWithParent(linux.IN_ATTRIB, 0, vfs.InodeEvent) + return nil +} + +// NewMemfd creates a new tmpfs regular file and file description that can back +// an anonymous fd created by memfd_create. +func NewMemfd(mount *vfs.Mount, creds *auth.Credentials, 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() + d.name = name + + // Per Linux, mm/shmem.c:__shmem_file_setup(), memfd files are set up with + // FMODE_READ | FMODE_WRITE. + var fd regularFileFD + 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 } diff --git a/pkg/sentry/fsimpl/tmpfs/tmpfs_test.go b/pkg/sentry/fsimpl/tmpfs/tmpfs_test.go new file mode 100644 index 000000000..a240fb276 --- /dev/null +++ b/pkg/sentry/fsimpl/tmpfs/tmpfs_test.go @@ -0,0 +1,156 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package tmpfs + +import ( + "fmt" + "sync/atomic" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/vfs" +) + +// nextFileID is used to generate unique file names. +var nextFileID int64 + +// newTmpfsRoot creates a new tmpfs mount, and returns the root. If the error +// is not nil, then cleanup should be called when the root is no longer needed. +func newTmpfsRoot(ctx context.Context) (*vfs.VirtualFilesystem, vfs.VirtualDentry, func(), error) { + creds := auth.CredentialsFromContext(ctx) + + vfsObj := &vfs.VirtualFilesystem{} + if err := vfsObj.Init(); err != nil { + return nil, vfs.VirtualDentry{}, nil, fmt.Errorf("VFS init: %v", err) + } + + vfsObj.MustRegisterFilesystemType("tmpfs", FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ + AllowUserMount: true, + }) + mntns, err := vfsObj.NewMountNamespace(ctx, creds, "", "tmpfs", &vfs.GetFilesystemOptions{}) + if err != nil { + return nil, vfs.VirtualDentry{}, nil, fmt.Errorf("failed to create tmpfs root mount: %v", err) + } + root := mntns.Root() + return vfsObj, root, func() { + root.DecRef() + mntns.DecRef() + }, nil +} + +// newFileFD creates a new file in a new tmpfs mount, and returns the FD. If +// the returned err is not nil, then cleanup should be called when the FD is no +// longer needed. +func newFileFD(ctx context.Context, mode linux.FileMode) (*vfs.FileDescription, func(), error) { + creds := auth.CredentialsFromContext(ctx) + vfsObj, root, cleanup, err := newTmpfsRoot(ctx) + if err != nil { + return nil, nil, err + } + + filename := fmt.Sprintf("tmpfs-test-file-%d", atomic.AddInt64(&nextFileID, 1)) + + // Create the file that will be write/read. + fd, err := vfsObj.OpenAt(ctx, creds, &vfs.PathOperation{ + Root: root, + Start: root, + Path: fspath.Parse(filename), + }, &vfs.OpenOptions{ + Flags: linux.O_RDWR | linux.O_CREAT | linux.O_EXCL, + Mode: linux.ModeRegular | mode, + }) + if err != nil { + cleanup() + return nil, nil, fmt.Errorf("failed to create file %q: %v", filename, err) + } + + return fd, cleanup, nil +} + +// newDirFD is like newFileFD, but for directories. +func newDirFD(ctx context.Context, mode linux.FileMode) (*vfs.FileDescription, func(), error) { + creds := auth.CredentialsFromContext(ctx) + vfsObj, root, cleanup, err := newTmpfsRoot(ctx) + if err != nil { + return nil, nil, err + } + + dirname := fmt.Sprintf("tmpfs-test-dir-%d", atomic.AddInt64(&nextFileID, 1)) + + // Create the dir. + if err := vfsObj.MkdirAt(ctx, creds, &vfs.PathOperation{ + Root: root, + Start: root, + Path: fspath.Parse(dirname), + }, &vfs.MkdirOptions{ + Mode: linux.ModeDirectory | mode, + }); err != nil { + cleanup() + return nil, nil, fmt.Errorf("failed to create directory %q: %v", dirname, err) + } + + // Open the dir and return it. + fd, err := vfsObj.OpenAt(ctx, creds, &vfs.PathOperation{ + Root: root, + Start: root, + Path: fspath.Parse(dirname), + }, &vfs.OpenOptions{ + Flags: linux.O_RDONLY | linux.O_DIRECTORY, + }) + if err != nil { + cleanup() + return nil, nil, fmt.Errorf("failed to open directory %q: %v", dirname, err) + } + + return fd, cleanup, nil +} + +// newPipeFD is like newFileFD, but for pipes. +func newPipeFD(ctx context.Context, mode linux.FileMode) (*vfs.FileDescription, func(), error) { + creds := auth.CredentialsFromContext(ctx) + vfsObj, root, cleanup, err := newTmpfsRoot(ctx) + if err != nil { + return nil, nil, err + } + + name := fmt.Sprintf("tmpfs-test-%d", atomic.AddInt64(&nextFileID, 1)) + + if err := vfsObj.MknodAt(ctx, creds, &vfs.PathOperation{ + Root: root, + Start: root, + Path: fspath.Parse(name), + }, &vfs.MknodOptions{ + Mode: linux.ModeNamedPipe | mode, + }); err != nil { + cleanup() + return nil, nil, fmt.Errorf("failed to create pipe %q: %v", name, err) + } + + fd, err := vfsObj.OpenAt(ctx, creds, &vfs.PathOperation{ + Root: root, + Start: root, + Path: fspath.Parse(name), + }, &vfs.OpenOptions{ + Flags: linux.O_RDWR, + }) + if err != nil { + cleanup() + return nil, nil, fmt.Errorf("failed to open pipe %q: %v", name, err) + } + + return fd, cleanup, nil +} diff --git a/pkg/sentry/hostfd/BUILD b/pkg/sentry/hostfd/BUILD new file mode 100644 index 000000000..364a78306 --- /dev/null +++ b/pkg/sentry/hostfd/BUILD @@ -0,0 +1,17 @@ +load("//tools:defs.bzl", "go_library") + +licenses(["notice"]) + +go_library( + name = "hostfd", + srcs = [ + "hostfd.go", + "hostfd_unsafe.go", + ], + visibility = ["//pkg/sentry:internal"], + deps = [ + "//pkg/safemem", + "//pkg/sync", + "@org_golang_x_sys//unix:go_default_library", + ], +) diff --git a/pkg/sentry/hostfd/hostfd.go b/pkg/sentry/hostfd/hostfd.go new file mode 100644 index 000000000..70dd9cafb --- /dev/null +++ b/pkg/sentry/hostfd/hostfd.go @@ -0,0 +1,84 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// Package hostfd provides efficient I/O with host file descriptors. +package hostfd + +import ( + "gvisor.dev/gvisor/pkg/safemem" + "gvisor.dev/gvisor/pkg/sync" +) + +// ReadWriterAt implements safemem.Reader and safemem.Writer by reading from +// and writing to a host file descriptor respectively. ReadWriterAts should be +// obtained by calling GetReadWriterAt. +// +// Clients should usually prefer to use Preadv2 and Pwritev2 directly. +type ReadWriterAt struct { + fd int32 + offset int64 + flags uint32 +} + +var rwpool = sync.Pool{ + New: func() interface{} { + return &ReadWriterAt{} + }, +} + +// GetReadWriterAt returns a ReadWriterAt that reads from / writes to the given +// host file descriptor, starting at the given offset and using the given +// preadv2(2)/pwritev2(2) flags. If offset is -1, the host file descriptor's +// offset is used instead. Users are responsible for ensuring that fd remains +// valid for the lifetime of the returned ReadWriterAt, and must call +// PutReadWriterAt when it is no longer needed. +func GetReadWriterAt(fd int32, offset int64, flags uint32) *ReadWriterAt { + rw := rwpool.Get().(*ReadWriterAt) + *rw = ReadWriterAt{ + fd: fd, + offset: offset, + flags: flags, + } + return rw +} + +// PutReadWriterAt releases a ReadWriterAt returned by a previous call to +// GetReadWriterAt that is no longer in use. +func PutReadWriterAt(rw *ReadWriterAt) { + rwpool.Put(rw) +} + +// ReadToBlocks implements safemem.Reader.ReadToBlocks. +func (rw *ReadWriterAt) ReadToBlocks(dsts safemem.BlockSeq) (uint64, error) { + if dsts.IsEmpty() { + return 0, nil + } + n, err := Preadv2(rw.fd, dsts, rw.offset, rw.flags) + if rw.offset >= 0 { + rw.offset += int64(n) + } + return n, err +} + +// WriteFromBlocks implements safemem.Writer.WriteFromBlocks. +func (rw *ReadWriterAt) WriteFromBlocks(srcs safemem.BlockSeq) (uint64, error) { + if srcs.IsEmpty() { + return 0, nil + } + n, err := Pwritev2(rw.fd, srcs, rw.offset, rw.flags) + if rw.offset >= 0 { + rw.offset += int64(n) + } + return n, err +} diff --git a/pkg/sentry/hostfd/hostfd_unsafe.go b/pkg/sentry/hostfd/hostfd_unsafe.go new file mode 100644 index 000000000..cd4dc67fb --- /dev/null +++ b/pkg/sentry/hostfd/hostfd_unsafe.go @@ -0,0 +1,85 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package hostfd + +import ( + "io" + "syscall" + "unsafe" + + "golang.org/x/sys/unix" + "gvisor.dev/gvisor/pkg/safemem" +) + +// Preadv2 reads up to dsts.NumBytes() bytes from host file descriptor fd into +// dsts. offset and flags are interpreted as for preadv2(2). +// +// Preconditions: !dsts.IsEmpty(). +func Preadv2(fd int32, dsts safemem.BlockSeq, offset int64, flags uint32) (uint64, error) { + // No buffering is necessary regardless of safecopy; host syscalls will + // return EFAULT if appropriate, instead of raising SIGBUS. + var ( + n uintptr + e syscall.Errno + ) + if flags == 0 && dsts.NumBlocks() == 1 { + // Use read() or pread() to avoid iovec allocation and copying. + dst := dsts.Head() + if offset == -1 { + n, _, e = syscall.Syscall(unix.SYS_READ, uintptr(fd), dst.Addr(), uintptr(dst.Len())) + } else { + n, _, e = syscall.Syscall6(unix.SYS_PREAD64, uintptr(fd), dst.Addr(), uintptr(dst.Len()), uintptr(offset), 0 /* pos_h */, 0 /* unused */) + } + } else { + iovs := safemem.IovecsFromBlockSeq(dsts) + n, _, e = syscall.Syscall6(unix.SYS_PREADV2, uintptr(fd), uintptr((unsafe.Pointer)(&iovs[0])), uintptr(len(iovs)), uintptr(offset), 0 /* pos_h */, uintptr(flags)) + } + if e != 0 { + return 0, e + } + if n == 0 { + return 0, io.EOF + } + return uint64(n), nil +} + +// Pwritev2 writes up to srcs.NumBytes() from srcs into host file descriptor +// fd. offset and flags are interpreted as for pwritev2(2). +// +// Preconditions: !srcs.IsEmpty(). +func Pwritev2(fd int32, srcs safemem.BlockSeq, offset int64, flags uint32) (uint64, error) { + // No buffering is necessary regardless of safecopy; host syscalls will + // return EFAULT if appropriate, instead of raising SIGBUS. + var ( + n uintptr + e syscall.Errno + ) + if flags == 0 && srcs.NumBlocks() == 1 { + // Use write() or pwrite() to avoid iovec allocation and copying. + src := srcs.Head() + if offset == -1 { + n, _, e = syscall.Syscall(unix.SYS_WRITE, uintptr(fd), src.Addr(), uintptr(src.Len())) + } else { + n, _, e = syscall.Syscall6(unix.SYS_PWRITE64, uintptr(fd), src.Addr(), uintptr(src.Len()), uintptr(offset), 0 /* pos_h */, 0 /* unused */) + } + } else { + iovs := safemem.IovecsFromBlockSeq(srcs) + n, _, e = syscall.Syscall6(unix.SYS_PWRITEV2, uintptr(fd), uintptr((unsafe.Pointer)(&iovs[0])), uintptr(len(iovs)), uintptr(offset), 0 /* pos_h */, uintptr(flags)) + } + if e != 0 { + return 0, e + } + return uint64(n), nil +} diff --git a/pkg/sentry/inet/BUILD b/pkg/sentry/inet/BUILD index 334432abf..07bf39fed 100644 --- a/pkg/sentry/inet/BUILD +++ b/pkg/sentry/inet/BUILD @@ -10,6 +10,7 @@ go_library( srcs = [ "context.go", "inet.go", + "namespace.go", "test_stack.go", ], deps = [ diff --git a/pkg/sentry/inet/namespace.go b/pkg/sentry/inet/namespace.go new file mode 100644 index 000000000..029af3025 --- /dev/null +++ b/pkg/sentry/inet/namespace.go @@ -0,0 +1,102 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package inet + +// Namespace represents a network namespace. See network_namespaces(7). +// +// +stateify savable +type Namespace struct { + // stack is the network stack implementation of this network namespace. + stack Stack `state:"nosave"` + + // creator allows kernel to create new network stack for network namespaces. + // If nil, no networking will function if network is namespaced. + // + // At afterLoad(), creator will be used to create network stack. Stateify + // needs to wait for this field to be loaded before calling afterLoad(). + creator NetworkStackCreator `state:"wait"` + + // isRoot indicates whether this is the root network namespace. + isRoot bool +} + +// NewRootNamespace creates the root network namespace, with creator +// allowing new network namespaces to be created. If creator is nil, no +// networking will function if the network is namespaced. +func NewRootNamespace(stack Stack, creator NetworkStackCreator) *Namespace { + return &Namespace{ + stack: stack, + creator: creator, + isRoot: true, + } +} + +// NewNamespace creates a new network namespace from the root. +func NewNamespace(root *Namespace) *Namespace { + n := &Namespace{ + creator: root.creator, + } + n.init() + return n +} + +// Stack returns the network stack of n. Stack may return nil if no network +// stack is configured. +func (n *Namespace) Stack() Stack { + return n.stack +} + +// IsRoot returns whether n is the root network namespace. +func (n *Namespace) IsRoot() bool { + return n.isRoot +} + +// RestoreRootStack restores the root network namespace with stack. This should +// only be called when restoring kernel. +func (n *Namespace) RestoreRootStack(stack Stack) { + if !n.isRoot { + panic("RestoreRootStack can only be called on root network namespace") + } + if n.stack != nil { + panic("RestoreRootStack called after a stack has already been set") + } + n.stack = stack +} + +func (n *Namespace) init() { + // Root network namespace will have stack assigned later. + if n.isRoot { + return + } + if n.creator != nil { + var err error + n.stack, err = n.creator.CreateStack() + if err != nil { + panic(err) + } + } +} + +// afterLoad is invoked by stateify. +func (n *Namespace) afterLoad() { + n.init() +} + +// NetworkStackCreator allows new instances of a network stack to be created. It +// is used by the kernel to create new network namespaces when requested. +type NetworkStackCreator interface { + // CreateStack creates a new network stack for a network namespace. + CreateStack() (Stack, error) +} diff --git a/pkg/sentry/kernel/BUILD b/pkg/sentry/kernel/BUILD index 2231d6973..a28eab8b8 100644 --- a/pkg/sentry/kernel/BUILD +++ b/pkg/sentry/kernel/BUILD @@ -157,6 +157,7 @@ go_library( "//pkg/context", "//pkg/cpuid", "//pkg/eventchannel", + "//pkg/fspath", "//pkg/log", "//pkg/metric", "//pkg/refs", @@ -167,6 +168,12 @@ go_library( "//pkg/sentry/fs", "//pkg/sentry/fs/lock", "//pkg/sentry/fs/timerfd", + "//pkg/sentry/fsbridge", + "//pkg/sentry/fsimpl/kernfs", + "//pkg/sentry/fsimpl/pipefs", + "//pkg/sentry/fsimpl/sockfs", + "//pkg/sentry/fsimpl/timerfd", + "//pkg/sentry/fsimpl/tmpfs", "//pkg/sentry/hostcpu", "//pkg/sentry/inet", "//pkg/sentry/kernel/auth", @@ -199,6 +206,7 @@ go_library( "//pkg/tcpip/stack", "//pkg/usermem", "//pkg/waiter", + "//tools/go_marshal/marshal", ], ) diff --git a/pkg/sentry/kernel/auth/credentials.go b/pkg/sentry/kernel/auth/credentials.go index e057d2c6d..6862f2ef5 100644 --- a/pkg/sentry/kernel/auth/credentials.go +++ b/pkg/sentry/kernel/auth/credentials.go @@ -232,3 +232,31 @@ func (c *Credentials) UseGID(gid GID) (KGID, error) { } return NoID, syserror.EPERM } + +// SetUID translates the provided uid to the root user namespace and updates c's +// uids to it. This performs no permissions or capabilities checks, the caller +// is responsible for ensuring the calling context is permitted to modify c. +func (c *Credentials) SetUID(uid UID) error { + kuid := c.UserNamespace.MapToKUID(uid) + if !kuid.Ok() { + return syserror.EINVAL + } + c.RealKUID = kuid + c.EffectiveKUID = kuid + c.SavedKUID = kuid + return nil +} + +// SetGID translates the provided gid to the root user namespace and updates c's +// gids to it. This performs no permissions or capabilities checks, the caller +// is responsible for ensuring the calling context is permitted to modify c. +func (c *Credentials) SetGID(gid GID) error { + kgid := c.UserNamespace.MapToKGID(gid) + if !kgid.Ok() { + return syserror.EINVAL + } + c.RealKGID = kgid + c.EffectiveKGID = kgid + c.SavedKGID = kgid + return nil +} diff --git a/pkg/sentry/kernel/epoll/BUILD b/pkg/sentry/kernel/epoll/BUILD index dedf0fa15..75eedd5a2 100644 --- a/pkg/sentry/kernel/epoll/BUILD +++ b/pkg/sentry/kernel/epoll/BUILD @@ -24,6 +24,7 @@ go_library( ], visibility = ["//pkg/sentry:internal"], deps = [ + "//pkg/abi/linux", "//pkg/context", "//pkg/refs", "//pkg/sentry/fs", diff --git a/pkg/sentry/kernel/epoll/epoll.go b/pkg/sentry/kernel/epoll/epoll.go index 8bffb78fc..3d78cd48f 100644 --- a/pkg/sentry/kernel/epoll/epoll.go +++ b/pkg/sentry/kernel/epoll/epoll.go @@ -20,6 +20,7 @@ import ( "fmt" "syscall" + "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/refs" "gvisor.dev/gvisor/pkg/sentry/fs" @@ -30,19 +31,6 @@ import ( "gvisor.dev/gvisor/pkg/waiter" ) -// Event describes the event mask that was observed and the user data to be -// returned when one of the events occurs. It has this format to match the linux -// format to avoid extra copying/allocation when writing events to userspace. -type Event struct { - // Events is the event mask containing the set of events that have been - // observed on an entry. - Events uint32 - - // Data is an opaque 64-bit value provided by the caller when adding the - // entry, and returned to the caller when the entry reports an event. - Data [2]int32 -} - // EntryFlags is a bitmask that holds an entry's flags. type EntryFlags int @@ -227,9 +215,9 @@ func (e *EventPoll) Readiness(mask waiter.EventMask) waiter.EventMask { } // ReadEvents returns up to max available events. -func (e *EventPoll) ReadEvents(max int) []Event { +func (e *EventPoll) ReadEvents(max int) []linux.EpollEvent { var local pollEntryList - var ret []Event + var ret []linux.EpollEvent e.listsMu.Lock() @@ -251,7 +239,7 @@ func (e *EventPoll) ReadEvents(max int) []Event { } // Add event to the array that will be returned to caller. - ret = append(ret, Event{ + ret = append(ret, linux.EpollEvent{ Events: uint32(ready), Data: entry.userData, }) @@ -296,8 +284,10 @@ func (*readyCallback) Callback(w *waiter.Entry) { e.waitingList.Remove(entry) e.readyList.PushBack(entry) entry.curList = &e.readyList + e.listsMu.Unlock() e.Notify(waiter.EventIn) + return } e.listsMu.Unlock() diff --git a/pkg/sentry/kernel/epoll/epoll_state.go b/pkg/sentry/kernel/epoll/epoll_state.go index a0d35d350..8e9f200d0 100644 --- a/pkg/sentry/kernel/epoll/epoll_state.go +++ b/pkg/sentry/kernel/epoll/epoll_state.go @@ -38,11 +38,14 @@ func (e *EventPoll) afterLoad() { } } - for it := e.waitingList.Front(); it != nil; it = it.Next() { - if it.id.File.Readiness(it.mask) != 0 { - e.waitingList.Remove(it) - e.readyList.PushBack(it) - it.curList = &e.readyList + for it := e.waitingList.Front(); it != nil; { + entry := it + it = it.Next() + + if entry.id.File.Readiness(entry.mask) != 0 { + e.waitingList.Remove(entry) + e.readyList.PushBack(entry) + entry.curList = &e.readyList e.Notify(waiter.EventIn) } } diff --git a/pkg/sentry/kernel/fd_table.go b/pkg/sentry/kernel/fd_table.go index 23b88f7a6..dbfcef0fa 100644 --- a/pkg/sentry/kernel/fd_table.go +++ b/pkg/sentry/kernel/fd_table.go @@ -152,7 +152,13 @@ func (f *FDTable) drop(file *fs.File) { // dropVFS2 drops the table reference. func (f *FDTable) dropVFS2(file *vfs.FileDescription) { // TODO(gvisor.dev/issue/1480): Release locks. - // TODO(gvisor.dev/issue/1479): Send inotify events. + + // Generate inotify events. + ev := uint32(linux.IN_CLOSE_NOWRITE) + if file.IsWritable() { + ev = linux.IN_CLOSE_WRITE + } + file.Dentry().InotifyWithParent(ev, 0, vfs.PathEvent) // Drop the table reference. file.DecRef() @@ -191,10 +197,12 @@ func (f *FDTable) Size() int { return int(size) } -// forEach iterates over all non-nil files. +// forEach iterates over all non-nil files in sorted order. // // It is the caller's responsibility to acquire an appropriate lock. func (f *FDTable) forEach(fn func(fd int32, file *fs.File, fileVFS2 *vfs.FileDescription, flags FDFlags)) { + // retries tracks the number of failed TryIncRef attempts for the same FD. + retries := 0 fd := int32(0) for { file, fileVFS2, flags, ok := f.getAll(fd) @@ -204,17 +212,26 @@ func (f *FDTable) forEach(fn func(fd int32, file *fs.File, fileVFS2 *vfs.FileDes switch { case file != nil: if !file.TryIncRef() { + retries++ + if retries > 1000 { + panic(fmt.Sprintf("File in FD table has been destroyed. FD: %d, File: %+v, FileOps: %+v", fd, file, file.FileOperations)) + } continue // Race caught. } fn(fd, file, nil, flags) file.DecRef() case fileVFS2 != nil: if !fileVFS2.TryIncRef() { + retries++ + if retries > 1000 { + panic(fmt.Sprintf("File in FD table has been destroyed. FD: %d, File: %+v, Impl: %+v", fd, fileVFS2, fileVFS2.Impl())) + } continue // Race caught. } fn(fd, nil, fileVFS2, flags) fileVFS2.DecRef() } + retries = 0 fd++ } } @@ -296,6 +313,105 @@ func (f *FDTable) NewFDs(ctx context.Context, fd int32, files []*fs.File, flags return fds, nil } +// NewFDsVFS2 allocates new FDs guaranteed to be the lowest number available +// greater than or equal to the fd parameter. All files will share the set +// flags. Success is guaranteed to be all or none. +func (f *FDTable) NewFDsVFS2(ctx context.Context, fd int32, files []*vfs.FileDescription, flags FDFlags) (fds []int32, err error) { + if fd < 0 { + // Don't accept negative FDs. + return nil, syscall.EINVAL + } + + // Default limit. + end := int32(math.MaxInt32) + + // Ensure we don't get past the provided limit. + if limitSet := limits.FromContext(ctx); limitSet != nil { + lim := limitSet.Get(limits.NumberOfFiles) + if lim.Cur != limits.Infinity { + end = int32(lim.Cur) + } + if fd >= end { + return nil, syscall.EMFILE + } + } + + f.mu.Lock() + defer f.mu.Unlock() + + // From f.next to find available fd. + if fd < f.next { + fd = f.next + } + + // Install all entries. + for i := fd; i < end && len(fds) < len(files); i++ { + if d, _, _ := f.getVFS2(i); d == nil { + f.setVFS2(i, files[len(fds)], flags) // Set the descriptor. + fds = append(fds, i) // Record the file descriptor. + } + } + + // Failure? Unwind existing FDs. + if len(fds) < len(files) { + for _, i := range fds { + f.setVFS2(i, nil, FDFlags{}) // Zap entry. + } + return nil, syscall.EMFILE + } + + if fd == f.next { + // Update next search start position. + f.next = fds[len(fds)-1] + 1 + } + + return fds, nil +} + +// NewFDVFS2 allocates a file descriptor greater than or equal to minfd for +// the given file description. If it succeeds, it takes a reference on file. +func (f *FDTable) NewFDVFS2(ctx context.Context, minfd int32, file *vfs.FileDescription, flags FDFlags) (int32, error) { + if minfd < 0 { + // Don't accept negative FDs. + return -1, syscall.EINVAL + } + + // Default limit. + end := int32(math.MaxInt32) + + // Ensure we don't get past the provided limit. + if limitSet := limits.FromContext(ctx); limitSet != nil { + lim := limitSet.Get(limits.NumberOfFiles) + if lim.Cur != limits.Infinity { + end = int32(lim.Cur) + } + if minfd >= end { + return -1, syscall.EMFILE + } + } + + f.mu.Lock() + defer f.mu.Unlock() + + // From f.next to find available fd. + fd := minfd + if fd < f.next { + fd = f.next + } + for fd < end { + if d, _, _ := f.getVFS2(fd); d == nil { + f.setVFS2(fd, file, flags) + if fd == f.next { + // Update next search start position. + f.next = fd + 1 + } + return fd, nil + } + fd++ + } + return -1, syscall.EMFILE +} + // NewFDAt sets the file reference for the given FD. If there is an active // reference for that FD, the ref count for that existing reference is // decremented. @@ -316,9 +432,6 @@ func (f *FDTable) newFDAt(ctx context.Context, fd int32, file *fs.File, fileVFS2 return syscall.EBADF } - f.mu.Lock() - defer f.mu.Unlock() - // Check the limit for the provided file. if limitSet := limits.FromContext(ctx); limitSet != nil { if lim := limitSet.Get(limits.NumberOfFiles); lim.Cur != limits.Infinity && uint64(fd) >= lim.Cur { @@ -327,6 +440,8 @@ func (f *FDTable) newFDAt(ctx context.Context, fd int32, file *fs.File, fileVFS2 } // Install the entry. + f.mu.Lock() + defer f.mu.Unlock() f.setAll(fd, file, fileVFS2, flags) return nil } @@ -404,7 +519,10 @@ func (f *FDTable) GetVFS2(fd int32) (*vfs.FileDescription, FDFlags) { } } -// GetFDs returns a list of valid fds. +// GetFDs returns a sorted list of valid fds. +// +// Precondition: The caller must be running on the task goroutine, or Task.mu +// must be locked. func (f *FDTable) GetFDs() []int32 { fds := make([]int32, 0, int(atomic.LoadInt32(&f.used))) f.forEach(func(fd int32, _ *fs.File, _ *vfs.FileDescription, _ FDFlags) { @@ -479,7 +597,9 @@ func (f *FDTable) Remove(fd int32) (*fs.File, *vfs.FileDescription) { case orig2 != nil: orig2.IncRef() } - f.setAll(fd, nil, nil, FDFlags{}) // Zap entry. + if orig != nil || orig2 != nil { + f.setAll(fd, nil, nil, FDFlags{}) // Zap entry. + } return orig, orig2 } diff --git a/pkg/sentry/kernel/fs_context.go b/pkg/sentry/kernel/fs_context.go index 2448c1d99..47f78df9a 100644 --- a/pkg/sentry/kernel/fs_context.go +++ b/pkg/sentry/kernel/fs_context.go @@ -19,6 +19,7 @@ import ( "gvisor.dev/gvisor/pkg/refs" "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/sync" ) @@ -37,10 +38,16 @@ type FSContext struct { // destroyed. root *fs.Dirent + // rootVFS2 is the filesystem root. + rootVFS2 vfs.VirtualDentry + // cwd is the current working directory. Will be nil iff the FSContext // has been destroyed. cwd *fs.Dirent + // cwdVFS2 is the current working directory. + cwdVFS2 vfs.VirtualDentry + // umask is the current file mode creation mask. When a thread using this // context invokes a syscall that creates a file, bits set in umask are // removed from the permissions that the file is created with. @@ -60,6 +67,19 @@ func newFSContext(root, cwd *fs.Dirent, umask uint) *FSContext { return &f } +// NewFSContextVFS2 returns a new filesystem context. +func NewFSContextVFS2(root, cwd vfs.VirtualDentry, umask uint) *FSContext { + root.IncRef() + cwd.IncRef() + f := FSContext{ + rootVFS2: root, + cwdVFS2: cwd, + umask: umask, + } + f.EnableLeakCheck("kernel.FSContext") + return &f +} + // destroy is the destructor for an FSContext. // // This will call DecRef on both root and cwd Dirents. If either call to @@ -75,11 +95,17 @@ func (f *FSContext) destroy() { f.mu.Lock() defer f.mu.Unlock() - f.root.DecRef() - f.root = nil - - f.cwd.DecRef() - f.cwd = nil + if VFS2Enabled { + f.rootVFS2.DecRef() + f.rootVFS2 = vfs.VirtualDentry{} + f.cwdVFS2.DecRef() + f.cwdVFS2 = vfs.VirtualDentry{} + } else { + f.root.DecRef() + f.root = nil + f.cwd.DecRef() + f.cwd = nil + } } // DecRef implements RefCounter.DecRef with destructor f.destroy. @@ -93,12 +119,21 @@ func (f *FSContext) DecRef() { func (f *FSContext) Fork() *FSContext { f.mu.Lock() defer f.mu.Unlock() - f.cwd.IncRef() - f.root.IncRef() + + if VFS2Enabled { + f.cwdVFS2.IncRef() + f.rootVFS2.IncRef() + } else { + f.cwd.IncRef() + f.root.IncRef() + } + return &FSContext{ - cwd: f.cwd, - root: f.root, - umask: f.umask, + cwd: f.cwd, + root: f.root, + cwdVFS2: f.cwdVFS2, + rootVFS2: f.rootVFS2, + umask: f.umask, } } @@ -109,12 +144,23 @@ func (f *FSContext) Fork() *FSContext { func (f *FSContext) WorkingDirectory() *fs.Dirent { f.mu.Lock() defer f.mu.Unlock() - if f.cwd != nil { - f.cwd.IncRef() - } + + f.cwd.IncRef() return f.cwd } +// WorkingDirectoryVFS2 returns the current working directory. +// +// This will return nil if called after destroy(), otherwise it will return a +// Dirent with a reference taken. +func (f *FSContext) WorkingDirectoryVFS2() vfs.VirtualDentry { + f.mu.Lock() + defer f.mu.Unlock() + + f.cwdVFS2.IncRef() + return f.cwdVFS2 +} + // SetWorkingDirectory sets the current working directory. // This will take an extra reference on the Dirent. // @@ -137,6 +183,20 @@ func (f *FSContext) SetWorkingDirectory(d *fs.Dirent) { old.DecRef() } +// SetWorkingDirectoryVFS2 sets the current working directory. +// This will take an extra reference on the VirtualDentry. +// +// This is not a valid call after destroy. +func (f *FSContext) SetWorkingDirectoryVFS2(d vfs.VirtualDentry) { + f.mu.Lock() + defer f.mu.Unlock() + + old := f.cwdVFS2 + f.cwdVFS2 = d + d.IncRef() + old.DecRef() +} + // RootDirectory returns the current filesystem root. // // This will return nil if called after destroy(), otherwise it will return a @@ -150,6 +210,18 @@ func (f *FSContext) RootDirectory() *fs.Dirent { return f.root } +// RootDirectoryVFS2 returns the current filesystem root. +// +// This will return nil if called after destroy(), otherwise it will return a +// Dirent with a reference taken. +func (f *FSContext) RootDirectoryVFS2() vfs.VirtualDentry { + f.mu.Lock() + defer f.mu.Unlock() + + f.rootVFS2.IncRef() + return f.rootVFS2 +} + // SetRootDirectory sets the root directory. // This will take an extra reference on the Dirent. // @@ -172,6 +244,28 @@ func (f *FSContext) SetRootDirectory(d *fs.Dirent) { old.DecRef() } +// SetRootDirectoryVFS2 sets the root directory. It takes a reference on vd. +// +// This is not a valid call after free. +func (f *FSContext) SetRootDirectoryVFS2(vd vfs.VirtualDentry) { + if !vd.Ok() { + panic("FSContext.SetRootDirectoryVFS2 called with zero-value VirtualDentry") + } + + f.mu.Lock() + + if !f.rootVFS2.Ok() { + f.mu.Unlock() + panic(fmt.Sprintf("FSContext.SetRootDirectoryVFS2(%v)) called after destroy", vd)) + } + + old := f.rootVFS2 + vd.IncRef() + f.rootVFS2 = vd + f.mu.Unlock() + old.DecRef() +} + // Umask returns the current umask. func (f *FSContext) Umask() uint { f.mu.Lock() diff --git a/pkg/sentry/kernel/kernel.go b/pkg/sentry/kernel/kernel.go index 3ee760ba2..5efeb3767 100644 --- a/pkg/sentry/kernel/kernel.go +++ b/pkg/sentry/kernel/kernel.go @@ -43,11 +43,17 @@ import ( "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/cpuid" "gvisor.dev/gvisor/pkg/eventchannel" + "gvisor.dev/gvisor/pkg/fspath" "gvisor.dev/gvisor/pkg/log" "gvisor.dev/gvisor/pkg/refs" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/fs" - "gvisor.dev/gvisor/pkg/sentry/fs/timerfd" + oldtimerfd "gvisor.dev/gvisor/pkg/sentry/fs/timerfd" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/pipefs" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/sockfs" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/timerfd" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/tmpfs" "gvisor.dev/gvisor/pkg/sentry/hostcpu" "gvisor.dev/gvisor/pkg/sentry/inet" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" @@ -71,6 +77,10 @@ import ( "gvisor.dev/gvisor/pkg/tcpip" ) +// VFS2Enabled is set to true when VFS2 is enabled. Added as a global for allow +// easy access everywhere. To be removed once VFS2 becomes the default. +var VFS2Enabled = false + // Kernel represents an emulated Linux kernel. It must be initialized by calling // Init() or LoadFrom(). // @@ -105,7 +115,7 @@ type Kernel struct { timekeeper *Timekeeper tasks *TaskSet rootUserNamespace *auth.UserNamespace - networkStack inet.Stack `state:"nosave"` + rootNetworkNamespace *inet.Namespace applicationCores uint useHostCores bool extraAuxv []arch.AuxEntry @@ -238,6 +248,33 @@ type Kernel struct { // SpecialOpts contains special kernel options. SpecialOpts + + // VFS keeps the filesystem state used across the kernel. + vfs vfs.VirtualFilesystem + + // hostMount is the Mount used for file descriptors that were imported + // from the host. + hostMount *vfs.Mount + + // pipeMount is the Mount used for pipes created by the pipe() and pipe2() + // syscalls (as opposed to named pipes created by mknod()). + pipeMount *vfs.Mount + + // shmMount is the Mount used for anonymous files created by the + // memfd_create() syscalls. It is analagous to Linux's shm_mnt. + shmMount *vfs.Mount + + // socketMount is the Mount used for sockets created by the socket() and + // socketpair() syscalls. There are several cases where a socket dentry will + // not be contained in socketMount: + // 1. Socket files created by mknod() + // 2. Socket fds imported from the host (Kernel.hostMount is used for these) + // 3. Socket files created by binding Unix sockets to a file path + socketMount *vfs.Mount + + // If set to true, report address space activation waits as if the task is in + // external wait so that the watchdog doesn't report the task stuck. + SleepForAddressSpaceActivation bool } // InitKernelArgs holds arguments to Init. @@ -251,8 +288,9 @@ type InitKernelArgs struct { // RootUserNamespace is the root user namespace. RootUserNamespace *auth.UserNamespace - // NetworkStack is the TCP/IP network stack. NetworkStack may be nil. - NetworkStack inet.Stack + // RootNetworkNamespace is the root network namespace. If nil, no networking + // will be available. + RootNetworkNamespace *inet.Namespace // ApplicationCores is the number of logical CPUs visible to sandboxed // applications. The set of logical CPU IDs is [0, ApplicationCores); thus @@ -297,6 +335,9 @@ func (k *Kernel) Init(args InitKernelArgs) error { if args.Timekeeper == nil { return fmt.Errorf("Timekeeper is nil") } + if args.Timekeeper.clocks == nil { + return fmt.Errorf("Must call Timekeeper.SetClocks() before Kernel.Init()") + } if args.RootUserNamespace == nil { return fmt.Errorf("RootUserNamespace is nil") } @@ -311,7 +352,10 @@ func (k *Kernel) Init(args InitKernelArgs) error { k.rootUTSNamespace = args.RootUTSNamespace k.rootIPCNamespace = args.RootIPCNamespace k.rootAbstractSocketNamespace = args.RootAbstractSocketNamespace - k.networkStack = args.NetworkStack + k.rootNetworkNamespace = args.RootNetworkNamespace + if k.rootNetworkNamespace == nil { + k.rootNetworkNamespace = inet.NewRootNamespace(nil, nil) + } k.applicationCores = args.ApplicationCores if args.UseHostCores { k.useHostCores = true @@ -331,6 +375,47 @@ func (k *Kernel) Init(args InitKernelArgs) error { k.monotonicClock = &timekeeperClock{tk: args.Timekeeper, c: sentrytime.Monotonic} k.futexes = futex.NewManager() k.netlinkPorts = port.New() + + if VFS2Enabled { + if err := k.vfs.Init(); err != nil { + return fmt.Errorf("failed to initialize VFS: %v", err) + } + + pipeFilesystem, err := pipefs.NewFilesystem(&k.vfs) + if err != nil { + return fmt.Errorf("failed to create pipefs filesystem: %v", err) + } + defer pipeFilesystem.DecRef() + pipeMount, err := k.vfs.NewDisconnectedMount(pipeFilesystem, nil, &vfs.MountOptions{}) + if err != nil { + return fmt.Errorf("failed to create pipefs mount: %v", err) + } + k.pipeMount = pipeMount + + tmpfsFilesystem, tmpfsRoot, err := tmpfs.NewFilesystem(k.SupervisorContext(), &k.vfs, auth.NewRootCredentials(k.rootUserNamespace)) + if err != nil { + return fmt.Errorf("failed to create tmpfs filesystem: %v", err) + } + defer tmpfsFilesystem.DecRef() + defer tmpfsRoot.DecRef() + shmMount, err := k.vfs.NewDisconnectedMount(tmpfsFilesystem, tmpfsRoot, &vfs.MountOptions{}) + if err != nil { + return fmt.Errorf("failed to create tmpfs mount: %v", err) + } + k.shmMount = shmMount + + socketFilesystem, err := sockfs.NewFilesystem(&k.vfs) + if err != nil { + return fmt.Errorf("failed to create sockfs filesystem: %v", err) + } + defer socketFilesystem.DecRef() + socketMount, err := k.vfs.NewDisconnectedMount(socketFilesystem, nil, &vfs.MountOptions{}) + if err != nil { + return fmt.Errorf("failed to create sockfs mount: %v", err) + } + k.socketMount = socketMount + } + return nil } @@ -450,6 +535,11 @@ func (k *Kernel) flushMountSourceRefs() error { // // Precondition: Must be called with the kernel paused. func (ts *TaskSet) forEachFDPaused(f func(*fs.File, *vfs.FileDescription) error) (err error) { + // TODO(gvisor.dev/issue/1663): Add save support for VFS2. + if VFS2Enabled { + return nil + } + ts.mu.RLock() defer ts.mu.RUnlock() for t := range ts.Root.tids { @@ -467,7 +557,7 @@ func (ts *TaskSet) forEachFDPaused(f func(*fs.File, *vfs.FileDescription) error) } func (ts *TaskSet) flushWritesToFiles(ctx context.Context) error { - // TODO(gvisor.dev/issues/1663): Add save support for VFS2. + // TODO(gvisor.dev/issue/1663): Add save support for VFS2. return ts.forEachFDPaused(func(file *fs.File, _ *vfs.FileDescription) error { if flags := file.Flags(); !flags.Write { return nil @@ -516,17 +606,32 @@ func (k *Kernel) invalidateUnsavableMappings(ctx context.Context) error { } func (ts *TaskSet) unregisterEpollWaiters() { + // TODO(gvisor.dev/issue/1663): Add save support for VFS2. + if VFS2Enabled { + return + } + ts.mu.RLock() defer ts.mu.RUnlock() + + // Tasks that belong to the same process could potentially point to the + // same FDTable. So we retain a map of processed ones to avoid + // processing the same FDTable multiple times. + processed := make(map[*FDTable]struct{}) for t := range ts.Root.tids { // We can skip locking Task.mu here since the kernel is paused. - if t.fdTable != nil { - t.fdTable.forEach(func(_ int32, file *fs.File, _ *vfs.FileDescription, _ FDFlags) { - if e, ok := file.FileOperations.(*epoll.EventPoll); ok { - e.UnregisterEpollWaiters() - } - }) + if t.fdTable == nil { + continue } + if _, ok := processed[t.fdTable]; ok { + continue + } + t.fdTable.forEach(func(_ int32, file *fs.File, _ *vfs.FileDescription, _ FDFlags) { + if e, ok := file.FileOperations.(*epoll.EventPoll); ok { + e.UnregisterEpollWaiters() + } + }) + processed[t.fdTable] = struct{}{} } } @@ -534,8 +639,6 @@ func (ts *TaskSet) unregisterEpollWaiters() { func (k *Kernel) LoadFrom(r io.Reader, net inet.Stack, clocks sentrytime.Clocks) error { loadStart := time.Now() - k.networkStack = net - initAppCores := k.applicationCores // Load the pre-saved CPUID FeatureSet. @@ -566,6 +669,10 @@ func (k *Kernel) LoadFrom(r io.Reader, net inet.Stack, clocks sentrytime.Clocks) log.Infof("Kernel load stats: %s", &stats) log.Infof("Kernel load took [%s].", time.Since(kernelStart)) + // rootNetworkNamespace should be populated after loading the state file. + // Restore the root network stack. + k.rootNetworkNamespace.RestoreRootStack(net) + // Load the memory file's state. memoryStart := time.Now() if err := k.mf.LoadFrom(k.SupervisorContext(), r); err != nil { @@ -624,7 +731,7 @@ type CreateProcessArgs struct { // File is a passed host FD pointing to a file to load as the init binary. // // This is checked if and only if Filename is "". - File *fs.File + File fsbridge.File // Argvv is a list of arguments. Argv []string @@ -673,6 +780,13 @@ type CreateProcessArgs struct { // increment it). MountNamespace *fs.MountNamespace + // MountNamespaceVFS2 optionally contains the mount namespace for this + // process. If nil, the init process's mount namespace is used. + // + // Anyone setting MountNamespaceVFS2 must donate a reference (i.e. + // increment it). + MountNamespaceVFS2 *vfs.MountNamespace + // ContainerID is the container that the process belongs to. ContainerID string } @@ -711,13 +825,26 @@ func (ctx *createProcessContext) Value(key interface{}) interface{} { return ctx.args.Credentials case fs.CtxRoot: if ctx.args.MountNamespace != nil { - // MountNamespace.Root() will take a reference on the root - // dirent for us. + // MountNamespace.Root() will take a reference on the root dirent for us. return ctx.args.MountNamespace.Root() } return nil + case vfs.CtxRoot: + if ctx.args.MountNamespaceVFS2 == nil { + return nil + } + // MountNamespaceVFS2.Root() takes a reference on the root dirent for us. + return ctx.args.MountNamespaceVFS2.Root() + case vfs.CtxMountNamespace: + if ctx.k.globalInit == nil { + return nil + } + // MountNamespaceVFS2 takes a reference for us. + return ctx.k.GlobalInit().Leader().MountNamespaceVFS2() case fs.CtxDirentCacheLimiter: return ctx.k.DirentCacheLimiter + case inet.CtxStack: + return ctx.k.RootNetworkNamespace().Stack() case ktime.CtxRealtimeClock: return ctx.k.RealtimeClock() case limits.CtxLimits: @@ -757,34 +884,77 @@ func (k *Kernel) CreateProcess(args CreateProcessArgs) (*ThreadGroup, ThreadID, defer k.extMu.Unlock() log.Infof("EXEC: %v", args.Argv) - // Grab the mount namespace. - mounts := args.MountNamespace - if mounts == nil { - mounts = k.GlobalInit().Leader().MountNamespace() - mounts.IncRef() - } - - tg := k.NewThreadGroup(mounts, args.PIDNamespace, NewSignalHandlers(), linux.SIGCHLD, args.Limits) ctx := args.NewContext(k) - // Get the root directory from the MountNamespace. - root := mounts.Root() - // The call to newFSContext below will take a reference on root, so we - // don't need to hold this one. - defer root.DecRef() - - // Grab the working directory. - remainingTraversals := uint(args.MaxSymlinkTraversals) - wd := root // Default. - if args.WorkingDirectory != "" { - var err error - wd, err = mounts.FindInode(ctx, root, nil, args.WorkingDirectory, &remainingTraversals) - if err != nil { - return nil, 0, fmt.Errorf("failed to find initial working directory %q: %v", args.WorkingDirectory, err) + var ( + opener fsbridge.Lookup + fsContext *FSContext + mntns *fs.MountNamespace + ) + + if VFS2Enabled { + mntnsVFS2 := args.MountNamespaceVFS2 + if mntnsVFS2 == nil { + // MountNamespaceVFS2 adds a reference to the namespace, which is + // transferred to the new process. + mntnsVFS2 = k.GlobalInit().Leader().MountNamespaceVFS2() + } + // Get the root directory from the MountNamespace. + root := args.MountNamespaceVFS2.Root() + // The call to newFSContext below will take a reference on root, so we + // don't need to hold this one. + defer root.DecRef() + + // Grab the working directory. + wd := root // Default. + if args.WorkingDirectory != "" { + pop := vfs.PathOperation{ + Root: root, + Start: wd, + Path: fspath.Parse(args.WorkingDirectory), + FollowFinalSymlink: true, + } + var err error + wd, err = k.VFS().GetDentryAt(ctx, args.Credentials, &pop, &vfs.GetDentryOptions{ + CheckSearchable: true, + }) + if err != nil { + return nil, 0, fmt.Errorf("failed to find initial working directory %q: %v", args.WorkingDirectory, err) + } + defer wd.DecRef() + } + opener = fsbridge.NewVFSLookup(mntnsVFS2, root, wd) + fsContext = NewFSContextVFS2(root, wd, args.Umask) + + } else { + mntns = args.MountNamespace + if mntns == nil { + mntns = k.GlobalInit().Leader().MountNamespace() + mntns.IncRef() } - defer wd.DecRef() + // Get the root directory from the MountNamespace. + root := mntns.Root() + // The call to newFSContext below will take a reference on root, so we + // don't need to hold this one. + defer root.DecRef() + + // Grab the working directory. + remainingTraversals := args.MaxSymlinkTraversals + wd := root // Default. + if args.WorkingDirectory != "" { + var err error + wd, err = mntns.FindInode(ctx, root, nil, args.WorkingDirectory, &remainingTraversals) + if err != nil { + return nil, 0, fmt.Errorf("failed to find initial working directory %q: %v", args.WorkingDirectory, err) + } + defer wd.DecRef() + } + opener = fsbridge.NewFSLookup(mntns, root, wd) + fsContext = newFSContext(root, wd, args.Umask) } + tg := k.NewThreadGroup(mntns, args.PIDNamespace, NewSignalHandlers(), linux.SIGCHLD, args.Limits) + // Check which file to start from. switch { case args.Filename != "": @@ -805,11 +975,9 @@ func (k *Kernel) CreateProcess(args CreateProcessArgs) (*ThreadGroup, ThreadID, } // Create a fresh task context. - remainingTraversals = uint(args.MaxSymlinkTraversals) + remainingTraversals := args.MaxSymlinkTraversals loadArgs := loader.LoadArgs{ - Mounts: mounts, - Root: root, - WorkingDirectory: wd, + Opener: opener, RemainingTraversals: &remainingTraversals, ResolveFinal: true, Filename: args.Filename, @@ -834,13 +1002,15 @@ func (k *Kernel) CreateProcess(args CreateProcessArgs) (*ThreadGroup, ThreadID, Kernel: k, ThreadGroup: tg, TaskContext: tc, - FSContext: newFSContext(root, wd, args.Umask), + FSContext: fsContext, FDTable: args.FDTable, Credentials: args.Credentials, + NetworkNamespace: k.RootNetworkNamespace(), AllowedCPUMask: sched.NewFullCPUSet(k.applicationCores), UTSNamespace: args.UTSNamespace, IPCNamespace: args.IPCNamespace, AbstractSocketNamespace: args.AbstractSocketNamespace, + MountNamespaceVFS2: args.MountNamespaceVFS2, ContainerID: args.ContainerID, } t, err := k.tasks.NewTask(config) @@ -923,9 +1093,15 @@ func (k *Kernel) pauseTimeLocked() { // This means we'll iterate FDTables shared by multiple tasks repeatedly, // but ktime.Timer.Pause is idempotent so this is harmless. if t.fdTable != nil { - t.fdTable.forEach(func(_ int32, file *fs.File, _ *vfs.FileDescription, _ FDFlags) { - if tfd, ok := file.FileOperations.(*timerfd.TimerOperations); ok { - tfd.PauseTimer() + t.fdTable.forEach(func(_ int32, file *fs.File, fd *vfs.FileDescription, _ FDFlags) { + if VFS2Enabled { + if tfd, ok := fd.Impl().(*timerfd.TimerFileDescription); ok { + tfd.PauseTimer() + } + } else { + if tfd, ok := file.FileOperations.(*oldtimerfd.TimerOperations); ok { + tfd.PauseTimer() + } } }) } @@ -953,9 +1129,15 @@ func (k *Kernel) resumeTimeLocked() { } } if t.fdTable != nil { - t.fdTable.forEach(func(_ int32, file *fs.File, _ *vfs.FileDescription, _ FDFlags) { - if tfd, ok := file.FileOperations.(*timerfd.TimerOperations); ok { - tfd.ResumeTimer() + t.fdTable.forEach(func(_ int32, file *fs.File, fd *vfs.FileDescription, _ FDFlags) { + if VFS2Enabled { + if tfd, ok := fd.Impl().(*timerfd.TimerFileDescription); ok { + tfd.ResumeTimer() + } + } else { + if tfd, ok := file.FileOperations.(*oldtimerfd.TimerOperations); ok { + tfd.ResumeTimer() + } } }) } @@ -1100,6 +1282,14 @@ func (k *Kernel) SendExternalSignal(info *arch.SignalInfo, context string) { k.sendExternalSignal(info, context) } +// SendExternalSignalThreadGroup injects a signal into an specific ThreadGroup. +// This function doesn't skip signals like SendExternalSignal does. +func (k *Kernel) SendExternalSignalThreadGroup(tg *ThreadGroup, info *arch.SignalInfo) error { + k.extMu.Lock() + defer k.extMu.Unlock() + return tg.SendSignal(info) +} + // SendContainerSignal sends the given signal to all processes inside the // namespace that match the given container ID. func (k *Kernel) SendContainerSignal(cid string, info *arch.SignalInfo) error { @@ -1178,10 +1368,9 @@ func (k *Kernel) RootAbstractSocketNamespace() *AbstractSocketNamespace { return k.rootAbstractSocketNamespace } -// NetworkStack returns the network stack. NetworkStack may return nil if no -// network stack is available. -func (k *Kernel) NetworkStack() inet.Stack { - return k.networkStack +// RootNetworkNamespace returns the root network namespace, always non-nil. +func (k *Kernel) RootNetworkNamespace() *inet.Namespace { + return k.rootNetworkNamespace } // GlobalInit returns the thread group with ID 1 in the root PID namespace, or @@ -1311,9 +1500,10 @@ func (k *Kernel) SupervisorContext() context.Context { // +stateify savable type SocketEntry struct { socketEntry - k *Kernel - Sock *refs.WeakRef - ID uint64 // Socket table entry number. + k *Kernel + Sock *refs.WeakRef + SockVFS2 *vfs.FileDescription + ID uint64 // Socket table entry number. } // WeakRefGone implements refs.WeakRefUser.WeakRefGone. @@ -1336,7 +1526,30 @@ func (k *Kernel) RecordSocket(sock *fs.File) { k.extMu.Unlock() } +// RecordSocketVFS2 adds a VFS2 socket to the system-wide socket table for +// tracking. +// +// Precondition: Caller must hold a reference to sock. +// +// Note that the socket table will not hold a reference on the +// vfs.FileDescription, because we do not support weak refs on VFS2 files. +func (k *Kernel) RecordSocketVFS2(sock *vfs.FileDescription) { + k.extMu.Lock() + id := k.nextSocketEntry + k.nextSocketEntry++ + s := &SocketEntry{ + k: k, + ID: id, + SockVFS2: sock, + } + k.sockets.PushBack(s) + k.extMu.Unlock() +} + // ListSockets returns a snapshot of all sockets. +// +// Callers of ListSockets() in VFS2 should use SocketEntry.SockVFS2.TryIncRef() +// to get a reference on a socket in the table. func (k *Kernel) ListSockets() []*SocketEntry { k.extMu.Lock() var socks []*SocketEntry @@ -1378,8 +1591,24 @@ func (ctx supervisorContext) Value(key interface{}) interface{} { return ctx.k.globalInit.mounts.Root() } return nil + case vfs.CtxRoot: + if ctx.k.globalInit == nil { + return vfs.VirtualDentry{} + } + mntns := ctx.k.GlobalInit().Leader().MountNamespaceVFS2() + defer mntns.DecRef() + // Root() takes a reference on the root dirent for us. + return mntns.Root() + case vfs.CtxMountNamespace: + if ctx.k.globalInit == nil { + return nil + } + // MountNamespaceVFS2() takes a reference for us. + return ctx.k.GlobalInit().Leader().MountNamespaceVFS2() case fs.CtxDirentCacheLimiter: return ctx.k.DirentCacheLimiter + case inet.CtxStack: + return ctx.k.RootNetworkNamespace().Stack() case ktime.CtxRealtimeClock: return ctx.k.RealtimeClock() case limits.CtxLimits: @@ -1423,3 +1652,36 @@ func (k *Kernel) EmitUnimplementedEvent(ctx context.Context) { Registers: t.Arch().StateData().Proto(), }) } + +// VFS returns the virtual filesystem for the kernel. +func (k *Kernel) VFS() *vfs.VirtualFilesystem { + return &k.vfs +} + +// SetHostMount sets the hostfs mount. +func (k *Kernel) SetHostMount(mnt *vfs.Mount) { + if k.hostMount != nil { + panic("Kernel.hostMount cannot be set more than once") + } + k.hostMount = mnt +} + +// HostMount returns the hostfs mount. +func (k *Kernel) HostMount() *vfs.Mount { + return k.hostMount +} + +// PipeMount returns the pipefs mount. +func (k *Kernel) PipeMount() *vfs.Mount { + return k.pipeMount +} + +// ShmMount returns the tmpfs mount. +func (k *Kernel) ShmMount() *vfs.Mount { + return k.shmMount +} + +// SocketMount returns the sockfs mount. +func (k *Kernel) SocketMount() *vfs.Mount { + return k.socketMount +} diff --git a/pkg/sentry/kernel/pipe/BUILD b/pkg/sentry/kernel/pipe/BUILD index 4c049d5b4..7bfa9075a 100644 --- a/pkg/sentry/kernel/pipe/BUILD +++ b/pkg/sentry/kernel/pipe/BUILD @@ -1,28 +1,14 @@ load("//tools:defs.bzl", "go_library", "go_test") -load("//tools/go_generics:defs.bzl", "go_template_instance") package(licenses = ["notice"]) -go_template_instance( - name = "buffer_list", - out = "buffer_list.go", - package = "pipe", - prefix = "buffer", - template = "//pkg/ilist:generic_list", - types = { - "Element": "*buffer", - "Linker": "*buffer", - }, -) - go_library( name = "pipe", srcs = [ - "buffer.go", - "buffer_list.go", "device.go", "node.go", "pipe.go", + "pipe_unsafe.go", "pipe_util.go", "reader.go", "reader_writer.go", @@ -33,6 +19,7 @@ go_library( deps = [ "//pkg/abi/linux", "//pkg/amutex", + "//pkg/buffer", "//pkg/context", "//pkg/safemem", "//pkg/sentry/arch", @@ -51,7 +38,6 @@ go_test( name = "pipe_test", size = "small", srcs = [ - "buffer_test.go", "node_test.go", "pipe_test.go", ], diff --git a/pkg/sentry/kernel/pipe/buffer.go b/pkg/sentry/kernel/pipe/buffer.go deleted file mode 100644 index fe3be5dbd..000000000 --- a/pkg/sentry/kernel/pipe/buffer.go +++ /dev/null @@ -1,115 +0,0 @@ -// Copyright 2018 The gVisor Authors. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -package pipe - -import ( - "io" - - "gvisor.dev/gvisor/pkg/safemem" - "gvisor.dev/gvisor/pkg/sync" -) - -// buffer encapsulates a queueable byte buffer. -// -// Note that the total size is slightly less than two pages. This -// is done intentionally to ensure that the buffer object aligns -// with runtime internals. We have no hard size or alignment -// requirements. This two page size will effectively minimize -// internal fragmentation, but still have a large enough chunk -// to limit excessive segmentation. -// -// +stateify savable -type buffer struct { - data [8144]byte - read int - write int - bufferEntry -} - -// Reset resets internal data. -// -// This must be called before use. -func (b *buffer) Reset() { - b.read = 0 - b.write = 0 -} - -// Empty indicates the buffer is empty. -// -// This indicates there is no data left to read. -func (b *buffer) Empty() bool { - return b.read == b.write -} - -// Full indicates the buffer is full. -// -// This indicates there is no capacity left to write. -func (b *buffer) Full() bool { - return b.write == len(b.data) -} - -// WriteFromBlocks implements safemem.Writer.WriteFromBlocks. -func (b *buffer) WriteFromBlocks(srcs safemem.BlockSeq) (uint64, error) { - dst := safemem.BlockSeqOf(safemem.BlockFromSafeSlice(b.data[b.write:])) - n, err := safemem.CopySeq(dst, srcs) - b.write += int(n) - return n, err -} - -// WriteFromReader writes to the buffer from an io.Reader. -func (b *buffer) WriteFromReader(r io.Reader, count int64) (int64, error) { - dst := b.data[b.write:] - if count < int64(len(dst)) { - dst = b.data[b.write:][:count] - } - n, err := r.Read(dst) - b.write += n - return int64(n), err -} - -// ReadToBlocks implements safemem.Reader.ReadToBlocks. -func (b *buffer) ReadToBlocks(dsts safemem.BlockSeq) (uint64, error) { - src := safemem.BlockSeqOf(safemem.BlockFromSafeSlice(b.data[b.read:b.write])) - n, err := safemem.CopySeq(dsts, src) - b.read += int(n) - return n, err -} - -// ReadToWriter reads from the buffer into an io.Writer. -func (b *buffer) ReadToWriter(w io.Writer, count int64, dup bool) (int64, error) { - src := b.data[b.read:b.write] - if count < int64(len(src)) { - src = b.data[b.read:][:count] - } - n, err := w.Write(src) - if !dup { - b.read += n - } - return int64(n), err -} - -// bufferPool is a pool for buffers. -var bufferPool = sync.Pool{ - New: func() interface{} { - return new(buffer) - }, -} - -// newBuffer grabs a new buffer from the pool. -func newBuffer() *buffer { - b := bufferPool.Get().(*buffer) - b.Reset() - return b -} diff --git a/pkg/sentry/kernel/pipe/pipe.go b/pkg/sentry/kernel/pipe/pipe.go index 08410283f..79645d7d2 100644 --- a/pkg/sentry/kernel/pipe/pipe.go +++ b/pkg/sentry/kernel/pipe/pipe.go @@ -20,6 +20,7 @@ import ( "sync/atomic" "syscall" + "gvisor.dev/gvisor/pkg/buffer" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sync" @@ -70,10 +71,10 @@ type Pipe struct { // mu protects all pipe internal state below. mu sync.Mutex `state:"nosave"` - // data is the buffer queue of pipe contents. + // view is the underlying set of buffers. // // This is protected by mu. - data bufferList + view buffer.View // max is the maximum size of the pipe in bytes. When this max has been // reached, writers will get EWOULDBLOCK. @@ -81,11 +82,6 @@ type Pipe struct { // This is protected by mu. max int64 - // size is the current size of the pipe in bytes. - // - // This is protected by mu. - size int64 - // hadWriter indicates if this pipe ever had a writer. Note that this // does not necessarily indicate there is *currently* a writer, just // that there has been a writer at some point since the pipe was @@ -196,7 +192,7 @@ type readOps struct { limit func(int64) // read performs the actual read operation. - read func(*buffer) (int64, error) + read func(*buffer.View) (int64, error) } // read reads data from the pipe into dst and returns the number of bytes @@ -211,82 +207,27 @@ func (p *Pipe) read(ctx context.Context, ops readOps) (int64, error) { p.mu.Lock() defer p.mu.Unlock() - - // Is the pipe empty? - if p.size == 0 { - if !p.HasWriters() { - // There are no writers, return EOF. - return 0, nil - } - return 0, syserror.ErrWouldBlock - } - - // Limit how much we consume. - if ops.left() > p.size { - ops.limit(p.size) - } - - done := int64(0) - for ops.left() > 0 { - // Pop the first buffer. - first := p.data.Front() - if first == nil { - break - } - - // Copy user data. - n, err := ops.read(first) - done += int64(n) - p.size -= n - - // Empty buffer? - if first.Empty() { - // Push to the free list. - p.data.Remove(first) - bufferPool.Put(first) - } - - // Handle errors. - if err != nil { - return done, err - } - } - - return done, nil + return p.readLocked(ctx, ops) } -// dup duplicates all data from this pipe into the given writer. -// -// There is no blocking behavior implemented here. The writer may propagate -// some blocking error. All the writes must be complete writes. -func (p *Pipe) dup(ctx context.Context, ops readOps) (int64, error) { - p.mu.Lock() - defer p.mu.Unlock() - +func (p *Pipe) readLocked(ctx context.Context, ops readOps) (int64, error) { // Is the pipe empty? - if p.size == 0 { + if p.view.Size() == 0 { if !p.HasWriters() { - // See above. + // There are no writers, return EOF. return 0, nil } return 0, syserror.ErrWouldBlock } // Limit how much we consume. - if ops.left() > p.size { - ops.limit(p.size) + if ops.left() > p.view.Size() { + ops.limit(p.view.Size()) } - done := int64(0) - for buf := p.data.Front(); buf != nil; buf = buf.Next() { - n, err := ops.read(buf) - done += n - if err != nil { - return done, err - } - } - - return done, nil + // Copy user data; the read op is responsible for trimming. + done, err := ops.read(&p.view) + return done, err } type writeOps struct { @@ -297,7 +238,7 @@ type writeOps struct { limit func(int64) // write should write to the provided buffer. - write func(*buffer) (int64, error) + write func(*buffer.View) (int64, error) } // write writes data from sv into the pipe and returns the number of bytes @@ -308,7 +249,10 @@ type writeOps struct { func (p *Pipe) write(ctx context.Context, ops writeOps) (int64, error) { p.mu.Lock() defer p.mu.Unlock() + return p.writeLocked(ctx, ops) +} +func (p *Pipe) writeLocked(ctx context.Context, ops writeOps) (int64, error) { // Can't write to a pipe with no readers. if !p.HasReaders() { return 0, syscall.EPIPE @@ -317,35 +261,28 @@ func (p *Pipe) write(ctx context.Context, ops writeOps) (int64, error) { // POSIX requires that a write smaller than atomicIOBytes (PIPE_BUF) be // atomic, but requires no atomicity for writes larger than this. wanted := ops.left() - if avail := p.max - p.size; wanted > avail { + avail := p.max - p.view.Size() + if wanted > avail { if wanted <= p.atomicIOBytes { return 0, syserror.ErrWouldBlock } ops.limit(avail) } - done := int64(0) - for ops.left() > 0 { - // Need a new buffer? - last := p.data.Back() - if last == nil || last.Full() { - // Add a new buffer to the data list. - last = newBuffer() - p.data.PushBack(last) - } - - // Copy user data. - n, err := ops.write(last) - done += int64(n) - p.size += n + // Copy user data. + done, err := ops.write(&p.view) + if err != nil { + return done, err + } - // Handle errors. - if err != nil { - return done, err - } + if done < avail { + // Non-failure, but short write. + return done, nil } - if wanted > done { - // Partial write due to full pipe. + if done < wanted { + // Partial write due to full pipe. Note that this could also be + // the short write case above, we would expect a second call + // and the write to return zero bytes in this case. return done, syserror.ErrWouldBlock } @@ -396,7 +333,7 @@ func (p *Pipe) HasWriters() bool { // Precondition: mu must be held. func (p *Pipe) rReadinessLocked() waiter.EventMask { ready := waiter.EventMask(0) - if p.HasReaders() && p.data.Front() != nil { + if p.HasReaders() && p.view.Size() != 0 { ready |= waiter.EventIn } if !p.HasWriters() && p.hadWriter { @@ -422,7 +359,7 @@ func (p *Pipe) rReadiness() waiter.EventMask { // Precondition: mu must be held. func (p *Pipe) wReadinessLocked() waiter.EventMask { ready := waiter.EventMask(0) - if p.HasWriters() && p.size < p.max { + if p.HasWriters() && p.view.Size() < p.max { ready |= waiter.EventOut } if !p.HasReaders() { @@ -451,7 +388,7 @@ func (p *Pipe) rwReadiness() waiter.EventMask { func (p *Pipe) queued() int64 { p.mu.Lock() defer p.mu.Unlock() - return p.size + return p.view.Size() } // FifoSize implements fs.FifoSizer.FifoSize. @@ -474,7 +411,7 @@ func (p *Pipe) SetFifoSize(size int64) (int64, error) { } p.mu.Lock() defer p.mu.Unlock() - if size < p.size { + if size < p.view.Size() { return 0, syserror.EBUSY } p.max = size diff --git a/pkg/sentry/kernel/pipe/buffer_test.go b/pkg/sentry/kernel/pipe/pipe_unsafe.go index 4d54b8b8f..dd60cba24 100644 --- a/pkg/sentry/kernel/pipe/buffer_test.go +++ b/pkg/sentry/kernel/pipe/pipe_unsafe.go @@ -15,18 +15,21 @@ package pipe import ( - "testing" "unsafe" - - "gvisor.dev/gvisor/pkg/usermem" ) -func TestBufferSize(t *testing.T) { - bufferSize := unsafe.Sizeof(buffer{}) - if bufferSize < usermem.PageSize { - t.Errorf("buffer is less than a page") - } - if bufferSize > (2 * usermem.PageSize) { - t.Errorf("buffer is greater than two pages") +// lockTwoPipes locks both x.mu and y.mu in an order that is guaranteed to be +// consistent for both lockTwoPipes(x, y) and lockTwoPipes(y, x), such that +// concurrent calls cannot deadlock. +// +// Preconditions: x != y. +func lockTwoPipes(x, y *Pipe) { + // Lock the two pipes in order of increasing address. + if uintptr(unsafe.Pointer(x)) < uintptr(unsafe.Pointer(y)) { + x.mu.Lock() + y.mu.Lock() + } else { + y.mu.Lock() + x.mu.Lock() } } diff --git a/pkg/sentry/kernel/pipe/pipe_util.go b/pkg/sentry/kernel/pipe/pipe_util.go index 80158239e..aacf28da2 100644 --- a/pkg/sentry/kernel/pipe/pipe_util.go +++ b/pkg/sentry/kernel/pipe/pipe_util.go @@ -21,6 +21,7 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/amutex" + "gvisor.dev/gvisor/pkg/buffer" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sync" @@ -49,9 +50,10 @@ func (p *Pipe) Read(ctx context.Context, dst usermem.IOSequence) (int64, error) limit: func(l int64) { dst = dst.TakeFirst64(l) }, - read: func(buf *buffer) (int64, error) { - n, err := dst.CopyOutFrom(ctx, buf) + read: func(view *buffer.View) (int64, error) { + n, err := dst.CopyOutFrom(ctx, view) dst = dst.DropFirst64(n) + view.TrimFront(n) return n, err }, }) @@ -70,16 +72,15 @@ func (p *Pipe) WriteTo(ctx context.Context, w io.Writer, count int64, dup bool) limit: func(l int64) { count = l }, - read: func(buf *buffer) (int64, error) { - n, err := buf.ReadToWriter(w, count, dup) + read: func(view *buffer.View) (int64, error) { + n, err := view.ReadToWriter(w, count) + if !dup { + view.TrimFront(n) + } count -= n return n, err }, } - if dup { - // There is no notification for dup operations. - return p.dup(ctx, ops) - } n, err := p.read(ctx, ops) if n > 0 { p.Notify(waiter.EventOut) @@ -96,8 +97,8 @@ func (p *Pipe) Write(ctx context.Context, src usermem.IOSequence) (int64, error) limit: func(l int64) { src = src.TakeFirst64(l) }, - write: func(buf *buffer) (int64, error) { - n, err := src.CopyInTo(ctx, buf) + write: func(view *buffer.View) (int64, error) { + n, err := src.CopyInTo(ctx, view) src = src.DropFirst64(n) return n, err }, @@ -117,8 +118,8 @@ func (p *Pipe) ReadFrom(ctx context.Context, r io.Reader, count int64) (int64, e limit: func(l int64) { count = l }, - write: func(buf *buffer) (int64, error) { - n, err := buf.WriteFromReader(r, count) + write: func(view *buffer.View) (int64, error) { + n, err := view.WriteFromReader(r, count) count -= n return n, err }, @@ -143,7 +144,7 @@ func (p *Pipe) Ioctl(ctx context.Context, io usermem.IO, args arch.SyscallArgume if v > math.MaxInt32 { v = math.MaxInt32 // Silently truncate. } - // Copy result to user-space. + // Copy result to userspace. _, err := usermem.CopyObjectOut(ctx, io, args[2].Pointer(), int32(v), usermem.IOOpts{ AddressSpaceActive: true, }) diff --git a/pkg/sentry/kernel/pipe/vfs.go b/pkg/sentry/kernel/pipe/vfs.go index a5675bd70..2602bed72 100644 --- a/pkg/sentry/kernel/pipe/vfs.go +++ b/pkg/sentry/kernel/pipe/vfs.go @@ -16,7 +16,9 @@ package pipe import ( "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/buffer" "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/safemem" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/sync" @@ -49,38 +51,42 @@ type VFSPipe struct { } // NewVFSPipe returns an initialized VFSPipe. -func NewVFSPipe(sizeBytes, atomicIOBytes int64) *VFSPipe { +func NewVFSPipe(isNamed bool, sizeBytes, atomicIOBytes int64) *VFSPipe { var vp VFSPipe - initPipe(&vp.pipe, true /* isNamed */, sizeBytes, atomicIOBytes) + initPipe(&vp.pipe, isNamed, sizeBytes, atomicIOBytes) return &vp } -// NewVFSPipeFD opens a named pipe. Named pipes have special blocking semantics -// during open: +// ReaderWriterPair returns read-only and write-only FDs for vp. // -// "Normally, opening the FIFO blocks until the other end is opened also. A -// process can open a FIFO in nonblocking mode. In this case, opening for -// read-only will succeed even if no-one has opened on the write side yet, -// opening for write-only will fail with ENXIO (no such device or address) -// unless the other end has already been opened. Under Linux, opening a FIFO -// for read and write will succeed both in blocking and nonblocking mode. POSIX -// leaves this behavior undefined. This can be used to open a FIFO for writing -// while there are no readers available." - fifo(7) -func (vp *VFSPipe) NewVFSPipeFD(ctx context.Context, vfsd *vfs.Dentry, vfsfd *vfs.FileDescription, flags uint32) (*VFSPipeFD, error) { +// Preconditions: statusFlags should not contain an open access mode. +func (vp *VFSPipe) ReaderWriterPair(mnt *vfs.Mount, vfsd *vfs.Dentry, statusFlags uint32) (*vfs.FileDescription, *vfs.FileDescription) { + return vp.newFD(mnt, vfsd, linux.O_RDONLY|statusFlags), vp.newFD(mnt, vfsd, linux.O_WRONLY|statusFlags) +} + +// Open opens the pipe represented by vp. +func (vp *VFSPipe) Open(ctx context.Context, mnt *vfs.Mount, vfsd *vfs.Dentry, statusFlags uint32) (*vfs.FileDescription, error) { vp.mu.Lock() defer vp.mu.Unlock() - readable := vfs.MayReadFileWithOpenFlags(flags) - writable := vfs.MayWriteFileWithOpenFlags(flags) + readable := vfs.MayReadFileWithOpenFlags(statusFlags) + writable := vfs.MayWriteFileWithOpenFlags(statusFlags) if !readable && !writable { return nil, syserror.EINVAL } - vfd, err := vp.open(vfsd, vfsfd, flags) - if err != nil { - return nil, err - } + fd := vp.newFD(mnt, vfsd, statusFlags) + // Named pipes have special blocking semantics during open: + // + // "Normally, opening the FIFO blocks until the other end is opened also. A + // process can open a FIFO in nonblocking mode. In this case, opening for + // read-only will succeed even if no-one has opened on the write side yet, + // opening for write-only will fail with ENXIO (no such device or address) + // unless the other end has already been opened. Under Linux, opening a + // FIFO for read and write will succeed both in blocking and nonblocking + // mode. POSIX leaves this behavior undefined. This can be used to open a + // FIFO for writing while there are no readers available." - fifo(7) switch { case readable && writable: // Pipes opened for read-write always succeed without blocking. @@ -89,23 +95,26 @@ func (vp *VFSPipe) NewVFSPipeFD(ctx context.Context, vfsd *vfs.Dentry, vfsfd *vf case readable: newHandleLocked(&vp.rWakeup) - // If this pipe is being opened as nonblocking and there's no + // If this pipe is being opened as blocking and there's no // writer, we have to wait for a writer to open the other end. - if flags&linux.O_NONBLOCK == 0 && !vp.pipe.HasWriters() && !waitFor(&vp.mu, &vp.wWakeup, ctx) { + if vp.pipe.isNamed && statusFlags&linux.O_NONBLOCK == 0 && !vp.pipe.HasWriters() && !waitFor(&vp.mu, &vp.wWakeup, ctx) { + fd.DecRef() return nil, syserror.EINTR } case writable: newHandleLocked(&vp.wWakeup) - if !vp.pipe.HasReaders() { - // Nonblocking, write-only opens fail with ENXIO when - // the read side isn't open yet. - if flags&linux.O_NONBLOCK != 0 { + if vp.pipe.isNamed && !vp.pipe.HasReaders() { + // Non-blocking, write-only opens fail with ENXIO when the read + // side isn't open yet. + if statusFlags&linux.O_NONBLOCK != 0 { + fd.DecRef() return nil, syserror.ENXIO } // Wait for a reader to open the other end. if !waitFor(&vp.mu, &vp.rWakeup, ctx) { + fd.DecRef() return nil, syserror.EINTR } } @@ -114,96 +123,95 @@ func (vp *VFSPipe) NewVFSPipeFD(ctx context.Context, vfsd *vfs.Dentry, vfsfd *vf panic("invalid pipe flags: must be readable, writable, or both") } - return vfd, nil + return fd, nil } // Preconditions: vp.mu must be held. -func (vp *VFSPipe) open(vfsd *vfs.Dentry, vfsfd *vfs.FileDescription, flags uint32) (*VFSPipeFD, error) { - var fd VFSPipeFD - fd.flags = flags - fd.readable = vfs.MayReadFileWithOpenFlags(flags) - fd.writable = vfs.MayWriteFileWithOpenFlags(flags) - fd.vfsfd = vfsfd - fd.pipe = &vp.pipe +func (vp *VFSPipe) newFD(mnt *vfs.Mount, vfsd *vfs.Dentry, statusFlags uint32) *vfs.FileDescription { + fd := &VFSPipeFD{ + pipe: &vp.pipe, + } + fd.vfsfd.Init(fd, statusFlags, mnt, vfsd, &vfs.FileDescriptionOptions{ + DenyPRead: true, + DenyPWrite: true, + UseDentryMetadata: true, + }) switch { - case fd.readable && fd.writable: + case fd.vfsfd.IsReadable() && fd.vfsfd.IsWritable(): vp.pipe.rOpen() vp.pipe.wOpen() - case fd.readable: + case fd.vfsfd.IsReadable(): vp.pipe.rOpen() - case fd.writable: + case fd.vfsfd.IsWritable(): vp.pipe.wOpen() default: panic("invalid pipe flags: must be readable, writable, or both") } - return &fd, nil + return &fd.vfsfd } -// VFSPipeFD implements a subset of vfs.FileDescriptionImpl for pipes. It is -// expected that filesystesm will use this in a struct implementing -// vfs.FileDescriptionImpl. +// VFSPipeFD implements vfs.FileDescriptionImpl for pipes. It also implements +// non-atomic usermem.IO methods, allowing it to be passed as usermem.IO to +// other FileDescriptions for splice(2) and tee(2). type VFSPipeFD struct { - pipe *Pipe - flags uint32 - readable bool - writable bool - vfsfd *vfs.FileDescription + vfsfd vfs.FileDescription + vfs.FileDescriptionDefaultImpl + vfs.DentryMetadataFileDescriptionImpl + + pipe *Pipe } // Release implements vfs.FileDescriptionImpl.Release. func (fd *VFSPipeFD) Release() { var event waiter.EventMask - if fd.readable { + if fd.vfsfd.IsReadable() { fd.pipe.rClose() - event |= waiter.EventIn + event |= waiter.EventOut } - if fd.writable { + if fd.vfsfd.IsWritable() { fd.pipe.wClose() - event |= waiter.EventOut + event |= waiter.EventIn | waiter.EventHUp } if event == 0 { panic("invalid pipe flags: must be readable, writable, or both") } - if fd.writable { - fd.vfsfd.VirtualDentry().Mount().EndWrite() - } - fd.pipe.Notify(event) } -// OnClose implements vfs.FileDescriptionImpl.OnClose. -func (fd *VFSPipeFD) OnClose(_ context.Context) error { - return nil +// Readiness implements waiter.Waitable.Readiness. +func (fd *VFSPipeFD) Readiness(mask waiter.EventMask) waiter.EventMask { + switch { + case fd.vfsfd.IsReadable() && fd.vfsfd.IsWritable(): + return fd.pipe.rwReadiness() + case fd.vfsfd.IsReadable(): + return fd.pipe.rReadiness() + case fd.vfsfd.IsWritable(): + return fd.pipe.wReadiness() + default: + panic("pipe FD is neither readable nor writable") + } } -// PRead implements vfs.FileDescriptionImpl.PRead. -func (fd *VFSPipeFD) PRead(_ context.Context, _ usermem.IOSequence, _ int64, _ vfs.ReadOptions) (int64, error) { - return 0, syserror.ESPIPE +// EventRegister implements waiter.Waitable.EventRegister. +func (fd *VFSPipeFD) EventRegister(e *waiter.Entry, mask waiter.EventMask) { + fd.pipe.EventRegister(e, mask) +} + +// EventUnregister implements waiter.Waitable.EventUnregister. +func (fd *VFSPipeFD) EventUnregister(e *waiter.Entry) { + fd.pipe.EventUnregister(e) } // Read implements vfs.FileDescriptionImpl.Read. func (fd *VFSPipeFD) Read(ctx context.Context, dst usermem.IOSequence, _ vfs.ReadOptions) (int64, error) { - if !fd.readable { - return 0, syserror.EINVAL - } - return fd.pipe.Read(ctx, dst) } -// PWrite implements vfs.FileDescriptionImpl.PWrite. -func (fd *VFSPipeFD) PWrite(_ context.Context, _ usermem.IOSequence, _ int64, _ vfs.WriteOptions) (int64, error) { - return 0, syserror.ESPIPE -} - // Write implements vfs.FileDescriptionImpl.Write. func (fd *VFSPipeFD) Write(ctx context.Context, src usermem.IOSequence, _ vfs.WriteOptions) (int64, error) { - if !fd.writable { - return 0, syserror.EINVAL - } - return fd.pipe.Write(ctx, src) } @@ -211,3 +219,230 @@ func (fd *VFSPipeFD) Write(ctx context.Context, src usermem.IOSequence, _ vfs.Wr func (fd *VFSPipeFD) Ioctl(ctx context.Context, uio usermem.IO, args arch.SyscallArguments) (uintptr, error) { return fd.pipe.Ioctl(ctx, uio, args) } + +// PipeSize implements fcntl(F_GETPIPE_SZ). +func (fd *VFSPipeFD) PipeSize() int64 { + // Inline Pipe.FifoSize() rather than calling it with nil Context and + // fs.File and ignoring the returned error (which is always nil). + fd.pipe.mu.Lock() + defer fd.pipe.mu.Unlock() + return fd.pipe.max +} + +// SetPipeSize implements fcntl(F_SETPIPE_SZ). +func (fd *VFSPipeFD) SetPipeSize(size int64) (int64, error) { + return fd.pipe.SetFifoSize(size) +} + +// IOSequence returns a useremm.IOSequence that reads up to count bytes from, +// or writes up to count bytes to, fd. +func (fd *VFSPipeFD) IOSequence(count int64) usermem.IOSequence { + return usermem.IOSequence{ + IO: fd, + Addrs: usermem.AddrRangeSeqOf(usermem.AddrRange{0, usermem.Addr(count)}), + } +} + +// CopyIn implements usermem.IO.CopyIn. +func (fd *VFSPipeFD) CopyIn(ctx context.Context, addr usermem.Addr, dst []byte, opts usermem.IOOpts) (int, error) { + origCount := int64(len(dst)) + n, err := fd.pipe.read(ctx, readOps{ + left: func() int64 { + return int64(len(dst)) + }, + limit: func(l int64) { + dst = dst[:l] + }, + read: func(view *buffer.View) (int64, error) { + n, err := view.ReadAt(dst, 0) + view.TrimFront(int64(n)) + return int64(n), err + }, + }) + if n > 0 { + fd.pipe.Notify(waiter.EventOut) + } + if err == nil && n != origCount { + return int(n), syserror.ErrWouldBlock + } + return int(n), err +} + +// CopyOut implements usermem.IO.CopyOut. +func (fd *VFSPipeFD) CopyOut(ctx context.Context, addr usermem.Addr, src []byte, opts usermem.IOOpts) (int, error) { + origCount := int64(len(src)) + n, err := fd.pipe.write(ctx, writeOps{ + left: func() int64 { + return int64(len(src)) + }, + limit: func(l int64) { + src = src[:l] + }, + write: func(view *buffer.View) (int64, error) { + view.Append(src) + return int64(len(src)), nil + }, + }) + if n > 0 { + fd.pipe.Notify(waiter.EventIn) + } + if err == nil && n != origCount { + return int(n), syserror.ErrWouldBlock + } + return int(n), err +} + +// ZeroOut implements usermem.IO.ZeroOut. +func (fd *VFSPipeFD) ZeroOut(ctx context.Context, addr usermem.Addr, toZero int64, opts usermem.IOOpts) (int64, error) { + origCount := toZero + n, err := fd.pipe.write(ctx, writeOps{ + left: func() int64 { + return toZero + }, + limit: func(l int64) { + toZero = l + }, + write: func(view *buffer.View) (int64, error) { + view.Grow(view.Size()+toZero, true /* zero */) + return toZero, nil + }, + }) + if n > 0 { + fd.pipe.Notify(waiter.EventIn) + } + if err == nil && n != origCount { + return n, syserror.ErrWouldBlock + } + return n, err +} + +// CopyInTo implements usermem.IO.CopyInTo. +func (fd *VFSPipeFD) CopyInTo(ctx context.Context, ars usermem.AddrRangeSeq, dst safemem.Writer, opts usermem.IOOpts) (int64, error) { + count := ars.NumBytes() + if count == 0 { + return 0, nil + } + origCount := count + n, err := fd.pipe.read(ctx, readOps{ + left: func() int64 { + return count + }, + limit: func(l int64) { + count = l + }, + read: func(view *buffer.View) (int64, error) { + n, err := view.ReadToSafememWriter(dst, uint64(count)) + view.TrimFront(int64(n)) + return int64(n), err + }, + }) + if n > 0 { + fd.pipe.Notify(waiter.EventOut) + } + if err == nil && n != origCount { + return n, syserror.ErrWouldBlock + } + return n, err +} + +// CopyOutFrom implements usermem.IO.CopyOutFrom. +func (fd *VFSPipeFD) CopyOutFrom(ctx context.Context, ars usermem.AddrRangeSeq, src safemem.Reader, opts usermem.IOOpts) (int64, error) { + count := ars.NumBytes() + if count == 0 { + return 0, nil + } + origCount := count + n, err := fd.pipe.write(ctx, writeOps{ + left: func() int64 { + return count + }, + limit: func(l int64) { + count = l + }, + write: func(view *buffer.View) (int64, error) { + n, err := view.WriteFromSafememReader(src, uint64(count)) + return int64(n), err + }, + }) + if n > 0 { + fd.pipe.Notify(waiter.EventIn) + } + if err == nil && n != origCount { + return n, syserror.ErrWouldBlock + } + return n, err +} + +// SwapUint32 implements usermem.IO.SwapUint32. +func (fd *VFSPipeFD) SwapUint32(ctx context.Context, addr usermem.Addr, new uint32, opts usermem.IOOpts) (uint32, error) { + // How did a pipe get passed as the virtual address space to futex(2)? + panic("VFSPipeFD.SwapUint32 called unexpectedly") +} + +// CompareAndSwapUint32 implements usermem.IO.CompareAndSwapUint32. +func (fd *VFSPipeFD) CompareAndSwapUint32(ctx context.Context, addr usermem.Addr, old, new uint32, opts usermem.IOOpts) (uint32, error) { + panic("VFSPipeFD.CompareAndSwapUint32 called unexpectedly") +} + +// LoadUint32 implements usermem.IO.LoadUint32. +func (fd *VFSPipeFD) LoadUint32(ctx context.Context, addr usermem.Addr, opts usermem.IOOpts) (uint32, error) { + panic("VFSPipeFD.LoadUint32 called unexpectedly") +} + +// Splice reads up to count bytes from src and writes them to dst. It returns +// the number of bytes moved. +// +// Preconditions: count > 0. +func Splice(ctx context.Context, dst, src *VFSPipeFD, count int64) (int64, error) { + return spliceOrTee(ctx, dst, src, count, true /* removeFromSrc */) +} + +// Tee reads up to count bytes from src and writes them to dst, without +// removing the read bytes from src. It returns the number of bytes copied. +// +// Preconditions: count > 0. +func Tee(ctx context.Context, dst, src *VFSPipeFD, count int64) (int64, error) { + return spliceOrTee(ctx, dst, src, count, false /* removeFromSrc */) +} + +// Preconditions: count > 0. +func spliceOrTee(ctx context.Context, dst, src *VFSPipeFD, count int64, removeFromSrc bool) (int64, error) { + if dst.pipe == src.pipe { + return 0, syserror.EINVAL + } + + lockTwoPipes(dst.pipe, src.pipe) + defer dst.pipe.mu.Unlock() + defer src.pipe.mu.Unlock() + + n, err := dst.pipe.writeLocked(ctx, writeOps{ + left: func() int64 { + return count + }, + limit: func(l int64) { + count = l + }, + write: func(dstView *buffer.View) (int64, error) { + return src.pipe.readLocked(ctx, readOps{ + left: func() int64 { + return count + }, + limit: func(l int64) { + count = l + }, + read: func(srcView *buffer.View) (int64, error) { + n, err := srcView.ReadToSafememWriter(dstView, uint64(count)) + if n > 0 && removeFromSrc { + srcView.TrimFront(int64(n)) + } + return int64(n), err + }, + }) + }, + }) + if n > 0 { + dst.pipe.Notify(waiter.EventIn) + src.pipe.Notify(waiter.EventOut) + } + return n, err +} diff --git a/pkg/sentry/kernel/ptrace.go b/pkg/sentry/kernel/ptrace.go index 35ad97d5d..e23e796ef 100644 --- a/pkg/sentry/kernel/ptrace.go +++ b/pkg/sentry/kernel/ptrace.go @@ -184,7 +184,6 @@ func (t *Task) CanTrace(target *Task, attach bool) bool { if targetCreds.PermittedCaps&^callerCreds.PermittedCaps != 0 { return false } - // TODO: Yama LSM return true } diff --git a/pkg/sentry/kernel/rseq.go b/pkg/sentry/kernel/rseq.go index efebfd872..18416643b 100644 --- a/pkg/sentry/kernel/rseq.go +++ b/pkg/sentry/kernel/rseq.go @@ -303,26 +303,14 @@ func (t *Task) rseqAddrInterrupt() { return } - buf = t.CopyScratchBuffer(linux.SizeOfRSeqCriticalSection) - if _, err := t.CopyInBytes(critAddr, buf); err != nil { + var cs linux.RSeqCriticalSection + if _, err := cs.CopyIn(t, critAddr); err != nil { t.Debugf("Failed to copy critical section from %#x for rseq: %v", critAddr, err) t.forceSignal(linux.SIGSEGV, false /* unconditional */) t.SendSignal(SignalInfoPriv(linux.SIGSEGV)) return } - // Manually marshal RSeqCriticalSection as this is in the hot path when - // rseq is enabled. It must be as fast as possible. - // - // TODO(b/130243041): Replace with go_marshal. - cs := linux.RSeqCriticalSection{ - Version: usermem.ByteOrder.Uint32(buf[0:4]), - Flags: usermem.ByteOrder.Uint32(buf[4:8]), - Start: usermem.ByteOrder.Uint64(buf[8:16]), - PostCommitOffset: usermem.ByteOrder.Uint64(buf[16:24]), - Abort: usermem.ByteOrder.Uint64(buf[24:32]), - } - if cs.Version != 0 { t.Debugf("Unknown version in %+v", cs) t.forceSignal(linux.SIGSEGV, false /* unconditional */) diff --git a/pkg/sentry/kernel/semaphore/semaphore.go b/pkg/sentry/kernel/semaphore/semaphore.go index 1000f3287..c00fa1138 100644 --- a/pkg/sentry/kernel/semaphore/semaphore.go +++ b/pkg/sentry/kernel/semaphore/semaphore.go @@ -554,6 +554,7 @@ func (s *sem) wakeWaiters() { for w := s.waiters.Front(); w != nil; { if s.value < w.value { // Still blocked, skip it. + w = w.Next() continue } w.ch <- struct{}{} diff --git a/pkg/sentry/kernel/sessions.go b/pkg/sentry/kernel/sessions.go index 047b5214d..0e19286de 100644 --- a/pkg/sentry/kernel/sessions.go +++ b/pkg/sentry/kernel/sessions.go @@ -246,7 +246,7 @@ func (pg *ProcessGroup) SendSignal(info *arch.SignalInfo) error { var lastErr error for tg := range tasks.Root.tgids { - if tg.ProcessGroup() == pg { + if tg.processGroup == pg { tg.signalHandlers.mu.Lock() infoCopy := *info if err := tg.leader.sendSignalLocked(&infoCopy, true /*group*/); err != nil { diff --git a/pkg/sentry/kernel/shm/shm.go b/pkg/sentry/kernel/shm/shm.go index 208569057..f66cfcc7f 100644 --- a/pkg/sentry/kernel/shm/shm.go +++ b/pkg/sentry/kernel/shm/shm.go @@ -461,7 +461,7 @@ func (s *Shm) AddMapping(ctx context.Context, _ memmap.MappingSpace, _ usermem.A func (s *Shm) RemoveMapping(ctx context.Context, _ memmap.MappingSpace, _ usermem.AddrRange, _ uint64, _ bool) { s.mu.Lock() defer s.mu.Unlock() - // TODO(b/38173783): RemoveMapping may be called during task exit, when ctx + // RemoveMapping may be called during task exit, when ctx // is context.Background. Gracefully handle missing clocks. Failing to // update the detach time in these cases is ok, since no one can observe the // omission. diff --git a/pkg/sentry/kernel/syscalls.go b/pkg/sentry/kernel/syscalls.go index 93c4fe969..413111faf 100644 --- a/pkg/sentry/kernel/syscalls.go +++ b/pkg/sentry/kernel/syscalls.go @@ -29,7 +29,7 @@ import ( // // The types below create fast lookup slices for all syscalls. This maximum // serves as a sanity check that we don't allocate huge slices for a very large -// syscall. +// syscall. This is checked during registration. const maxSyscallNum = 2000 // SyscallSupportLevel is a syscall support levels. @@ -209,65 +209,71 @@ type Stracer interface { // SyscallEnter is called on syscall entry. // // The returned private data is passed to SyscallExit. - // - // TODO(gvisor.dev/issue/155): remove kernel imports from the strace - // package so that the type can be used directly. SyscallEnter(t *Task, sysno uintptr, args arch.SyscallArguments, flags uint32) interface{} // SyscallExit is called on syscall exit. SyscallExit(context interface{}, t *Task, sysno, rval uintptr, err error) } -// SyscallTable is a lookup table of system calls. Critically, a SyscallTable -// is *immutable*. In order to make supporting suspend and resume sane, they -// must be uniquely registered and may not change during operation. +// SyscallTable is a lookup table of system calls. // -// +stateify savable +// Note that a SyscallTable is not savable directly. Instead, they are saved as +// an OS/Arch pair and lookup happens again on restore. type SyscallTable struct { // OS is the operating system that this syscall table implements. - OS abi.OS `state:"wait"` + OS abi.OS // Arch is the architecture that this syscall table targets. - Arch arch.Arch `state:"wait"` + Arch arch.Arch // The OS version that this syscall table implements. - Version Version `state:"manual"` + Version Version // AuditNumber is a numeric constant that represents the syscall table. If // non-zero, auditNumber must be one of the AUDIT_ARCH_* values defined by // linux/audit.h. - AuditNumber uint32 `state:"manual"` + AuditNumber uint32 // Table is the collection of functions. - Table map[uintptr]Syscall `state:"manual"` + Table map[uintptr]Syscall // lookup is a fixed-size array that holds the syscalls (indexed by // their numbers). It is used for fast look ups. - lookup []SyscallFn `state:"manual"` + lookup []SyscallFn // Emulate is a collection of instruction addresses to emulate. The // keys are addresses, and the values are system call numbers. - Emulate map[usermem.Addr]uintptr `state:"manual"` + Emulate map[usermem.Addr]uintptr // The function to call in case of a missing system call. - Missing MissingFn `state:"manual"` + Missing MissingFn // Stracer traces this syscall table. - Stracer Stracer `state:"manual"` + Stracer Stracer // External is used to handle an external callback. - External func(*Kernel) `state:"manual"` + External func(*Kernel) // ExternalFilterBefore is called before External is called before the syscall is executed. // External is not called if it returns false. - ExternalFilterBefore func(*Task, uintptr, arch.SyscallArguments) bool `state:"manual"` + ExternalFilterBefore func(*Task, uintptr, arch.SyscallArguments) bool // ExternalFilterAfter is called before External is called after the syscall is executed. // External is not called if it returns false. - ExternalFilterAfter func(*Task, uintptr, arch.SyscallArguments) bool `state:"manual"` + ExternalFilterAfter func(*Task, uintptr, arch.SyscallArguments) bool // FeatureEnable stores the strace and one-shot enable bits. - FeatureEnable SyscallFlagsTable `state:"manual"` + FeatureEnable SyscallFlagsTable +} + +// MaxSysno returns the largest system call number. +func (s *SyscallTable) MaxSysno() (max uintptr) { + for num := range s.Table { + if num > max { + max = num + } + } + return max } // allSyscallTables contains all known tables. @@ -290,6 +296,20 @@ func LookupSyscallTable(os abi.OS, a arch.Arch) (*SyscallTable, bool) { // RegisterSyscallTable registers a new syscall table for use by a Kernel. func RegisterSyscallTable(s *SyscallTable) { + if max := s.MaxSysno(); max > maxSyscallNum { + panic(fmt.Sprintf("SyscallTable %+v contains too large syscall number %d", s, max)) + } + if _, ok := LookupSyscallTable(s.OS, s.Arch); ok { + panic(fmt.Sprintf("Duplicate SyscallTable registered for OS %v Arch %v", s.OS, s.Arch)) + } + allSyscallTables = append(allSyscallTables, s) + s.Init() +} + +// Init initializes the system call table. +// +// This should normally be called only during registration. +func (s *SyscallTable) Init() { if s.Table == nil { // Ensure non-nil lookup table. s.Table = make(map[uintptr]Syscall) @@ -299,35 +319,16 @@ func RegisterSyscallTable(s *SyscallTable) { s.Emulate = make(map[usermem.Addr]uintptr) } - var max uintptr - for num := range s.Table { - if num > max { - max = num - } - } - - if max > maxSyscallNum { - panic(fmt.Sprintf("SyscallTable %+v contains too large syscall number %d", s, max)) - } - - s.lookup = make([]SyscallFn, max+1) + max := s.MaxSysno() // Checked during RegisterSyscallTable. // Initialize the fast-lookup table. + s.lookup = make([]SyscallFn, max+1) for num, sc := range s.Table { s.lookup[num] = sc.Fn } + // Initialize all features. s.FeatureEnable.init(s.Table, max) - - if _, ok := LookupSyscallTable(s.OS, s.Arch); ok { - panic(fmt.Sprintf("Duplicate SyscallTable registered for OS %v Arch %v", s.OS, s.Arch)) - } - - // Save a reference to this table. - // - // This is required for a Kernel to find the table and for save/restore - // operations below. - allSyscallTables = append(allSyscallTables, s) } // Lookup returns the syscall implementation, if one exists. diff --git a/pkg/sentry/kernel/syscalls_state.go b/pkg/sentry/kernel/syscalls_state.go index 00358326b..90f890495 100644 --- a/pkg/sentry/kernel/syscalls_state.go +++ b/pkg/sentry/kernel/syscalls_state.go @@ -14,16 +14,34 @@ package kernel -import "fmt" +import ( + "fmt" -// afterLoad is invoked by stateify. -func (s *SyscallTable) afterLoad() { - otherTable, ok := LookupSyscallTable(s.OS, s.Arch) - if !ok { - // Couldn't find a reference? - panic(fmt.Sprintf("syscall table not found for OS %v Arch %v", s.OS, s.Arch)) + "gvisor.dev/gvisor/pkg/abi" + "gvisor.dev/gvisor/pkg/sentry/arch" +) + +// syscallTableInfo is used to reload the SyscallTable. +// +// +stateify savable +type syscallTableInfo struct { + OS abi.OS + Arch arch.Arch +} + +// saveSt saves the SyscallTable. +func (tc *TaskContext) saveSt() syscallTableInfo { + return syscallTableInfo{ + OS: tc.st.OS, + Arch: tc.st.Arch, } +} - // Copy the table. - *s = *otherTable +// loadSt loads the SyscallTable. +func (tc *TaskContext) loadSt(sti syscallTableInfo) { + st, ok := LookupSyscallTable(sti.OS, sti.Arch) + if !ok { + panic(fmt.Sprintf("syscall table not found for OS %v, Arch %v", sti.OS, sti.Arch)) + } + tc.st = st // Save the table reference. } diff --git a/pkg/sentry/kernel/task.go b/pkg/sentry/kernel/task.go index 981e8c7fe..f48247c94 100644 --- a/pkg/sentry/kernel/task.go +++ b/pkg/sentry/kernel/task.go @@ -37,6 +37,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/usage" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" "gvisor.dev/gvisor/pkg/waiter" ) @@ -424,6 +425,11 @@ type Task struct { // abstractSockets is protected by mu. abstractSockets *AbstractSocketNamespace + // mountNamespaceVFS2 is the task's mount namespace. + // + // It is protected by mu. It is owned by the task goroutine. + mountNamespaceVFS2 *vfs.MountNamespace + // parentDeathSignal is sent to this task's thread group when its parent exits. // // parentDeathSignal is protected by mu. @@ -478,16 +484,13 @@ type Task struct { // bit. // // numaPolicy and numaNodeMask are protected by mu. - numaPolicy int32 + numaPolicy linux.NumaPolicy numaNodeMask uint64 - // If netns is true, the task is in a non-root network namespace. Network - // namespaces aren't currently implemented in full; being in a network - // namespace simply prevents the task from observing any network devices - // (including loopback) or using abstract socket addresses (see unix(7)). + // netns is the task's network namespace. netns is never nil. // - // netns is protected by mu. netns is owned by the task goroutine. - netns bool + // netns is protected by mu. + netns *inet.Namespace // If rseqPreempted is true, before the next call to p.Switch(), // interrupt rseq critical regions as defined by rseqAddr and @@ -638,6 +641,11 @@ func (t *Task) Value(key interface{}) interface{} { return int32(t.ThreadGroup().ID()) case fs.CtxRoot: return t.fsContext.RootDirectory() + case vfs.CtxRoot: + return t.fsContext.RootDirectoryVFS2() + case vfs.CtxMountNamespace: + t.mountNamespaceVFS2.IncRef() + return t.mountNamespaceVFS2 case fs.CtxDirentCacheLimiter: return t.k.DirentCacheLimiter case inet.CtxStack: @@ -701,6 +709,14 @@ func (t *Task) SyscallRestartBlock() SyscallRestartBlock { // Preconditions: The caller must be running on the task goroutine, or t.mu // must be locked. func (t *Task) IsChrooted() bool { + if VFS2Enabled { + realRoot := t.mountNamespaceVFS2.Root() + defer realRoot.DecRef() + root := t.fsContext.RootDirectoryVFS2() + defer root.DecRef() + return root != realRoot + } + realRoot := t.tg.mounts.Root() defer realRoot.DecRef() root := t.fsContext.RootDirectory() @@ -761,6 +777,15 @@ func (t *Task) NewFDs(fd int32, files []*fs.File, flags FDFlags) ([]int32, error return t.fdTable.NewFDs(t, fd, files, flags) } +// NewFDsVFS2 is a convenience wrapper for t.FDTable().NewFDsVFS2. +// +// This automatically passes the task as the context. +// +// Precondition: same as FDTable. +func (t *Task) NewFDsVFS2(fd int32, files []*vfs.FileDescription, flags FDFlags) ([]int32, error) { + return t.fdTable.NewFDsVFS2(t, fd, files, flags) +} + // NewFDFrom is a convenience wrapper for t.FDTable().NewFDs with a single file. // // This automatically passes the task as the context. @@ -774,6 +799,15 @@ func (t *Task) NewFDFrom(fd int32, file *fs.File, flags FDFlags) (int32, error) return fds[0], nil } +// NewFDFromVFS2 is a convenience wrapper for t.FDTable().NewFDVFS2. +// +// This automatically passes the task as the context. +// +// Precondition: same as FDTable.Get. +func (t *Task) NewFDFromVFS2(fd int32, file *vfs.FileDescription, flags FDFlags) (int32, error) { + return t.fdTable.NewFDVFS2(t, fd, file, flags) +} + // NewFDAt is a convenience wrapper for t.FDTable().NewFDAt. // // This automatically passes the task as the context. @@ -783,6 +817,15 @@ func (t *Task) NewFDAt(fd int32, file *fs.File, flags FDFlags) error { return t.fdTable.NewFDAt(t, fd, file, flags) } +// NewFDAtVFS2 is a convenience wrapper for t.FDTable().NewFDAtVFS2. +// +// This automatically passes the task as the context. +// +// Precondition: same as FDTable. +func (t *Task) NewFDAtVFS2(fd int32, file *vfs.FileDescription, flags FDFlags) error { + return t.fdTable.NewFDAtVFS2(t, fd, file, flags) +} + // WithMuLocked executes f with t.mu locked. func (t *Task) WithMuLocked(f func(*Task)) { t.mu.Lock() @@ -796,6 +839,15 @@ func (t *Task) MountNamespace() *fs.MountNamespace { return t.tg.mounts } +// MountNamespaceVFS2 returns t's MountNamespace. A reference is taken on the +// returned mount namespace. +func (t *Task) MountNamespaceVFS2() *vfs.MountNamespace { + t.mu.Lock() + defer t.mu.Unlock() + t.mountNamespaceVFS2.IncRef() + return t.mountNamespaceVFS2 +} + // AbstractSockets returns t's AbstractSocketNamespace. func (t *Task) AbstractSockets() *AbstractSocketNamespace { return t.abstractSockets @@ -805,3 +857,30 @@ func (t *Task) AbstractSockets() *AbstractSocketNamespace { func (t *Task) ContainerID() string { return t.containerID } + +// OOMScoreAdj gets the task's thread group's OOM score adjustment. +func (t *Task) OOMScoreAdj() int32 { + return atomic.LoadInt32(&t.tg.oomScoreAdj) +} + +// SetOOMScoreAdj sets the task's thread group's OOM score adjustment. The +// value should be between -1000 and 1000 inclusive. +func (t *Task) SetOOMScoreAdj(adj int32) error { + if adj > 1000 || adj < -1000 { + return syserror.EINVAL + } + atomic.StoreInt32(&t.tg.oomScoreAdj, adj) + return nil +} + +// UID returns t's uid. +// TODO(gvisor.dev/issue/170): This method is not namespaced yet. +func (t *Task) UID() uint32 { + return uint32(t.Credentials().EffectiveKUID) +} + +// GID returns t's gid. +// TODO(gvisor.dev/issue/170): This method is not namespaced yet. +func (t *Task) GID() uint32 { + return uint32(t.Credentials().EffectiveKGID) +} diff --git a/pkg/sentry/kernel/task_clone.go b/pkg/sentry/kernel/task_clone.go index 53d4d211b..e1ecca99e 100644 --- a/pkg/sentry/kernel/task_clone.go +++ b/pkg/sentry/kernel/task_clone.go @@ -15,8 +15,11 @@ package kernel import ( + "sync/atomic" + "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/bpf" + "gvisor.dev/gvisor/pkg/sentry/inet" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" ) @@ -54,8 +57,7 @@ type SharingOptions struct { NewUserNamespace bool // If NewNetworkNamespace is true, the task should have an independent - // network namespace. (Note that network namespaces are not really - // implemented; see comment on Task.netns for details.) + // network namespace. NewNetworkNamespace bool // If NewFiles is true, the task should use an independent file descriptor @@ -199,6 +201,17 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) { ipcns = NewIPCNamespace(userns) } + netns := t.NetworkNamespace() + if opts.NewNetworkNamespace { + netns = inet.NewNamespace(netns) + } + + // TODO(b/63601033): Implement CLONE_NEWNS. + mntnsVFS2 := t.mountNamespaceVFS2 + if mntnsVFS2 != nil { + mntnsVFS2.IncRef() + } + tc, err := t.tc.Fork(t, t.k, !opts.NewAddressSpace) if err != nil { return 0, nil, err @@ -241,12 +254,15 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) { rseqAddr := usermem.Addr(0) rseqSignature := uint32(0) if opts.NewThreadGroup { - tg.mounts.IncRef() + if tg.mounts != nil { + tg.mounts.IncRef() + } sh := t.tg.signalHandlers if opts.NewSignalHandlers { sh = sh.Fork() } tg = t.k.NewThreadGroup(tg.mounts, pidns, sh, opts.TerminationSignal, tg.limits.GetCopy()) + tg.oomScoreAdj = atomic.LoadInt32(&t.tg.oomScoreAdj) rseqAddr = t.rseqAddr rseqSignature = t.rseqSignature } @@ -260,11 +276,12 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) { FDTable: fdTable, Credentials: creds, Niceness: t.Niceness(), - NetworkNamespaced: t.netns, + NetworkNamespace: netns, AllowedCPUMask: t.CPUMask(), UTSNamespace: utsns, IPCNamespace: ipcns, AbstractSocketNamespace: t.abstractSockets, + MountNamespaceVFS2: mntnsVFS2, RSeqAddr: rseqAddr, RSeqSignature: rseqSignature, ContainerID: t.ContainerID(), @@ -274,9 +291,6 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) { } else { cfg.InheritParent = t } - if opts.NewNetworkNamespace { - cfg.NetworkNamespaced = true - } nt, err := t.tg.pidns.owner.NewTask(cfg) if err != nil { if opts.NewThreadGroup { @@ -473,7 +487,7 @@ func (t *Task) Unshare(opts *SharingOptions) error { t.mu.Unlock() return syserror.EPERM } - t.netns = true + t.netns = inet.NewNamespace(t.netns) } if opts.NewUTSNamespace { if !haveCapSysAdmin { diff --git a/pkg/sentry/kernel/task_context.go b/pkg/sentry/kernel/task_context.go index 2d6e7733c..9fa528384 100644 --- a/pkg/sentry/kernel/task_context.go +++ b/pkg/sentry/kernel/task_context.go @@ -49,7 +49,7 @@ type TaskContext struct { fu *futex.Manager // st is the task's syscall table. - st *SyscallTable + st *SyscallTable `state:".(syscallTableInfo)"` } // release releases all resources held by the TaskContext. release is called by @@ -58,7 +58,6 @@ func (tc *TaskContext) release() { // Nil out pointers so that if the task is saved after release, it doesn't // follow the pointers to possibly now-invalid objects. if tc.MemoryManager != nil { - // TODO(b/38173783) tc.MemoryManager.DecUsers(context.Background()) tc.MemoryManager = nil } @@ -136,11 +135,11 @@ func (t *Task) Stack() *arch.Stack { func (k *Kernel) LoadTaskImage(ctx context.Context, args loader.LoadArgs) (*TaskContext, *syserr.Error) { // If File is not nil, we should load that instead of resolving Filename. if args.File != nil { - args.Filename = args.File.MappedName(ctx) + args.Filename = args.File.PathnameWithDeleted(ctx) } // Prepare a new user address space to load into. - m := mm.NewMemoryManager(k, k) + m := mm.NewMemoryManager(k, k, k.SleepForAddressSpaceActivation) defer m.DecUsers(ctx) args.MemoryManager = m diff --git a/pkg/sentry/kernel/task_exec.go b/pkg/sentry/kernel/task_exec.go index 8f57a34a6..9b69f3cbe 100644 --- a/pkg/sentry/kernel/task_exec.go +++ b/pkg/sentry/kernel/task_exec.go @@ -198,6 +198,10 @@ func (r *runSyscallAfterExecStop) execute(t *Task) taskRunState { t.tg.oldRSeqCritical.Store(&OldRSeqCriticalRegion{}) t.tg.pidns.owner.mu.Unlock() + oldFDTable := t.fdTable + t.fdTable = t.fdTable.Fork() + oldFDTable.DecRef() + // Remove FDs with the CloseOnExec flag set. t.fdTable.RemoveIf(func(_ *fs.File, _ *vfs.FileDescription, flags FDFlags) bool { return flags.CloseOnExec @@ -220,7 +224,7 @@ func (r *runSyscallAfterExecStop) execute(t *Task) taskRunState { t.mu.Unlock() t.unstopVforkParent() // NOTE(b/30316266): All locks must be dropped prior to calling Activate. - t.MemoryManager().Activate() + t.MemoryManager().Activate(t) t.ptraceExec(oldTID) return (*runSyscallExit)(nil) diff --git a/pkg/sentry/kernel/task_exit.go b/pkg/sentry/kernel/task_exit.go index 435761e5a..c4ade6e8e 100644 --- a/pkg/sentry/kernel/task_exit.go +++ b/pkg/sentry/kernel/task_exit.go @@ -269,6 +269,13 @@ func (*runExitMain) execute(t *Task) taskRunState { t.fsContext.DecRef() t.fdTable.DecRef() + t.mu.Lock() + if t.mountNamespaceVFS2 != nil { + t.mountNamespaceVFS2.DecRef() + t.mountNamespaceVFS2 = nil + } + t.mu.Unlock() + // If this is the last task to exit from the thread group, release the // thread group's resources. if lastExiter { diff --git a/pkg/sentry/kernel/task_identity.go b/pkg/sentry/kernel/task_identity.go index ce3e6ef28..0325967e4 100644 --- a/pkg/sentry/kernel/task_identity.go +++ b/pkg/sentry/kernel/task_identity.go @@ -455,7 +455,7 @@ func (t *Task) SetKeepCaps(k bool) { t.creds.Store(creds) } -// updateCredsForExec updates t.creds to reflect an execve(). +// updateCredsForExecLocked updates t.creds to reflect an execve(). // // NOTE(b/30815691): We currently do not implement privileged executables // (set-user/group-ID bits and file capabilities). This allows us to make a lot diff --git a/pkg/sentry/kernel/task_log.go b/pkg/sentry/kernel/task_log.go index 41259210c..eeccaa197 100644 --- a/pkg/sentry/kernel/task_log.go +++ b/pkg/sentry/kernel/task_log.go @@ -32,21 +32,21 @@ const ( // Infof logs an formatted info message by calling log.Infof. func (t *Task) Infof(fmt string, v ...interface{}) { if log.IsLogging(log.Info) { - log.Infof(t.logPrefix.Load().(string)+fmt, v...) + log.InfofAtDepth(1, t.logPrefix.Load().(string)+fmt, v...) } } // Warningf logs a warning string by calling log.Warningf. func (t *Task) Warningf(fmt string, v ...interface{}) { if log.IsLogging(log.Warning) { - log.Warningf(t.logPrefix.Load().(string)+fmt, v...) + log.WarningfAtDepth(1, t.logPrefix.Load().(string)+fmt, v...) } } // Debugf creates a debug string that includes the task ID. func (t *Task) Debugf(fmt string, v ...interface{}) { if log.IsLogging(log.Debug) { - log.Debugf(t.logPrefix.Load().(string)+fmt, v...) + log.DebugfAtDepth(1, t.logPrefix.Load().(string)+fmt, v...) } } @@ -198,18 +198,11 @@ func (t *Task) traceExecEvent(tc *TaskContext) { if !trace.IsEnabled() { return } - d := tc.MemoryManager.Executable() - if d == nil { + file := tc.MemoryManager.Executable() + if file == nil { trace.Logf(t.traceContext, traceCategory, "exec: << unknown >>") return } - defer d.DecRef() - root := t.fsContext.RootDirectory() - if root == nil { - trace.Logf(t.traceContext, traceCategory, "exec: << no root directory >>") - return - } - defer root.DecRef() - n, _ := d.FullName(root) - trace.Logf(t.traceContext, traceCategory, "exec: %s", n) + defer file.DecRef() + trace.Logf(t.traceContext, traceCategory, "exec: %s", file.PathnameWithDeleted(t)) } diff --git a/pkg/sentry/kernel/task_net.go b/pkg/sentry/kernel/task_net.go index 172a31e1d..f7711232c 100644 --- a/pkg/sentry/kernel/task_net.go +++ b/pkg/sentry/kernel/task_net.go @@ -22,14 +22,23 @@ import ( func (t *Task) IsNetworkNamespaced() bool { t.mu.Lock() defer t.mu.Unlock() - return t.netns + return !t.netns.IsRoot() } // NetworkContext returns the network stack used by the task. NetworkContext // may return nil if no network stack is available. +// +// TODO(gvisor.dev/issue/1833): Migrate callers of this method to +// NetworkNamespace(). func (t *Task) NetworkContext() inet.Stack { - if t.IsNetworkNamespaced() { - return nil - } - return t.k.networkStack + t.mu.Lock() + defer t.mu.Unlock() + return t.netns.Stack() +} + +// NetworkNamespace returns the network namespace observed by the task. +func (t *Task) NetworkNamespace() *inet.Namespace { + t.mu.Lock() + defer t.mu.Unlock() + return t.netns } diff --git a/pkg/sentry/kernel/task_run.go b/pkg/sentry/kernel/task_run.go index 5568c91bc..d654dd997 100644 --- a/pkg/sentry/kernel/task_run.go +++ b/pkg/sentry/kernel/task_run.go @@ -96,6 +96,7 @@ func (t *Task) run(threadID uintptr) { t.tg.liveGoroutines.Done() t.tg.pidns.owner.liveGoroutines.Done() t.tg.pidns.owner.runningGoroutines.Done() + t.p.Release() // Keep argument alive because stack trace for dead variables may not be correct. runtime.KeepAlive(threadID) @@ -126,13 +127,39 @@ func (t *Task) doStop() { } } +func (*runApp) handleCPUIDInstruction(t *Task) error { + if len(arch.CPUIDInstruction) == 0 { + // CPUID emulation isn't supported, but this code can be + // executed, because the ptrace platform returns + // ErrContextSignalCPUID on page faults too. Look at + // pkg/sentry/platform/ptrace/ptrace.go:context.Switch for more + // details. + return platform.ErrContextSignal + } + // Is this a CPUID instruction? + region := trace.StartRegion(t.traceContext, cpuidRegion) + expected := arch.CPUIDInstruction[:] + found := make([]byte, len(expected)) + _, err := t.CopyIn(usermem.Addr(t.Arch().IP()), &found) + if err == nil && bytes.Equal(expected, found) { + // Skip the cpuid instruction. + t.Arch().CPUIDEmulate(t) + t.Arch().SetIP(t.Arch().IP() + uintptr(len(expected))) + region.End() + + return nil + } + region.End() // Not an actual CPUID, but required copy-in. + return platform.ErrContextSignal +} + // The runApp state checks for interrupts before executing untrusted // application code. // // +stateify savable type runApp struct{} -func (*runApp) execute(t *Task) taskRunState { +func (app *runApp) execute(t *Task) taskRunState { if t.interrupted() { // Checkpointing instructs tasks to stop by sending an interrupt, so we // must check for stops before entering runInterrupt (instead of @@ -237,21 +264,10 @@ func (*runApp) execute(t *Task) taskRunState { return (*runApp)(nil) case platform.ErrContextSignalCPUID: - // Is this a CPUID instruction? - region := trace.StartRegion(t.traceContext, cpuidRegion) - expected := arch.CPUIDInstruction[:] - found := make([]byte, len(expected)) - _, err := t.CopyIn(usermem.Addr(t.Arch().IP()), &found) - if err == nil && bytes.Equal(expected, found) { - // Skip the cpuid instruction. - t.Arch().CPUIDEmulate(t) - t.Arch().SetIP(t.Arch().IP() + uintptr(len(expected))) - region.End() - + if err := app.handleCPUIDInstruction(t); err == nil { // Resume execution. return (*runApp)(nil) } - region.End() // Not an actual CPUID, but required copy-in. // The instruction at the given RIP was not a CPUID, and we // fallthrough to the default signal deliver behavior below. @@ -338,7 +354,7 @@ func (*runApp) execute(t *Task) taskRunState { default: // What happened? Can't continue. t.Warningf("Unexpected SwitchToApp error: %v", err) - t.PrepareExit(ExitStatus{Code: t.ExtractErrno(err, -1)}) + t.PrepareExit(ExitStatus{Code: ExtractErrno(err, -1)}) return (*runExit)(nil) } } diff --git a/pkg/sentry/kernel/task_sched.go b/pkg/sentry/kernel/task_sched.go index 8b148db35..09366b60c 100644 --- a/pkg/sentry/kernel/task_sched.go +++ b/pkg/sentry/kernel/task_sched.go @@ -653,14 +653,14 @@ func (t *Task) SetNiceness(n int) { } // NumaPolicy returns t's current numa policy. -func (t *Task) NumaPolicy() (policy int32, nodeMask uint64) { +func (t *Task) NumaPolicy() (policy linux.NumaPolicy, nodeMask uint64) { t.mu.Lock() defer t.mu.Unlock() return t.numaPolicy, t.numaNodeMask } // SetNumaPolicy sets t's numa policy. -func (t *Task) SetNumaPolicy(policy int32, nodeMask uint64) { +func (t *Task) SetNumaPolicy(policy linux.NumaPolicy, nodeMask uint64) { t.mu.Lock() defer t.mu.Unlock() t.numaPolicy = policy diff --git a/pkg/sentry/kernel/task_signals.go b/pkg/sentry/kernel/task_signals.go index 8802db142..79766cafe 100644 --- a/pkg/sentry/kernel/task_signals.go +++ b/pkg/sentry/kernel/task_signals.go @@ -174,7 +174,7 @@ func (t *Task) deliverSignal(info *arch.SignalInfo, act arch.SignalAct) taskRunS fallthrough case (sre == ERESTARTSYS && !act.IsRestart()): t.Debugf("Not restarting syscall %d after errno %d: interrupted by signal %d", t.Arch().SyscallNo(), sre, info.Signo) - t.Arch().SetReturn(uintptr(-t.ExtractErrno(syserror.EINTR, -1))) + t.Arch().SetReturn(uintptr(-ExtractErrno(syserror.EINTR, -1))) default: t.Debugf("Restarting syscall %d after errno %d: interrupted by signal %d", t.Arch().SyscallNo(), sre, info.Signo) t.Arch().RestartSyscall() @@ -263,6 +263,19 @@ func (t *Task) deliverSignalToHandler(info *arch.SignalInfo, act arch.SignalAct) if t.haveSavedSignalMask { mask = t.savedSignalMask } + + // Set up the restorer. + // x86-64 should always uses SA_RESTORER, but this flag is optional on other platforms. + // Please see the linux code as reference: + // linux/arch/x86/kernel/signal.c:__setup_rt_frame() + // If SA_RESTORER is not configured, we can use the sigreturn trampolines + // the vdso provides instead. + // Please see the linux code as reference: + // linux/arch/arm64/kernel/signal.c:setup_return() + if act.Flags&linux.SA_RESTORER == 0 { + act.Restorer = t.MemoryManager().VDSOSigReturn() + } + if err := t.Arch().SignalSetup(st, &act, info, &alt, mask); err != nil { return err } @@ -513,8 +526,6 @@ func (t *Task) canReceiveSignalLocked(sig linux.Signal) bool { if t.stop != nil { return false } - // - TODO(b/38173783): No special case for when t is also the sending task, - // because the identity of the sender is unknown. // - Do not choose tasks that have already been interrupted, as they may be // busy handling another signal. if len(t.interruptChan) != 0 { @@ -705,7 +716,7 @@ func (tg *ThreadGroup) SetSignalAct(sig linux.Signal, actptr *arch.SignalAct) (a func (t *Task) CopyOutSignalAct(addr usermem.Addr, s *arch.SignalAct) error { n := t.Arch().NewSignalAct() n.SerializeFrom(s) - _, err := t.CopyOut(addr, n) + _, err := n.CopyOut(t, addr) return err } @@ -714,7 +725,7 @@ func (t *Task) CopyOutSignalAct(addr usermem.Addr, s *arch.SignalAct) error { func (t *Task) CopyInSignalAct(addr usermem.Addr) (arch.SignalAct, error) { n := t.Arch().NewSignalAct() var s arch.SignalAct - if _, err := t.CopyIn(addr, n); err != nil { + if _, err := n.CopyIn(t, addr); err != nil { return s, err } n.DeserializeTo(&s) @@ -726,7 +737,7 @@ func (t *Task) CopyInSignalAct(addr usermem.Addr) (arch.SignalAct, error) { func (t *Task) CopyOutSignalStack(addr usermem.Addr, s *arch.SignalStack) error { n := t.Arch().NewSignalStack() n.SerializeFrom(s) - _, err := t.CopyOut(addr, n) + _, err := n.CopyOut(t, addr) return err } @@ -735,7 +746,7 @@ func (t *Task) CopyOutSignalStack(addr usermem.Addr, s *arch.SignalStack) error func (t *Task) CopyInSignalStack(addr usermem.Addr) (arch.SignalStack, error) { n := t.Arch().NewSignalStack() var s arch.SignalStack - if _, err := t.CopyIn(addr, n); err != nil { + if _, err := n.CopyIn(t, addr); err != nil { return s, err } n.DeserializeTo(&s) diff --git a/pkg/sentry/kernel/task_start.go b/pkg/sentry/kernel/task_start.go index de838beef..8485fb4b6 100644 --- a/pkg/sentry/kernel/task_start.go +++ b/pkg/sentry/kernel/task_start.go @@ -17,10 +17,12 @@ package kernel import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/inet" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/kernel/futex" "gvisor.dev/gvisor/pkg/sentry/kernel/sched" "gvisor.dev/gvisor/pkg/sentry/usage" + "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" ) @@ -64,9 +66,8 @@ type TaskConfig struct { // Niceness is the niceness of the new task. Niceness int - // If NetworkNamespaced is true, the new task should observe a non-root - // network namespace. - NetworkNamespaced bool + // NetworkNamespace is the network namespace to be used for the new task. + NetworkNamespace *inet.Namespace // AllowedCPUMask contains the cpus that this task can run on. AllowedCPUMask sched.CPUSet @@ -80,6 +81,9 @@ type TaskConfig struct { // AbstractSocketNamespace is the AbstractSocketNamespace of the new task. AbstractSocketNamespace *AbstractSocketNamespace + // MountNamespaceVFS2 is the MountNamespace of the new task. + MountNamespaceVFS2 *vfs.MountNamespace + // RSeqAddr is a pointer to the the userspace linux.RSeq structure. RSeqAddr usermem.Addr @@ -100,6 +104,9 @@ func (ts *TaskSet) NewTask(cfg *TaskConfig) (*Task, error) { cfg.TaskContext.release() cfg.FSContext.DecRef() cfg.FDTable.DecRef() + if cfg.MountNamespaceVFS2 != nil { + cfg.MountNamespaceVFS2.DecRef() + } return nil, err } return t, nil @@ -116,28 +123,29 @@ func (ts *TaskSet) newTask(cfg *TaskConfig) (*Task, error) { parent: cfg.Parent, children: make(map[*Task]struct{}), }, - runState: (*runApp)(nil), - interruptChan: make(chan struct{}, 1), - signalMask: cfg.SignalMask, - signalStack: arch.SignalStack{Flags: arch.SignalStackFlagDisable}, - tc: *tc, - fsContext: cfg.FSContext, - fdTable: cfg.FDTable, - p: cfg.Kernel.Platform.NewContext(), - k: cfg.Kernel, - ptraceTracees: make(map[*Task]struct{}), - allowedCPUMask: cfg.AllowedCPUMask.Copy(), - ioUsage: &usage.IO{}, - niceness: cfg.Niceness, - netns: cfg.NetworkNamespaced, - utsns: cfg.UTSNamespace, - ipcns: cfg.IPCNamespace, - abstractSockets: cfg.AbstractSocketNamespace, - rseqCPU: -1, - rseqAddr: cfg.RSeqAddr, - rseqSignature: cfg.RSeqSignature, - futexWaiter: futex.NewWaiter(), - containerID: cfg.ContainerID, + runState: (*runApp)(nil), + interruptChan: make(chan struct{}, 1), + signalMask: cfg.SignalMask, + signalStack: arch.SignalStack{Flags: arch.SignalStackFlagDisable}, + tc: *tc, + fsContext: cfg.FSContext, + fdTable: cfg.FDTable, + p: cfg.Kernel.Platform.NewContext(), + k: cfg.Kernel, + ptraceTracees: make(map[*Task]struct{}), + allowedCPUMask: cfg.AllowedCPUMask.Copy(), + ioUsage: &usage.IO{}, + niceness: cfg.Niceness, + netns: cfg.NetworkNamespace, + utsns: cfg.UTSNamespace, + ipcns: cfg.IPCNamespace, + abstractSockets: cfg.AbstractSocketNamespace, + mountNamespaceVFS2: cfg.MountNamespaceVFS2, + rseqCPU: -1, + rseqAddr: cfg.RSeqAddr, + rseqSignature: cfg.RSeqSignature, + futexWaiter: futex.NewWaiter(), + containerID: cfg.ContainerID, } t.creds.Store(cfg.Credentials) t.endStopCond.L = &t.tg.signalHandlers.mu diff --git a/pkg/sentry/kernel/task_syscall.go b/pkg/sentry/kernel/task_syscall.go index d555d69a8..a5903b0b5 100644 --- a/pkg/sentry/kernel/task_syscall.go +++ b/pkg/sentry/kernel/task_syscall.go @@ -194,6 +194,19 @@ func (t *Task) executeSyscall(sysno uintptr, args arch.SyscallArguments) (rval u // // The syscall path is very hot; avoid defer. func (t *Task) doSyscall() taskRunState { + // Save value of the register which is clobbered in the following + // t.Arch().SetReturn(-ENOSYS) operation. This is dedicated to arm64. + // + // On x86, register rax was shared by syscall number and return + // value, and at the entry of the syscall handler, the rax was + // saved to regs.orig_rax which was exposed to userspace. + // But on arm64, syscall number was passed through X8, and the X0 + // was shared by the first syscall argument and return value. The + // X0 was saved to regs.orig_x0 which was not exposed to userspace. + // So we have to do the same operation here to save the X0 value + // into the task context. + t.Arch().SyscallSaveOrig() + sysno := t.Arch().SyscallNo() args := t.Arch().SyscallArgs() @@ -269,6 +282,7 @@ func (*runSyscallAfterSyscallEnterStop) execute(t *Task) taskRunState { return (*runSyscallExit)(nil) } args := t.Arch().SyscallArgs() + return t.doSyscallInvoke(sysno, args) } @@ -298,7 +312,7 @@ func (t *Task) doSyscallInvoke(sysno uintptr, args arch.SyscallArguments) taskRu return ctrl.next } } else if err != nil { - t.Arch().SetReturn(uintptr(-t.ExtractErrno(err, int(sysno)))) + t.Arch().SetReturn(uintptr(-ExtractErrno(err, int(sysno)))) t.haveSyscallReturn = true } else { t.Arch().SetReturn(rval) @@ -417,7 +431,7 @@ func (t *Task) doVsyscallInvoke(sysno uintptr, args arch.SyscallArguments, calle // A return is not emulated in this case. return (*runApp)(nil) } - t.Arch().SetReturn(uintptr(-t.ExtractErrno(err, int(sysno)))) + t.Arch().SetReturn(uintptr(-ExtractErrno(err, int(sysno)))) } t.Arch().SetIP(t.Arch().Value(caller)) t.Arch().SetStack(t.Arch().Stack() + uintptr(t.Arch().Width())) @@ -427,7 +441,7 @@ func (t *Task) doVsyscallInvoke(sysno uintptr, args arch.SyscallArguments, calle // ExtractErrno extracts an integer error number from the error. // The syscall number is purely for context in the error case. Use -1 if // syscall number is unknown. -func (t *Task) ExtractErrno(err error, sysno int) int { +func ExtractErrno(err error, sysno int) int { switch err := err.(type) { case nil: return 0 @@ -441,11 +455,11 @@ func (t *Task) ExtractErrno(err error, sysno int) int { // handled (and the SIGBUS is delivered). return int(syscall.EFAULT) case *os.PathError: - return t.ExtractErrno(err.Err, sysno) + return ExtractErrno(err.Err, sysno) case *os.LinkError: - return t.ExtractErrno(err.Err, sysno) + return ExtractErrno(err.Err, sysno) case *os.SyscallError: - return t.ExtractErrno(err.Err, sysno) + return ExtractErrno(err.Err, sysno) default: if errno, ok := syserror.TranslateError(err); ok { return int(errno) diff --git a/pkg/sentry/kernel/task_usermem.go b/pkg/sentry/kernel/task_usermem.go index 2bf3ce8a8..b02044ad2 100644 --- a/pkg/sentry/kernel/task_usermem.go +++ b/pkg/sentry/kernel/task_usermem.go @@ -30,7 +30,7 @@ var MAX_RW_COUNT = int(usermem.Addr(math.MaxInt32).RoundDown()) // Activate ensures that the task has an active address space. func (t *Task) Activate() { if mm := t.MemoryManager(); mm != nil { - if err := mm.Activate(); err != nil { + if err := mm.Activate(t); err != nil { panic("unable to activate mm: " + err.Error()) } } diff --git a/pkg/sentry/kernel/thread_group.go b/pkg/sentry/kernel/thread_group.go index 768e958d2..52849f5b3 100644 --- a/pkg/sentry/kernel/thread_group.go +++ b/pkg/sentry/kernel/thread_group.go @@ -254,9 +254,16 @@ type ThreadGroup struct { // // tty is protected by the signal mutex. tty *TTY + + // oomScoreAdj is the thread group's OOM score adjustment. This is + // currently not used but is maintained for consistency. + // TODO(gvisor.dev/issue/1967) + // + // oomScoreAdj is accessed using atomic memory operations. + oomScoreAdj int32 } -// NewThreadGroup returns a new, empty thread group in PID namespace ns. The +// NewThreadGroup returns a new, empty thread group in PID namespace pidns. The // thread group leader will send its parent terminationSignal when it exits. // The new thread group isn't visible to the system until a task has been // created inside of it by a successful call to TaskSet.NewTask. @@ -317,7 +324,9 @@ func (tg *ThreadGroup) release() { for _, it := range its { it.DestroyTimer() } - tg.mounts.DecRef() + if tg.mounts != nil { + tg.mounts.DecRef() + } } // forEachChildThreadGroupLocked indicates over all child ThreadGroups. diff --git a/pkg/sentry/kernel/time/time.go b/pkg/sentry/kernel/time/time.go index 706de83ef..e959700f2 100644 --- a/pkg/sentry/kernel/time/time.go +++ b/pkg/sentry/kernel/time/time.go @@ -245,7 +245,7 @@ type Clock interface { type WallRateClock struct{} // WallTimeUntil implements Clock.WallTimeUntil. -func (WallRateClock) WallTimeUntil(t, now Time) time.Duration { +func (*WallRateClock) WallTimeUntil(t, now Time) time.Duration { return t.Sub(now) } @@ -254,16 +254,16 @@ func (WallRateClock) WallTimeUntil(t, now Time) time.Duration { type NoClockEvents struct{} // Readiness implements waiter.Waitable.Readiness. -func (NoClockEvents) Readiness(mask waiter.EventMask) waiter.EventMask { +func (*NoClockEvents) Readiness(mask waiter.EventMask) waiter.EventMask { return 0 } // EventRegister implements waiter.Waitable.EventRegister. -func (NoClockEvents) EventRegister(e *waiter.Entry, mask waiter.EventMask) { +func (*NoClockEvents) EventRegister(e *waiter.Entry, mask waiter.EventMask) { } // EventUnregister implements waiter.Waitable.EventUnregister. -func (NoClockEvents) EventUnregister(e *waiter.Entry) { +func (*NoClockEvents) EventUnregister(e *waiter.Entry) { } // ClockEventsQueue implements waiter.Waitable by wrapping waiter.Queue and @@ -273,7 +273,7 @@ type ClockEventsQueue struct { } // Readiness implements waiter.Waitable.Readiness. -func (ClockEventsQueue) Readiness(mask waiter.EventMask) waiter.EventMask { +func (*ClockEventsQueue) Readiness(mask waiter.EventMask) waiter.EventMask { return 0 } diff --git a/pkg/sentry/kernel/timekeeper.go b/pkg/sentry/kernel/timekeeper.go index dc99301de..da0ea7bb5 100644 --- a/pkg/sentry/kernel/timekeeper.go +++ b/pkg/sentry/kernel/timekeeper.go @@ -16,6 +16,7 @@ package kernel import ( "fmt" + "sync/atomic" "time" "gvisor.dev/gvisor/pkg/log" @@ -48,6 +49,9 @@ type Timekeeper struct { // It is set only once, by SetClocks. monotonicOffset int64 `state:"nosave"` + // monotonicLowerBound is the lowerBound for monotonic time. + monotonicLowerBound int64 `state:"nosave"` + // restored, if non-nil, indicates that this Timekeeper was restored // from a state file. The clocks are not set until restored is closed. restored chan struct{} `state:"nosave"` @@ -271,6 +275,21 @@ func (t *Timekeeper) GetTime(c sentrytime.ClockID) (int64, error) { now, err := t.clocks.GetTime(c) if err == nil && c == sentrytime.Monotonic { now += t.monotonicOffset + for { + // It's possible that the clock is shaky. This may be due to + // platform issues, e.g. the KVM platform relies on the guest + // TSC and host TSC, which may not be perfectly in sync. To + // work around this issue, ensure that the monotonic time is + // always bounded by the last time read. + oldLowerBound := atomic.LoadInt64(&t.monotonicLowerBound) + if now < oldLowerBound { + now = oldLowerBound + break + } + if atomic.CompareAndSwapInt64(&t.monotonicLowerBound, oldLowerBound, now) { + break + } + } } return now, err } diff --git a/pkg/sentry/loader/BUILD b/pkg/sentry/loader/BUILD index 23790378a..c6aa65f28 100644 --- a/pkg/sentry/loader/BUILD +++ b/pkg/sentry/loader/BUILD @@ -33,6 +33,7 @@ go_library( "//pkg/sentry/fs", "//pkg/sentry/fs/anon", "//pkg/sentry/fs/fsutil", + "//pkg/sentry/fsbridge", "//pkg/sentry/kernel/auth", "//pkg/sentry/limits", "//pkg/sentry/memmap", @@ -40,6 +41,7 @@ go_library( "//pkg/sentry/pgalloc", "//pkg/sentry/uniqueid", "//pkg/sentry/usage", + "//pkg/sentry/vfs", "//pkg/syserr", "//pkg/syserror", "//pkg/usermem", diff --git a/pkg/sentry/loader/elf.go b/pkg/sentry/loader/elf.go index 122ed05c2..616fafa2c 100644 --- a/pkg/sentry/loader/elf.go +++ b/pkg/sentry/loader/elf.go @@ -27,7 +27,7 @@ import ( "gvisor.dev/gvisor/pkg/cpuid" "gvisor.dev/gvisor/pkg/log" "gvisor.dev/gvisor/pkg/sentry/arch" - "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" "gvisor.dev/gvisor/pkg/sentry/limits" "gvisor.dev/gvisor/pkg/sentry/memmap" "gvisor.dev/gvisor/pkg/sentry/mm" @@ -97,11 +97,11 @@ type elfInfo struct { // accepts from the ELF, and it doesn't parse unnecessary parts of the file. // // ctx may be nil if f does not need it. -func parseHeader(ctx context.Context, f *fs.File) (elfInfo, error) { +func parseHeader(ctx context.Context, f fsbridge.File) (elfInfo, error) { // Check ident first; it will tell us the endianness of the rest of the // structs. var ident [elf.EI_NIDENT]byte - _, err := readFull(ctx, f, usermem.BytesIOSequence(ident[:]), 0) + _, err := f.ReadFull(ctx, usermem.BytesIOSequence(ident[:]), 0) if err != nil { log.Infof("Error reading ELF ident: %v", err) // The entire ident array always exists. @@ -137,7 +137,7 @@ func parseHeader(ctx context.Context, f *fs.File) (elfInfo, error) { var hdr elf.Header64 hdrBuf := make([]byte, header64Size) - _, err = readFull(ctx, f, usermem.BytesIOSequence(hdrBuf), 0) + _, err = f.ReadFull(ctx, usermem.BytesIOSequence(hdrBuf), 0) if err != nil { log.Infof("Error reading ELF header: %v", err) // The entire header always exists. @@ -187,7 +187,7 @@ func parseHeader(ctx context.Context, f *fs.File) (elfInfo, error) { } phdrBuf := make([]byte, totalPhdrSize) - _, err = readFull(ctx, f, usermem.BytesIOSequence(phdrBuf), int64(hdr.Phoff)) + _, err = f.ReadFull(ctx, usermem.BytesIOSequence(phdrBuf), int64(hdr.Phoff)) if err != nil { log.Infof("Error reading ELF phdrs: %v", err) // If phdrs were specified, they should all exist. @@ -227,7 +227,7 @@ func parseHeader(ctx context.Context, f *fs.File) (elfInfo, error) { // mapSegment maps a phdr into the Task. offset is the offset to apply to // phdr.Vaddr. -func mapSegment(ctx context.Context, m *mm.MemoryManager, f *fs.File, phdr *elf.ProgHeader, offset usermem.Addr) error { +func mapSegment(ctx context.Context, m *mm.MemoryManager, f fsbridge.File, phdr *elf.ProgHeader, offset usermem.Addr) error { // We must make a page-aligned mapping. adjust := usermem.Addr(phdr.Vaddr).PageOffset() @@ -395,7 +395,7 @@ type loadedELF struct { // // Preconditions: // * f is an ELF file -func loadParsedELF(ctx context.Context, m *mm.MemoryManager, f *fs.File, info elfInfo, sharedLoadOffset usermem.Addr) (loadedELF, error) { +func loadParsedELF(ctx context.Context, m *mm.MemoryManager, f fsbridge.File, info elfInfo, sharedLoadOffset usermem.Addr) (loadedELF, error) { first := true var start, end usermem.Addr var interpreter string @@ -431,7 +431,7 @@ func loadParsedELF(ctx context.Context, m *mm.MemoryManager, f *fs.File, info el } path := make([]byte, phdr.Filesz) - _, err := readFull(ctx, f, usermem.BytesIOSequence(path), int64(phdr.Off)) + _, err := f.ReadFull(ctx, usermem.BytesIOSequence(path), int64(phdr.Off)) if err != nil { // If an interpreter was specified, it should exist. ctx.Infof("Error reading PT_INTERP path: %v", err) @@ -564,7 +564,7 @@ func loadParsedELF(ctx context.Context, m *mm.MemoryManager, f *fs.File, info el // Preconditions: // * f is an ELF file // * f is the first ELF loaded into m -func loadInitialELF(ctx context.Context, m *mm.MemoryManager, fs *cpuid.FeatureSet, f *fs.File) (loadedELF, arch.Context, error) { +func loadInitialELF(ctx context.Context, m *mm.MemoryManager, fs *cpuid.FeatureSet, f fsbridge.File) (loadedELF, arch.Context, error) { info, err := parseHeader(ctx, f) if err != nil { ctx.Infof("Failed to parse initial ELF: %v", err) @@ -602,7 +602,7 @@ func loadInitialELF(ctx context.Context, m *mm.MemoryManager, fs *cpuid.FeatureS // // Preconditions: // * f is an ELF file -func loadInterpreterELF(ctx context.Context, m *mm.MemoryManager, f *fs.File, initial loadedELF) (loadedELF, error) { +func loadInterpreterELF(ctx context.Context, m *mm.MemoryManager, f fsbridge.File, initial loadedELF) (loadedELF, error) { info, err := parseHeader(ctx, f) if err != nil { if err == syserror.ENOEXEC { @@ -649,16 +649,14 @@ func loadELF(ctx context.Context, args LoadArgs) (loadedELF, arch.Context, error // Refresh the traversal limit. *args.RemainingTraversals = linux.MaxSymlinkTraversals args.Filename = bin.interpreter - d, i, err := openPath(ctx, args) + intFile, err := openPath(ctx, args) if err != nil { ctx.Infof("Error opening interpreter %s: %v", bin.interpreter, err) return loadedELF{}, nil, err } - defer i.DecRef() - // We don't need the Dirent. - d.DecRef() + defer intFile.DecRef() - interp, err = loadInterpreterELF(ctx, args.MemoryManager, i, bin) + interp, err = loadInterpreterELF(ctx, args.MemoryManager, intFile, bin) if err != nil { ctx.Infof("Error loading interpreter: %v", err) return loadedELF{}, nil, err diff --git a/pkg/sentry/loader/interpreter.go b/pkg/sentry/loader/interpreter.go index 098a45d36..3886b4d33 100644 --- a/pkg/sentry/loader/interpreter.go +++ b/pkg/sentry/loader/interpreter.go @@ -19,7 +19,7 @@ import ( "io" "gvisor.dev/gvisor/pkg/context" - "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" ) @@ -37,9 +37,9 @@ const ( ) // parseInterpreterScript returns the interpreter path and argv. -func parseInterpreterScript(ctx context.Context, filename string, f *fs.File, argv []string) (newpath string, newargv []string, err error) { +func parseInterpreterScript(ctx context.Context, filename string, f fsbridge.File, argv []string) (newpath string, newargv []string, err error) { line := make([]byte, interpMaxLineLength) - n, err := readFull(ctx, f, usermem.BytesIOSequence(line), 0) + n, err := f.ReadFull(ctx, usermem.BytesIOSequence(line), 0) // Short read is OK. if err != nil && err != io.ErrUnexpectedEOF { if err == io.EOF { diff --git a/pkg/sentry/loader/loader.go b/pkg/sentry/loader/loader.go index 9a613d6b7..88449fe95 100644 --- a/pkg/sentry/loader/loader.go +++ b/pkg/sentry/loader/loader.go @@ -20,7 +20,6 @@ import ( "fmt" "io" "path" - "strings" "gvisor.dev/gvisor/pkg/abi" "gvisor.dev/gvisor/pkg/abi/linux" @@ -29,8 +28,10 @@ import ( "gvisor.dev/gvisor/pkg/rand" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/mm" + "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserr" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" @@ -41,16 +42,6 @@ type LoadArgs struct { // MemoryManager is the memory manager to load the executable into. MemoryManager *mm.MemoryManager - // Mounts is the mount namespace in which to look up Filename. - Mounts *fs.MountNamespace - - // Root is the root directory under which to look up Filename. - Root *fs.Dirent - - // WorkingDirectory is the working directory under which to look up - // Filename. - WorkingDirectory *fs.Dirent - // RemainingTraversals is the maximum number of symlinks to follow to // resolve Filename. This counter is passed by reference to keep it // updated throughout the call stack. @@ -65,7 +56,12 @@ type LoadArgs struct { // File is an open fs.File object of the executable. If File is not // nil, then File will be loaded and Filename will be ignored. - File *fs.File + // + // The caller is responsible for checking that the user can execute this file. + File fsbridge.File + + // Opener is used to open the executable file when 'File' is nil. + Opener fsbridge.Lookup // CloseOnExec indicates that the executable (or one of its parent // directories) was opened with O_CLOEXEC. If the executable is an @@ -106,103 +102,32 @@ func readFull(ctx context.Context, f *fs.File, dst usermem.IOSequence, offset in // installed in the Task FDTable. The caller takes ownership of both. // // args.Filename must be a readable, executable, regular file. -func openPath(ctx context.Context, args LoadArgs) (*fs.Dirent, *fs.File, error) { +func openPath(ctx context.Context, args LoadArgs) (fsbridge.File, error) { if args.Filename == "" { ctx.Infof("cannot open empty name") - return nil, nil, syserror.ENOENT - } - - var d *fs.Dirent - var err error - if args.ResolveFinal { - d, err = args.Mounts.FindInode(ctx, args.Root, args.WorkingDirectory, args.Filename, args.RemainingTraversals) - } else { - d, err = args.Mounts.FindLink(ctx, args.Root, args.WorkingDirectory, args.Filename, args.RemainingTraversals) - } - if err != nil { - return nil, nil, err - } - // Defer a DecRef for the sake of failure cases. - defer d.DecRef() - - if !args.ResolveFinal && fs.IsSymlink(d.Inode.StableAttr) { - return nil, nil, syserror.ELOOP - } - - if err := checkPermission(ctx, d); err != nil { - return nil, nil, err - } - - // If they claim it's a directory, then make sure. - // - // N.B. we reject directories below, but we must first reject - // non-directories passed as directories. - if strings.HasSuffix(args.Filename, "/") && !fs.IsDir(d.Inode.StableAttr) { - return nil, nil, syserror.ENOTDIR - } - - if err := checkIsRegularFile(ctx, d, args.Filename); err != nil { - return nil, nil, err - } - - f, err := d.Inode.GetFile(ctx, d, fs.FileFlags{Read: true}) - if err != nil { - return nil, nil, err - } - // Defer a DecRef for the sake of failure cases. - defer f.DecRef() - - if err := checkPread(ctx, f, args.Filename); err != nil { - return nil, nil, err - } - - d.IncRef() - f.IncRef() - return d, f, err -} - -// checkFile performs checks on a file to be executed. -func checkFile(ctx context.Context, f *fs.File, filename string) error { - if err := checkPermission(ctx, f.Dirent); err != nil { - return err + return nil, syserror.ENOENT } - if err := checkIsRegularFile(ctx, f.Dirent, filename); err != nil { - return err - } - - return checkPread(ctx, f, filename) -} - -// checkPermission checks whether the file is readable and executable. -func checkPermission(ctx context.Context, d *fs.Dirent) error { - perms := fs.PermMask{ - // TODO(gvisor.dev/issue/160): Linux requires only execute - // permission, not read. However, our backing filesystems may - // prevent us from reading the file without read permission. - // - // Additionally, a task with a non-readable executable has - // additional constraints on access via ptrace and procfs. - Read: true, - Execute: true, + // TODO(gvisor.dev/issue/160): Linux requires only execute permission, + // not read. However, our backing filesystems may prevent us from reading + // the file without read permission. Additionally, a task with a + // non-readable executable has additional constraints on access via + // ptrace and procfs. + opts := vfs.OpenOptions{ + Flags: linux.O_RDONLY, + FileExec: true, } - return d.Inode.CheckPermission(ctx, perms) + return args.Opener.OpenPath(ctx, args.Filename, opts, args.RemainingTraversals, args.ResolveFinal) } // checkIsRegularFile prevents us from trying to execute a directory, pipe, etc. -func checkIsRegularFile(ctx context.Context, d *fs.Dirent, filename string) error { - attr := d.Inode.StableAttr - if !fs.IsRegular(attr) { - ctx.Infof("%s is not regular: %v", filename, attr) - return syserror.EACCES +func checkIsRegularFile(ctx context.Context, file fsbridge.File, filename string) error { + t, err := file.Type(ctx) + if err != nil { + return err } - return nil -} - -// checkPread checks whether we can read the file at arbitrary offsets. -func checkPread(ctx context.Context, f *fs.File, filename string) error { - if !f.Flags().Pread { - ctx.Infof("%s cannot be read at an offset: %+v", filename, f.Flags()) + if t != linux.ModeRegular { + ctx.Infof("%q is not a regular file: %v", filename, t) return syserror.EACCES } return nil @@ -224,8 +149,10 @@ const ( maxLoaderAttempts = 6 ) -// loadExecutable loads an executable that is pointed to by args.File. If nil, -// the path args.Filename is resolved and loaded. If the executable is an +// loadExecutable loads an executable that is pointed to by args.File. The +// caller is responsible for checking that the user can execute this file. +// If nil, the path args.Filename is resolved and loaded (check that the user +// can execute this file is done here in this case). If the executable is an // interpreter script rather than an ELF, the binary of the corresponding // interpreter will be loaded. // @@ -234,37 +161,27 @@ const ( // * arch.Context matching the binary arch // * fs.Dirent of the binary file // * Possibly updated args.Argv -func loadExecutable(ctx context.Context, args LoadArgs) (loadedELF, arch.Context, *fs.Dirent, []string, error) { +func loadExecutable(ctx context.Context, args LoadArgs) (loadedELF, arch.Context, fsbridge.File, []string, error) { for i := 0; i < maxLoaderAttempts; i++ { - var ( - d *fs.Dirent - err error - ) if args.File == nil { - d, args.File, err = openPath(ctx, args) - // We will return d in the successful case, but defer a DecRef for the - // sake of intermediate loops and failure cases. - if d != nil { - defer d.DecRef() - } - if args.File != nil { - defer args.File.DecRef() + var err error + args.File, err = openPath(ctx, args) + if err != nil { + ctx.Infof("Error opening %s: %v", args.Filename, err) + return loadedELF{}, nil, nil, nil, err } + // Ensure file is release in case the code loops or errors out. + defer args.File.DecRef() } else { - d = args.File.Dirent - d.IncRef() - defer d.DecRef() - err = checkFile(ctx, args.File, args.Filename) - } - if err != nil { - ctx.Infof("Error opening %s: %v", args.Filename, err) - return loadedELF{}, nil, nil, nil, err + if err := checkIsRegularFile(ctx, args.File, args.Filename); err != nil { + return loadedELF{}, nil, nil, nil, err + } } // Check the header. Is this an ELF or interpreter script? var hdr [4]uint8 // N.B. We assume that reading from a regular file cannot block. - _, err = readFull(ctx, args.File, usermem.BytesIOSequence(hdr[:]), 0) + _, err := args.File.ReadFull(ctx, usermem.BytesIOSequence(hdr[:]), 0) // Allow unexpected EOF, as a valid executable could be only three bytes // (e.g., #!a). if err != nil && err != io.ErrUnexpectedEOF { @@ -281,9 +198,10 @@ func loadExecutable(ctx context.Context, args LoadArgs) (loadedELF, arch.Context ctx.Infof("Error loading ELF: %v", err) return loadedELF{}, nil, nil, nil, err } - // An ELF is always terminal. Hold on to d. - d.IncRef() - return loaded, ac, d, args.Argv, err + // An ELF is always terminal. Hold on to file. + args.File.IncRef() + return loaded, ac, args.File, args.Argv, err + case bytes.Equal(hdr[:2], []byte(interpreterScriptMagic)): if args.CloseOnExec { return loadedELF{}, nil, nil, nil, syserror.ENOENT @@ -295,6 +213,7 @@ func loadExecutable(ctx context.Context, args LoadArgs) (loadedELF, arch.Context } // Refresh the traversal limit for the interpreter. *args.RemainingTraversals = linux.MaxSymlinkTraversals + default: ctx.Infof("Unknown magic: %v", hdr) return loadedELF{}, nil, nil, nil, syserror.ENOEXEC @@ -317,11 +236,11 @@ func loadExecutable(ctx context.Context, args LoadArgs) (loadedELF, arch.Context // * Load is called on the Task goroutine. func Load(ctx context.Context, args LoadArgs, extraAuxv []arch.AuxEntry, vdso *VDSO) (abi.OS, arch.Context, string, *syserr.Error) { // Load the executable itself. - loaded, ac, d, newArgv, err := loadExecutable(ctx, args) + loaded, ac, file, newArgv, err := loadExecutable(ctx, args) if err != nil { return 0, nil, "", syserr.NewDynamic(fmt.Sprintf("Failed to load %s: %v", args.Filename, err), syserr.FromError(err).ToLinux()) } - defer d.DecRef() + defer file.DecRef() // Load the VDSO. vdsoAddr, err := loadVDSO(ctx, args.MemoryManager, vdso, loaded) @@ -390,7 +309,16 @@ func Load(ctx context.Context, args LoadArgs, extraAuxv []arch.AuxEntry, vdso *V m.SetEnvvStart(sl.EnvvStart) m.SetEnvvEnd(sl.EnvvEnd) m.SetAuxv(auxv) - m.SetExecutable(d) + m.SetExecutable(file) + + symbolValue, err := getSymbolValueFromVDSO("rt_sigreturn") + if err != nil { + return 0, nil, "", syserr.NewDynamic(fmt.Sprintf("Failed to find rt_sigreturn in vdso: %v", err), syserr.FromError(err).ToLinux()) + } + + // Found rt_sigretrun. + addr := uint64(vdsoAddr) + symbolValue - vdsoPrelink + m.SetVDSOSigReturn(addr) ac.SetIP(uintptr(loaded.entry)) ac.SetStack(uintptr(stack.Bottom)) diff --git a/pkg/sentry/loader/vdso.go b/pkg/sentry/loader/vdso.go index 52f446ed7..00977fc08 100644 --- a/pkg/sentry/loader/vdso.go +++ b/pkg/sentry/loader/vdso.go @@ -15,9 +15,11 @@ package loader import ( + "bytes" "debug/elf" "fmt" "io" + "strings" "gvisor.dev/gvisor/pkg/abi" "gvisor.dev/gvisor/pkg/context" @@ -27,6 +29,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fs/anon" "gvisor.dev/gvisor/pkg/sentry/fs/fsutil" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" "gvisor.dev/gvisor/pkg/sentry/memmap" "gvisor.dev/gvisor/pkg/sentry/mm" "gvisor.dev/gvisor/pkg/sentry/pgalloc" @@ -37,6 +40,8 @@ import ( "gvisor.dev/gvisor/pkg/waiter" ) +const vdsoPrelink = 0xffffffffff700000 + type fileContext struct { context.Context } @@ -69,6 +74,8 @@ type byteReader struct { var _ fs.FileOperations = (*byteReader)(nil) // newByteReaderFile creates a fake file to read data from. +// +// TODO(gvisor.dev/issue/1623): Convert to VFS2. func newByteReaderFile(ctx context.Context, data []byte) *fs.File { // Create a fake inode. inode := fs.NewInode( @@ -123,7 +130,7 @@ func (b *byteReader) Write(ctx context.Context, file *fs.File, src usermem.IOSeq // * PT_LOAD segments don't extend beyond the end of the file. // // ctx may be nil if f does not need it. -func validateVDSO(ctx context.Context, f *fs.File, size uint64) (elfInfo, error) { +func validateVDSO(ctx context.Context, f fsbridge.File, size uint64) (elfInfo, error) { info, err := parseHeader(ctx, f) if err != nil { log.Infof("Unable to parse VDSO header: %v", err) @@ -218,10 +225,31 @@ type VDSO struct { phdrs []elf.ProgHeader `state:".([]elfProgHeader)"` } +// getSymbolValueFromVDSO returns the specific symbol value in vdso.so. +func getSymbolValueFromVDSO(symbol string) (uint64, error) { + f, err := elf.NewFile(bytes.NewReader(vdsoBin)) + if err != nil { + return 0, err + } + syms, err := f.Symbols() + if err != nil { + return 0, err + } + + for _, sym := range syms { + if elf.ST_BIND(sym.Info) != elf.STB_LOCAL && sym.Section != elf.SHN_UNDEF { + if strings.Contains(sym.Name, symbol) { + return sym.Value, nil + } + } + } + return 0, fmt.Errorf("no %v in vdso.so", symbol) +} + // PrepareVDSO validates the system VDSO and returns a VDSO, containing the // param page for updating by the kernel. func PrepareVDSO(ctx context.Context, mfp pgalloc.MemoryFileProvider) (*VDSO, error) { - vdsoFile := newByteReaderFile(ctx, vdsoBin) + vdsoFile := fsbridge.NewFSFile(newByteReaderFile(ctx, vdsoBin)) // First make sure the VDSO is valid. vdsoFile does not use ctx, so a // nil context can be passed. diff --git a/pkg/sentry/mm/BUILD b/pkg/sentry/mm/BUILD index e5729ced5..a036ce53c 100644 --- a/pkg/sentry/mm/BUILD +++ b/pkg/sentry/mm/BUILD @@ -25,6 +25,7 @@ go_template_instance( out = "vma_set.go", consts = { "minDegree": "8", + "trackGaps": "1", }, imports = { "usermem": "gvisor.dev/gvisor/pkg/usermem", @@ -105,8 +106,8 @@ go_library( "//pkg/safecopy", "//pkg/safemem", "//pkg/sentry/arch", - "//pkg/sentry/fs", "//pkg/sentry/fs/proc/seqfile", + "//pkg/sentry/fsbridge", "//pkg/sentry/kernel/auth", "//pkg/sentry/kernel/futex", "//pkg/sentry/kernel/shm", diff --git a/pkg/sentry/mm/address_space.go b/pkg/sentry/mm/address_space.go index e58a63deb..5c667117c 100644 --- a/pkg/sentry/mm/address_space.go +++ b/pkg/sentry/mm/address_space.go @@ -18,7 +18,7 @@ import ( "fmt" "sync/atomic" - "gvisor.dev/gvisor/pkg/atomicbitops" + "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/platform" "gvisor.dev/gvisor/pkg/usermem" ) @@ -39,11 +39,18 @@ func (mm *MemoryManager) AddressSpace() platform.AddressSpace { // // When this MemoryManager is no longer needed by a task, it should call // Deactivate to release the reference. -func (mm *MemoryManager) Activate() error { +func (mm *MemoryManager) Activate(ctx context.Context) error { // Fast path: the MemoryManager already has an active // platform.AddressSpace, and we just need to indicate that we need it too. - if atomicbitops.IncUnlessZeroInt32(&mm.active) { - return nil + for { + active := atomic.LoadInt32(&mm.active) + if active == 0 { + // Fall back to the slow path. + break + } + if atomic.CompareAndSwapInt32(&mm.active, active, active+1) { + return nil + } } for { @@ -85,16 +92,20 @@ func (mm *MemoryManager) Activate() error { if as == nil { // AddressSpace is unavailable, we must wait. // - // activeMu must not be held while waiting, as the user - // of the address space we are waiting on may attempt - // to take activeMu. - // - // Don't call UninterruptibleSleepStart to register the - // wait to allow the watchdog stuck task to trigger in - // case a process is starved waiting for the address - // space. + // activeMu must not be held while waiting, as the user of the address + // space we are waiting on may attempt to take activeMu. mm.activeMu.Unlock() + + sleep := mm.p.CooperativelySchedulesAddressSpace() && mm.sleepForActivation + if sleep { + // Mark this task sleeping while waiting for the address space to + // prevent the watchdog from reporting it as a stuck task. + ctx.UninterruptibleSleepStart(false) + } <-c + if sleep { + ctx.UninterruptibleSleepFinish(false) + } continue } @@ -118,8 +129,15 @@ func (mm *MemoryManager) Activate() error { func (mm *MemoryManager) Deactivate() { // Fast path: this is not the last goroutine to deactivate the // MemoryManager. - if atomicbitops.DecUnlessOneInt32(&mm.active) { - return + for { + active := atomic.LoadInt32(&mm.active) + if active == 1 { + // Fall back to the slow path. + break + } + if atomic.CompareAndSwapInt32(&mm.active, active, active-1) { + return + } } mm.activeMu.Lock() @@ -183,8 +201,10 @@ func (mm *MemoryManager) mapASLocked(pseg pmaIterator, ar usermem.AddrRange, pre if pma.needCOW { perms.Write = false } - if err := mm.as.MapFile(pmaMapAR.Start, pma.file, pseg.fileRangeOf(pmaMapAR), perms, precommit); err != nil { - return err + if perms.Any() { // MapFile precondition + if err := mm.as.MapFile(pmaMapAR.Start, pma.file, pseg.fileRangeOf(pmaMapAR), perms, precommit); err != nil { + return err + } } pseg = pseg.NextSegment() } diff --git a/pkg/sentry/mm/aio_context.go b/pkg/sentry/mm/aio_context.go index cb29d94b0..379148903 100644 --- a/pkg/sentry/mm/aio_context.go +++ b/pkg/sentry/mm/aio_context.go @@ -59,25 +59,27 @@ func (a *aioManager) newAIOContext(events uint32, id uint64) bool { } a.contexts[id] = &AIOContext{ - done: make(chan struct{}, 1), + requestReady: make(chan struct{}, 1), maxOutstanding: events, } return true } -// destroyAIOContext destroys an asynchronous I/O context. +// destroyAIOContext destroys an asynchronous I/O context. It doesn't wait for +// for pending requests to complete. Returns the destroyed AIOContext so it can +// be drained. // -// False is returned if the context does not exist. -func (a *aioManager) destroyAIOContext(id uint64) bool { +// Nil is returned if the context does not exist. +func (a *aioManager) destroyAIOContext(id uint64) *AIOContext { a.mu.Lock() defer a.mu.Unlock() ctx, ok := a.contexts[id] if !ok { - return false + return nil } delete(a.contexts, id) ctx.destroy() - return true + return ctx } // lookupAIOContext looks up the given context. @@ -102,8 +104,8 @@ type ioResult struct { // // +stateify savable type AIOContext struct { - // done is the notification channel used for all requests. - done chan struct{} `state:"nosave"` + // requestReady is the notification channel used for all requests. + requestReady chan struct{} `state:"nosave"` // mu protects below. mu sync.Mutex `state:"nosave"` @@ -129,8 +131,14 @@ func (ctx *AIOContext) destroy() { ctx.mu.Lock() defer ctx.mu.Unlock() ctx.dead = true - if ctx.outstanding == 0 { - close(ctx.done) + ctx.checkForDone() +} + +// Preconditions: ctx.mu must be held by caller. +func (ctx *AIOContext) checkForDone() { + if ctx.dead && ctx.outstanding == 0 { + close(ctx.requestReady) + ctx.requestReady = nil } } @@ -154,11 +162,12 @@ func (ctx *AIOContext) PopRequest() (interface{}, bool) { // Is there anything ready? if e := ctx.results.Front(); e != nil { - ctx.results.Remove(e) - ctx.outstanding-- - if ctx.outstanding == 0 && ctx.dead { - close(ctx.done) + if ctx.outstanding == 0 { + panic("AIOContext outstanding is going negative") } + ctx.outstanding-- + ctx.results.Remove(e) + ctx.checkForDone() return e.data, true } return nil, false @@ -172,26 +181,58 @@ func (ctx *AIOContext) FinishRequest(data interface{}) { // Push to the list and notify opportunistically. The channel notify // here is guaranteed to be safe because outstanding must be non-zero. - // The done channel is only closed when outstanding reaches zero. + // The requestReady channel is only closed when outstanding reaches zero. ctx.results.PushBack(&ioResult{data: data}) select { - case ctx.done <- struct{}{}: + case ctx.requestReady <- struct{}{}: default: } } // WaitChannel returns a channel that is notified when an AIO request is -// completed. -// -// The boolean return value indicates whether or not the context is active. -func (ctx *AIOContext) WaitChannel() (chan struct{}, bool) { +// completed. Returns nil if the context is destroyed and there are no more +// outstanding requests. +func (ctx *AIOContext) WaitChannel() chan struct{} { ctx.mu.Lock() defer ctx.mu.Unlock() - if ctx.outstanding == 0 && ctx.dead { - return nil, false + return ctx.requestReady +} + +// Dead returns true if the context has been destroyed. +func (ctx *AIOContext) Dead() bool { + ctx.mu.Lock() + defer ctx.mu.Unlock() + return ctx.dead +} + +// CancelPendingRequest forgets about a request that hasn't yet completed. +func (ctx *AIOContext) CancelPendingRequest() { + ctx.mu.Lock() + defer ctx.mu.Unlock() + + if ctx.outstanding == 0 { + panic("AIOContext outstanding is going negative") } - return ctx.done, true + ctx.outstanding-- + ctx.checkForDone() +} + +// Drain drops all completed requests. Pending requests remain untouched. +func (ctx *AIOContext) Drain() { + ctx.mu.Lock() + defer ctx.mu.Unlock() + + if ctx.outstanding == 0 { + return + } + size := uint32(ctx.results.Len()) + if ctx.outstanding < size { + panic("AIOContext outstanding is going negative") + } + ctx.outstanding -= size + ctx.results.Reset() + ctx.checkForDone() } // aioMappable implements memmap.MappingIdentity and memmap.Mappable for AIO @@ -332,9 +373,9 @@ func (mm *MemoryManager) NewAIOContext(ctx context.Context, events uint32) (uint Length: aioRingBufferSize, MappingIdentity: m, Mappable: m, - // TODO(fvoznika): Linux does "do_mmap_pgoff(..., PROT_READ | - // PROT_WRITE, ...)" in fs/aio.c:aio_setup_ring(); why do we make this - // mapping read-only? + // Linux uses "do_mmap_pgoff(..., PROT_READ | PROT_WRITE, ...)" in + // fs/aio.c:aio_setup_ring(). Since we don't implement AIO_RING_MAGIC, + // user mode should not write to this page. Perms: usermem.Read, MaxPerms: usermem.Read, }) @@ -349,11 +390,11 @@ func (mm *MemoryManager) NewAIOContext(ctx context.Context, events uint32) (uint return id, nil } -// DestroyAIOContext destroys an asynchronous I/O context. It returns false if -// the context does not exist. -func (mm *MemoryManager) DestroyAIOContext(ctx context.Context, id uint64) bool { +// DestroyAIOContext destroys an asynchronous I/O context. It returns the +// destroyed context. nil if the context does not exist. +func (mm *MemoryManager) DestroyAIOContext(ctx context.Context, id uint64) *AIOContext { if _, ok := mm.LookupAIOContext(ctx, id); !ok { - return false + return nil } // Only unmaps after it assured that the address is a valid aio context to diff --git a/pkg/sentry/mm/aio_context_state.go b/pkg/sentry/mm/aio_context_state.go index c37fc9f7b..3dabac1af 100644 --- a/pkg/sentry/mm/aio_context_state.go +++ b/pkg/sentry/mm/aio_context_state.go @@ -16,5 +16,5 @@ package mm // afterLoad is invoked by stateify. func (a *AIOContext) afterLoad() { - a.done = make(chan struct{}, 1) + a.requestReady = make(chan struct{}, 1) } diff --git a/pkg/sentry/mm/lifecycle.go b/pkg/sentry/mm/lifecycle.go index 47b8fbf43..aac56679b 100644 --- a/pkg/sentry/mm/lifecycle.go +++ b/pkg/sentry/mm/lifecycle.go @@ -18,7 +18,6 @@ import ( "fmt" "sync/atomic" - "gvisor.dev/gvisor/pkg/atomicbitops" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/limits" @@ -29,16 +28,17 @@ import ( ) // NewMemoryManager returns a new MemoryManager with no mappings and 1 user. -func NewMemoryManager(p platform.Platform, mfp pgalloc.MemoryFileProvider) *MemoryManager { +func NewMemoryManager(p platform.Platform, mfp pgalloc.MemoryFileProvider, sleepForActivation bool) *MemoryManager { return &MemoryManager{ - p: p, - mfp: mfp, - haveASIO: p.SupportsAddressSpaceIO(), - privateRefs: &privateRefs{}, - users: 1, - auxv: arch.Auxv{}, - dumpability: UserDumpable, - aioManager: aioManager{contexts: make(map[uint64]*AIOContext)}, + p: p, + mfp: mfp, + haveASIO: p.SupportsAddressSpaceIO(), + privateRefs: &privateRefs{}, + users: 1, + auxv: arch.Auxv{}, + dumpability: UserDumpable, + aioManager: aioManager{contexts: make(map[uint64]*AIOContext)}, + sleepForActivation: sleepForActivation, } } @@ -80,9 +80,11 @@ func (mm *MemoryManager) Fork(ctx context.Context) (*MemoryManager, error) { envv: mm.envv, auxv: append(arch.Auxv(nil), mm.auxv...), // IncRef'd below, once we know that there isn't an error. - executable: mm.executable, - dumpability: mm.dumpability, - aioManager: aioManager{contexts: make(map[uint64]*AIOContext)}, + executable: mm.executable, + dumpability: mm.dumpability, + aioManager: aioManager{contexts: make(map[uint64]*AIOContext)}, + sleepForActivation: mm.sleepForActivation, + vdsoSigReturnAddr: mm.vdsoSigReturnAddr, } // Copy vmas. @@ -229,7 +231,15 @@ func (mm *MemoryManager) Fork(ctx context.Context) (*MemoryManager, error) { // IncUsers increments mm's user count and returns true. If the user count is // already 0, IncUsers does nothing and returns false. func (mm *MemoryManager) IncUsers() bool { - return atomicbitops.IncUnlessZeroInt32(&mm.users) + for { + users := atomic.LoadInt32(&mm.users) + if users == 0 { + return false + } + if atomic.CompareAndSwapInt32(&mm.users, users, users+1) { + return true + } + } } // DecUsers decrements mm's user count. If the user count reaches 0, all diff --git a/pkg/sentry/mm/metadata.go b/pkg/sentry/mm/metadata.go index f550acae0..28e5057f7 100644 --- a/pkg/sentry/mm/metadata.go +++ b/pkg/sentry/mm/metadata.go @@ -16,7 +16,7 @@ package mm import ( "gvisor.dev/gvisor/pkg/sentry/arch" - "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" "gvisor.dev/gvisor/pkg/usermem" ) @@ -132,7 +132,7 @@ func (mm *MemoryManager) SetAuxv(auxv arch.Auxv) { // // An additional reference will be taken in the case of a non-nil executable, // which must be released by the caller. -func (mm *MemoryManager) Executable() *fs.Dirent { +func (mm *MemoryManager) Executable() fsbridge.File { mm.metadataMu.Lock() defer mm.metadataMu.Unlock() @@ -147,15 +147,15 @@ func (mm *MemoryManager) Executable() *fs.Dirent { // SetExecutable sets the executable. // // This takes a reference on d. -func (mm *MemoryManager) SetExecutable(d *fs.Dirent) { +func (mm *MemoryManager) SetExecutable(file fsbridge.File) { mm.metadataMu.Lock() // Grab a new reference. - d.IncRef() + file.IncRef() // Set the executable. orig := mm.executable - mm.executable = d + mm.executable = file mm.metadataMu.Unlock() @@ -167,3 +167,17 @@ func (mm *MemoryManager) SetExecutable(d *fs.Dirent) { orig.DecRef() } } + +// VDSOSigReturn returns the address of vdso_sigreturn. +func (mm *MemoryManager) VDSOSigReturn() uint64 { + mm.metadataMu.Lock() + defer mm.metadataMu.Unlock() + return mm.vdsoSigReturnAddr +} + +// SetVDSOSigReturn sets the address of vdso_sigreturn. +func (mm *MemoryManager) SetVDSOSigReturn(addr uint64) { + mm.metadataMu.Lock() + defer mm.metadataMu.Unlock() + mm.vdsoSigReturnAddr = addr +} diff --git a/pkg/sentry/mm/mm.go b/pkg/sentry/mm/mm.go index 09e582dd3..6db7c3d40 100644 --- a/pkg/sentry/mm/mm.go +++ b/pkg/sentry/mm/mm.go @@ -35,9 +35,10 @@ package mm import ( + "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/safemem" "gvisor.dev/gvisor/pkg/sentry/arch" - "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" "gvisor.dev/gvisor/pkg/sentry/memmap" "gvisor.dev/gvisor/pkg/sentry/pgalloc" "gvisor.dev/gvisor/pkg/sentry/platform" @@ -215,7 +216,7 @@ type MemoryManager struct { // is not nil, it holds a reference on the Dirent. // // executable is protected by metadataMu. - executable *fs.Dirent + executable fsbridge.File // dumpability describes if and how this MemoryManager may be dumped to // userspace. @@ -226,6 +227,14 @@ type MemoryManager struct { // aioManager keeps track of AIOContexts used for async IOs. AIOManager // must be cloned when CLONE_VM is used. aioManager aioManager + + // sleepForActivation indicates whether the task should report to be sleeping + // before trying to activate the address space. When set to true, delays in + // activation are not reported as stuck tasks by the watchdog. + sleepForActivation bool + + // vdsoSigReturnAddr is the address of 'vdso_sigreturn'. + vdsoSigReturnAddr uint64 } // vma represents a virtual memory area. @@ -278,7 +287,7 @@ type vma struct { mlockMode memmap.MLockMode // numaPolicy is the NUMA policy for this vma set by mbind(). - numaPolicy int32 + numaPolicy linux.NumaPolicy // numaNodemask is the NUMA nodemask for this vma set by mbind(). numaNodemask uint64 diff --git a/pkg/sentry/mm/mm_test.go b/pkg/sentry/mm/mm_test.go index edacca741..fdc308542 100644 --- a/pkg/sentry/mm/mm_test.go +++ b/pkg/sentry/mm/mm_test.go @@ -31,7 +31,7 @@ import ( func testMemoryManager(ctx context.Context) *MemoryManager { p := platform.FromContext(ctx) mfp := pgalloc.MemoryFileProviderFromContext(ctx) - mm := NewMemoryManager(p, mfp) + mm := NewMemoryManager(p, mfp, false) mm.layout = arch.MmapLayout{ MinAddr: p.MinUserAddress(), MaxAddr: p.MaxUserAddress(), diff --git a/pkg/sentry/mm/procfs.go b/pkg/sentry/mm/procfs.go index 1ab92f046..6efe5102b 100644 --- a/pkg/sentry/mm/procfs.go +++ b/pkg/sentry/mm/procfs.go @@ -148,7 +148,7 @@ func (mm *MemoryManager) appendVMAMapsEntryLocked(ctx context.Context, vseg vmaI // Do not include the guard page: fs/proc/task_mmu.c:show_map_vma() => // stack_guard_page_start(). - fmt.Fprintf(b, "%08x-%08x %s%s %08x %02x:%02x %d ", + lineLen, _ := fmt.Fprintf(b, "%08x-%08x %s%s %08x %02x:%02x %d ", vseg.Start(), vseg.End(), vma.realPerms, private, vma.off, devMajor, devMinor, ino) // Figure out our filename or hint. @@ -165,7 +165,7 @@ func (mm *MemoryManager) appendVMAMapsEntryLocked(ctx context.Context, vseg vmaI } if s != "" { // Per linux, we pad until the 74th character. - if pad := 73 - b.Len(); pad > 0 { + if pad := 73 - lineLen; pad > 0 { b.WriteString(strings.Repeat(" ", pad)) } b.WriteString(s) diff --git a/pkg/sentry/mm/syscalls.go b/pkg/sentry/mm/syscalls.go index c5dfa5972..3f496aa9f 100644 --- a/pkg/sentry/mm/syscalls.go +++ b/pkg/sentry/mm/syscalls.go @@ -974,7 +974,7 @@ func (mm *MemoryManager) MLockAll(ctx context.Context, opts MLockAllOpts) error } // NumaPolicy implements the semantics of Linux's get_mempolicy(MPOL_F_ADDR). -func (mm *MemoryManager) NumaPolicy(addr usermem.Addr) (int32, uint64, error) { +func (mm *MemoryManager) NumaPolicy(addr usermem.Addr) (linux.NumaPolicy, uint64, error) { mm.mappingMu.RLock() defer mm.mappingMu.RUnlock() vseg := mm.vmas.FindSegment(addr) @@ -986,7 +986,7 @@ func (mm *MemoryManager) NumaPolicy(addr usermem.Addr) (int32, uint64, error) { } // SetNumaPolicy implements the semantics of Linux's mbind(). -func (mm *MemoryManager) SetNumaPolicy(addr usermem.Addr, length uint64, policy int32, nodemask uint64) error { +func (mm *MemoryManager) SetNumaPolicy(addr usermem.Addr, length uint64, policy linux.NumaPolicy, nodemask uint64) error { if !addr.IsPageAligned() { return syserror.EINVAL } diff --git a/pkg/sentry/mm/vma.go b/pkg/sentry/mm/vma.go index 9a14e69e6..16d8207e9 100644 --- a/pkg/sentry/mm/vma.go +++ b/pkg/sentry/mm/vma.go @@ -195,7 +195,7 @@ func (mm *MemoryManager) applicationAddrRange() usermem.AddrRange { // Preconditions: mm.mappingMu must be locked. func (mm *MemoryManager) findLowestAvailableLocked(length, alignment uint64, bounds usermem.AddrRange) (usermem.Addr, error) { - for gap := mm.vmas.LowerBoundGap(bounds.Start); gap.Ok() && gap.Start() < bounds.End; gap = gap.NextGap() { + for gap := mm.vmas.LowerBoundGap(bounds.Start); gap.Ok() && gap.Start() < bounds.End; gap = gap.NextLargeEnoughGap(usermem.Addr(length)) { if gr := gap.availableRange().Intersect(bounds); uint64(gr.Length()) >= length { // Can we shift up to match the alignment? if offset := uint64(gr.Start) % alignment; offset != 0 { @@ -214,7 +214,7 @@ func (mm *MemoryManager) findLowestAvailableLocked(length, alignment uint64, bou // Preconditions: mm.mappingMu must be locked. func (mm *MemoryManager) findHighestAvailableLocked(length, alignment uint64, bounds usermem.AddrRange) (usermem.Addr, error) { - for gap := mm.vmas.UpperBoundGap(bounds.End); gap.Ok() && gap.End() > bounds.Start; gap = gap.PrevGap() { + for gap := mm.vmas.UpperBoundGap(bounds.End); gap.Ok() && gap.End() > bounds.Start; gap = gap.PrevLargeEnoughGap(usermem.Addr(length)) { if gr := gap.availableRange().Intersect(bounds); uint64(gr.Length()) >= length { // Can we shift down to match the alignment? start := gr.End - usermem.Addr(length) diff --git a/pkg/sentry/pgalloc/BUILD b/pkg/sentry/pgalloc/BUILD index 1eeb9f317..a9836ba71 100644 --- a/pkg/sentry/pgalloc/BUILD +++ b/pkg/sentry/pgalloc/BUILD @@ -33,6 +33,7 @@ go_template_instance( out = "usage_set.go", consts = { "minDegree": "10", + "trackGaps": "1", }, imports = { "platform": "gvisor.dev/gvisor/pkg/sentry/platform", @@ -48,6 +49,26 @@ go_template_instance( }, ) +go_template_instance( + name = "reclaim_set", + out = "reclaim_set.go", + consts = { + "minDegree": "10", + }, + imports = { + "platform": "gvisor.dev/gvisor/pkg/sentry/platform", + }, + package = "pgalloc", + prefix = "reclaim", + template = "//pkg/segment:generic_set", + types = { + "Key": "uint64", + "Range": "platform.FileRange", + "Value": "reclaimSetValue", + "Functions": "reclaimSetFunctions", + }, +) + go_library( name = "pgalloc", srcs = [ @@ -56,6 +77,7 @@ go_library( "evictable_range_set.go", "pgalloc.go", "pgalloc_unsafe.go", + "reclaim_set.go", "save_restore.go", "usage_set.go", ], diff --git a/pkg/sentry/pgalloc/pgalloc.go b/pkg/sentry/pgalloc/pgalloc.go index 577e9306a..c8d9facc2 100644 --- a/pkg/sentry/pgalloc/pgalloc.go +++ b/pkg/sentry/pgalloc/pgalloc.go @@ -108,12 +108,6 @@ type MemoryFile struct { usageSwapped uint64 usageLast time.Time - // minUnallocatedPage is the minimum page that may be unallocated. - // i.e., there are no unallocated pages below minUnallocatedPage. - // - // minUnallocatedPage is protected by mu. - minUnallocatedPage uint64 - // fileSize is the size of the backing memory file in bytes. fileSize is // always a power-of-two multiple of chunkSize. // @@ -146,11 +140,9 @@ type MemoryFile struct { // is protected by mu. reclaimable bool - // minReclaimablePage is the minimum page that may be reclaimable. - // i.e., all reclaimable pages are >= minReclaimablePage. - // - // minReclaimablePage is protected by mu. - minReclaimablePage uint64 + // relcaim is the collection of regions for reclaim. relcaim is protected + // by mu. + reclaim reclaimSet // reclaimCond is signaled (with mu locked) when reclaimable or destroyed // transitions from false to true. @@ -180,6 +172,11 @@ type MemoryFileOpts struct { // notifications to determine when eviction is necessary. This option has // no effect unless DelayedEviction is DelayedEvictionEnabled. UseHostMemcgPressure bool + + // If ManualZeroing is true, MemoryFile must not assume that new pages + // obtained from the host are zero-filled, such that MemoryFile must manually + // zero newly-allocated pages. + ManualZeroing bool } // DelayedEvictionType is the type of MemoryFileOpts.DelayedEviction. @@ -268,12 +265,10 @@ type evictableMemoryUserInfo struct { } const ( - chunkShift = 24 - chunkSize = 1 << chunkShift // 16 MB + chunkShift = 30 + chunkSize = 1 << chunkShift // 1 GB chunkMask = chunkSize - 1 - initialSize = chunkSize - // maxPage is the highest 64-bit page. maxPage = math.MaxUint64 &^ (usermem.PageSize - 1) ) @@ -297,19 +292,12 @@ func NewMemoryFile(file *os.File, opts MemoryFileOpts) (*MemoryFile, error) { if err := file.Truncate(0); err != nil { return nil, err } - if err := file.Truncate(initialSize); err != nil { - return nil, err - } f := &MemoryFile{ - opts: opts, - fileSize: initialSize, - file: file, - // No pages are reclaimable. DecRef will always be able to - // decrease minReclaimablePage from this point. - minReclaimablePage: maxPage, - evictable: make(map[EvictableMemoryUser]*evictableMemoryUserInfo), + opts: opts, + file: file, + evictable: make(map[EvictableMemoryUser]*evictableMemoryUserInfo), } - f.mappings.Store(make([]uintptr, initialSize/chunkSize)) + f.mappings.Store(make([]uintptr, 0)) f.reclaimCond.L = &f.mu if f.opts.DelayedEviction == DelayedEvictionEnabled && f.opts.UseHostMemcgPressure { @@ -399,39 +387,37 @@ func (f *MemoryFile) Allocate(length uint64, kind usage.MemoryKind) (platform.Fi alignment = usermem.HugePageSize } - start, minUnallocatedPage := findUnallocatedRange(&f.usage, f.minUnallocatedPage, length, alignment) - end := start + length - // File offsets are int64s. Since length must be strictly positive, end - // cannot legitimately be 0. - if end < start || int64(end) <= 0 { + // Find a range in the underlying file. + fr, ok := findAvailableRange(&f.usage, f.fileSize, length, alignment) + if !ok { return platform.FileRange{}, syserror.ENOMEM } - // Expand the file if needed. Double the file size on each expansion; - // uncommitted pages have effectively no cost. - fileSize := f.fileSize - for int64(end) > fileSize { - if fileSize >= 2*fileSize { - // fileSize overflow. - return platform.FileRange{}, syserror.ENOMEM - } - fileSize *= 2 - } - if fileSize > f.fileSize { - if err := f.file.Truncate(fileSize); err != nil { + // Expand the file if needed. Note that findAvailableRange will + // appropriately double the fileSize when required. + if int64(fr.End) > f.fileSize { + if err := f.file.Truncate(int64(fr.End)); err != nil { return platform.FileRange{}, err } - f.fileSize = fileSize + f.fileSize = int64(fr.End) f.mappingsMu.Lock() oldMappings := f.mappings.Load().([]uintptr) - newMappings := make([]uintptr, fileSize>>chunkShift) + newMappings := make([]uintptr, f.fileSize>>chunkShift) copy(newMappings, oldMappings) f.mappings.Store(newMappings) f.mappingsMu.Unlock() } // Mark selected pages as in use. - fr := platform.FileRange{start, end} + if f.opts.ManualZeroing { + if err := f.forEachMappingSlice(fr, func(bs []byte) { + for i := range bs { + bs[i] = 0 + } + }); err != nil { + return platform.FileRange{}, err + } + } if !f.usage.Add(fr, usageInfo{ kind: kind, refs: 1, @@ -439,49 +425,71 @@ func (f *MemoryFile) Allocate(length uint64, kind usage.MemoryKind) (platform.Fi panic(fmt.Sprintf("allocating %v: failed to insert into usage set:\n%v", fr, &f.usage)) } - if minUnallocatedPage < start { - f.minUnallocatedPage = minUnallocatedPage - } else { - // start was the first unallocated page. The next must be - // somewhere beyond end. - f.minUnallocatedPage = end - } - return fr, nil } -// findUnallocatedRange returns the first unallocated page in usage of the -// specified length and alignment beginning at page start and the first single -// unallocated page. -func findUnallocatedRange(usage *usageSet, start, length, alignment uint64) (uint64, uint64) { - // Only searched until the first page is found. - firstPage := start - foundFirstPage := false - alignMask := alignment - 1 - for seg := usage.LowerBoundSegment(start); seg.Ok(); seg = seg.NextSegment() { - r := seg.Range() +// findAvailableRange returns an available range in the usageSet. +// +// Note that scanning for available slots takes place from end first backwards, +// then forwards. This heuristic has important consequence for how sequential +// mappings can be merged in the host VMAs, given that addresses for both +// application and sentry mappings are allocated top-down (from higher to +// lower addresses). The file is also grown expoentially in order to create +// space for mappings to be allocated downwards. +// +// Precondition: alignment must be a power of 2. +func findAvailableRange(usage *usageSet, fileSize int64, length, alignment uint64) (platform.FileRange, bool) { + alignmentMask := alignment - 1 + for gap := usage.UpperBoundGap(uint64(fileSize)); gap.Ok(); gap = gap.PrevLargeEnoughGap(length) { + // Start searching only at end of file. + end := gap.End() + if end > uint64(fileSize) { + end = uint64(fileSize) + } - if !foundFirstPage && r.Start > firstPage { - foundFirstPage = true + // Start at the top and align downwards. + start := end - length + if start > end { + break // Underflow. } + start &^= alignmentMask - if start >= r.End { - // start was rounded up to an alignment boundary from the end - // of a previous segment and is now beyond r.End. + // Is the gap still sufficient? + if start < gap.Start() { continue } - // This segment represents allocated or reclaimable pages; only the - // range from start to the segment's beginning is allocatable, and the - // next allocatable range begins after the segment. - if r.Start > start && r.Start-start >= length { - break + + // Allocate in the given gap. + return platform.FileRange{start, start + length}, true + } + + // Check that it's possible to fit this allocation at the end of a file of any size. + min := usage.LastGap().Start() + min = (min + alignmentMask) &^ alignmentMask + if min+length < min { + // Overflow. + return platform.FileRange{}, false + } + + // Determine the minimum file size required to fit this allocation at its end. + for { + if fileSize >= 2*fileSize { + // Is this because it's initially empty? + if fileSize == 0 { + fileSize += chunkSize + } else { + // fileSize overflow. + return platform.FileRange{}, false + } + } else { + // Double the current fileSize. + fileSize *= 2 } - start = (r.End + alignMask) &^ alignMask - if !foundFirstPage { - firstPage = r.End + start := (uint64(fileSize) - length) &^ alignmentMask + if start >= min { + return platform.FileRange{start, start + length}, true } } - return start, firstPage } // AllocateAndFill allocates memory of the given kind and fills it by calling @@ -602,6 +610,7 @@ func (f *MemoryFile) DecRef(fr platform.FileRange) { } val.refs-- if val.refs == 0 { + f.reclaim.Add(seg.Range(), reclaimSetValue{}) freed = true // Reclassify memory as System, until it's freed by the reclaim // goroutine. @@ -614,10 +623,6 @@ func (f *MemoryFile) DecRef(fr platform.FileRange) { f.usage.MergeAdjacent(fr) if freed { - if fr.Start < f.minReclaimablePage { - // We've freed at least one lower page. - f.minReclaimablePage = fr.Start - } f.reclaimable = true f.reclaimCond.Signal() } @@ -1016,6 +1021,7 @@ func (f *MemoryFile) String() string { // for allocation. func (f *MemoryFile) runReclaim() { for { + // N.B. We must call f.markReclaimed on the returned FrameRange. fr, ok := f.findReclaimable() if !ok { break @@ -1071,6 +1077,10 @@ func (f *MemoryFile) runReclaim() { } } +// findReclaimable finds memory that has been marked for reclaim. +// +// Note that there returned range will be removed from tracking. It +// must be reclaimed (removed from f.usage) at this point. func (f *MemoryFile) findReclaimable() (platform.FileRange, bool) { f.mu.Lock() defer f.mu.Unlock() @@ -1089,18 +1099,15 @@ func (f *MemoryFile) findReclaimable() (platform.FileRange, bool) { } f.reclaimCond.Wait() } - // Allocate returns the first usable range in offset order and is - // currently a linear scan, so reclaiming from the beginning of the - // file minimizes the expected latency of Allocate. - for seg := f.usage.LowerBoundSegment(f.minReclaimablePage); seg.Ok(); seg = seg.NextSegment() { - if seg.ValuePtr().refs == 0 { - f.minReclaimablePage = seg.End() - return seg.Range(), true - } + // Allocate works from the back of the file inwards, so reclaim + // preserves this order to minimize the cost of the search. + if seg := f.reclaim.LastSegment(); seg.Ok() { + fr := seg.Range() + f.reclaim.Remove(seg) + return fr, true } - // No pages are reclaimable. + // Nothing is reclaimable. f.reclaimable = false - f.minReclaimablePage = maxPage } } @@ -1108,8 +1115,8 @@ func (f *MemoryFile) markReclaimed(fr platform.FileRange) { f.mu.Lock() defer f.mu.Unlock() seg := f.usage.FindSegment(fr.Start) - // All of fr should be mapped to a single uncommitted reclaimable segment - // accounted to System. + // All of fr should be mapped to a single uncommitted reclaimable + // segment accounted to System. if !seg.Ok() { panic(fmt.Sprintf("reclaimed pages %v include unreferenced pages:\n%v", fr, &f.usage)) } @@ -1123,14 +1130,10 @@ func (f *MemoryFile) markReclaimed(fr platform.FileRange) { }); got != want { panic(fmt.Sprintf("reclaimed pages %v in segment %v has incorrect state %v, wanted %v:\n%v", fr, seg.Range(), got, want, &f.usage)) } - // Deallocate reclaimed pages. Even though all of seg is reclaimable, the - // caller of markReclaimed may not have decommitted it, so we can only mark - // fr as reclaimed. + // Deallocate reclaimed pages. Even though all of seg is reclaimable, + // the caller of markReclaimed may not have decommitted it, so we can + // only mark fr as reclaimed. f.usage.Remove(f.usage.Isolate(seg, fr)) - if fr.Start < f.minUnallocatedPage { - // We've deallocated at least one lower page. - f.minUnallocatedPage = fr.Start - } } // StartEvictions requests that f evict all evictable allocations. It does not @@ -1241,3 +1244,27 @@ func (evictableRangeSetFunctions) Merge(_ EvictableRange, _ evictableRangeSetVal func (evictableRangeSetFunctions) Split(_ EvictableRange, _ evictableRangeSetValue, _ uint64) (evictableRangeSetValue, evictableRangeSetValue) { return evictableRangeSetValue{}, evictableRangeSetValue{} } + +// reclaimSetValue is the value type of reclaimSet. +type reclaimSetValue struct{} + +type reclaimSetFunctions struct{} + +func (reclaimSetFunctions) MinKey() uint64 { + return 0 +} + +func (reclaimSetFunctions) MaxKey() uint64 { + return math.MaxUint64 +} + +func (reclaimSetFunctions) ClearValue(val *reclaimSetValue) { +} + +func (reclaimSetFunctions) Merge(_ platform.FileRange, _ reclaimSetValue, _ platform.FileRange, _ reclaimSetValue) (reclaimSetValue, bool) { + return reclaimSetValue{}, true +} + +func (reclaimSetFunctions) Split(_ platform.FileRange, _ reclaimSetValue, _ uint64) (reclaimSetValue, reclaimSetValue) { + return reclaimSetValue{}, reclaimSetValue{} +} diff --git a/pkg/sentry/pgalloc/pgalloc_test.go b/pkg/sentry/pgalloc/pgalloc_test.go index 293f22c6b..b5b68eb52 100644 --- a/pkg/sentry/pgalloc/pgalloc_test.go +++ b/pkg/sentry/pgalloc/pgalloc_test.go @@ -23,39 +23,49 @@ import ( const ( page = usermem.PageSize hugepage = usermem.HugePageSize + topPage = (1 << 63) - page ) func TestFindUnallocatedRange(t *testing.T) { for _, test := range []struct { - desc string - usage *usageSegmentDataSlices - start uint64 - length uint64 - alignment uint64 - unallocated uint64 - minUnallocated uint64 + desc string + usage *usageSegmentDataSlices + fileSize int64 + length uint64 + alignment uint64 + start uint64 + expectFail bool }{ { - desc: "Initial allocation succeeds", - usage: &usageSegmentDataSlices{}, - start: 0, - length: page, - alignment: page, - unallocated: 0, - minUnallocated: 0, + desc: "Initial allocation succeeds", + usage: &usageSegmentDataSlices{}, + length: page, + alignment: page, + start: chunkSize - page, // Grows by chunkSize, allocate down. }, { - desc: "Allocation begins at start of file", + desc: "Allocation finds empty space at start of file", usage: &usageSegmentDataSlices{ Start: []uint64{page}, End: []uint64{2 * page}, Values: []usageInfo{{refs: 1}}, }, - start: 0, - length: page, - alignment: page, - unallocated: 0, - minUnallocated: 0, + fileSize: 2 * page, + length: page, + alignment: page, + start: 0, + }, + { + desc: "Allocation finds empty space at end of file", + usage: &usageSegmentDataSlices{ + Start: []uint64{0}, + End: []uint64{page}, + Values: []usageInfo{{refs: 1}}, + }, + fileSize: 2 * page, + length: page, + alignment: page, + start: page, }, { desc: "In-use frames are not allocatable", @@ -64,11 +74,10 @@ func TestFindUnallocatedRange(t *testing.T) { End: []uint64{page, 2 * page}, Values: []usageInfo{{refs: 1}, {refs: 2}}, }, - start: 0, - length: page, - alignment: page, - unallocated: 2 * page, - minUnallocated: 2 * page, + fileSize: 2 * page, + length: page, + alignment: page, + start: 3 * page, // Double fileSize, allocate top-down. }, { desc: "Reclaimable frames are not allocatable", @@ -77,11 +86,10 @@ func TestFindUnallocatedRange(t *testing.T) { End: []uint64{page, 2 * page, 3 * page}, Values: []usageInfo{{refs: 1}, {refs: 0}, {refs: 1}}, }, - start: 0, - length: page, - alignment: page, - unallocated: 3 * page, - minUnallocated: 3 * page, + fileSize: 3 * page, + length: page, + alignment: page, + start: 5 * page, // Double fileSize, grow down. }, { desc: "Gaps between in-use frames are allocatable", @@ -90,11 +98,10 @@ func TestFindUnallocatedRange(t *testing.T) { End: []uint64{page, 3 * page}, Values: []usageInfo{{refs: 1}, {refs: 1}}, }, - start: 0, - length: page, - alignment: page, - unallocated: page, - minUnallocated: page, + fileSize: 3 * page, + length: page, + alignment: page, + start: page, }, { desc: "Inadequately-sized gaps are rejected", @@ -103,14 +110,13 @@ func TestFindUnallocatedRange(t *testing.T) { End: []uint64{page, 3 * page}, Values: []usageInfo{{refs: 1}, {refs: 1}}, }, - start: 0, - length: 2 * page, - alignment: page, - unallocated: 3 * page, - minUnallocated: page, + fileSize: 3 * page, + length: 2 * page, + alignment: page, + start: 4 * page, // Double fileSize, grow down. }, { - desc: "Hugepage alignment is honored", + desc: "Alignment is honored at end of file", usage: &usageSegmentDataSlices{ Start: []uint64{0, hugepage + page}, // Hugepage-sized gap here that shouldn't be allocated from @@ -118,37 +124,95 @@ func TestFindUnallocatedRange(t *testing.T) { End: []uint64{page, hugepage + 2*page}, Values: []usageInfo{{refs: 1}, {refs: 1}}, }, - start: 0, - length: hugepage, - alignment: hugepage, - unallocated: 2 * hugepage, - minUnallocated: page, + fileSize: hugepage + 2*page, + length: hugepage, + alignment: hugepage, + start: 3 * hugepage, // Double fileSize until alignment is satisfied, grow down. + }, + { + desc: "Alignment is honored before end of file", + usage: &usageSegmentDataSlices{ + Start: []uint64{0, 2*hugepage + page}, + // Page will need to be shifted down from top. + End: []uint64{page, 2*hugepage + 2*page}, + Values: []usageInfo{{refs: 1}, {refs: 1}}, + }, + fileSize: 2*hugepage + 2*page, + length: hugepage, + alignment: hugepage, + start: hugepage, }, { - desc: "Pages before start ignored", + desc: "Allocations are compact if possible", usage: &usageSegmentDataSlices{ Start: []uint64{page, 3 * page}, End: []uint64{2 * page, 4 * page}, Values: []usageInfo{{refs: 1}, {refs: 2}}, }, - start: page, - length: page, - alignment: page, - unallocated: 2 * page, - minUnallocated: 2 * page, + fileSize: 4 * page, + length: page, + alignment: page, + start: 2 * page, + }, + { + desc: "Top-down allocation within one gap", + usage: &usageSegmentDataSlices{ + Start: []uint64{page, 4 * page, 7 * page}, + End: []uint64{2 * page, 5 * page, 8 * page}, + Values: []usageInfo{{refs: 1}, {refs: 2}, {refs: 1}}, + }, + fileSize: 8 * page, + length: page, + alignment: page, + start: 6 * page, + }, + { + desc: "Top-down allocation between multiple gaps", + usage: &usageSegmentDataSlices{ + Start: []uint64{page, 3 * page, 5 * page}, + End: []uint64{2 * page, 4 * page, 6 * page}, + Values: []usageInfo{{refs: 1}, {refs: 2}, {refs: 1}}, + }, + fileSize: 6 * page, + length: page, + alignment: page, + start: 4 * page, }, { - desc: "start may be in the middle of segment", + desc: "Top-down allocation with large top gap", usage: &usageSegmentDataSlices{ - Start: []uint64{0, 3 * page}, + Start: []uint64{page, 3 * page}, End: []uint64{2 * page, 4 * page}, Values: []usageInfo{{refs: 1}, {refs: 2}}, }, - start: page, - length: page, - alignment: page, - unallocated: 2 * page, - minUnallocated: 2 * page, + fileSize: 8 * page, + length: page, + alignment: page, + start: 7 * page, + }, + { + desc: "Gaps found with possible overflow", + usage: &usageSegmentDataSlices{ + Start: []uint64{page, topPage - page}, + End: []uint64{2 * page, topPage}, + Values: []usageInfo{{refs: 1}, {refs: 1}}, + }, + fileSize: topPage, + length: page, + alignment: page, + start: topPage - 2*page, + }, + { + desc: "Overflow detected", + usage: &usageSegmentDataSlices{ + Start: []uint64{page}, + End: []uint64{topPage}, + Values: []usageInfo{{refs: 1}}, + }, + fileSize: topPage, + length: 2 * page, + alignment: page, + expectFail: true, }, } { t.Run(test.desc, func(t *testing.T) { @@ -156,12 +220,18 @@ func TestFindUnallocatedRange(t *testing.T) { if err := usage.ImportSortedSlices(test.usage); err != nil { t.Fatalf("Failed to initialize usage from %v: %v", test.usage, err) } - unallocated, minUnallocated := findUnallocatedRange(&usage, test.start, test.length, test.alignment) - if unallocated != test.unallocated { - t.Errorf("findUnallocatedRange(%v, %x, %x, %x): got unallocated %x, wanted %x", test.usage, test.start, test.length, test.alignment, unallocated, test.unallocated) + fr, ok := findAvailableRange(&usage, test.fileSize, test.length, test.alignment) + if !test.expectFail && !ok { + t.Fatalf("findAvailableRange(%v, %x, %x, %x): got %x, false wanted %x, true", test.usage, test.fileSize, test.length, test.alignment, fr.Start, test.start) + } + if test.expectFail && ok { + t.Fatalf("findAvailableRange(%v, %x, %x, %x): got %x, true wanted %x, false", test.usage, test.fileSize, test.length, test.alignment, fr.Start, test.start) + } + if ok && fr.Start != test.start { + t.Errorf("findAvailableRange(%v, %x, %x, %x): got start=%x, wanted %x", test.usage, test.fileSize, test.length, test.alignment, fr.Start, test.start) } - if minUnallocated != test.minUnallocated { - t.Errorf("findUnallocatedRange(%v, %x, %x, %x): got minUnallocated %x, wanted %x", test.usage, test.start, test.length, test.alignment, minUnallocated, test.minUnallocated) + if ok && fr.End != test.start+test.length { + t.Errorf("findAvailableRange(%v, %x, %x, %x): got end=%x, wanted %x", test.usage, test.fileSize, test.length, test.alignment, fr.End, test.start+test.length) } }) } diff --git a/pkg/sentry/platform/kvm/BUILD b/pkg/sentry/platform/kvm/BUILD index 159f7eafd..4792454c4 100644 --- a/pkg/sentry/platform/kvm/BUILD +++ b/pkg/sentry/platform/kvm/BUILD @@ -6,8 +6,8 @@ go_library( name = "kvm", srcs = [ "address_space.go", - "allocator.go", "bluepill.go", + "bluepill_allocator.go", "bluepill_amd64.go", "bluepill_amd64.s", "bluepill_amd64_unsafe.go", diff --git a/pkg/sentry/platform/kvm/address_space.go b/pkg/sentry/platform/kvm/address_space.go index be213bfe8..faf1d5e1c 100644 --- a/pkg/sentry/platform/kvm/address_space.go +++ b/pkg/sentry/platform/kvm/address_space.go @@ -26,16 +26,15 @@ import ( // dirtySet tracks vCPUs for invalidation. type dirtySet struct { - vCPUs []uint64 + vCPUMasks []uint64 } // forEach iterates over all CPUs in the dirty set. +// +//go:nosplit func (ds *dirtySet) forEach(m *machine, fn func(c *vCPU)) { - m.mu.RLock() - defer m.mu.RUnlock() - - for index := range ds.vCPUs { - mask := atomic.SwapUint64(&ds.vCPUs[index], 0) + for index := range ds.vCPUMasks { + mask := atomic.SwapUint64(&ds.vCPUMasks[index], 0) if mask != 0 { for bit := 0; bit < 64; bit++ { if mask&(1<<uint64(bit)) == 0 { @@ -54,7 +53,7 @@ func (ds *dirtySet) mark(c *vCPU) bool { index := uint64(c.id) / 64 bit := uint64(1) << uint(c.id%64) - oldValue := atomic.LoadUint64(&ds.vCPUs[index]) + oldValue := atomic.LoadUint64(&ds.vCPUMasks[index]) if oldValue&bit != 0 { return false // Not clean. } @@ -62,7 +61,7 @@ func (ds *dirtySet) mark(c *vCPU) bool { // Set the bit unilaterally, and ensure that a flush takes place. Note // that it's possible for races to occur here, but since the flush is // taking place long after these lines there's no race in practice. - atomicbitops.OrUint64(&ds.vCPUs[index], bit) + atomicbitops.OrUint64(&ds.vCPUMasks[index], bit) return true // Previously clean. } @@ -113,7 +112,12 @@ type hostMapEntry struct { length uintptr } -func (as *addressSpace) mapHost(addr usermem.Addr, m hostMapEntry, at usermem.AccessType) (inv bool) { +// mapLocked maps the given host entry. +// +// +checkescape:hard,stack +// +//go:nosplit +func (as *addressSpace) mapLocked(addr usermem.Addr, m hostMapEntry, at usermem.AccessType) (inv bool) { for m.length > 0 { physical, length, ok := translateToPhysical(m.addr) if !ok { @@ -133,18 +137,10 @@ func (as *addressSpace) mapHost(addr usermem.Addr, m hostMapEntry, at usermem.Ac // important; if the pagetable mappings were installed before // ensuring the physical pages were available, then some other // thread could theoretically access them. - // - // Due to the way KVM's shadow paging implementation works, - // modifications to the page tables while in host mode may not - // be trapped, leading to the shadow pages being out of sync. - // Therefore, we need to ensure that we are in guest mode for - // page table modifications. See the call to bluepill, below. - as.machine.retryInGuest(func() { - inv = as.pageTables.Map(addr, length, pagetables.MapOpts{ - AccessType: at, - User: true, - }, physical) || inv - }) + inv = as.pageTables.Map(addr, length, pagetables.MapOpts{ + AccessType: at, + User: true, + }, physical) || inv m.addr += length m.length -= length addr += usermem.Addr(length) @@ -176,6 +172,10 @@ func (as *addressSpace) MapFile(addr usermem.Addr, f platform.File, fr platform. return err } + // See block in mapLocked. + as.pageTables.Allocator.(*allocator).cpu = as.machine.Get() + defer as.machine.Put(as.pageTables.Allocator.(*allocator).cpu) + // Map the mappings in the sentry's address space (guest physical memory) // into the application's address space (guest virtual memory). inv := false @@ -190,7 +190,12 @@ func (as *addressSpace) MapFile(addr usermem.Addr, f platform.File, fr platform. _ = s[i] // Touch to commit. } } - prev := as.mapHost(addr, hostMapEntry{ + + // See bluepill_allocator.go. + bluepill(as.pageTables.Allocator.(*allocator).cpu) + + // Perform the mapping. + prev := as.mapLocked(addr, hostMapEntry{ addr: b.Addr(), length: uintptr(b.Len()), }, at) @@ -204,17 +209,27 @@ func (as *addressSpace) MapFile(addr usermem.Addr, f platform.File, fr platform. return nil } +// unmapLocked is an escape-checked wrapped around Unmap. +// +// +checkescape:hard,stack +// +//go:nosplit +func (as *addressSpace) unmapLocked(addr usermem.Addr, length uint64) bool { + return as.pageTables.Unmap(addr, uintptr(length)) +} + // Unmap unmaps the given range by calling pagetables.PageTables.Unmap. func (as *addressSpace) Unmap(addr usermem.Addr, length uint64) { as.mu.Lock() defer as.mu.Unlock() - // See above re: retryInGuest. - var prev bool - as.machine.retryInGuest(func() { - prev = as.pageTables.Unmap(addr, uintptr(length)) || prev - }) - if prev { + // See above & bluepill_allocator.go. + as.pageTables.Allocator.(*allocator).cpu = as.machine.Get() + defer as.machine.Put(as.pageTables.Allocator.(*allocator).cpu) + bluepill(as.pageTables.Allocator.(*allocator).cpu) + + if prev := as.unmapLocked(addr, length); prev { + // Invalidate all active vCPUs. as.invalidate() // Recycle any freed intermediate pages. @@ -227,7 +242,7 @@ func (as *addressSpace) Release() { as.Unmap(0, ^uint64(0)) // Free all pages from the allocator. - as.pageTables.Allocator.(allocator).base.Drain() + as.pageTables.Allocator.(*allocator).base.Drain() // Drop all cached machine references. as.machine.dropPageTables(as.pageTables) diff --git a/pkg/sentry/platform/kvm/bluepill.go b/pkg/sentry/platform/kvm/bluepill.go index 35cd55fef..4b23f7803 100644 --- a/pkg/sentry/platform/kvm/bluepill.go +++ b/pkg/sentry/platform/kvm/bluepill.go @@ -81,12 +81,6 @@ func (c *vCPU) die(context *arch.SignalContext64, msg string) { // Save the death message, which will be thrown. c.dieState.message = msg - // Reload all registers to have an accurate stack trace when we return - // to host mode. This means that the stack should be unwound correctly. - if errno := c.getUserRegisters(&c.dieState.guestRegs); errno != 0 { - throw(msg) - } - // Setup the trampoline. dieArchSetup(c, context, &c.dieState.guestRegs) } diff --git a/pkg/sentry/platform/kvm/allocator.go b/pkg/sentry/platform/kvm/bluepill_allocator.go index 3f35414bb..9485e1301 100644 --- a/pkg/sentry/platform/kvm/allocator.go +++ b/pkg/sentry/platform/kvm/bluepill_allocator.go @@ -21,56 +21,80 @@ import ( ) type allocator struct { - base *pagetables.RuntimeAllocator + base pagetables.RuntimeAllocator + + // cpu must be set prior to any pagetable operation. + // + // Due to the way KVM's shadow paging implementation works, + // modifications to the page tables while in host mode may not be + // trapped, leading to the shadow pages being out of sync. Therefore, + // we need to ensure that we are in guest mode for page table + // modifications. See the call to bluepill, below. + cpu *vCPU } // newAllocator is used to define the allocator. -func newAllocator() allocator { - return allocator{ - base: pagetables.NewRuntimeAllocator(), - } +func newAllocator() *allocator { + a := new(allocator) + a.base.Init() + return a } // NewPTEs implements pagetables.Allocator.NewPTEs. // +// +checkescape:all +// //go:nosplit -func (a allocator) NewPTEs() *pagetables.PTEs { - return a.base.NewPTEs() +func (a *allocator) NewPTEs() *pagetables.PTEs { + ptes := a.base.NewPTEs() // escapes: bluepill below. + if a.cpu != nil { + bluepill(a.cpu) + } + return ptes } // PhysicalFor returns the physical address for a set of PTEs. // +// +checkescape:all +// //go:nosplit -func (a allocator) PhysicalFor(ptes *pagetables.PTEs) uintptr { +func (a *allocator) PhysicalFor(ptes *pagetables.PTEs) uintptr { virtual := a.base.PhysicalFor(ptes) physical, _, ok := translateToPhysical(virtual) if !ok { - panic(fmt.Sprintf("PhysicalFor failed for %p", ptes)) + panic(fmt.Sprintf("PhysicalFor failed for %p", ptes)) // escapes: panic. } return physical } // LookupPTEs implements pagetables.Allocator.LookupPTEs. // +// +checkescape:all +// //go:nosplit -func (a allocator) LookupPTEs(physical uintptr) *pagetables.PTEs { +func (a *allocator) LookupPTEs(physical uintptr) *pagetables.PTEs { virtualStart, physicalStart, _, ok := calculateBluepillFault(physical, physicalRegions) if !ok { - panic(fmt.Sprintf("LookupPTEs failed for 0x%x", physical)) + panic(fmt.Sprintf("LookupPTEs failed for 0x%x", physical)) // escapes: panic. } return a.base.LookupPTEs(virtualStart + (physical - physicalStart)) } // FreePTEs implements pagetables.Allocator.FreePTEs. // +// +checkescape:all +// //go:nosplit -func (a allocator) FreePTEs(ptes *pagetables.PTEs) { - a.base.FreePTEs(ptes) +func (a *allocator) FreePTEs(ptes *pagetables.PTEs) { + a.base.FreePTEs(ptes) // escapes: bluepill below. + if a.cpu != nil { + bluepill(a.cpu) + } } // Recycle implements pagetables.Allocator.Recycle. // //go:nosplit -func (a allocator) Recycle() { +func (a *allocator) Recycle() { a.base.Recycle() } diff --git a/pkg/sentry/platform/kvm/bluepill_amd64.go b/pkg/sentry/platform/kvm/bluepill_amd64.go index 133c2203d..ddc1554d5 100644 --- a/pkg/sentry/platform/kvm/bluepill_amd64.go +++ b/pkg/sentry/platform/kvm/bluepill_amd64.go @@ -63,6 +63,8 @@ func bluepillArchEnter(context *arch.SignalContext64) *vCPU { // KernelSyscall handles kernel syscalls. // +// +checkescape:all +// //go:nosplit func (c *vCPU) KernelSyscall() { regs := c.Registers() @@ -72,13 +74,15 @@ func (c *vCPU) KernelSyscall() { // We only trigger a bluepill entry in the bluepill function, and can // therefore be guaranteed that there is no floating point state to be // loaded on resuming from halt. We only worry about saving on exit. - ring0.SaveFloatingPoint((*byte)(c.floatingPointState)) + ring0.SaveFloatingPoint((*byte)(c.floatingPointState)) // escapes: no. ring0.Halt() - ring0.WriteFS(uintptr(regs.Fs_base)) // Reload host segment. + ring0.WriteFS(uintptr(regs.Fs_base)) // escapes: no, reload host segment. } // KernelException handles kernel exceptions. // +// +checkescape:all +// //go:nosplit func (c *vCPU) KernelException(vector ring0.Vector) { regs := c.Registers() @@ -89,9 +93,9 @@ func (c *vCPU) KernelException(vector ring0.Vector) { regs.Rip = 0 } // See above. - ring0.SaveFloatingPoint((*byte)(c.floatingPointState)) + ring0.SaveFloatingPoint((*byte)(c.floatingPointState)) // escapes: no. ring0.Halt() - ring0.WriteFS(uintptr(regs.Fs_base)) // Reload host segment. + ring0.WriteFS(uintptr(regs.Fs_base)) // escapes: no; reload host segment. } // bluepillArchExit is called during bluepillEnter. diff --git a/pkg/sentry/platform/kvm/bluepill_amd64_unsafe.go b/pkg/sentry/platform/kvm/bluepill_amd64_unsafe.go index a63a6a071..99cac665d 100644 --- a/pkg/sentry/platform/kvm/bluepill_amd64_unsafe.go +++ b/pkg/sentry/platform/kvm/bluepill_amd64_unsafe.go @@ -31,6 +31,12 @@ import ( // //go:nosplit func dieArchSetup(c *vCPU, context *arch.SignalContext64, guestRegs *userRegs) { + // Reload all registers to have an accurate stack trace when we return + // to host mode. This means that the stack should be unwound correctly. + if errno := c.getUserRegisters(&c.dieState.guestRegs); errno != 0 { + throw(c.dieState.message) + } + // If the vCPU is in user mode, we set the stack to the stored stack // value in the vCPU itself. We don't want to unwind the user stack. if guestRegs.RFLAGS&ring0.UserFlagsSet == ring0.UserFlagsSet { diff --git a/pkg/sentry/platform/kvm/bluepill_arm64.go b/pkg/sentry/platform/kvm/bluepill_arm64.go index 552341721..83643c602 100644 --- a/pkg/sentry/platform/kvm/bluepill_arm64.go +++ b/pkg/sentry/platform/kvm/bluepill_arm64.go @@ -54,26 +54,60 @@ func bluepillArchExit(c *vCPU, context *arch.SignalContext64) { context.Pstate = regs.Pstate context.Pstate &^= uint64(ring0.UserFlagsClear) context.Pstate |= ring0.UserFlagsSet + + lazyVfp := c.GetLazyVFP() + if lazyVfp != 0 { + fpsimd := fpsimdPtr((*byte)(c.floatingPointState)) + context.Fpsimd64.Fpsr = fpsimd.Fpsr + context.Fpsimd64.Fpcr = fpsimd.Fpcr + context.Fpsimd64.Vregs = fpsimd.Vregs + } } // KernelSyscall handles kernel syscalls. // +// +checkescape:all +// //go:nosplit func (c *vCPU) KernelSyscall() { regs := c.Registers() if regs.Regs[8] != ^uint64(0) { regs.Pc -= 4 // Rewind. } + + vfpEnable := ring0.CPACREL1() + if vfpEnable != 0 { + fpsimd := fpsimdPtr((*byte)(c.floatingPointState)) + fpcr := ring0.GetFPCR() + fpsr := ring0.GetFPSR() + fpsimd.Fpcr = uint32(fpcr) + fpsimd.Fpsr = uint32(fpsr) + ring0.SaveVRegs((*byte)(c.floatingPointState)) + } + ring0.Halt() } // KernelException handles kernel exceptions. // +// +checkescape:all +// //go:nosplit func (c *vCPU) KernelException(vector ring0.Vector) { regs := c.Registers() if vector == ring0.Vector(bounce) { regs.Pc = 0 } + + vfpEnable := ring0.CPACREL1() + if vfpEnable != 0 { + fpsimd := fpsimdPtr((*byte)(c.floatingPointState)) + fpcr := ring0.GetFPCR() + fpsr := ring0.GetFPSR() + fpsimd.Fpcr = uint32(fpcr) + fpsimd.Fpsr = uint32(fpsr) + ring0.SaveVRegs((*byte)(c.floatingPointState)) + } + ring0.Halt() } diff --git a/pkg/sentry/platform/kvm/bluepill_arm64.s b/pkg/sentry/platform/kvm/bluepill_arm64.s index c61700892..04efa0147 100644 --- a/pkg/sentry/platform/kvm/bluepill_arm64.s +++ b/pkg/sentry/platform/kvm/bluepill_arm64.s @@ -82,6 +82,8 @@ fallback: // dieTrampoline: see bluepill.go, bluepill_arm64_unsafe.go for documentation. TEXT ·dieTrampoline(SB),NOSPLIT,$0 - // TODO(gvisor.dev/issue/1249): dieTrampoline supporting for Arm64. - MOVD R9, 8(RSP) - BL ·dieHandler(SB) + // R0: Fake the old PC as caller + // R1: First argument (vCPU) + MOVD.P R1, 8(RSP) // R1: First argument (vCPU) + MOVD.P R0, 8(RSP) // R0: Fake the old PC as caller + B ·dieHandler(SB) diff --git a/pkg/sentry/platform/kvm/bluepill_arm64_unsafe.go b/pkg/sentry/platform/kvm/bluepill_arm64_unsafe.go index 2f02c03cf..4ca2b7717 100644 --- a/pkg/sentry/platform/kvm/bluepill_arm64_unsafe.go +++ b/pkg/sentry/platform/kvm/bluepill_arm64_unsafe.go @@ -17,10 +17,47 @@ package kvm import ( + "unsafe" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/platform/ring0" ) +// fpsimdPtr returns a fpsimd64 for the given address. +// +//go:nosplit +func fpsimdPtr(addr *byte) *arch.FpsimdContext { + return (*arch.FpsimdContext)(unsafe.Pointer(addr)) +} + +// dieArchSetup initialies the state for dieTrampoline. +// +// The arm64 dieTrampoline requires the vCPU to be set in R1, and the last PC +// to be in R0. The trampoline then simulates a call to dieHandler from the +// provided PC. +// //go:nosplit func dieArchSetup(c *vCPU, context *arch.SignalContext64, guestRegs *userRegs) { - // TODO(gvisor.dev/issue/1249): dieTrampoline supporting for Arm64. + // If the vCPU is in user mode, we set the stack to the stored stack + // value in the vCPU itself. We don't want to unwind the user stack. + if guestRegs.Regs.Pstate&ring0.PSR_MODE_MASK == ring0.PSR_MODE_EL0t { + regs := c.CPU.Registers() + context.Regs[0] = regs.Regs[0] + context.Sp = regs.Sp + context.Regs[29] = regs.Regs[29] // stack base address + } else { + context.Regs[0] = guestRegs.Regs.Pc + context.Sp = guestRegs.Regs.Sp + context.Regs[29] = guestRegs.Regs.Regs[29] + context.Pstate = guestRegs.Regs.Pstate + } + context.Regs[1] = uint64(uintptr(unsafe.Pointer(c))) + context.Pc = uint64(dieTrampolineAddr) +} + +// bluepillArchFpContext returns the arch-specific fpsimd context. +// +//go:nosplit +func bluepillArchFpContext(context unsafe.Pointer) *arch.FpsimdContext { + return &((*arch.SignalContext64)(context).Fpsimd64) } diff --git a/pkg/sentry/platform/kvm/bluepill_unsafe.go b/pkg/sentry/platform/kvm/bluepill_unsafe.go index 9add7c944..c025aa0bb 100644 --- a/pkg/sentry/platform/kvm/bluepill_unsafe.go +++ b/pkg/sentry/platform/kvm/bluepill_unsafe.go @@ -13,7 +13,7 @@ // limitations under the License. // +build go1.12 -// +build !go1.15 +// +build !go1.16 // Check go:linkname function signatures when updating Go version. @@ -64,6 +64,8 @@ func bluepillArchContext(context unsafe.Pointer) *arch.SignalContext64 { // signal stack. It should only execute raw system calls and functions that are // explicitly marked go:nosplit. // +// +checkescape:all +// //go:nosplit func bluepillHandler(context unsafe.Pointer) { // Sanitize the registers; interrupts must always be disabled. @@ -82,7 +84,8 @@ func bluepillHandler(context unsafe.Pointer) { } for { - switch _, _, errno := syscall.RawSyscall(syscall.SYS_IOCTL, uintptr(c.fd), _KVM_RUN, 0); errno { + _, _, errno := syscall.RawSyscall(syscall.SYS_IOCTL, uintptr(c.fd), _KVM_RUN, 0) // escapes: no. + switch errno { case 0: // Expected case. case syscall.EINTR: // First, we process whatever pending signal @@ -90,7 +93,7 @@ func bluepillHandler(context unsafe.Pointer) { // currently, all signals are masked and the signal // must have been delivered directly to this thread. timeout := syscall.Timespec{} - sig, _, errno := syscall.RawSyscall6( + sig, _, errno := syscall.RawSyscall6( // escapes: no. syscall.SYS_RT_SIGTIMEDWAIT, uintptr(unsafe.Pointer(&bounceSignalMask)), 0, // siginfo. @@ -125,7 +128,7 @@ func bluepillHandler(context unsafe.Pointer) { // MMIO exit we receive EFAULT from the run ioctl. We // always inject an NMI here since we may be in kernel // mode and have interrupts disabled. - if _, _, errno := syscall.RawSyscall( + if _, _, errno := syscall.RawSyscall( // escapes: no. syscall.SYS_IOCTL, uintptr(c.fd), _KVM_NMI, 0); errno != 0 { diff --git a/pkg/sentry/platform/kvm/context.go b/pkg/sentry/platform/kvm/context.go index c769ac7b4..6507121ea 100644 --- a/pkg/sentry/platform/kvm/context.go +++ b/pkg/sentry/platform/kvm/context.go @@ -85,3 +85,6 @@ func (c *context) Switch(as platform.AddressSpace, ac arch.Context, _ int32) (*a func (c *context) Interrupt() { c.interrupt.NotifyInterrupt() } + +// Release implements platform.Context.Release(). +func (c *context) Release() {} diff --git a/pkg/sentry/platform/kvm/kvm.go b/pkg/sentry/platform/kvm/kvm.go index 972ba85c3..ae813e24e 100644 --- a/pkg/sentry/platform/kvm/kvm.go +++ b/pkg/sentry/platform/kvm/kvm.go @@ -27,6 +27,38 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) +// userMemoryRegion is a region of physical memory. +// +// This mirrors kvm_memory_region. +type userMemoryRegion struct { + slot uint32 + flags uint32 + guestPhysAddr uint64 + memorySize uint64 + userspaceAddr uint64 +} + +// runData is the run structure. This may be mapped for synchronous register +// access (although that doesn't appear to be supported by my kernel at least). +// +// This mirrors kvm_run. +type runData struct { + requestInterruptWindow uint8 + _ [7]uint8 + + exitReason uint32 + readyForInterruptInjection uint8 + ifFlag uint8 + _ [2]uint8 + + cr8 uint64 + apicBase uint64 + + // This is the union data for exits. Interpretation depends entirely on + // the exitReason above (see vCPU code for more information). + data [32]uint64 +} + // KVM represents a lightweight VM context. type KVM struct { platform.NoCPUPreemptionDetection @@ -159,6 +191,11 @@ func (*constructor) OpenDevice() (*os.File, error) { return OpenDevice() } +// Flags implements platform.Constructor.Flags(). +func (*constructor) Requirements() platform.Requirements { + return platform.Requirements{} +} + func init() { platform.Register("kvm", &constructor{}) } diff --git a/pkg/sentry/platform/kvm/kvm_amd64.go b/pkg/sentry/platform/kvm/kvm_amd64.go index c5a6f9c7d..093497bc4 100644 --- a/pkg/sentry/platform/kvm/kvm_amd64.go +++ b/pkg/sentry/platform/kvm/kvm_amd64.go @@ -21,17 +21,6 @@ import ( "gvisor.dev/gvisor/pkg/sentry/platform/ring0" ) -// userMemoryRegion is a region of physical memory. -// -// This mirrors kvm_memory_region. -type userMemoryRegion struct { - slot uint32 - flags uint32 - guestPhysAddr uint64 - memorySize uint64 - userspaceAddr uint64 -} - // userRegs represents KVM user registers. // // This mirrors kvm_regs. @@ -169,27 +158,6 @@ type modelControlRegisters struct { entries [16]modelControlRegister } -// runData is the run structure. This may be mapped for synchronous register -// access (although that doesn't appear to be supported by my kernel at least). -// -// This mirrors kvm_run. -type runData struct { - requestInterruptWindow uint8 - _ [7]uint8 - - exitReason uint32 - readyForInterruptInjection uint8 - ifFlag uint8 - _ [2]uint8 - - cr8 uint64 - apicBase uint64 - - // This is the union data for exits. Interpretation depends entirely on - // the exitReason above (see vCPU code for more information). - data [32]uint64 -} - // cpuidEntry is a single CPUID entry. // // This mirrors kvm_cpuid_entry2. diff --git a/pkg/sentry/platform/kvm/kvm_arm64.go b/pkg/sentry/platform/kvm/kvm_arm64.go index 2319c86d3..29d457a7e 100644 --- a/pkg/sentry/platform/kvm/kvm_arm64.go +++ b/pkg/sentry/platform/kvm/kvm_arm64.go @@ -17,20 +17,10 @@ package kvm import ( - "syscall" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/platform/ring0" ) -// userMemoryRegion is a region of physical memory. -// -// This mirrors kvm_memory_region. -type userMemoryRegion struct { - slot uint32 - flags uint32 - guestPhysAddr uint64 - memorySize uint64 - userspaceAddr uint64 -} - type kvmOneReg struct { id uint64 addr uint64 @@ -46,38 +36,17 @@ type userFpsimdState struct { } type userRegs struct { - Regs syscall.PtraceRegs + Regs arch.Registers sp_el1 uint64 elr_el1 uint64 spsr [KVM_NR_SPSR]uint64 fpRegs userFpsimdState } -// runData is the run structure. This may be mapped for synchronous register -// access (although that doesn't appear to be supported by my kernel at least). -// -// This mirrors kvm_run. -type runData struct { - requestInterruptWindow uint8 - _ [7]uint8 - - exitReason uint32 - readyForInterruptInjection uint8 - ifFlag uint8 - _ [2]uint8 - - cr8 uint64 - apicBase uint64 - - // This is the union data for exits. Interpretation depends entirely on - // the exitReason above (see vCPU code for more information). - data [32]uint64 -} - // updateGlobalOnce does global initialization. It has to be called only once. func updateGlobalOnce(fd int) error { physicalInit() err := updateSystemValues(int(fd)) - updateVectorTable() + ring0.Init() return err } diff --git a/pkg/sentry/platform/kvm/kvm_const_arm64.go b/pkg/sentry/platform/kvm/kvm_const_arm64.go index 5a74c6e36..531ae8b1e 100644 --- a/pkg/sentry/platform/kvm/kvm_const_arm64.go +++ b/pkg/sentry/platform/kvm/kvm_const_arm64.go @@ -19,30 +19,31 @@ const ( _KVM_GET_ONE_REG = 0x4010aeab _KVM_SET_ONE_REG = 0x4010aeac - _KVM_ARM_PREFERRED_TARGET = 0x8020aeaf - _KVM_ARM_VCPU_INIT = 0x4020aeae - _KVM_ARM64_REGS_PSTATE = 0x6030000000100042 - _KVM_ARM64_REGS_SP_EL1 = 0x6030000000100044 - _KVM_ARM64_REGS_R0 = 0x6030000000100000 - _KVM_ARM64_REGS_R1 = 0x6030000000100002 - _KVM_ARM64_REGS_R2 = 0x6030000000100004 - _KVM_ARM64_REGS_R3 = 0x6030000000100006 - _KVM_ARM64_REGS_R8 = 0x6030000000100010 - _KVM_ARM64_REGS_R18 = 0x6030000000100024 - _KVM_ARM64_REGS_PC = 0x6030000000100040 - _KVM_ARM64_REGS_MAIR_EL1 = 0x603000000013c510 - _KVM_ARM64_REGS_TCR_EL1 = 0x603000000013c102 - _KVM_ARM64_REGS_TTBR0_EL1 = 0x603000000013c100 - _KVM_ARM64_REGS_TTBR1_EL1 = 0x603000000013c101 - _KVM_ARM64_REGS_SCTLR_EL1 = 0x603000000013c080 - _KVM_ARM64_REGS_CPACR_EL1 = 0x603000000013c082 - _KVM_ARM64_REGS_VBAR_EL1 = 0x603000000013c600 + _KVM_ARM_TARGET_GENERIC_V8 = 5 + _KVM_ARM_PREFERRED_TARGET = 0x8020aeaf + _KVM_ARM_VCPU_INIT = 0x4020aeae + _KVM_ARM64_REGS_PSTATE = 0x6030000000100042 + _KVM_ARM64_REGS_SP_EL1 = 0x6030000000100044 + _KVM_ARM64_REGS_R0 = 0x6030000000100000 + _KVM_ARM64_REGS_R1 = 0x6030000000100002 + _KVM_ARM64_REGS_R2 = 0x6030000000100004 + _KVM_ARM64_REGS_R3 = 0x6030000000100006 + _KVM_ARM64_REGS_R8 = 0x6030000000100010 + _KVM_ARM64_REGS_R18 = 0x6030000000100024 + _KVM_ARM64_REGS_PC = 0x6030000000100040 + _KVM_ARM64_REGS_MAIR_EL1 = 0x603000000013c510 + _KVM_ARM64_REGS_TCR_EL1 = 0x603000000013c102 + _KVM_ARM64_REGS_TTBR0_EL1 = 0x603000000013c100 + _KVM_ARM64_REGS_TTBR1_EL1 = 0x603000000013c101 + _KVM_ARM64_REGS_SCTLR_EL1 = 0x603000000013c080 + _KVM_ARM64_REGS_CPACR_EL1 = 0x603000000013c082 + _KVM_ARM64_REGS_VBAR_EL1 = 0x603000000013c600 ) // Arm64: Architectural Feature Access Control Register EL1. const ( - _FPEN_NOTRAP = 0x3 - _FPEN_SHIFT = 0x20 + _FPEN_NOTRAP = 3 + _FPEN_SHIFT = 20 ) // Arm64: System Control Register EL1. diff --git a/pkg/sentry/platform/kvm/kvm_test.go b/pkg/sentry/platform/kvm/kvm_test.go index c42752d50..6c8f4fa28 100644 --- a/pkg/sentry/platform/kvm/kvm_test.go +++ b/pkg/sentry/platform/kvm/kvm_test.go @@ -117,10 +117,10 @@ func TestKernelFloatingPoint(t *testing.T) { }) } -func applicationTest(t testHarness, useHostMappings bool, target func(), fn func(*vCPU, *syscall.PtraceRegs, *pagetables.PageTables) bool) { +func applicationTest(t testHarness, useHostMappings bool, target func(), fn func(*vCPU, *arch.Registers, *pagetables.PageTables) bool) { // Initialize registers & page tables. var ( - regs syscall.PtraceRegs + regs arch.Registers pt *pagetables.PageTables ) testutil.SetTestTarget(®s, target) @@ -154,7 +154,7 @@ func applicationTest(t testHarness, useHostMappings bool, target func(), fn func } func TestApplicationSyscall(t *testing.T) { - applicationTest(t, true, testutil.SyscallLoop, func(c *vCPU, regs *syscall.PtraceRegs, pt *pagetables.PageTables) bool { + applicationTest(t, true, testutil.SyscallLoop, func(c *vCPU, regs *arch.Registers, pt *pagetables.PageTables) bool { var si arch.SignalInfo if _, err := c.SwitchToUser(ring0.SwitchOpts{ Registers: regs, @@ -168,7 +168,7 @@ func TestApplicationSyscall(t *testing.T) { } return false }) - applicationTest(t, true, testutil.SyscallLoop, func(c *vCPU, regs *syscall.PtraceRegs, pt *pagetables.PageTables) bool { + applicationTest(t, true, testutil.SyscallLoop, func(c *vCPU, regs *arch.Registers, pt *pagetables.PageTables) bool { var si arch.SignalInfo if _, err := c.SwitchToUser(ring0.SwitchOpts{ Registers: regs, @@ -184,7 +184,7 @@ func TestApplicationSyscall(t *testing.T) { } func TestApplicationFault(t *testing.T) { - applicationTest(t, true, testutil.Touch, func(c *vCPU, regs *syscall.PtraceRegs, pt *pagetables.PageTables) bool { + applicationTest(t, true, testutil.Touch, func(c *vCPU, regs *arch.Registers, pt *pagetables.PageTables) bool { testutil.SetTouchTarget(regs, nil) // Cause fault. var si arch.SignalInfo if _, err := c.SwitchToUser(ring0.SwitchOpts{ @@ -199,7 +199,7 @@ func TestApplicationFault(t *testing.T) { } return false }) - applicationTest(t, true, testutil.Touch, func(c *vCPU, regs *syscall.PtraceRegs, pt *pagetables.PageTables) bool { + applicationTest(t, true, testutil.Touch, func(c *vCPU, regs *arch.Registers, pt *pagetables.PageTables) bool { testutil.SetTouchTarget(regs, nil) // Cause fault. var si arch.SignalInfo if _, err := c.SwitchToUser(ring0.SwitchOpts{ @@ -216,7 +216,7 @@ func TestApplicationFault(t *testing.T) { } func TestRegistersSyscall(t *testing.T) { - applicationTest(t, true, testutil.TwiddleRegsSyscall, func(c *vCPU, regs *syscall.PtraceRegs, pt *pagetables.PageTables) bool { + applicationTest(t, true, testutil.TwiddleRegsSyscall, func(c *vCPU, regs *arch.Registers, pt *pagetables.PageTables) bool { testutil.SetTestRegs(regs) // Fill values for all registers. for { var si arch.SignalInfo @@ -239,7 +239,7 @@ func TestRegistersSyscall(t *testing.T) { } func TestRegistersFault(t *testing.T) { - applicationTest(t, true, testutil.TwiddleRegsFault, func(c *vCPU, regs *syscall.PtraceRegs, pt *pagetables.PageTables) bool { + applicationTest(t, true, testutil.TwiddleRegsFault, func(c *vCPU, regs *arch.Registers, pt *pagetables.PageTables) bool { testutil.SetTestRegs(regs) // Fill values for all registers. for { var si arch.SignalInfo @@ -263,7 +263,7 @@ func TestRegistersFault(t *testing.T) { } func TestSegments(t *testing.T) { - applicationTest(t, true, testutil.TwiddleSegments, func(c *vCPU, regs *syscall.PtraceRegs, pt *pagetables.PageTables) bool { + applicationTest(t, true, testutil.TwiddleSegments, func(c *vCPU, regs *arch.Registers, pt *pagetables.PageTables) bool { testutil.SetTestSegments(regs) for { var si arch.SignalInfo @@ -287,7 +287,7 @@ func TestSegments(t *testing.T) { } func TestBounce(t *testing.T) { - applicationTest(t, true, testutil.SpinLoop, func(c *vCPU, regs *syscall.PtraceRegs, pt *pagetables.PageTables) bool { + applicationTest(t, true, testutil.SpinLoop, func(c *vCPU, regs *arch.Registers, pt *pagetables.PageTables) bool { go func() { time.Sleep(time.Millisecond) c.BounceToKernel() @@ -302,7 +302,7 @@ func TestBounce(t *testing.T) { } return false }) - applicationTest(t, true, testutil.SpinLoop, func(c *vCPU, regs *syscall.PtraceRegs, pt *pagetables.PageTables) bool { + applicationTest(t, true, testutil.SpinLoop, func(c *vCPU, regs *arch.Registers, pt *pagetables.PageTables) bool { go func() { time.Sleep(time.Millisecond) c.BounceToKernel() @@ -321,7 +321,7 @@ func TestBounce(t *testing.T) { } func TestBounceStress(t *testing.T) { - applicationTest(t, true, testutil.SpinLoop, func(c *vCPU, regs *syscall.PtraceRegs, pt *pagetables.PageTables) bool { + applicationTest(t, true, testutil.SpinLoop, func(c *vCPU, regs *arch.Registers, pt *pagetables.PageTables) bool { randomSleep := func() { // O(hundreds of microseconds) is appropriate to ensure // different overlaps and different schedules. @@ -357,7 +357,7 @@ func TestBounceStress(t *testing.T) { func TestInvalidate(t *testing.T) { var data uintptr // Used below. - applicationTest(t, true, testutil.Touch, func(c *vCPU, regs *syscall.PtraceRegs, pt *pagetables.PageTables) bool { + applicationTest(t, true, testutil.Touch, func(c *vCPU, regs *arch.Registers, pt *pagetables.PageTables) bool { testutil.SetTouchTarget(regs, &data) // Read legitimate value. for { var si arch.SignalInfo @@ -398,7 +398,7 @@ func IsFault(err error, si *arch.SignalInfo) bool { } func TestEmptyAddressSpace(t *testing.T) { - applicationTest(t, false, testutil.SyscallLoop, func(c *vCPU, regs *syscall.PtraceRegs, pt *pagetables.PageTables) bool { + applicationTest(t, false, testutil.SyscallLoop, func(c *vCPU, regs *arch.Registers, pt *pagetables.PageTables) bool { var si arch.SignalInfo if _, err := c.SwitchToUser(ring0.SwitchOpts{ Registers: regs, @@ -412,7 +412,7 @@ func TestEmptyAddressSpace(t *testing.T) { } return false }) - applicationTest(t, false, testutil.SyscallLoop, func(c *vCPU, regs *syscall.PtraceRegs, pt *pagetables.PageTables) bool { + applicationTest(t, false, testutil.SyscallLoop, func(c *vCPU, regs *arch.Registers, pt *pagetables.PageTables) bool { var si arch.SignalInfo if _, err := c.SwitchToUser(ring0.SwitchOpts{ Registers: regs, @@ -471,7 +471,7 @@ func BenchmarkApplicationSyscall(b *testing.B) { i int // Iteration includes machine.Get() / machine.Put(). a int // Count for ErrContextInterrupt. ) - applicationTest(b, true, testutil.SyscallLoop, func(c *vCPU, regs *syscall.PtraceRegs, pt *pagetables.PageTables) bool { + applicationTest(b, true, testutil.SyscallLoop, func(c *vCPU, regs *arch.Registers, pt *pagetables.PageTables) bool { var si arch.SignalInfo if _, err := c.SwitchToUser(ring0.SwitchOpts{ Registers: regs, @@ -493,7 +493,7 @@ func BenchmarkApplicationSyscall(b *testing.B) { func BenchmarkKernelSyscall(b *testing.B) { // Note that the target passed here is irrelevant, we never execute SwitchToUser. - applicationTest(b, true, testutil.Getpid, func(c *vCPU, regs *syscall.PtraceRegs, pt *pagetables.PageTables) bool { + applicationTest(b, true, testutil.Getpid, func(c *vCPU, regs *arch.Registers, pt *pagetables.PageTables) bool { // iteration does not include machine.Get() / machine.Put(). for i := 0; i < b.N; i++ { testutil.Getpid() @@ -508,7 +508,7 @@ func BenchmarkWorldSwitchToUserRoundtrip(b *testing.B) { i int a int ) - applicationTest(b, true, testutil.SyscallLoop, func(c *vCPU, regs *syscall.PtraceRegs, pt *pagetables.PageTables) bool { + applicationTest(b, true, testutil.SyscallLoop, func(c *vCPU, regs *arch.Registers, pt *pagetables.PageTables) bool { var si arch.SignalInfo if _, err := c.SwitchToUser(ring0.SwitchOpts{ Registers: regs, diff --git a/pkg/sentry/platform/kvm/machine.go b/pkg/sentry/platform/kvm/machine.go index 8076c7529..6c54712d1 100644 --- a/pkg/sentry/platform/kvm/machine.go +++ b/pkg/sentry/platform/kvm/machine.go @@ -52,16 +52,19 @@ type machine struct { // available is notified when vCPUs are available. available sync.Cond - // vCPUs are the machine vCPUs. + // vCPUsByTID are the machine vCPUs. // // These are populated dynamically. - vCPUs map[uint64]*vCPU + vCPUsByTID map[uint64]*vCPU // vCPUsByID are the machine vCPUs, can be indexed by the vCPU's ID. - vCPUsByID map[int]*vCPU + vCPUsByID []*vCPU // maxVCPUs is the maximum number of vCPUs supported by the machine. maxVCPUs int + + // nextID is the next vCPU ID. + nextID uint32 } const ( @@ -137,9 +140,8 @@ type dieState struct { // // Precondition: mu must be held. func (m *machine) newVCPU() *vCPU { - id := len(m.vCPUs) - // Create the vCPU. + id := int(atomic.AddUint32(&m.nextID, 1) - 1) fd, _, errno := syscall.RawSyscall(syscall.SYS_IOCTL, uintptr(m.fd), _KVM_CREATE_VCPU, uintptr(id)) if errno != 0 { panic(fmt.Sprintf("error creating new vCPU: %v", errno)) @@ -176,11 +178,7 @@ func (m *machine) newVCPU() *vCPU { // newMachine returns a new VM context. func newMachine(vm int) (*machine, error) { // Create the machine. - m := &machine{ - fd: vm, - vCPUs: make(map[uint64]*vCPU), - vCPUsByID: make(map[int]*vCPU), - } + m := &machine{fd: vm} m.available.L = &m.mu m.kernel.Init(ring0.KernelOpts{ PageTables: pagetables.New(newAllocator()), @@ -194,6 +192,10 @@ func newMachine(vm int) (*machine, error) { } log.Debugf("The maximum number of vCPUs is %d.", m.maxVCPUs) + // Create the vCPUs map/slices. + m.vCPUsByTID = make(map[uint64]*vCPU) + m.vCPUsByID = make([]*vCPU, m.maxVCPUs) + // Apply the physical mappings. Note that these mappings may point to // guest physical addresses that are not actually available. These // physical pages are mapped on demand, see kernel_unsafe.go. @@ -274,6 +276,8 @@ func newMachine(vm int) (*machine, error) { // not available. This attempts to be efficient for calls in the hot path. // // This panics on error. +// +//go:nosplit func (m *machine) mapPhysical(physical, length uintptr, phyRegions []physicalRegion, flags uint32) { for end := physical + length; physical < end; { _, physicalStart, length, ok := calculateBluepillFault(physical, phyRegions) @@ -304,7 +308,11 @@ func (m *machine) Destroy() { runtime.SetFinalizer(m, nil) // Destroy vCPUs. - for _, c := range m.vCPUs { + for _, c := range m.vCPUsByID { + if c == nil { + continue + } + // Ensure the vCPU is not still running in guest mode. This is // possible iff teardown has been done by other threads, and // somehow a single thread has not executed any system calls. @@ -329,13 +337,15 @@ func (m *machine) Destroy() { } // Get gets an available vCPU. +// +// This will return with the OS thread locked. func (m *machine) Get() *vCPU { + m.mu.RLock() runtime.LockOSThread() tid := procid.Current() - m.mu.RLock() // Check for an exact match. - if c := m.vCPUs[tid]; c != nil { + if c := m.vCPUsByTID[tid]; c != nil { c.lock() m.mu.RUnlock() return c @@ -343,15 +353,29 @@ func (m *machine) Get() *vCPU { // The happy path failed. We now proceed to acquire an exclusive lock // (because the vCPU map may change), and scan all available vCPUs. + // In this case, we first unlock the OS thread. Otherwise, if mu is + // not available, the current system thread will be parked and a new + // system thread spawned. We avoid this situation by simply refreshing + // tid after relocking the system thread. m.mu.RUnlock() + runtime.UnlockOSThread() m.mu.Lock() + runtime.LockOSThread() + tid = procid.Current() + + // Recheck for an exact match. + if c := m.vCPUsByTID[tid]; c != nil { + c.lock() + m.mu.Unlock() + return c + } for { // Scan for an available vCPU. - for origTID, c := range m.vCPUs { + for origTID, c := range m.vCPUsByTID { if atomic.CompareAndSwapUint32(&c.state, vCPUReady, vCPUUser) { - delete(m.vCPUs, origTID) - m.vCPUs[tid] = c + delete(m.vCPUsByTID, origTID) + m.vCPUsByTID[tid] = c m.mu.Unlock() c.loadSegments(tid) return c @@ -359,17 +383,17 @@ func (m *machine) Get() *vCPU { } // Create a new vCPU (maybe). - if len(m.vCPUs) < m.maxVCPUs { + if int(m.nextID) < m.maxVCPUs { c := m.newVCPU() c.lock() - m.vCPUs[tid] = c + m.vCPUsByTID[tid] = c m.mu.Unlock() c.loadSegments(tid) return c } // Scan for something not in user mode. - for origTID, c := range m.vCPUs { + for origTID, c := range m.vCPUsByTID { if !atomic.CompareAndSwapUint32(&c.state, vCPUGuest, vCPUGuest|vCPUWaiter) { continue } @@ -387,8 +411,8 @@ func (m *machine) Get() *vCPU { } // Steal the vCPU. - delete(m.vCPUs, origTID) - m.vCPUs[tid] = c + delete(m.vCPUsByTID, origTID) + m.vCPUsByTID[tid] = c m.mu.Unlock() c.loadSegments(tid) return c @@ -415,7 +439,7 @@ func (m *machine) Put(c *vCPU) { // newDirtySet returns a new dirty set. func (m *machine) newDirtySet() *dirtySet { return &dirtySet{ - vCPUs: make([]uint64, (m.maxVCPUs+63)/64, (m.maxVCPUs+63)/64), + vCPUMasks: make([]uint64, (m.maxVCPUs+63)/64, (m.maxVCPUs+63)/64), } } diff --git a/pkg/sentry/platform/kvm/machine_amd64.go b/pkg/sentry/platform/kvm/machine_amd64.go index 923ce3909..acc823ba6 100644 --- a/pkg/sentry/platform/kvm/machine_amd64.go +++ b/pkg/sentry/platform/kvm/machine_amd64.go @@ -51,9 +51,10 @@ func (m *machine) initArchState() error { recover() debug.SetPanicOnFault(old) }() - m.retryInGuest(func() { - ring0.SetCPUIDFaulting(true) - }) + c := m.Get() + defer m.Put(c) + bluepill(c) + ring0.SetCPUIDFaulting(true) return nil } @@ -89,8 +90,8 @@ func (m *machine) dropPageTables(pt *pagetables.PageTables) { defer m.mu.Unlock() // Clear from all PCIDs. - for _, c := range m.vCPUs { - if c.PCIDs != nil { + for _, c := range m.vCPUsByID { + if c != nil && c.PCIDs != nil { c.PCIDs.Drop(pt) } } @@ -335,29 +336,6 @@ func (c *vCPU) SwitchToUser(switchOpts ring0.SwitchOpts, info *arch.SignalInfo) } } -// retryInGuest runs the given function in guest mode. -// -// If the function does not complete in guest mode (due to execution of a -// system call due to a GC stall, for example), then it will be retried. The -// given function must be idempotent as a result of the retry mechanism. -func (m *machine) retryInGuest(fn func()) { - c := m.Get() - defer m.Put(c) - for { - c.ClearErrorCode() // See below. - bluepill(c) // Force guest mode. - fn() // Execute the given function. - _, user := c.ErrorCode() - if user { - // If user is set, then we haven't bailed back to host - // mode via a kernel exception or system call. We - // consider the full function to have executed in guest - // mode and we can return. - break - } - } -} - // On x86 platform, the flags for "setMemoryRegion" can always be set as 0. // There is no need to return read-only physicalRegions. func rdonlyRegionsForSetMem() (phyRegions []physicalRegion) { diff --git a/pkg/sentry/platform/kvm/machine_amd64_unsafe.go b/pkg/sentry/platform/kvm/machine_amd64_unsafe.go index 7156c245f..290f035dd 100644 --- a/pkg/sentry/platform/kvm/machine_amd64_unsafe.go +++ b/pkg/sentry/platform/kvm/machine_amd64_unsafe.go @@ -154,7 +154,7 @@ func (c *vCPU) setUserRegisters(uregs *userRegs) error { // //go:nosplit func (c *vCPU) getUserRegisters(uregs *userRegs) syscall.Errno { - if _, _, errno := syscall.RawSyscall( + if _, _, errno := syscall.RawSyscall( // escapes: no. syscall.SYS_IOCTL, uintptr(c.fd), _KVM_GET_REGS, diff --git a/pkg/sentry/platform/kvm/machine_arm64.go b/pkg/sentry/platform/kvm/machine_arm64.go index 09552837a..750751aa3 100644 --- a/pkg/sentry/platform/kvm/machine_arm64.go +++ b/pkg/sentry/platform/kvm/machine_arm64.go @@ -28,6 +28,10 @@ type vCPUArchState struct { // // This starts above fixedKernelPCID. PCIDs *pagetables.PCIDs + + // floatingPointState is the floating point state buffer used in guest + // to host transitions. See usage in bluepill_arm64.go. + floatingPointState *arch.FloatingPointData } const ( @@ -56,6 +60,12 @@ func rdonlyRegionsForSetMem() (phyRegions []physicalRegion) { if !vr.accessType.Write && vr.accessType.Read { rdonlyRegions = append(rdonlyRegions, vr.region) } + + // TODO(gvisor.dev/issue/2686): PROT_NONE should be specially treated. + // Workaround: treated as rdonly temporarily. + if !vr.accessType.Write && !vr.accessType.Read && !vr.accessType.Execute { + rdonlyRegions = append(rdonlyRegions, vr.region) + } }) for _, r := range rdonlyRegions { diff --git a/pkg/sentry/platform/kvm/machine_arm64_unsafe.go b/pkg/sentry/platform/kvm/machine_arm64_unsafe.go index 1c8384e6b..3c02cef7c 100644 --- a/pkg/sentry/platform/kvm/machine_arm64_unsafe.go +++ b/pkg/sentry/platform/kvm/machine_arm64_unsafe.go @@ -29,30 +29,6 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) -// setMemoryRegion initializes a region. -// -// This may be called from bluepillHandler, and therefore returns an errno -// directly (instead of wrapping in an error) to avoid allocations. -// -//go:nosplit -func (m *machine) setMemoryRegion(slot int, physical, length, virtual uintptr) syscall.Errno { - userRegion := userMemoryRegion{ - slot: uint32(slot), - flags: 0, - guestPhysAddr: uint64(physical), - memorySize: uint64(length), - userspaceAddr: uint64(virtual), - } - - // Set the region. - _, _, errno := syscall.RawSyscall( - syscall.SYS_IOCTL, - uintptr(m.fd), - _KVM_SET_USER_MEMORY_REGION, - uintptr(unsafe.Pointer(&userRegion))) - return errno -} - type kvmVcpuInit struct { target uint32 features [7]uint32 @@ -72,69 +48,6 @@ func (m *machine) initArchState() error { return nil } -func getPageWithReflect(p uintptr) []byte { - return (*(*[0xFFFFFF]byte)(unsafe.Pointer(p & ^uintptr(syscall.Getpagesize()-1))))[:syscall.Getpagesize()] -} - -// Work around: move ring0.Vectors() into a specific address with 11-bits alignment. -// -// According to the design documentation of Arm64, -// the start address of exception vector table should be 11-bits aligned. -// Please see the code in linux kernel as reference: arch/arm64/kernel/entry.S -// But, we can't align a function's start address to a specific address by using golang. -// We have raised this question in golang community: -// https://groups.google.com/forum/m/#!topic/golang-dev/RPj90l5x86I -// This function will be removed when golang supports this feature. -// -// There are 2 jobs were implemented in this function: -// 1, move the start address of exception vector table into the specific address. -// 2, modify the offset of each instruction. -func updateVectorTable() { - fromLocation := reflect.ValueOf(ring0.Vectors).Pointer() - offset := fromLocation & (1<<11 - 1) - if offset != 0 { - offset = 1<<11 - offset - } - - toLocation := fromLocation + offset - page := getPageWithReflect(toLocation) - if err := syscall.Mprotect(page, syscall.PROT_READ|syscall.PROT_WRITE|syscall.PROT_EXEC); err != nil { - panic(err) - } - - page = getPageWithReflect(toLocation + 4096) - if err := syscall.Mprotect(page, syscall.PROT_READ|syscall.PROT_WRITE|syscall.PROT_EXEC); err != nil { - panic(err) - } - - // Move exception-vector-table into the specific address. - var entry *uint32 - var entryFrom *uint32 - for i := 1; i <= 0x800; i++ { - entry = (*uint32)(unsafe.Pointer(toLocation + 0x800 - uintptr(i))) - entryFrom = (*uint32)(unsafe.Pointer(fromLocation + 0x800 - uintptr(i))) - *entry = *entryFrom - } - - // The offset from the address of each unconditionally branch is changed. - // We should modify the offset of each instruction. - nums := []uint32{0x0, 0x80, 0x100, 0x180, 0x200, 0x280, 0x300, 0x380, 0x400, 0x480, 0x500, 0x580, 0x600, 0x680, 0x700, 0x780} - for _, num := range nums { - entry = (*uint32)(unsafe.Pointer(toLocation + uintptr(num))) - *entry = *entry - (uint32)(offset/4) - } - - page = getPageWithReflect(toLocation) - if err := syscall.Mprotect(page, syscall.PROT_READ|syscall.PROT_EXEC); err != nil { - panic(err) - } - - page = getPageWithReflect(toLocation + 4096) - if err := syscall.Mprotect(page, syscall.PROT_READ|syscall.PROT_EXEC); err != nil { - panic(err) - } -} - // initArchState initializes architecture-specific state. func (c *vCPU) initArchState() error { var ( @@ -147,6 +60,7 @@ func (c *vCPU) initArchState() error { reg.addr = uint64(reflect.ValueOf(&data).Pointer()) regGet.addr = uint64(reflect.ValueOf(&dataGet).Pointer()) + vcpuInit.target = _KVM_ARM_TARGET_GENERIC_V8 vcpuInit.features[0] |= (1 << _KVM_ARM_VCPU_PSCI_0_2) if _, _, errno := syscall.RawSyscall( syscall.SYS_IOCTL, @@ -158,7 +72,8 @@ func (c *vCPU) initArchState() error { // cpacr_el1 reg.id = _KVM_ARM64_REGS_CPACR_EL1 - data = (_FPEN_NOTRAP << _FPEN_SHIFT) + // It is off by default, and it is turned on only when in use. + data = 0 // Disable fpsimd. if err := c.setOneRegister(®); err != nil { return err } @@ -250,6 +165,7 @@ func (c *vCPU) initArchState() error { return err } + c.floatingPointState = arch.NewFloatingPointData() return nil } diff --git a/pkg/sentry/platform/kvm/machine_unsafe.go b/pkg/sentry/platform/kvm/machine_unsafe.go index f04be2ab5..9f86f6a7a 100644 --- a/pkg/sentry/platform/kvm/machine_unsafe.go +++ b/pkg/sentry/platform/kvm/machine_unsafe.go @@ -13,7 +13,7 @@ // limitations under the License. // +build go1.12 -// +build !go1.15 +// +build !go1.16 // Check go:linkname function signatures when updating Go version. @@ -115,7 +115,7 @@ func (a *atomicAddressSpace) get() *addressSpace { // //go:nosplit func (c *vCPU) notify() { - _, _, errno := syscall.RawSyscall6( + _, _, errno := syscall.RawSyscall6( // escapes: no. syscall.SYS_FUTEX, uintptr(unsafe.Pointer(&c.state)), linux.FUTEX_WAKE|linux.FUTEX_PRIVATE_FLAG, diff --git a/pkg/sentry/platform/kvm/testutil/BUILD b/pkg/sentry/platform/kvm/testutil/BUILD index f7605df8a..f7feb8683 100644 --- a/pkg/sentry/platform/kvm/testutil/BUILD +++ b/pkg/sentry/platform/kvm/testutil/BUILD @@ -13,4 +13,5 @@ go_library( "testutil_arm64.s", ], visibility = ["//pkg/sentry/platform/kvm:__pkg__"], + deps = ["//pkg/sentry/arch"], ) diff --git a/pkg/sentry/platform/kvm/testutil/testutil_amd64.go b/pkg/sentry/platform/kvm/testutil/testutil_amd64.go index 4c108abbf..8048eedec 100644 --- a/pkg/sentry/platform/kvm/testutil/testutil_amd64.go +++ b/pkg/sentry/platform/kvm/testutil/testutil_amd64.go @@ -18,19 +18,20 @@ package testutil import ( "reflect" - "syscall" + + "gvisor.dev/gvisor/pkg/sentry/arch" ) // TwiddleSegments reads segments into known registers. func TwiddleSegments() // SetTestTarget sets the rip appropriately. -func SetTestTarget(regs *syscall.PtraceRegs, fn func()) { +func SetTestTarget(regs *arch.Registers, fn func()) { regs.Rip = uint64(reflect.ValueOf(fn).Pointer()) } // SetTouchTarget sets rax appropriately. -func SetTouchTarget(regs *syscall.PtraceRegs, target *uintptr) { +func SetTouchTarget(regs *arch.Registers, target *uintptr) { if target != nil { regs.Rax = uint64(reflect.ValueOf(target).Pointer()) } else { @@ -39,12 +40,12 @@ func SetTouchTarget(regs *syscall.PtraceRegs, target *uintptr) { } // RewindSyscall rewinds a syscall RIP. -func RewindSyscall(regs *syscall.PtraceRegs) { +func RewindSyscall(regs *arch.Registers) { regs.Rip -= 2 } // SetTestRegs initializes registers to known values. -func SetTestRegs(regs *syscall.PtraceRegs) { +func SetTestRegs(regs *arch.Registers) { regs.R15 = 0x15 regs.R14 = 0x14 regs.R13 = 0x13 @@ -64,7 +65,7 @@ func SetTestRegs(regs *syscall.PtraceRegs) { } // CheckTestRegs checks that registers were twiddled per TwiddleRegs. -func CheckTestRegs(regs *syscall.PtraceRegs, full bool) (err error) { +func CheckTestRegs(regs *arch.Registers, full bool) (err error) { if need := ^uint64(0x15); regs.R15 != need { err = addRegisterMismatch(err, "R15", regs.R15, need) } @@ -121,13 +122,13 @@ var fsData uint64 = 0x55 var gsData uint64 = 0x85 // SetTestSegments initializes segments to known values. -func SetTestSegments(regs *syscall.PtraceRegs) { +func SetTestSegments(regs *arch.Registers) { regs.Fs_base = uint64(reflect.ValueOf(&fsData).Pointer()) regs.Gs_base = uint64(reflect.ValueOf(&gsData).Pointer()) } // CheckTestSegments checks that registers were twiddled per TwiddleSegments. -func CheckTestSegments(regs *syscall.PtraceRegs) (err error) { +func CheckTestSegments(regs *arch.Registers) (err error) { if regs.Rax != fsData { err = addRegisterMismatch(err, "Rax", regs.Rax, fsData) } diff --git a/pkg/sentry/platform/kvm/testutil/testutil_arm64.go b/pkg/sentry/platform/kvm/testutil/testutil_arm64.go index 40b2e4acc..ca902c8c1 100644 --- a/pkg/sentry/platform/kvm/testutil/testutil_arm64.go +++ b/pkg/sentry/platform/kvm/testutil/testutil_arm64.go @@ -19,16 +19,17 @@ package testutil import ( "fmt" "reflect" - "syscall" + + "gvisor.dev/gvisor/pkg/sentry/arch" ) // SetTestTarget sets the rip appropriately. -func SetTestTarget(regs *syscall.PtraceRegs, fn func()) { +func SetTestTarget(regs *arch.Registers, fn func()) { regs.Pc = uint64(reflect.ValueOf(fn).Pointer()) } // SetTouchTarget sets rax appropriately. -func SetTouchTarget(regs *syscall.PtraceRegs, target *uintptr) { +func SetTouchTarget(regs *arch.Registers, target *uintptr) { if target != nil { regs.Regs[8] = uint64(reflect.ValueOf(target).Pointer()) } else { @@ -37,19 +38,19 @@ func SetTouchTarget(regs *syscall.PtraceRegs, target *uintptr) { } // RewindSyscall rewinds a syscall RIP. -func RewindSyscall(regs *syscall.PtraceRegs) { +func RewindSyscall(regs *arch.Registers) { regs.Pc -= 4 } // SetTestRegs initializes registers to known values. -func SetTestRegs(regs *syscall.PtraceRegs) { +func SetTestRegs(regs *arch.Registers) { for i := 0; i <= 30; i++ { regs.Regs[i] = uint64(i) + 1 } } // CheckTestRegs checks that registers were twiddled per TwiddleRegs. -func CheckTestRegs(regs *syscall.PtraceRegs, full bool) (err error) { +func CheckTestRegs(regs *arch.Registers, full bool) (err error) { for i := 0; i <= 30; i++ { if need := ^uint64(i + 1); regs.Regs[i] != need { err = addRegisterMismatch(err, fmt.Sprintf("R%d", i), regs.Regs[i], need) diff --git a/pkg/sentry/platform/platform.go b/pkg/sentry/platform/platform.go index 2ca696382..171513f3f 100644 --- a/pkg/sentry/platform/platform.go +++ b/pkg/sentry/platform/platform.go @@ -148,6 +148,9 @@ type Context interface { // Interrupt interrupts a concurrent call to Switch(), causing it to return // ErrContextInterrupt. Interrupt() + + // Release() releases any resources associated with this context. + Release() } var ( @@ -353,10 +356,28 @@ func (fr FileRange) String() string { return fmt.Sprintf("[%#x, %#x)", fr.Start, fr.End) } +// Requirements is used to specify platform specific requirements. +type Requirements struct { + // RequiresCurrentPIDNS indicates that the sandbox has to be started in the + // current pid namespace. + RequiresCurrentPIDNS bool + // RequiresCapSysPtrace indicates that the sandbox has to be started with + // the CAP_SYS_PTRACE capability. + RequiresCapSysPtrace bool +} + // Constructor represents a platform type. type Constructor interface { + // New returns a new platform instance. + // + // Arguments: + // + // * deviceFile - the device file (e.g. /dev/kvm for the KVM platform). New(deviceFile *os.File) (Platform, error) OpenDevice() (*os.File, error) + + // Requirements returns platform specific requirements. + Requirements() Requirements } // platforms contains all available platform types. diff --git a/pkg/sentry/platform/ptrace/BUILD b/pkg/sentry/platform/ptrace/BUILD index 95abd321e..30402c2df 100644 --- a/pkg/sentry/platform/ptrace/BUILD +++ b/pkg/sentry/platform/ptrace/BUILD @@ -9,6 +9,7 @@ go_library( "ptrace.go", "ptrace_amd64.go", "ptrace_arm64.go", + "ptrace_arm64_unsafe.go", "ptrace_unsafe.go", "stub_amd64.s", "stub_arm64.s", diff --git a/pkg/sentry/platform/ptrace/ptrace.go b/pkg/sentry/platform/ptrace/ptrace.go index 03adb624b..08d055e05 100644 --- a/pkg/sentry/platform/ptrace/ptrace.go +++ b/pkg/sentry/platform/ptrace/ptrace.go @@ -177,6 +177,9 @@ func (c *context) Interrupt() { c.interrupt.NotifyInterrupt() } +// Release implements platform.Context.Release(). +func (c *context) Release() {} + // PTrace represents a collection of ptrace subprocesses. type PTrace struct { platform.MMapMinAddr @@ -248,6 +251,16 @@ func (*constructor) OpenDevice() (*os.File, error) { return nil, nil } +// Flags implements platform.Constructor.Flags(). +func (*constructor) Requirements() platform.Requirements { + // TODO(b/75837838): Also set a new PID namespace so that we limit + // access to other host processes. + return platform.Requirements{ + RequiresCapSysPtrace: true, + RequiresCurrentPIDNS: true, + } +} + func init() { platform.Register("ptrace", &constructor{}) } diff --git a/pkg/sentry/platform/ptrace/ptrace_amd64.go b/pkg/sentry/platform/ptrace/ptrace_amd64.go index db0212538..3b9a870a5 100644 --- a/pkg/sentry/platform/ptrace/ptrace_amd64.go +++ b/pkg/sentry/platform/ptrace/ptrace_amd64.go @@ -15,9 +15,8 @@ package ptrace import ( - "syscall" - "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/arch" ) // fpRegSet returns the GETREGSET/SETREGSET register set type to be used. @@ -28,6 +27,20 @@ func fpRegSet(useXsave bool) uintptr { return linux.NT_PRFPREG } -func stackPointer(r *syscall.PtraceRegs) uintptr { +func stackPointer(r *arch.Registers) uintptr { return uintptr(r.Rsp) } + +// x86 use the fs_base register to store the TLS pointer which can be +// get/set in "func (t *thread) get/setRegs(regs *arch.Registers)". +// So both of the get/setTLS() operations are noop here. + +// getTLS gets the thread local storage register. +func (t *thread) getTLS(tls *uint64) error { + return nil +} + +// setTLS sets the thread local storage register. +func (t *thread) setTLS(tls *uint64) error { + return nil +} diff --git a/pkg/sentry/platform/ptrace/ptrace_arm64.go b/pkg/sentry/platform/ptrace/ptrace_arm64.go index 4db28c534..5c869926a 100644 --- a/pkg/sentry/platform/ptrace/ptrace_arm64.go +++ b/pkg/sentry/platform/ptrace/ptrace_arm64.go @@ -15,9 +15,8 @@ package ptrace import ( - "syscall" - "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/arch" ) // fpRegSet returns the GETREGSET/SETREGSET register set type to be used. @@ -25,6 +24,6 @@ func fpRegSet(_ bool) uintptr { return linux.NT_PRFPREG } -func stackPointer(r *syscall.PtraceRegs) uintptr { +func stackPointer(r *arch.Registers) uintptr { return uintptr(r.Sp) } diff --git a/pkg/sentry/platform/ptrace/ptrace_arm64_unsafe.go b/pkg/sentry/platform/ptrace/ptrace_arm64_unsafe.go new file mode 100644 index 000000000..32b8a6be9 --- /dev/null +++ b/pkg/sentry/platform/ptrace/ptrace_arm64_unsafe.go @@ -0,0 +1,62 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// +build arm64 + +package ptrace + +import ( + "syscall" + "unsafe" + + "gvisor.dev/gvisor/pkg/abi/linux" +) + +// getTLS gets the thread local storage register. +func (t *thread) getTLS(tls *uint64) error { + iovec := syscall.Iovec{ + Base: (*byte)(unsafe.Pointer(tls)), + Len: uint64(unsafe.Sizeof(*tls)), + } + _, _, errno := syscall.RawSyscall6( + syscall.SYS_PTRACE, + syscall.PTRACE_GETREGSET, + uintptr(t.tid), + linux.NT_ARM_TLS, + uintptr(unsafe.Pointer(&iovec)), + 0, 0) + if errno != 0 { + return errno + } + return nil +} + +// setTLS sets the thread local storage register. +func (t *thread) setTLS(tls *uint64) error { + iovec := syscall.Iovec{ + Base: (*byte)(unsafe.Pointer(tls)), + Len: uint64(unsafe.Sizeof(*tls)), + } + _, _, errno := syscall.RawSyscall6( + syscall.SYS_PTRACE, + syscall.PTRACE_SETREGSET, + uintptr(t.tid), + linux.NT_ARM_TLS, + uintptr(unsafe.Pointer(&iovec)), + 0, 0) + if errno != 0 { + return errno + } + return nil +} diff --git a/pkg/sentry/platform/ptrace/ptrace_unsafe.go b/pkg/sentry/platform/ptrace/ptrace_unsafe.go index 6c0ed7b3e..8b72d24e8 100644 --- a/pkg/sentry/platform/ptrace/ptrace_unsafe.go +++ b/pkg/sentry/platform/ptrace/ptrace_unsafe.go @@ -24,7 +24,7 @@ import ( ) // getRegs gets the general purpose register set. -func (t *thread) getRegs(regs *syscall.PtraceRegs) error { +func (t *thread) getRegs(regs *arch.Registers) error { iovec := syscall.Iovec{ Base: (*byte)(unsafe.Pointer(regs)), Len: uint64(unsafe.Sizeof(*regs)), @@ -43,7 +43,7 @@ func (t *thread) getRegs(regs *syscall.PtraceRegs) error { } // setRegs sets the general purpose register set. -func (t *thread) setRegs(regs *syscall.PtraceRegs) error { +func (t *thread) setRegs(regs *arch.Registers) error { iovec := syscall.Iovec{ Base: (*byte)(unsafe.Pointer(regs)), Len: uint64(unsafe.Sizeof(*regs)), diff --git a/pkg/sentry/platform/ptrace/subprocess.go b/pkg/sentry/platform/ptrace/subprocess.go index 31b7cec53..2389423b0 100644 --- a/pkg/sentry/platform/ptrace/subprocess.go +++ b/pkg/sentry/platform/ptrace/subprocess.go @@ -63,7 +63,7 @@ type thread struct { // initRegs are the initial registers for the first thread. // // These are used for the register set for system calls. - initRegs syscall.PtraceRegs + initRegs arch.Registers } // threadPool is a collection of threads. @@ -317,7 +317,7 @@ const ( ) func (t *thread) dumpAndPanic(message string) { - var regs syscall.PtraceRegs + var regs arch.Registers message += "\n" if err := t.getRegs(®s); err == nil { message += dumpRegs(®s) @@ -332,7 +332,7 @@ func (t *thread) unexpectedStubExit() { msg, err := t.getEventMessage() status := syscall.WaitStatus(msg) if status.Signaled() && status.Signal() == syscall.SIGKILL { - // SIGKILL can be only sent by an user or OOM-killer. In both + // SIGKILL can be only sent by a user or OOM-killer. In both // these cases, we don't need to panic. There is no reasons to // think that something wrong in gVisor. log.Warningf("The ptrace stub process %v has been killed by SIGKILL.", t.tgid) @@ -423,7 +423,7 @@ func (t *thread) init() { // This is _not_ for use by application system calls, rather it is for use when // a system call must be injected into the remote context (e.g. mmap, munmap). // Note that clones are handled separately. -func (t *thread) syscall(regs *syscall.PtraceRegs) (uintptr, error) { +func (t *thread) syscall(regs *arch.Registers) (uintptr, error) { // Set registers. if err := t.setRegs(regs); err != nil { panic(fmt.Sprintf("ptrace set regs failed: %v", err)) @@ -461,7 +461,7 @@ func (t *thread) syscall(regs *syscall.PtraceRegs) (uintptr, error) { // syscallIgnoreInterrupt ignores interrupts on the system call thread and // restarts the syscall if the kernel indicates that should happen. func (t *thread) syscallIgnoreInterrupt( - initRegs *syscall.PtraceRegs, + initRegs *arch.Registers, sysno uintptr, args ...arch.SyscallArgument) (uintptr, error) { for { @@ -506,6 +506,9 @@ func (s *subprocess) switchToApp(c *context, ac arch.Context) bool { regs := &ac.StateData().Regs t.resetSysemuRegs(regs) + // Extract TLS register + tls := uint64(ac.TLS()) + // Check for interrupts, and ensure that future interrupts will signal t. if !c.interrupt.Enable(t) { // Pending interrupt; simulate. @@ -526,6 +529,9 @@ func (s *subprocess) switchToApp(c *context, ac arch.Context) bool { if err := t.setFPRegs(fpState, uint64(fpLen), useXsave); err != nil { panic(fmt.Sprintf("ptrace set fpregs (%+v) failed: %v", fpState, err)) } + if err := t.setTLS(&tls); err != nil { + panic(fmt.Sprintf("ptrace set tls (%+v) failed: %v", tls, err)) + } for { // Start running until the next system call. @@ -555,6 +561,12 @@ func (s *subprocess) switchToApp(c *context, ac arch.Context) bool { if err := t.getFPRegs(fpState, uint64(fpLen), useXsave); err != nil { panic(fmt.Sprintf("ptrace get fpregs failed: %v", err)) } + if err := t.getTLS(&tls); err != nil { + panic(fmt.Sprintf("ptrace get tls failed: %v", err)) + } + if !ac.SetTLS(uintptr(tls)) { + panic(fmt.Sprintf("tls value %v is invalid", tls)) + } // Is it a system call? if sig == (syscallEvent | syscall.SIGTRAP) { diff --git a/pkg/sentry/platform/ptrace/subprocess_amd64.go b/pkg/sentry/platform/ptrace/subprocess_amd64.go index e99798c56..84b699f0d 100644 --- a/pkg/sentry/platform/ptrace/subprocess_amd64.go +++ b/pkg/sentry/platform/ptrace/subprocess_amd64.go @@ -21,6 +21,7 @@ import ( "strings" "syscall" + "golang.org/x/sys/unix" "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/seccomp" "gvisor.dev/gvisor/pkg/sentry/arch" @@ -40,7 +41,7 @@ const ( // resetSysemuRegs sets up emulation registers. // // This should be called prior to calling sysemu. -func (t *thread) resetSysemuRegs(regs *syscall.PtraceRegs) { +func (t *thread) resetSysemuRegs(regs *arch.Registers) { regs.Cs = t.initRegs.Cs regs.Ss = t.initRegs.Ss regs.Ds = t.initRegs.Ds @@ -52,7 +53,7 @@ func (t *thread) resetSysemuRegs(regs *syscall.PtraceRegs) { // createSyscallRegs sets up syscall registers. // // This should be called to generate registers for a system call. -func createSyscallRegs(initRegs *syscall.PtraceRegs, sysno uintptr, args ...arch.SyscallArgument) syscall.PtraceRegs { +func createSyscallRegs(initRegs *arch.Registers, sysno uintptr, args ...arch.SyscallArgument) arch.Registers { // Copy initial registers. regs := *initRegs @@ -81,18 +82,18 @@ func createSyscallRegs(initRegs *syscall.PtraceRegs, sysno uintptr, args ...arch } // isSingleStepping determines if the registers indicate single-stepping. -func isSingleStepping(regs *syscall.PtraceRegs) bool { +func isSingleStepping(regs *arch.Registers) bool { return (regs.Eflags & arch.X86TrapFlag) != 0 } // updateSyscallRegs updates registers after finishing sysemu. -func updateSyscallRegs(regs *syscall.PtraceRegs) { +func updateSyscallRegs(regs *arch.Registers) { // Ptrace puts -ENOSYS in rax on syscall-enter-stop. regs.Rax = regs.Orig_rax } // syscallReturnValue extracts a sensible return from registers. -func syscallReturnValue(regs *syscall.PtraceRegs) (uintptr, error) { +func syscallReturnValue(regs *arch.Registers) (uintptr, error) { rval := int64(regs.Rax) if rval < 0 { return 0, syscall.Errno(-rval) @@ -100,7 +101,7 @@ func syscallReturnValue(regs *syscall.PtraceRegs) (uintptr, error) { return uintptr(rval), nil } -func dumpRegs(regs *syscall.PtraceRegs) string { +func dumpRegs(regs *arch.Registers) string { var m strings.Builder fmt.Fprintf(&m, "Registers:\n") @@ -142,7 +143,7 @@ func (t *thread) adjustInitRegsRip() { } // Pass the expected PPID to the child via R15 when creating stub process. -func initChildProcessPPID(initregs *syscall.PtraceRegs, ppid int32) { +func initChildProcessPPID(initregs *arch.Registers, ppid int32) { initregs.R15 = uint64(ppid) // Rbx has to be set to 1 when creating stub process. initregs.Rbx = 1 @@ -155,7 +156,7 @@ func initChildProcessPPID(initregs *syscall.PtraceRegs, ppid int32) { // // Note that this should only be called after verifying that the signalInfo has // been generated by the kernel. -func patchSignalInfo(regs *syscall.PtraceRegs, signalInfo *arch.SignalInfo) { +func patchSignalInfo(regs *arch.Registers, signalInfo *arch.SignalInfo) { if linux.Signal(signalInfo.Signo) == linux.SIGSYS { signalInfo.Signo = int32(linux.SIGSEGV) @@ -183,13 +184,76 @@ func enableCpuidFault() { // appendArchSeccompRules append architecture specific seccomp rules when creating BPF program. // Ref attachedThread() for more detail. -func appendArchSeccompRules(rules []seccomp.RuleSet) []seccomp.RuleSet { - return append(rules, seccomp.RuleSet{ - Rules: seccomp.SyscallRules{ - syscall.SYS_ARCH_PRCTL: []seccomp.Rule{ - {seccomp.AllowValue(linux.ARCH_SET_CPUID), seccomp.AllowValue(0)}, +func appendArchSeccompRules(rules []seccomp.RuleSet, defaultAction linux.BPFAction) []seccomp.RuleSet { + rules = append(rules, + // Rules for trapping vsyscall access. + seccomp.RuleSet{ + Rules: seccomp.SyscallRules{ + syscall.SYS_GETTIMEOFDAY: {}, + syscall.SYS_TIME: {}, + unix.SYS_GETCPU: {}, // SYS_GETCPU was not defined in package syscall on amd64. }, - }, - Action: linux.SECCOMP_RET_ALLOW, - }) + Action: linux.SECCOMP_RET_TRAP, + Vsyscall: true, + }) + if defaultAction != linux.SECCOMP_RET_ALLOW { + rules = append(rules, + seccomp.RuleSet{ + Rules: seccomp.SyscallRules{ + syscall.SYS_ARCH_PRCTL: []seccomp.Rule{ + {seccomp.AllowValue(linux.ARCH_SET_CPUID), seccomp.AllowValue(0)}, + }, + }, + Action: linux.SECCOMP_RET_ALLOW, + }) + } + return rules +} + +// probeSeccomp returns true iff seccomp is run after ptrace notifications, +// which is generally the case for kernel version >= 4.8. This check is dynamic +// because kernels have be backported behavior. +// +// See createStub for more information. +// +// Precondition: the runtime OS thread must be locked. +func probeSeccomp() bool { + // Create a completely new, destroyable process. + t, err := attachedThread(0, linux.SECCOMP_RET_ERRNO) + if err != nil { + panic(fmt.Sprintf("seccomp probe failed: %v", err)) + } + defer t.destroy() + + // Set registers to the yield system call. This call is not allowed + // by the filters specified in the attachThread function. + regs := createSyscallRegs(&t.initRegs, syscall.SYS_SCHED_YIELD) + if err := t.setRegs(®s); err != nil { + panic(fmt.Sprintf("ptrace set regs failed: %v", err)) + } + + for { + // Attempt an emulation. + if _, _, errno := syscall.RawSyscall6(syscall.SYS_PTRACE, unix.PTRACE_SYSEMU, uintptr(t.tid), 0, 0, 0, 0); errno != 0 { + panic(fmt.Sprintf("ptrace syscall-enter failed: %v", errno)) + } + + sig := t.wait(stopped) + if sig == (syscallEvent | syscall.SIGTRAP) { + // Did the seccomp errno hook already run? This would + // indicate that seccomp is first in line and we're + // less than 4.8. + if err := t.getRegs(®s); err != nil { + panic(fmt.Sprintf("ptrace get-regs failed: %v", err)) + } + if _, err := syscallReturnValue(®s); err == nil { + // The seccomp errno mode ran first, and reset + // the error in the registers. + return false + } + // The seccomp hook did not run yet, and therefore it + // is safe to use RET_KILL mode for dispatched calls. + return true + } + } } diff --git a/pkg/sentry/platform/ptrace/subprocess_arm64.go b/pkg/sentry/platform/ptrace/subprocess_arm64.go index 7b975137f..bd618fae8 100644 --- a/pkg/sentry/platform/ptrace/subprocess_arm64.go +++ b/pkg/sentry/platform/ptrace/subprocess_arm64.go @@ -41,13 +41,13 @@ const ( // resetSysemuRegs sets up emulation registers. // // This should be called prior to calling sysemu. -func (t *thread) resetSysemuRegs(regs *syscall.PtraceRegs) { +func (t *thread) resetSysemuRegs(regs *arch.Registers) { } // createSyscallRegs sets up syscall registers. // // This should be called to generate registers for a system call. -func createSyscallRegs(initRegs *syscall.PtraceRegs, sysno uintptr, args ...arch.SyscallArgument) syscall.PtraceRegs { +func createSyscallRegs(initRegs *arch.Registers, sysno uintptr, args ...arch.SyscallArgument) arch.Registers { // Copy initial registers (Pc, Sp, etc.). regs := *initRegs @@ -78,7 +78,7 @@ func createSyscallRegs(initRegs *syscall.PtraceRegs, sysno uintptr, args ...arch } // isSingleStepping determines if the registers indicate single-stepping. -func isSingleStepping(regs *syscall.PtraceRegs) bool { +func isSingleStepping(regs *arch.Registers) bool { // Refer to the ARM SDM D2.12.3: software step state machine // return (regs.Pstate.SS == 1) && (MDSCR_EL1.SS == 1). // @@ -89,13 +89,13 @@ func isSingleStepping(regs *syscall.PtraceRegs) bool { } // updateSyscallRegs updates registers after finishing sysemu. -func updateSyscallRegs(regs *syscall.PtraceRegs) { +func updateSyscallRegs(regs *arch.Registers) { // No special work is necessary. return } // syscallReturnValue extracts a sensible return from registers. -func syscallReturnValue(regs *syscall.PtraceRegs) (uintptr, error) { +func syscallReturnValue(regs *arch.Registers) (uintptr, error) { rval := int64(regs.Regs[0]) if rval < 0 { return 0, syscall.Errno(-rval) @@ -103,7 +103,7 @@ func syscallReturnValue(regs *syscall.PtraceRegs) (uintptr, error) { return uintptr(rval), nil } -func dumpRegs(regs *syscall.PtraceRegs) string { +func dumpRegs(regs *arch.Registers) string { var m strings.Builder fmt.Fprintf(&m, "Registers:\n") @@ -125,7 +125,7 @@ func (t *thread) adjustInitRegsRip() { } // Pass the expected PPID to the child via X7 when creating stub process -func initChildProcessPPID(initregs *syscall.PtraceRegs, ppid int32) { +func initChildProcessPPID(initregs *arch.Registers, ppid int32) { initregs.Regs[7] = uint64(ppid) // R9 has to be set to 1 when creating stub process. initregs.Regs[9] = 1 @@ -138,7 +138,7 @@ func initChildProcessPPID(initregs *syscall.PtraceRegs, ppid int32) { // // Note that this should only be called after verifying that the signalInfo has // been generated by the kernel. -func patchSignalInfo(regs *syscall.PtraceRegs, signalInfo *arch.SignalInfo) { +func patchSignalInfo(regs *arch.Registers, signalInfo *arch.SignalInfo) { if linux.Signal(signalInfo.Signo) == linux.SIGSYS { signalInfo.Signo = int32(linux.SIGSEGV) @@ -160,6 +160,15 @@ func enableCpuidFault() { // appendArchSeccompRules append architecture specific seccomp rules when creating BPF program. // Ref attachedThread() for more detail. -func appendArchSeccompRules(rules []seccomp.RuleSet) []seccomp.RuleSet { +func appendArchSeccompRules(rules []seccomp.RuleSet, defaultAction linux.BPFAction) []seccomp.RuleSet { return rules } + +// probeSeccomp returns true if seccomp is run after ptrace notifications, +// which is generally the case for kernel version >= 4.8. +// +// On arm64, the support of PTRACE_SYSEMU was added in the 5.3 kernel, so +// probeSeccomp can always return true. +func probeSeccomp() bool { + return true +} diff --git a/pkg/sentry/platform/ptrace/subprocess_linux.go b/pkg/sentry/platform/ptrace/subprocess_linux.go index 74968dfdf..2ce528601 100644 --- a/pkg/sentry/platform/ptrace/subprocess_linux.go +++ b/pkg/sentry/platform/ptrace/subprocess_linux.go @@ -20,7 +20,6 @@ import ( "fmt" "syscall" - "golang.org/x/sys/unix" "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/log" "gvisor.dev/gvisor/pkg/procid" @@ -30,54 +29,6 @@ import ( const syscallEvent syscall.Signal = 0x80 -// probeSeccomp returns true iff seccomp is run after ptrace notifications, -// which is generally the case for kernel version >= 4.8. This check is dynamic -// because kernels have be backported behavior. -// -// See createStub for more information. -// -// Precondition: the runtime OS thread must be locked. -func probeSeccomp() bool { - // Create a completely new, destroyable process. - t, err := attachedThread(0, linux.SECCOMP_RET_ERRNO) - if err != nil { - panic(fmt.Sprintf("seccomp probe failed: %v", err)) - } - defer t.destroy() - - // Set registers to the yield system call. This call is not allowed - // by the filters specified in the attachThread function. - regs := createSyscallRegs(&t.initRegs, syscall.SYS_SCHED_YIELD) - if err := t.setRegs(®s); err != nil { - panic(fmt.Sprintf("ptrace set regs failed: %v", err)) - } - - for { - // Attempt an emulation. - if _, _, errno := syscall.RawSyscall6(syscall.SYS_PTRACE, unix.PTRACE_SYSEMU, uintptr(t.tid), 0, 0, 0, 0); errno != 0 { - panic(fmt.Sprintf("ptrace syscall-enter failed: %v", errno)) - } - - sig := t.wait(stopped) - if sig == (syscallEvent | syscall.SIGTRAP) { - // Did the seccomp errno hook already run? This would - // indicate that seccomp is first in line and we're - // less than 4.8. - if err := t.getRegs(®s); err != nil { - panic(fmt.Sprintf("ptrace get-regs failed: %v", err)) - } - if _, err := syscallReturnValue(®s); err == nil { - // The seccomp errno mode ran first, and reset - // the error in the registers. - return false - } - // The seccomp hook did not run yet, and therefore it - // is safe to use RET_KILL mode for dispatched calls. - return true - } - } -} - // createStub creates a fresh stub processes. // // Precondition: the runtime OS thread must be locked. @@ -123,18 +74,7 @@ func attachedThread(flags uintptr, defaultAction linux.BPFAction) (*thread, erro // stub and all its children. This is used to create child stubs // (below), so we must include the ability to fork, but otherwise lock // down available calls only to what is needed. - rules := []seccomp.RuleSet{ - // Rules for trapping vsyscall access. - { - Rules: seccomp.SyscallRules{ - syscall.SYS_GETTIMEOFDAY: {}, - syscall.SYS_TIME: {}, - unix.SYS_GETCPU: {}, // SYS_GETCPU was not defined in package syscall on amd64. - }, - Action: linux.SECCOMP_RET_TRAP, - Vsyscall: true, - }, - } + rules := []seccomp.RuleSet{} if defaultAction != linux.SECCOMP_RET_ALLOW { rules = append(rules, seccomp.RuleSet{ Rules: seccomp.SyscallRules{ @@ -173,9 +113,8 @@ func attachedThread(flags uintptr, defaultAction linux.BPFAction) (*thread, erro }, Action: linux.SECCOMP_RET_ALLOW, }) - - rules = appendArchSeccompRules(rules) } + rules = appendArchSeccompRules(rules, defaultAction) instrs, err := seccomp.BuildProgram(rules, defaultAction) if err != nil { return nil, err diff --git a/pkg/sentry/platform/ptrace/subprocess_unsafe.go b/pkg/sentry/platform/ptrace/subprocess_unsafe.go index 2ae6b9f9d..0bee995e4 100644 --- a/pkg/sentry/platform/ptrace/subprocess_unsafe.go +++ b/pkg/sentry/platform/ptrace/subprocess_unsafe.go @@ -13,7 +13,7 @@ // limitations under the License. // +build go1.12 -// +build !go1.15 +// +build !go1.16 // Check go:linkname function signatures when updating Go version. diff --git a/pkg/sentry/platform/ring0/BUILD b/pkg/sentry/platform/ring0/BUILD index 934b6fbcd..679b287c3 100644 --- a/pkg/sentry/platform/ring0/BUILD +++ b/pkg/sentry/platform/ring0/BUILD @@ -72,11 +72,14 @@ go_library( "lib_amd64.s", "lib_arm64.go", "lib_arm64.s", + "lib_arm64_unsafe.go", "ring0.go", ], visibility = ["//pkg/sentry:internal"], deps = [ "//pkg/cpuid", + "//pkg/safecopy", + "//pkg/sentry/arch", "//pkg/sentry/platform/ring0/pagetables", "//pkg/usermem", ], diff --git a/pkg/sentry/platform/ring0/aarch64.go b/pkg/sentry/platform/ring0/aarch64.go index f6da41c27..8122ac6e2 100644 --- a/pkg/sentry/platform/ring0/aarch64.go +++ b/pkg/sentry/platform/ring0/aarch64.go @@ -27,26 +27,27 @@ const ( _PTE_PGT_BASE = 0x7000 _PTE_PGT_SIZE = 0x1000 - _PSR_MODE_EL0t = 0x0 - _PSR_MODE_EL1t = 0x4 - _PSR_MODE_EL1h = 0x5 - _PSR_EL_MASK = 0xf - - _PSR_D_BIT = 0x200 - _PSR_A_BIT = 0x100 - _PSR_I_BIT = 0x80 - _PSR_F_BIT = 0x40 + _PSR_D_BIT = 0x00000200 + _PSR_A_BIT = 0x00000100 + _PSR_I_BIT = 0x00000080 + _PSR_F_BIT = 0x00000040 ) const ( + // PSR bits + PSR_MODE_EL0t = 0x00000000 + PSR_MODE_EL1t = 0x00000004 + PSR_MODE_EL1h = 0x00000005 + PSR_MODE_MASK = 0x0000000f + // KernelFlagsSet should always be set in the kernel. - KernelFlagsSet = _PSR_MODE_EL1h + KernelFlagsSet = PSR_MODE_EL1h // UserFlagsSet are always set in userspace. - UserFlagsSet = _PSR_MODE_EL0t + UserFlagsSet = PSR_MODE_EL0t - KernelFlagsClear = _PSR_EL_MASK - UserFlagsClear = _PSR_EL_MASK + KernelFlagsClear = PSR_MODE_MASK + UserFlagsClear = PSR_MODE_MASK PsrDefaultSet = _PSR_D_BIT | _PSR_A_BIT | _PSR_I_BIT | _PSR_F_BIT ) diff --git a/pkg/sentry/platform/ring0/defs.go b/pkg/sentry/platform/ring0/defs.go index 86fd5ed58..e6daf24df 100644 --- a/pkg/sentry/platform/ring0/defs.go +++ b/pkg/sentry/platform/ring0/defs.go @@ -15,8 +15,7 @@ package ring0 import ( - "syscall" - + "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/platform/ring0/pagetables" ) @@ -72,7 +71,7 @@ type CPU struct { // registers is a set of registers; these may be used on kernel system // calls and exceptions via the Registers function. - registers syscall.PtraceRegs + registers arch.Registers // hooks are kernel hooks. hooks Hooks @@ -83,14 +82,14 @@ type CPU struct { // This is explicitly safe to call during KernelException and KernelSyscall. // //go:nosplit -func (c *CPU) Registers() *syscall.PtraceRegs { +func (c *CPU) Registers() *arch.Registers { return &c.registers } // SwitchOpts are passed to the Switch function. type SwitchOpts struct { // Registers are the user register state. - Registers *syscall.PtraceRegs + Registers *arch.Registers // FloatingPointState is a byte pointer where floating point state is // saved and restored. diff --git a/pkg/sentry/platform/ring0/defs_arm64.go b/pkg/sentry/platform/ring0/defs_arm64.go index 1583dda12..0e2ab716c 100644 --- a/pkg/sentry/platform/ring0/defs_arm64.go +++ b/pkg/sentry/platform/ring0/defs_arm64.go @@ -124,6 +124,12 @@ func (c *CPU) SetAppAddr(value uintptr) { c.appAddr = value } +// GetLazyVFP returns the value of cpacr_el1. +//go:nosplit +func (c *CPU) GetLazyVFP() (value uintptr) { + return c.lazyVFP +} + // SwitchArchOpts are embedded in SwitchOpts. type SwitchArchOpts struct { // UserASID indicates that the application ASID to be used on switch, diff --git a/pkg/sentry/platform/ring0/entry_amd64.go b/pkg/sentry/platform/ring0/entry_amd64.go index a5ce67885..7fa43c2f5 100644 --- a/pkg/sentry/platform/ring0/entry_amd64.go +++ b/pkg/sentry/platform/ring0/entry_amd64.go @@ -17,7 +17,7 @@ package ring0 import ( - "syscall" + "gvisor.dev/gvisor/pkg/sentry/arch" ) // This is an assembly function. @@ -41,7 +41,7 @@ func swapgs() // 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, *syscall.PtraceRegs) Vector +func sysret(*CPU, *arch.Registers) Vector // "iret is the cadillac of CPL switching." // @@ -50,7 +50,7 @@ func sysret(*CPU, *syscall.PtraceRegs) 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, *syscall.PtraceRegs) Vector +func iret(*CPU, *arch.Registers) Vector // exception is the generic exception entry. // diff --git a/pkg/sentry/platform/ring0/entry_arm64.s b/pkg/sentry/platform/ring0/entry_arm64.s index baa6c4910..db6465663 100644 --- a/pkg/sentry/platform/ring0/entry_arm64.s +++ b/pkg/sentry/platform/ring0/entry_arm64.s @@ -25,10 +25,14 @@ // not available for calls. // +// ERET returns using the ELR and SPSR for the current exception level. #define ERET() \ WORD $0xd69f03e0 +// RSV_REG is a register that holds el1 information temporarily. #define RSV_REG R18_PLATFORM + +// RSV_REG_APP is a register that holds el0 information temporarily. #define RSV_REG_APP R9 #define FPEN_NOTRAP 0x3 @@ -36,6 +40,12 @@ #define FPEN_ENABLE (FPEN_NOTRAP << FPEN_SHIFT) +// Saves a register set. +// +// This is a macro because it may need to executed in contents where a stack is +// not available for calls. +// +// The following registers are not saved: R9, R18. #define REGISTERS_SAVE(reg, offset) \ MOVD R0, offset+PTRACE_R0(reg); \ MOVD R1, offset+PTRACE_R1(reg); \ @@ -67,6 +77,12 @@ MOVD R29, offset+PTRACE_R29(reg); \ MOVD R30, offset+PTRACE_R30(reg); +// Loads a register set. +// +// This is a macro because it may need to executed in contents where a stack is +// not available for calls. +// +// The following registers are not loaded: R9, R18. #define REGISTERS_LOAD(reg, offset) \ MOVD offset+PTRACE_R0(reg), R0; \ MOVD offset+PTRACE_R1(reg), R1; \ @@ -98,7 +114,7 @@ MOVD offset+PTRACE_R29(reg), R29; \ MOVD offset+PTRACE_R30(reg), R30; -//NOP +// NOP-s #define nop31Instructions() \ WORD $0xd503201f; \ WORD $0xd503201f; \ @@ -254,6 +270,7 @@ #define ESR_ELx_WFx_ISS_WFE (UL(1) << 0) #define ESR_ELx_xVC_IMM_MASK ((1UL << 16) - 1) +// LOAD_KERNEL_ADDRESS loads a kernel address. #define LOAD_KERNEL_ADDRESS(from, to) \ MOVD from, to; \ ORR $0xffff000000000000, to, to; @@ -263,15 +280,18 @@ LOAD_KERNEL_ADDRESS(CPU_SELF(from), RSV_REG); \ MOVD $CPU_STACK_TOP(RSV_REG), RSV_REG; \ MOVD RSV_REG, RSP; \ + WORD $0xd538d092; \ //MRS TPIDR_EL1, R18 ISB $15; \ DSB $15; +// SWITCH_TO_APP_PAGETABLE sets a new pagetable for a container application. #define SWITCH_TO_APP_PAGETABLE(from) \ MOVD CPU_TTBR0_APP(from), RSV_REG; \ WORD $0xd5182012; \ // MSR R18, TTBR0_EL1 ISB $15; \ DSB $15; +// SWITCH_TO_KVM_PAGETABLE sets the kvm pagetable. #define SWITCH_TO_KVM_PAGETABLE(from) \ MOVD CPU_TTBR0_KVM(from), RSV_REG; \ WORD $0xd5182012; \ // MSR R18, TTBR0_EL1 @@ -294,6 +314,7 @@ WORD $0xd5181040; \ //MSR R0, CPACR_EL1 ISB $15; +// KERNEL_ENTRY_FROM_EL0 is the entry code of the vcpu from el0 to el1. #define KERNEL_ENTRY_FROM_EL0 \ SUB $16, RSP, RSP; \ // step1, save r18, r9 into kernel temporary stack. STP (RSV_REG, RSV_REG_APP), 16*0(RSP); \ @@ -315,19 +336,22 @@ WORD $0xd5384103; \ // MRS SP_EL0, R3 MOVD R3, PTRACE_SP(RSV_REG_APP); +// KERNEL_ENTRY_FROM_EL1 is the entry code of the vcpu from el1 to el1. #define KERNEL_ENTRY_FROM_EL1 \ WORD $0xd538d092; \ //MRS TPIDR_EL1, R18 - REGISTERS_SAVE(RSV_REG, CPU_REGISTERS); \ // save sentry context + REGISTERS_SAVE(RSV_REG, CPU_REGISTERS); \ // Save sentry context. MOVD RSV_REG_APP, CPU_REGISTERS+PTRACE_R9(RSV_REG); \ WORD $0xd5384004; \ // MRS SPSR_EL1, R4 MOVD R4, CPU_REGISTERS+PTRACE_PSTATE(RSV_REG); \ MRS ELR_EL1, R4; \ MOVD R4, CPU_REGISTERS+PTRACE_PC(RSV_REG); \ MOVD RSP, R4; \ - MOVD R4, CPU_REGISTERS+PTRACE_SP(RSV_REG); + MOVD R4, CPU_REGISTERS+PTRACE_SP(RSV_REG); \ + LOAD_KERNEL_STACK(RSV_REG); // Load the temporary stack. +// Halt halts execution. TEXT ·Halt(SB),NOSPLIT,$0 - // clear bluepill. + // Clear bluepill. WORD $0xd538d092 //MRS TPIDR_EL1, R18 CMP RSV_REG, R9 BNE mmio_exit @@ -341,8 +365,22 @@ mmio_exit: // MMIO_EXIT. MOVD $0, R9 MOVD R0, 0xffff000000001000(R9) - B ·kernelExitToEl1(SB) + RET + +// HaltAndResume halts execution and point the pointer to the resume function. +TEXT ·HaltAndResume(SB),NOSPLIT,$0 + BL ·Halt(SB) + B ·kernelExitToEl1(SB) // Resume. + +// HaltEl1SvcAndResume calls Hooks.KernelSyscall and resume. +TEXT ·HaltEl1SvcAndResume(SB),NOSPLIT,$0 + WORD $0xd538d092 // MRS TPIDR_EL1, R18 + MOVD CPU_SELF(RSV_REG), R3 // Load vCPU. + MOVD R3, 8(RSP) // First argument (vCPU). + CALL ·kernelSyscall(SB) // Call the trampoline. + B ·kernelExitToEl1(SB) // Resume. +// Shutdown stops the guest. TEXT ·Shutdown(SB),NOSPLIT,$0 // PSCI EVENT. MOVD $0x84000009, R0 @@ -356,6 +394,8 @@ TEXT ·Current(SB),NOSPLIT,$0-8 #define STACK_FRAME_SIZE 16 +// kernelExitToEl0 is the entrypoint for application in guest_el0. +// Prepare the vcpu environment for container application. TEXT ·kernelExitToEl0(SB),NOSPLIT,$0 // Step1, save sentry context into memory. REGISTERS_SAVE(RSV_REG, CPU_REGISTERS) @@ -426,9 +466,26 @@ TEXT ·kernelExitToEl0(SB),NOSPLIT,$0 ERET() +// kernelExitToEl1 is the entrypoint for sentry in guest_el1. +// Prepare the vcpu environment for sentry. TEXT ·kernelExitToEl1(SB),NOSPLIT,$0 + WORD $0xd538d092 //MRS TPIDR_EL1, R18 + + MOVD CPU_REGISTERS+PTRACE_PSTATE(RSV_REG), R1 + WORD $0xd5184001 //MSR R1, SPSR_EL1 + + MOVD CPU_REGISTERS+PTRACE_PC(RSV_REG), R1 + MSR R1, ELR_EL1 + + MOVD CPU_REGISTERS+PTRACE_SP(RSV_REG), R1 + MOVD R1, RSP + + REGISTERS_LOAD(RSV_REG, CPU_REGISTERS) + MOVD CPU_REGISTERS+PTRACE_R9(RSV_REG), RSV_REG_APP + ERET() +// Start is the CPU entrypoint. TEXT ·Start(SB),NOSPLIT,$0 IRQ_DISABLE MOVD R8, RSV_REG @@ -437,18 +494,23 @@ TEXT ·Start(SB),NOSPLIT,$0 B ·kernelExitToEl1(SB) +// El1_sync_invalid is the handler for an invalid EL1_sync. TEXT ·El1_sync_invalid(SB),NOSPLIT,$0 B ·Shutdown(SB) +// El1_irq_invalid is the handler for an invalid El1_irq. TEXT ·El1_irq_invalid(SB),NOSPLIT,$0 B ·Shutdown(SB) +// El1_fiq_invalid is the handler for an invalid El1_fiq. TEXT ·El1_fiq_invalid(SB),NOSPLIT,$0 B ·Shutdown(SB) +// El1_error_invalid is the handler for an invalid El1_error. TEXT ·El1_error_invalid(SB),NOSPLIT,$0 B ·Shutdown(SB) +// El1_sync is the handler for El1_sync. TEXT ·El1_sync(SB),NOSPLIT,$0 KERNEL_ENTRY_FROM_EL1 WORD $0xd5385219 // MRS ESR_EL1, R25 @@ -484,10 +546,10 @@ el1_da: MOVD $PageFault, R3 MOVD R3, CPU_VECTOR_CODE(RSV_REG) - B ·Halt(SB) + B ·HaltAndResume(SB) el1_ia: - B ·Halt(SB) + B ·HaltAndResume(SB) el1_sp_pc: B ·Shutdown(SB) @@ -496,7 +558,9 @@ el1_undef: B ·Shutdown(SB) el1_svc: - B ·Halt(SB) + MOVD $0, CPU_ERROR_CODE(RSV_REG) + MOVD $0, CPU_ERROR_TYPE(RSV_REG) + B ·HaltEl1SvcAndResume(SB) el1_dbg: B ·Shutdown(SB) @@ -508,15 +572,19 @@ el1_fpsimd_acc: el1_invalid: B ·Shutdown(SB) +// El1_irq is the handler for El1_irq. TEXT ·El1_irq(SB),NOSPLIT,$0 B ·Shutdown(SB) +// El1_fiq is the handler for El1_fiq. TEXT ·El1_fiq(SB),NOSPLIT,$0 B ·Shutdown(SB) +// El1_error is the handler for El1_error. TEXT ·El1_error(SB),NOSPLIT,$0 B ·Shutdown(SB) +// El0_sync is the handler for El0_sync. TEXT ·El0_sync(SB),NOSPLIT,$0 KERNEL_ENTRY_FROM_EL0 WORD $0xd5385219 // MRS ESR_EL1, R25 @@ -554,7 +622,7 @@ el0_svc: MOVD $Syscall, R3 MOVD R3, CPU_VECTOR_CODE(RSV_REG) - B ·Halt(SB) + B ·HaltAndResume(SB) el0_da: WORD $0xd538d092 //MRS TPIDR_EL1, R18 @@ -568,7 +636,7 @@ el0_da: MOVD $PageFault, R3 MOVD R3, CPU_VECTOR_CODE(RSV_REG) - B ·Halt(SB) + B ·HaltAndResume(SB) el0_ia: B ·Shutdown(SB) @@ -613,7 +681,7 @@ TEXT ·El0_error(SB),NOSPLIT,$0 MOVD $VirtualizationException, R3 MOVD R3, CPU_VECTOR_CODE(RSV_REG) - B ·Halt(SB) + B ·HaltAndResume(SB) TEXT ·El0_sync_invalid(SB),NOSPLIT,$0 B ·Shutdown(SB) @@ -627,6 +695,7 @@ TEXT ·El0_fiq_invalid(SB),NOSPLIT,$0 TEXT ·El0_error_invalid(SB),NOSPLIT,$0 B ·Shutdown(SB) +// Vectors implements exception vector table. TEXT ·Vectors(SB),NOSPLIT,$0 B ·El1_sync_invalid(SB) nop31Instructions() diff --git a/pkg/sentry/platform/ring0/gen_offsets/BUILD b/pkg/sentry/platform/ring0/gen_offsets/BUILD index 4cae10459..549f3d228 100644 --- a/pkg/sentry/platform/ring0/gen_offsets/BUILD +++ b/pkg/sentry/platform/ring0/gen_offsets/BUILD @@ -27,6 +27,7 @@ go_binary( visibility = ["//pkg/sentry/platform/ring0:__pkg__"], deps = [ "//pkg/cpuid", + "//pkg/sentry/arch", "//pkg/sentry/platform/ring0/pagetables", "//pkg/usermem", ], diff --git a/pkg/sentry/platform/ring0/kernel.go b/pkg/sentry/platform/ring0/kernel.go index 900c0bba7..021693791 100644 --- a/pkg/sentry/platform/ring0/kernel.go +++ b/pkg/sentry/platform/ring0/kernel.go @@ -31,23 +31,39 @@ type defaultHooks struct{} // KernelSyscall implements Hooks.KernelSyscall. // +// +checkescape:all +// //go:nosplit -func (defaultHooks) KernelSyscall() { Halt() } +func (defaultHooks) KernelSyscall() { + Halt() +} // KernelException implements Hooks.KernelException. // +// +checkescape:all +// //go:nosplit -func (defaultHooks) KernelException(Vector) { Halt() } +func (defaultHooks) KernelException(Vector) { + Halt() +} // kernelSyscall is a trampoline. // +// +checkescape:hard,stack +// //go:nosplit -func kernelSyscall(c *CPU) { c.hooks.KernelSyscall() } +func kernelSyscall(c *CPU) { + c.hooks.KernelSyscall() +} // kernelException is a trampoline. // +// +checkescape:hard,stack +// //go:nosplit -func kernelException(c *CPU, vector Vector) { c.hooks.KernelException(vector) } +func kernelException(c *CPU, vector Vector) { + c.hooks.KernelException(vector) +} // Init initializes a new CPU. // diff --git a/pkg/sentry/platform/ring0/kernel_amd64.go b/pkg/sentry/platform/ring0/kernel_amd64.go index 0feff8778..d37981dbf 100644 --- a/pkg/sentry/platform/ring0/kernel_amd64.go +++ b/pkg/sentry/platform/ring0/kernel_amd64.go @@ -178,6 +178,8 @@ func IsCanonical(addr uint64) bool { // // Precondition: the Rip, Rsp, Fs and Gs registers must be canonical. // +// +checkescape:all +// //go:nosplit func (c *CPU) SwitchToUser(switchOpts SwitchOpts) (vector Vector) { userCR3 := switchOpts.PageTables.CR3(!switchOpts.Flush, switchOpts.UserPCID) @@ -192,9 +194,9 @@ func (c *CPU) SwitchToUser(switchOpts SwitchOpts) (vector Vector) { // Perform the switch. swapgs() // GS will be swapped on return. - WriteFS(uintptr(regs.Fs_base)) // Set application FS. - WriteGS(uintptr(regs.Gs_base)) // Set application GS. - LoadFloatingPoint(switchOpts.FloatingPointState) // Copy in floating point. + 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 { @@ -204,8 +206,8 @@ func (c *CPU) SwitchToUser(switchOpts SwitchOpts) (vector Vector) { } writeCR3(uintptr(kernelCR3)) // Return to kernel address space. jumpToUser() // Return to lower half. - SaveFloatingPoint(switchOpts.FloatingPointState) // Copy out floating point. - WriteFS(uintptr(c.registers.Fs_base)) // Restore kernel FS. + SaveFloatingPoint(switchOpts.FloatingPointState) // escapes: no. Copy out floating point. + WriteFS(uintptr(c.registers.Fs_base)) // escapes: no. Restore kernel FS. return } diff --git a/pkg/sentry/platform/ring0/kernel_arm64.go b/pkg/sentry/platform/ring0/kernel_arm64.go index c3d341998..ccacaea6b 100644 --- a/pkg/sentry/platform/ring0/kernel_arm64.go +++ b/pkg/sentry/platform/ring0/kernel_arm64.go @@ -16,6 +16,14 @@ package ring0 +// HaltAndResume halts execution and point the pointer to the resume function. +//go:nosplit +func HaltAndResume() + +// HaltEl1SvcAndResume calls Hooks.KernelSyscall and resume. +//go:nosplit +func HaltEl1SvcAndResume() + // init initializes architecture-specific state. func (k *Kernel) init(opts KernelOpts) { // Save the root page tables. diff --git a/pkg/sentry/platform/ring0/lib_arm64.go b/pkg/sentry/platform/ring0/lib_arm64.go index af075aae4..a6345010d 100644 --- a/pkg/sentry/platform/ring0/lib_arm64.go +++ b/pkg/sentry/platform/ring0/lib_arm64.go @@ -20,13 +20,13 @@ package ring0 func CPACREL1() (value uintptr) // FPCR returns the value of FPCR register. -func FPCR() (value uintptr) +func GetFPCR() (value uintptr) // SetFPCR writes the FPCR value. func SetFPCR(value uintptr) // FPSR returns the value of FPSR register. -func FPSR() (value uintptr) +func GetFPSR() (value uintptr) // SetFPSR writes the FPSR value. func SetFPSR(value uintptr) @@ -37,3 +37,16 @@ func SaveVRegs(*byte) // LoadVRegs loads V0-V31 registers. func LoadVRegs(*byte) + +// GetTLS returns the value of TPIDR_EL0 register. +func GetTLS() (value uint64) + +// SetTLS writes the TPIDR_EL0 value. +func SetTLS(value uint64) + +// Init sets function pointers based on architectural features. +// +// This must be called prior to using ring0. +func Init() { + rewriteVectors() +} diff --git a/pkg/sentry/platform/ring0/lib_arm64.s b/pkg/sentry/platform/ring0/lib_arm64.s index 0e6a6235b..b63e14b41 100644 --- a/pkg/sentry/platform/ring0/lib_arm64.s +++ b/pkg/sentry/platform/ring0/lib_arm64.s @@ -15,6 +15,16 @@ #include "funcdata.h" #include "textflag.h" +TEXT ·GetTLS(SB),NOSPLIT,$0-8 + MRS TPIDR_EL0, R1 + MOVD R1, ret+0(FP) + RET + +TEXT ·SetTLS(SB),NOSPLIT,$0-8 + MOVD addr+0(FP), R1 + MSR R1, TPIDR_EL0 + RET + TEXT ·CPACREL1(SB),NOSPLIT,$0-8 WORD $0xd5381041 // MRS CPACR_EL1, R1 MOVD R1, ret+0(FP) diff --git a/pkg/sentry/platform/ring0/lib_arm64_unsafe.go b/pkg/sentry/platform/ring0/lib_arm64_unsafe.go new file mode 100644 index 000000000..c05166fea --- /dev/null +++ b/pkg/sentry/platform/ring0/lib_arm64_unsafe.go @@ -0,0 +1,108 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// +build arm64 + +package ring0 + +import ( + "reflect" + "syscall" + "unsafe" + + "gvisor.dev/gvisor/pkg/safecopy" + "gvisor.dev/gvisor/pkg/usermem" +) + +const ( + nopInstruction = 0xd503201f + instSize = unsafe.Sizeof(uint32(0)) + vectorsRawLen = 0x800 +) + +func unsafeSlice(addr uintptr, length int) (slice []uint32) { + hdr := (*reflect.SliceHeader)(unsafe.Pointer(&slice)) + hdr.Data = addr + hdr.Len = length / int(instSize) + hdr.Cap = length / int(instSize) + return slice +} + +// Work around: move ring0.Vectors() into a specific address with 11-bits alignment. +// +// According to the design documentation of Arm64, +// the start address of exception vector table should be 11-bits aligned. +// Please see the code in linux kernel as reference: arch/arm64/kernel/entry.S +// But, we can't align a function's start address to a specific address by using golang. +// We have raised this question in golang community: +// https://groups.google.com/forum/m/#!topic/golang-dev/RPj90l5x86I +// This function will be removed when golang supports this feature. +// +// There are 2 jobs were implemented in this function: +// 1, move the start address of exception vector table into the specific address. +// 2, modify the offset of each instruction. +func rewriteVectors() { + vectorsBegin := reflect.ValueOf(Vectors).Pointer() + + // The exception-vector-table is required to be 11-bits aligned. + // And the size is 0x800. + // Please see the documentation as reference: + // https://developer.arm.com/docs/100933/0100/aarch64-exception-vector-table + // + // But, golang does not allow to set a function's address to a specific value. + // So, for gvisor, I defined the size of exception-vector-table as 4K, + // filled the 2nd 2K part with NOP-s. + // So that, I can safely move the 1st 2K part into the address with 11-bits alignment. + // + // So, the prerequisite for this function to work correctly is: + // vectorsSafeLen >= 0x1000 + // vectorsRawLen = 0x800 + vectorsSafeLen := int(safecopy.FindEndAddress(vectorsBegin) - vectorsBegin) + if vectorsSafeLen < 2*vectorsRawLen { + panic("Can't update vectors") + } + + vectorsSafeTable := unsafeSlice(vectorsBegin, vectorsSafeLen) // Now a []uint32 + vectorsRawLen32 := vectorsRawLen / int(instSize) + + offset := vectorsBegin & (1<<11 - 1) + if offset != 0 { + offset = 1<<11 - offset + } + + pageBegin := (vectorsBegin + offset) & ^uintptr(usermem.PageSize-1) + + _, _, errno := syscall.Syscall(syscall.SYS_MPROTECT, uintptr(pageBegin), uintptr(usermem.PageSize), uintptr(syscall.PROT_READ|syscall.PROT_WRITE|syscall.PROT_EXEC)) + if errno != 0 { + panic(errno.Error()) + } + + offset = offset / instSize // By index, not bytes. + // Move exception-vector-table into the specific address, should uses memmove here. + for i := 1; i <= vectorsRawLen32; i++ { + vectorsSafeTable[int(offset)+vectorsRawLen32-i] = vectorsSafeTable[vectorsRawLen32-i] + } + + // Adjust branch since instruction was moved forward. + for i := 0; i < vectorsRawLen32; i++ { + if vectorsSafeTable[int(offset)+i] != nopInstruction { + vectorsSafeTable[int(offset)+i] -= uint32(offset) + } + } + + _, _, errno = syscall.Syscall(syscall.SYS_MPROTECT, uintptr(pageBegin), uintptr(usermem.PageSize), uintptr(syscall.PROT_READ|syscall.PROT_EXEC)) + if errno != 0 { + panic(errno.Error()) + } +} diff --git a/pkg/sentry/platform/ring0/offsets_amd64.go b/pkg/sentry/platform/ring0/offsets_amd64.go index 85cc3fdad..b8ab120a0 100644 --- a/pkg/sentry/platform/ring0/offsets_amd64.go +++ b/pkg/sentry/platform/ring0/offsets_amd64.go @@ -20,7 +20,8 @@ import ( "fmt" "io" "reflect" - "syscall" + + "gvisor.dev/gvisor/pkg/sentry/arch" ) // Emit prints architecture-specific offsets. @@ -64,7 +65,7 @@ func Emit(w io.Writer) { fmt.Fprintf(w, "#define SyscallInt80 0x%02x\n", SyscallInt80) fmt.Fprintf(w, "#define Syscall 0x%02x\n", Syscall) - p := &syscall.PtraceRegs{} + p := &arch.Registers{} fmt.Fprintf(w, "\n// Ptrace registers.\n") fmt.Fprintf(w, "#define PTRACE_R15 0x%02x\n", reflect.ValueOf(&p.R15).Pointer()-reflect.ValueOf(p).Pointer()) fmt.Fprintf(w, "#define PTRACE_R14 0x%02x\n", reflect.ValueOf(&p.R14).Pointer()-reflect.ValueOf(p).Pointer()) diff --git a/pkg/sentry/platform/ring0/offsets_arm64.go b/pkg/sentry/platform/ring0/offsets_arm64.go index 057fb5c69..f3de962f0 100644 --- a/pkg/sentry/platform/ring0/offsets_arm64.go +++ b/pkg/sentry/platform/ring0/offsets_arm64.go @@ -20,7 +20,8 @@ import ( "fmt" "io" "reflect" - "syscall" + + "gvisor.dev/gvisor/pkg/sentry/arch" ) // Emit prints architecture-specific offsets. @@ -87,7 +88,7 @@ func Emit(w io.Writer) { fmt.Fprintf(w, "#define Syscall 0x%02x\n", Syscall) fmt.Fprintf(w, "#define VirtualizationException 0x%02x\n", VirtualizationException) - p := &syscall.PtraceRegs{} + p := &arch.Registers{} fmt.Fprintf(w, "\n// Ptrace registers.\n") fmt.Fprintf(w, "#define PTRACE_R0 0x%02x\n", reflect.ValueOf(&p.Regs[0]).Pointer()-reflect.ValueOf(p).Pointer()) fmt.Fprintf(w, "#define PTRACE_R1 0x%02x\n", reflect.ValueOf(&p.Regs[1]).Pointer()-reflect.ValueOf(p).Pointer()) diff --git a/pkg/sentry/platform/ring0/pagetables/BUILD b/pkg/sentry/platform/ring0/pagetables/BUILD index 971eed7fa..16d5f478b 100644 --- a/pkg/sentry/platform/ring0/pagetables/BUILD +++ b/pkg/sentry/platform/ring0/pagetables/BUILD @@ -7,7 +7,7 @@ go_template( name = "generic_walker", srcs = select_arch( amd64 = ["walker_amd64.go"], - arm64 = ["walker_amd64.go"], + arm64 = ["walker_arm64.go"], ), opt_types = [ "Visitor", @@ -80,6 +80,9 @@ go_library( "pagetables_amd64.go", "pagetables_arm64.go", "pagetables_x86.go", + "pcids.go", + "pcids_aarch64.go", + "pcids_aarch64.s", "pcids_x86.go", "walker_amd64.go", "walker_arm64.go", diff --git a/pkg/sentry/platform/ring0/pagetables/allocator.go b/pkg/sentry/platform/ring0/pagetables/allocator.go index 23fd5c352..8d75b7599 100644 --- a/pkg/sentry/platform/ring0/pagetables/allocator.go +++ b/pkg/sentry/platform/ring0/pagetables/allocator.go @@ -53,9 +53,14 @@ type RuntimeAllocator struct { // NewRuntimeAllocator returns an allocator that uses runtime allocation. func NewRuntimeAllocator() *RuntimeAllocator { - return &RuntimeAllocator{ - used: make(map[*PTEs]struct{}), - } + r := new(RuntimeAllocator) + r.Init() + return r +} + +// Init initializes a RuntimeAllocator. +func (r *RuntimeAllocator) Init() { + r.used = make(map[*PTEs]struct{}) } // Recycle returns freed pages to the pool. diff --git a/pkg/sentry/platform/ring0/pagetables/pagetables.go b/pkg/sentry/platform/ring0/pagetables/pagetables.go index 87e88e97d..7f18ac296 100644 --- a/pkg/sentry/platform/ring0/pagetables/pagetables.go +++ b/pkg/sentry/platform/ring0/pagetables/pagetables.go @@ -86,6 +86,8 @@ func (*mapVisitor) requiresSplit() bool { return true } // // Precondition: addr & length must be page-aligned, their sum must not overflow. // +// +checkescape:hard,stack +// //go:nosplit func (p *PageTables) Map(addr usermem.Addr, length uintptr, opts MapOpts, physical uintptr) bool { if !opts.AccessType.Any() { @@ -128,6 +130,8 @@ func (v *unmapVisitor) visit(start uintptr, pte *PTE, align uintptr) { // // Precondition: addr & length must be page-aligned. // +// +checkescape:hard,stack +// //go:nosplit func (p *PageTables) Unmap(addr usermem.Addr, length uintptr) bool { w := unmapWalker{ @@ -162,6 +166,8 @@ func (v *emptyVisitor) visit(start uintptr, pte *PTE, align uintptr) { // // Precondition: addr & length must be page-aligned. // +// +checkescape:hard,stack +// //go:nosplit func (p *PageTables) IsEmpty(addr usermem.Addr, length uintptr) bool { w := emptyWalker{ @@ -197,6 +203,8 @@ func (*lookupVisitor) requiresSplit() bool { return false } // Lookup returns the physical address for the given virtual address. // +// +checkescape:hard,stack +// //go:nosplit func (p *PageTables) Lookup(addr usermem.Addr) (physical uintptr, opts MapOpts) { mask := uintptr(usermem.PageSize - 1) diff --git a/pkg/sentry/platform/ring0/pagetables/pagetables_x86.go b/pkg/sentry/platform/ring0/pagetables/pagetables_x86.go index dcf061df9..157438d9b 100644 --- a/pkg/sentry/platform/ring0/pagetables/pagetables_x86.go +++ b/pkg/sentry/platform/ring0/pagetables/pagetables_x86.go @@ -12,7 +12,7 @@ // See the License for the specific language governing permissions and // limitations under the License. -// +build i386 amd64 +// +build 386 amd64 package pagetables diff --git a/pkg/sentry/platform/ring0/pagetables/pcids.go b/pkg/sentry/platform/ring0/pagetables/pcids.go new file mode 100644 index 000000000..964496aac --- /dev/null +++ b/pkg/sentry/platform/ring0/pagetables/pcids.go @@ -0,0 +1,104 @@ +// Copyright 2018 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package pagetables + +import ( + "gvisor.dev/gvisor/pkg/sync" +) + +// PCIDs is a simple PCID database. +// +// This is not protected by locks and is thus suitable for use only with a +// single CPU at a time. +type PCIDs struct { + // mu protects below. + mu sync.Mutex + + // cache are the assigned page tables. + cache map[*PageTables]uint16 + + // avail are available PCIDs. + avail []uint16 +} + +// NewPCIDs returns a new PCID database. +// +// start is the first index to assign. Typically this will be one, as the zero +// pcid will always be flushed on transition (see pagetables_x86.go). This may +// be more than one if specific PCIDs are reserved. +// +// Nil is returned iff the start and size are out of range. +func NewPCIDs(start, size uint16) *PCIDs { + if start+uint16(size) > limitPCID { + return nil // See comment. + } + p := &PCIDs{ + cache: make(map[*PageTables]uint16), + } + for pcid := start; pcid < start+size; pcid++ { + p.avail = append(p.avail, pcid) + } + return p +} + +// Assign assigns a PCID to the given PageTables. +// +// This may overwrite any previous assignment provided. If this in the case, +// true is returned to indicate that the PCID should be flushed. +func (p *PCIDs) Assign(pt *PageTables) (uint16, bool) { + p.mu.Lock() + if pcid, ok := p.cache[pt]; ok { + p.mu.Unlock() + return pcid, false // No flush. + } + + // Is there something available? + if len(p.avail) > 0 { + pcid := p.avail[len(p.avail)-1] + p.avail = p.avail[:len(p.avail)-1] + p.cache[pt] = pcid + + // We need to flush because while this is in the available + // pool, it may have been used previously. + p.mu.Unlock() + return pcid, true + } + + // Evict an existing table. + for old, pcid := range p.cache { + delete(p.cache, old) + p.cache[pt] = pcid + + // A flush is definitely required in this case, these page + // tables may still be active. (They will just be assigned some + // other PCID if and when they hit the given CPU again.) + p.mu.Unlock() + return pcid, true + } + + // No PCID. + p.mu.Unlock() + return 0, false +} + +// Drop drops references to a set of page tables. +func (p *PCIDs) Drop(pt *PageTables) { + p.mu.Lock() + if pcid, ok := p.cache[pt]; ok { + delete(p.cache, pt) + p.avail = append(p.avail, pcid) + } + p.mu.Unlock() +} diff --git a/pkg/sentry/syscalls/linux/vfs2/linux64.go b/pkg/sentry/platform/ring0/pagetables/pcids_aarch64.go index 19ee36081..fbfd41d83 100644 --- a/pkg/sentry/syscalls/linux/vfs2/linux64.go +++ b/pkg/sentry/platform/ring0/pagetables/pcids_aarch64.go @@ -12,5 +12,21 @@ // See the License for the specific language governing permissions and // limitations under the License. -// Package vfs2 provides syscall implementations that use VFS2. -package vfs2 +// +build arm64 + +package pagetables + +// limitPCID is the maximum value of PCIDs. +// +// In VMSAv8-64, the PCID(ASID) size is an IMPLEMENTATION DEFINED choice +// of 8 bits or 16 bits, and ID_AA64MMFR0_EL1.ASIDBits identifies the +// supported size. When an implementation supports a 16-bit ASID, TCR_ELx.AS +// selects whether the top 8 bits of the ASID are used. +var limitPCID uint16 + +// GetASIDBits return the system ASID bits, 8 or 16 bits. +func GetASIDBits() uint8 + +func init() { + limitPCID = uint16(1)<<GetASIDBits() - 1 +} diff --git a/pkg/sentry/platform/ring0/pagetables/pcids_aarch64.s b/pkg/sentry/platform/ring0/pagetables/pcids_aarch64.s new file mode 100644 index 000000000..e9d62d768 --- /dev/null +++ b/pkg/sentry/platform/ring0/pagetables/pcids_aarch64.s @@ -0,0 +1,45 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// +build arm64 + +#include "funcdata.h" +#include "textflag.h" + +#define ID_AA64MMFR0_ASIDBITS_SHIFT 4 +#define ID_AA64MMFR0_ASIDBITS_16 2 +#define TCR_EL1_AS_BIT 36 + +// GetASIDBits return the system ASID bits, 8 or 16 bits. +// +// func GetASIDBits() uint8 +TEXT ·GetASIDBits(SB),NOSPLIT,$0-1 + // First, check whether 16bits ASID is supported. + // ID_AA64MMFR0_EL1.ASIDBITS[7:4] == 0010. + WORD $0xd5380700 // MRS ID_AA64MMFR0_EL1, R0 + UBFX $ID_AA64MMFR0_ASIDBITS_SHIFT, R0, $4, R0 + CMPW $ID_AA64MMFR0_ASIDBITS_16, R0 + BNE bits_8 + + // Second, check whether 16bits ASID is enabled. + // TCR_EL1.AS[36] == 1. + WORD $0xd5382040 // MRS TCR_EL1, R0 + TBZ $TCR_EL1_AS_BIT, R0, bits_8 + MOVD $16, R0 + B done +bits_8: + MOVD $8, R0 +done: + MOVB R0, ret+0(FP) + RET diff --git a/pkg/sentry/platform/ring0/pagetables/pcids_x86.go b/pkg/sentry/platform/ring0/pagetables/pcids_x86.go index e199bae18..91fc5e8dd 100644 --- a/pkg/sentry/platform/ring0/pagetables/pcids_x86.go +++ b/pkg/sentry/platform/ring0/pagetables/pcids_x86.go @@ -1,4 +1,4 @@ -// Copyright 2018 The gVisor Authors. +// Copyright 2020 The gVisor Authors. // // Licensed under the Apache License, Version 2.0 (the "License"); // you may not use this file except in compliance with the License. @@ -16,94 +16,5 @@ package pagetables -import ( - "gvisor.dev/gvisor/pkg/sync" -) - -// limitPCID is the number of valid PCIDs. -const limitPCID = 4096 - -// PCIDs is a simple PCID database. -// -// This is not protected by locks and is thus suitable for use only with a -// single CPU at a time. -type PCIDs struct { - // mu protects below. - mu sync.Mutex - - // cache are the assigned page tables. - cache map[*PageTables]uint16 - - // avail are available PCIDs. - avail []uint16 -} - -// NewPCIDs returns a new PCID database. -// -// start is the first index to assign. Typically this will be one, as the zero -// pcid will always be flushed on transition (see pagetables_x86.go). This may -// be more than one if specific PCIDs are reserved. -// -// Nil is returned iff the start and size are out of range. -func NewPCIDs(start, size uint16) *PCIDs { - if start+uint16(size) >= limitPCID { - return nil // See comment. - } - p := &PCIDs{ - cache: make(map[*PageTables]uint16), - } - for pcid := start; pcid < start+size; pcid++ { - p.avail = append(p.avail, pcid) - } - return p -} - -// Assign assigns a PCID to the given PageTables. -// -// This may overwrite any previous assignment provided. If this in the case, -// true is returned to indicate that the PCID should be flushed. -func (p *PCIDs) Assign(pt *PageTables) (uint16, bool) { - p.mu.Lock() - if pcid, ok := p.cache[pt]; ok { - p.mu.Unlock() - return pcid, false // No flush. - } - - // Is there something available? - if len(p.avail) > 0 { - pcid := p.avail[len(p.avail)-1] - p.avail = p.avail[:len(p.avail)-1] - p.cache[pt] = pcid - - // We need to flush because while this is in the available - // pool, it may have been used previously. - p.mu.Unlock() - return pcid, true - } - - // Evict an existing table. - for old, pcid := range p.cache { - delete(p.cache, old) - p.cache[pt] = pcid - - // A flush is definitely required in this case, these page - // tables may still be active. (They will just be assigned some - // other PCID if and when they hit the given CPU again.) - p.mu.Unlock() - return pcid, true - } - - // No PCID. - p.mu.Unlock() - return 0, false -} - -// Drop drops references to a set of page tables. -func (p *PCIDs) Drop(pt *PageTables) { - p.mu.Lock() - if pcid, ok := p.cache[pt]; ok { - delete(p.cache, pt) - p.avail = append(p.avail, pcid) - } - p.mu.Unlock() -} +// limitPCID is the maximum value of valid PCIDs. +const limitPCID = 4095 diff --git a/pkg/sentry/platform/ring0/x86.go b/pkg/sentry/platform/ring0/x86.go index 5f80d64e8..9da0ea685 100644 --- a/pkg/sentry/platform/ring0/x86.go +++ b/pkg/sentry/platform/ring0/x86.go @@ -12,7 +12,7 @@ // See the License for the specific language governing permissions and // limitations under the License. -// +build i386 amd64 +// +build 386 amd64 package ring0 diff --git a/pkg/sentry/sighandling/sighandling.go b/pkg/sentry/sighandling/sighandling.go index ba1f9043d..83195d5a1 100644 --- a/pkg/sentry/sighandling/sighandling.go +++ b/pkg/sentry/sighandling/sighandling.go @@ -85,6 +85,11 @@ func StartSignalForwarding(handler func(linux.Signal)) func() { for sig := 1; sig <= numSignals+1; sig++ { sigchan := make(chan os.Signal, 1) sigchans = append(sigchans, sigchan) + + // SIGURG is used by Go's runtime scheduler. + if sig == int(linux.SIGURG) { + continue + } signal.Notify(sigchan, syscall.Signal(sig)) } // Start up our listener. diff --git a/pkg/sentry/socket/BUILD b/pkg/sentry/socket/BUILD index 611fa22c3..c40c6d673 100644 --- a/pkg/sentry/socket/BUILD +++ b/pkg/sentry/socket/BUILD @@ -16,6 +16,7 @@ go_library( "//pkg/sentry/kernel", "//pkg/sentry/kernel/time", "//pkg/sentry/socket/unix/transport", + "//pkg/sentry/vfs", "//pkg/syserr", "//pkg/tcpip", "//pkg/usermem", diff --git a/pkg/sentry/socket/control/BUILD b/pkg/sentry/socket/control/BUILD index 4d42d29cb..ca16d0381 100644 --- a/pkg/sentry/socket/control/BUILD +++ b/pkg/sentry/socket/control/BUILD @@ -4,7 +4,10 @@ package(licenses = ["notice"]) go_library( name = "control", - srcs = ["control.go"], + srcs = [ + "control.go", + "control_vfs2.go", + ], imports = [ "gvisor.dev/gvisor/pkg/sentry/fs", ], @@ -18,6 +21,7 @@ go_library( "//pkg/sentry/kernel/auth", "//pkg/sentry/socket", "//pkg/sentry/socket/unix/transport", + "//pkg/sentry/vfs", "//pkg/syserror", "//pkg/tcpip", "//pkg/usermem", diff --git a/pkg/sentry/socket/control/control.go b/pkg/sentry/socket/control/control.go index 4667373d2..8b439a078 100644 --- a/pkg/sentry/socket/control/control.go +++ b/pkg/sentry/socket/control/control.go @@ -41,6 +41,8 @@ type SCMCredentials interface { Credentials(t *kernel.Task) (kernel.ThreadID, auth.UID, auth.GID) } +// LINT.IfChange + // SCMRights represents a SCM_RIGHTS socket control message. type SCMRights interface { transport.RightsControlMessage @@ -142,6 +144,8 @@ func PackRights(t *kernel.Task, rights SCMRights, cloexec bool, buf []byte, flag return putCmsg(buf, flags, linux.SCM_RIGHTS, align, fds) } +// LINT.ThenChange(./control_vfs2.go) + // scmCredentials represents an SCM_CREDENTIALS socket control message. // // +stateify savable @@ -329,7 +333,7 @@ func PackTOS(t *kernel.Task, tos uint8, buf []byte) []byte { } // PackTClass packs an IPV6_TCLASS socket control message. -func PackTClass(t *kernel.Task, tClass int32, buf []byte) []byte { +func PackTClass(t *kernel.Task, tClass uint32, buf []byte) []byte { return putCmsgStruct( buf, linux.SOL_IPV6, @@ -537,11 +541,19 @@ func Parse(t *kernel.Task, socketOrEndpoint interface{}, buf []byte) (socket.Con } if len(fds) > 0 { - rights, err := NewSCMRights(t, fds) - if err != nil { - return socket.ControlMessages{}, err + if kernel.VFS2Enabled { + rights, err := NewSCMRightsVFS2(t, fds) + if err != nil { + return socket.ControlMessages{}, err + } + cmsgs.Unix.Rights = rights + } else { + rights, err := NewSCMRights(t, fds) + if err != nil { + return socket.ControlMessages{}, err + } + cmsgs.Unix.Rights = rights } - cmsgs.Unix.Rights = rights } return cmsgs, nil @@ -566,6 +578,8 @@ func MakeCreds(t *kernel.Task) SCMCredentials { return &scmCredentials{t, tcred.EffectiveKUID, tcred.EffectiveKGID} } +// LINT.IfChange + // New creates default control messages if needed. func New(t *kernel.Task, socketOrEndpoint interface{}, rights SCMRights) transport.ControlMessages { return transport.ControlMessages{ @@ -573,3 +587,5 @@ func New(t *kernel.Task, socketOrEndpoint interface{}, rights SCMRights) transpo Rights: rights, } } + +// LINT.ThenChange(./control_vfs2.go) diff --git a/pkg/sentry/socket/control/control_vfs2.go b/pkg/sentry/socket/control/control_vfs2.go new file mode 100644 index 000000000..fd08179be --- /dev/null +++ b/pkg/sentry/socket/control/control_vfs2.go @@ -0,0 +1,131 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package control + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" +) + +// SCMRightsVFS2 represents a SCM_RIGHTS socket control message. +type SCMRightsVFS2 interface { + transport.RightsControlMessage + + // Files returns up to max RightsFiles. + // + // Returned files are consumed and ownership is transferred to the caller. + // Subsequent calls to Files will return the next files. + Files(ctx context.Context, max int) (rf RightsFilesVFS2, truncated bool) +} + +// RightsFiles represents a SCM_RIGHTS socket control message. A reference is +// maintained for each vfs.FileDescription and is release either when an FD is created or +// when the Release method is called. +type RightsFilesVFS2 []*vfs.FileDescription + +// NewSCMRightsVFS2 creates a new SCM_RIGHTS socket control message +// representation using local sentry FDs. +func NewSCMRightsVFS2(t *kernel.Task, fds []int32) (SCMRightsVFS2, error) { + files := make(RightsFilesVFS2, 0, len(fds)) + for _, fd := range fds { + file := t.GetFileVFS2(fd) + if file == nil { + files.Release() + return nil, syserror.EBADF + } + files = append(files, file) + } + return &files, nil +} + +// Files implements SCMRights.Files. +func (fs *RightsFilesVFS2) Files(ctx context.Context, max int) (RightsFilesVFS2, bool) { + n := max + var trunc bool + if l := len(*fs); n > l { + n = l + } else if n < l { + trunc = true + } + rf := (*fs)[:n] + *fs = (*fs)[n:] + return rf, trunc +} + +// Clone implements transport.RightsControlMessage.Clone. +func (fs *RightsFilesVFS2) Clone() transport.RightsControlMessage { + nfs := append(RightsFilesVFS2(nil), *fs...) + for _, nf := range nfs { + nf.IncRef() + } + return &nfs +} + +// Release implements transport.RightsControlMessage.Release. +func (fs *RightsFilesVFS2) Release() { + for _, f := range *fs { + f.DecRef() + } + *fs = nil +} + +// rightsFDsVFS2 gets up to the specified maximum number of FDs. +func rightsFDsVFS2(t *kernel.Task, rights SCMRightsVFS2, cloexec bool, max int) ([]int32, bool) { + files, trunc := rights.Files(t, max) + fds := make([]int32, 0, len(files)) + for i := 0; i < max && len(files) > 0; i++ { + fd, err := t.NewFDFromVFS2(0, files[0], kernel.FDFlags{ + CloseOnExec: cloexec, + }) + files[0].DecRef() + files = files[1:] + if err != nil { + t.Warningf("Error inserting FD: %v", err) + // This is what Linux does. + break + } + + fds = append(fds, int32(fd)) + } + return fds, trunc +} + +// PackRightsVFS2 packs as many FDs as will fit into the unused capacity of buf. +func PackRightsVFS2(t *kernel.Task, rights SCMRightsVFS2, cloexec bool, buf []byte, flags int) ([]byte, int) { + maxFDs := (cap(buf) - len(buf) - linux.SizeOfControlMessageHeader) / 4 + // Linux does not return any FDs if none fit. + if maxFDs <= 0 { + flags |= linux.MSG_CTRUNC + return buf, flags + } + fds, trunc := rightsFDsVFS2(t, rights, cloexec, maxFDs) + if trunc { + flags |= linux.MSG_CTRUNC + } + align := t.Arch().Width() + return putCmsg(buf, flags, linux.SCM_RIGHTS, align, fds) +} + +// NewVFS2 creates default control messages if needed. +func NewVFS2(t *kernel.Task, socketOrEndpoint interface{}, rights SCMRightsVFS2) transport.ControlMessages { + return transport.ControlMessages{ + Credentials: makeCreds(t, socketOrEndpoint), + Rights: rights, + } +} diff --git a/pkg/sentry/socket/hostinet/BUILD b/pkg/sentry/socket/hostinet/BUILD index 023bad156..e82d6cd1e 100644 --- a/pkg/sentry/socket/hostinet/BUILD +++ b/pkg/sentry/socket/hostinet/BUILD @@ -10,6 +10,7 @@ go_library( "save_restore.go", "socket.go", "socket_unsafe.go", + "socket_vfs2.go", "sockopt_impl.go", "stack.go", ], @@ -25,11 +26,14 @@ go_library( "//pkg/sentry/device", "//pkg/sentry/fs", "//pkg/sentry/fs/fsutil", + "//pkg/sentry/fsimpl/sockfs", + "//pkg/sentry/hostfd", "//pkg/sentry/inet", "//pkg/sentry/kernel", "//pkg/sentry/kernel/time", "//pkg/sentry/socket", "//pkg/sentry/socket/control", + "//pkg/sentry/vfs", "//pkg/syserr", "//pkg/syserror", "//pkg/tcpip/stack", diff --git a/pkg/sentry/socket/hostinet/socket.go b/pkg/sentry/socket/hostinet/socket.go index 22f78d2e2..c11e82c10 100644 --- a/pkg/sentry/socket/hostinet/socket.go +++ b/pkg/sentry/socket/hostinet/socket.go @@ -25,6 +25,7 @@ import ( "gvisor.dev/gvisor/pkg/fdnotifier" "gvisor.dev/gvisor/pkg/log" "gvisor.dev/gvisor/pkg/safemem" + "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fs/fsutil" "gvisor.dev/gvisor/pkg/sentry/kernel" @@ -49,6 +50,8 @@ const ( maxControlLen = 1024 ) +// LINT.IfChange + // socketOperations implements fs.FileOperations and socket.Socket for a socket // implemented using a host socket. type socketOperations struct { @@ -59,23 +62,37 @@ type socketOperations struct { fsutil.FileNoSplice `state:"nosave"` fsutil.FileNoopFlush `state:"nosave"` fsutil.FileUseInodeUnstableAttr `state:"nosave"` + + socketOpsCommon +} + +// socketOpsCommon contains the socket operations common to VFS1 and VFS2. +// +// +stateify savable +type socketOpsCommon struct { socket.SendReceiveTimeout family int // Read-only. stype linux.SockType // Read-only. protocol int // Read-only. - fd int // must be O_NONBLOCK queue waiter.Queue + + // fd is the host socket fd. It must have O_NONBLOCK, so that operations + // will return EWOULDBLOCK instead of blocking on the host. This allows us to + // handle blocking behavior independently in the sentry. + fd int } var _ = socket.Socket(&socketOperations{}) func newSocketFile(ctx context.Context, family int, stype linux.SockType, protocol int, fd int, nonblock bool) (*fs.File, *syserr.Error) { s := &socketOperations{ - family: family, - stype: stype, - protocol: protocol, - fd: fd, + socketOpsCommon: socketOpsCommon{ + family: family, + stype: stype, + protocol: protocol, + fd: fd, + }, } if err := fdnotifier.AddFD(int32(fd), &s.queue); err != nil { return nil, syserr.FromError(err) @@ -86,28 +103,33 @@ func newSocketFile(ctx context.Context, family int, stype linux.SockType, protoc } // Release implements fs.FileOperations.Release. -func (s *socketOperations) Release() { +func (s *socketOpsCommon) Release() { fdnotifier.RemoveFD(int32(s.fd)) syscall.Close(s.fd) } // Readiness implements waiter.Waitable.Readiness. -func (s *socketOperations) Readiness(mask waiter.EventMask) waiter.EventMask { +func (s *socketOpsCommon) Readiness(mask waiter.EventMask) waiter.EventMask { return fdnotifier.NonBlockingPoll(int32(s.fd), mask) } // EventRegister implements waiter.Waitable.EventRegister. -func (s *socketOperations) EventRegister(e *waiter.Entry, mask waiter.EventMask) { +func (s *socketOpsCommon) EventRegister(e *waiter.Entry, mask waiter.EventMask) { s.queue.EventRegister(e, mask) fdnotifier.UpdateFD(int32(s.fd)) } // EventUnregister implements waiter.Waitable.EventUnregister. -func (s *socketOperations) EventUnregister(e *waiter.Entry) { +func (s *socketOpsCommon) EventUnregister(e *waiter.Entry) { s.queue.EventUnregister(e) fdnotifier.UpdateFD(int32(s.fd)) } +// Ioctl implements fs.FileOperations.Ioctl. +func (s *socketOperations) Ioctl(ctx context.Context, _ *fs.File, io usermem.IO, args arch.SyscallArguments) (uintptr, error) { + return ioctl(ctx, s.fd, io, args) +} + // Read implements fs.FileOperations.Read. func (s *socketOperations) Read(ctx context.Context, _ *fs.File, dst usermem.IOSequence, _ int64) (int64, error) { n, err := dst.CopyOutFrom(ctx, safemem.ReaderFunc(func(dsts safemem.BlockSeq) (uint64, error) { @@ -155,7 +177,7 @@ func (s *socketOperations) Write(ctx context.Context, _ *fs.File, src usermem.IO } // Connect implements socket.Socket.Connect. -func (s *socketOperations) Connect(t *kernel.Task, sockaddr []byte, blocking bool) *syserr.Error { +func (s *socketOpsCommon) Connect(t *kernel.Task, sockaddr []byte, blocking bool) *syserr.Error { if len(sockaddr) > sizeofSockaddr { sockaddr = sockaddr[:sizeofSockaddr] } @@ -195,7 +217,7 @@ func (s *socketOperations) Connect(t *kernel.Task, sockaddr []byte, blocking boo } // Accept implements socket.Socket.Accept. -func (s *socketOperations) Accept(t *kernel.Task, peerRequested bool, flags int, blocking bool) (int32, linux.SockAddr, uint32, *syserr.Error) { +func (s *socketOpsCommon) Accept(t *kernel.Task, peerRequested bool, flags int, blocking bool) (int32, linux.SockAddr, uint32, *syserr.Error) { var peerAddr linux.SockAddr var peerAddrBuf []byte var peerAddrlen uint32 @@ -209,7 +231,7 @@ func (s *socketOperations) Accept(t *kernel.Task, peerRequested bool, flags int, } // Conservatively ignore all flags specified by the application and add - // SOCK_NONBLOCK since socketOperations requires it. + // SOCK_NONBLOCK since socketOpsCommon requires it. fd, syscallErr := accept4(s.fd, peerAddrPtr, peerAddrlenPtr, syscall.SOCK_NONBLOCK|syscall.SOCK_CLOEXEC) if blocking { var ch chan struct{} @@ -235,23 +257,41 @@ func (s *socketOperations) Accept(t *kernel.Task, peerRequested bool, flags int, return 0, peerAddr, peerAddrlen, syserr.FromError(syscallErr) } - f, err := newSocketFile(t, s.family, s.stype, s.protocol, fd, flags&syscall.SOCK_NONBLOCK != 0) - if err != nil { - syscall.Close(fd) - return 0, nil, 0, err - } - defer f.DecRef() + var ( + kfd int32 + kerr error + ) + if kernel.VFS2Enabled { + f, err := newVFS2Socket(t, s.family, s.stype, s.protocol, fd, uint32(flags&syscall.SOCK_NONBLOCK)) + if err != nil { + syscall.Close(fd) + return 0, nil, 0, err + } + defer f.DecRef() - kfd, kerr := t.NewFDFrom(0, f, kernel.FDFlags{ - CloseOnExec: flags&syscall.SOCK_CLOEXEC != 0, - }) - t.Kernel().RecordSocket(f) + kfd, kerr = t.NewFDFromVFS2(0, f, kernel.FDFlags{ + CloseOnExec: flags&syscall.SOCK_CLOEXEC != 0, + }) + t.Kernel().RecordSocketVFS2(f) + } else { + f, err := newSocketFile(t, s.family, s.stype, s.protocol, fd, flags&syscall.SOCK_NONBLOCK != 0) + if err != nil { + syscall.Close(fd) + return 0, nil, 0, err + } + defer f.DecRef() + + kfd, kerr = t.NewFDFrom(0, f, kernel.FDFlags{ + CloseOnExec: flags&syscall.SOCK_CLOEXEC != 0, + }) + t.Kernel().RecordSocket(f) + } return kfd, peerAddr, peerAddrlen, syserr.FromError(kerr) } // Bind implements socket.Socket.Bind. -func (s *socketOperations) Bind(t *kernel.Task, sockaddr []byte) *syserr.Error { +func (s *socketOpsCommon) Bind(t *kernel.Task, sockaddr []byte) *syserr.Error { if len(sockaddr) > sizeofSockaddr { sockaddr = sockaddr[:sizeofSockaddr] } @@ -264,12 +304,12 @@ func (s *socketOperations) Bind(t *kernel.Task, sockaddr []byte) *syserr.Error { } // Listen implements socket.Socket.Listen. -func (s *socketOperations) Listen(t *kernel.Task, backlog int) *syserr.Error { +func (s *socketOpsCommon) Listen(t *kernel.Task, backlog int) *syserr.Error { return syserr.FromError(syscall.Listen(s.fd, backlog)) } // Shutdown implements socket.Socket.Shutdown. -func (s *socketOperations) Shutdown(t *kernel.Task, how int) *syserr.Error { +func (s *socketOpsCommon) Shutdown(t *kernel.Task, how int) *syserr.Error { switch how { case syscall.SHUT_RD, syscall.SHUT_WR, syscall.SHUT_RDWR: return syserr.FromError(syscall.Shutdown(s.fd, how)) @@ -279,7 +319,7 @@ func (s *socketOperations) Shutdown(t *kernel.Task, how int) *syserr.Error { } // GetSockOpt implements socket.Socket.GetSockOpt. -func (s *socketOperations) GetSockOpt(t *kernel.Task, level int, name int, outPtr usermem.Addr, outLen int) (interface{}, *syserr.Error) { +func (s *socketOpsCommon) GetSockOpt(t *kernel.Task, level int, name int, outPtr usermem.Addr, outLen int) (interface{}, *syserr.Error) { if outLen < 0 { return nil, syserr.ErrInvalidArgument } @@ -328,7 +368,7 @@ func (s *socketOperations) GetSockOpt(t *kernel.Task, level int, name int, outPt } // SetSockOpt implements socket.Socket.SetSockOpt. -func (s *socketOperations) SetSockOpt(t *kernel.Task, level int, name int, opt []byte) *syserr.Error { +func (s *socketOpsCommon) SetSockOpt(t *kernel.Task, level int, name int, opt []byte) *syserr.Error { // Whitelist options and constrain option length. optlen := setSockOptLen(t, level, name) switch level { @@ -374,7 +414,7 @@ func (s *socketOperations) SetSockOpt(t *kernel.Task, level int, name int, opt [ } // RecvMsg implements socket.Socket.RecvMsg. -func (s *socketOperations) RecvMsg(t *kernel.Task, dst usermem.IOSequence, flags int, haveDeadline bool, deadline ktime.Time, senderRequested bool, controlLen uint64) (int, int, linux.SockAddr, uint32, socket.ControlMessages, *syserr.Error) { +func (s *socketOpsCommon) RecvMsg(t *kernel.Task, dst usermem.IOSequence, flags int, haveDeadline bool, deadline ktime.Time, senderRequested bool, controlLen uint64) (int, int, linux.SockAddr, uint32, socket.ControlMessages, *syserr.Error) { // Whitelist flags. // // FIXME(jamieliu): We can't support MSG_ERRQUEUE because it uses ancillary @@ -496,7 +536,7 @@ func (s *socketOperations) RecvMsg(t *kernel.Task, dst usermem.IOSequence, flags } // SendMsg implements socket.Socket.SendMsg. -func (s *socketOperations) SendMsg(t *kernel.Task, src usermem.IOSequence, to []byte, flags int, haveDeadline bool, deadline ktime.Time, controlMessages socket.ControlMessages) (int, *syserr.Error) { +func (s *socketOpsCommon) SendMsg(t *kernel.Task, src usermem.IOSequence, to []byte, flags int, haveDeadline bool, deadline ktime.Time, controlMessages socket.ControlMessages) (int, *syserr.Error) { // Whitelist flags. if flags&^(syscall.MSG_DONTWAIT|syscall.MSG_EOR|syscall.MSG_FASTOPEN|syscall.MSG_MORE|syscall.MSG_NOSIGNAL) != 0 { return 0, syserr.ErrInvalidArgument @@ -515,7 +555,7 @@ func (s *socketOperations) SendMsg(t *kernel.Task, src usermem.IOSequence, to [] if uint64(src.NumBytes()) != srcs.NumBytes() { return 0, nil } - if srcs.IsEmpty() { + if srcs.IsEmpty() && len(controlBuf) == 0 { return 0, nil } @@ -585,7 +625,7 @@ func translateIOSyscallError(err error) error { } // State implements socket.Socket.State. -func (s *socketOperations) State() uint32 { +func (s *socketOpsCommon) State() uint32 { info := linux.TCPInfo{} buf, err := getsockopt(s.fd, syscall.SOL_TCP, syscall.TCP_INFO, linux.SizeOfTCPInfo) if err != nil { @@ -607,7 +647,7 @@ func (s *socketOperations) State() uint32 { } // Type implements socket.Socket.Type. -func (s *socketOperations) Type() (family int, skType linux.SockType, protocol int) { +func (s *socketOpsCommon) Type() (family int, skType linux.SockType, protocol int) { return s.family, s.stype, s.protocol } @@ -663,8 +703,11 @@ func (p *socketProvider) Pair(t *kernel.Task, stype linux.SockType, protocol int return nil, nil, nil } +// LINT.ThenChange(./socket_vfs2.go) + func init() { for _, family := range []int{syscall.AF_INET, syscall.AF_INET6} { socket.RegisterProvider(family, &socketProvider{family}) + socket.RegisterProviderVFS2(family, &socketProviderVFS2{}) } } diff --git a/pkg/sentry/socket/hostinet/socket_unsafe.go b/pkg/sentry/socket/hostinet/socket_unsafe.go index cd67234d2..3f420c2ec 100644 --- a/pkg/sentry/socket/hostinet/socket_unsafe.go +++ b/pkg/sentry/socket/hostinet/socket_unsafe.go @@ -21,7 +21,6 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/arch" - "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/socket" "gvisor.dev/gvisor/pkg/syserr" @@ -54,12 +53,11 @@ func writev(fd int, srcs []syscall.Iovec) (uint64, error) { return uint64(n), nil } -// Ioctl implements fs.FileOperations.Ioctl. -func (s *socketOperations) Ioctl(ctx context.Context, _ *fs.File, io usermem.IO, args arch.SyscallArguments) (uintptr, error) { +func ioctl(ctx context.Context, fd int, io usermem.IO, args arch.SyscallArguments) (uintptr, error) { switch cmd := uintptr(args[1].Int()); cmd { case syscall.TIOCINQ, syscall.TIOCOUTQ: var val int32 - if _, _, errno := syscall.Syscall(syscall.SYS_IOCTL, uintptr(s.fd), cmd, uintptr(unsafe.Pointer(&val))); errno != 0 { + if _, _, errno := syscall.Syscall(syscall.SYS_IOCTL, uintptr(fd), cmd, uintptr(unsafe.Pointer(&val))); errno != 0 { return 0, translateIOSyscallError(errno) } var buf [4]byte @@ -93,7 +91,7 @@ func getsockopt(fd int, level, name int, optlen int) ([]byte, error) { } // GetSockName implements socket.Socket.GetSockName. -func (s *socketOperations) GetSockName(t *kernel.Task) (linux.SockAddr, uint32, *syserr.Error) { +func (s *socketOpsCommon) GetSockName(t *kernel.Task) (linux.SockAddr, uint32, *syserr.Error) { addr := make([]byte, sizeofSockaddr) addrlen := uint32(len(addr)) _, _, errno := syscall.Syscall(syscall.SYS_GETSOCKNAME, uintptr(s.fd), uintptr(unsafe.Pointer(&addr[0])), uintptr(unsafe.Pointer(&addrlen))) @@ -104,7 +102,7 @@ func (s *socketOperations) GetSockName(t *kernel.Task) (linux.SockAddr, uint32, } // GetPeerName implements socket.Socket.GetPeerName. -func (s *socketOperations) GetPeerName(t *kernel.Task) (linux.SockAddr, uint32, *syserr.Error) { +func (s *socketOpsCommon) GetPeerName(t *kernel.Task) (linux.SockAddr, uint32, *syserr.Error) { addr := make([]byte, sizeofSockaddr) addrlen := uint32(len(addr)) _, _, errno := syscall.Syscall(syscall.SYS_GETPEERNAME, uintptr(s.fd), uintptr(unsafe.Pointer(&addr[0])), uintptr(unsafe.Pointer(&addrlen))) diff --git a/pkg/sentry/socket/hostinet/socket_vfs2.go b/pkg/sentry/socket/hostinet/socket_vfs2.go new file mode 100644 index 000000000..677743113 --- /dev/null +++ b/pkg/sentry/socket/hostinet/socket_vfs2.go @@ -0,0 +1,184 @@ +// 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 hostinet + +import ( + "syscall" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/fdnotifier" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/sockfs" + "gvisor.dev/gvisor/pkg/sentry/hostfd" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/socket" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserr" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" + "gvisor.dev/gvisor/pkg/waiter" +) + +type socketVFS2 struct { + vfsfd vfs.FileDescription + vfs.FileDescriptionDefaultImpl + + // We store metadata for hostinet sockets internally. Technically, we should + // access metadata (e.g. through stat, chmod) on the host for correctness, + // but this is not very useful for inet socket fds, which do not belong to a + // concrete file anyway. + vfs.DentryMetadataFileDescriptionImpl + + socketOpsCommon +} + +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) + + s := &socketVFS2{ + socketOpsCommon: socketOpsCommon{ + family: family, + stype: stype, + protocol: protocol, + fd: fd, + }, + } + if err := fdnotifier.AddFD(int32(fd), &s.queue); err != nil { + return nil, syserr.FromError(err) + } + vfsfd := &s.vfsfd + if err := vfsfd.Init(s, linux.O_RDWR|(flags&linux.O_NONBLOCK), mnt, d, &vfs.FileDescriptionOptions{ + DenyPRead: true, + DenyPWrite: true, + UseDentryMetadata: true, + }); err != nil { + return nil, syserr.FromError(err) + } + return vfsfd, nil +} + +// Readiness implements waiter.Waitable.Readiness. +func (s *socketVFS2) Readiness(mask waiter.EventMask) waiter.EventMask { + return s.socketOpsCommon.Readiness(mask) +} + +// EventRegister implements waiter.Waitable.EventRegister. +func (s *socketVFS2) EventRegister(e *waiter.Entry, mask waiter.EventMask) { + s.socketOpsCommon.EventRegister(e, mask) +} + +// EventUnregister implements waiter.Waitable.EventUnregister. +func (s *socketVFS2) EventUnregister(e *waiter.Entry) { + s.socketOpsCommon.EventUnregister(e) +} + +// Ioctl implements vfs.FileDescriptionImpl. +func (s *socketVFS2) Ioctl(ctx context.Context, uio usermem.IO, args arch.SyscallArguments) (uintptr, error) { + return ioctl(ctx, s.fd, uio, args) +} + +// PRead implements vfs.FileDescriptionImpl. +func (s *socketVFS2) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { + return 0, syserror.ESPIPE +} + +// Read implements vfs.FileDescriptionImpl. +func (s *socketVFS2) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { + // All flags other than RWF_NOWAIT should be ignored. + // TODO(gvisor.dev/issue/2601): Support RWF_NOWAIT. + if opts.Flags != 0 { + return 0, syserror.EOPNOTSUPP + } + + reader := hostfd.GetReadWriterAt(int32(s.fd), -1, opts.Flags) + n, err := dst.CopyOutFrom(ctx, reader) + hostfd.PutReadWriterAt(reader) + return int64(n), err +} + +// PWrite implements vfs.FileDescriptionImpl. +func (s *socketVFS2) PWrite(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) { + return 0, syserror.ESPIPE +} + +// Write implements vfs.FileDescriptionImpl. +func (s *socketVFS2) Write(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) { + // All flags other than RWF_NOWAIT should be ignored. + // TODO(gvisor.dev/issue/2601): Support RWF_NOWAIT. + if opts.Flags != 0 { + return 0, syserror.EOPNOTSUPP + } + + writer := hostfd.GetReadWriterAt(int32(s.fd), -1, opts.Flags) + n, err := src.CopyInTo(ctx, writer) + hostfd.PutReadWriterAt(writer) + return int64(n), err +} + +type socketProviderVFS2 struct { + family int +} + +// Socket implements socket.ProviderVFS2.Socket. +func (p *socketProviderVFS2) Socket(t *kernel.Task, stypeflags linux.SockType, protocol int) (*vfs.FileDescription, *syserr.Error) { + // Check that we are using the host network stack. + stack := t.NetworkContext() + if stack == nil { + return nil, nil + } + if _, ok := stack.(*Stack); !ok { + return nil, nil + } + + // Only accept TCP and UDP. + stype := stypeflags & linux.SOCK_TYPE_MASK + switch stype { + case syscall.SOCK_STREAM: + switch protocol { + case 0, syscall.IPPROTO_TCP: + // ok + default: + return nil, nil + } + case syscall.SOCK_DGRAM: + switch protocol { + case 0, syscall.IPPROTO_UDP: + // ok + default: + return nil, nil + } + default: + return nil, nil + } + + // Conservatively ignore all flags specified by the application and add + // SOCK_NONBLOCK since socketOperations requires it. Pass a protocol of 0 + // to simplify the syscall filters, since 0 and IPPROTO_* are equivalent. + fd, err := syscall.Socket(p.family, int(stype)|syscall.SOCK_NONBLOCK|syscall.SOCK_CLOEXEC, 0) + if err != nil { + return nil, syserr.FromError(err) + } + return newVFS2Socket(t, p.family, stype, protocol, fd, uint32(stypeflags&syscall.SOCK_NONBLOCK)) +} + +// Pair implements socket.Provider.Pair. +func (p *socketProviderVFS2) Pair(t *kernel.Task, stype linux.SockType, protocol int) (*vfs.FileDescription, *vfs.FileDescription, *syserr.Error) { + // Not supported by AF_INET/AF_INET6. + return nil, nil, nil +} diff --git a/pkg/sentry/socket/netfilter/BUILD b/pkg/sentry/socket/netfilter/BUILD index c91ec7494..721094bbf 100644 --- a/pkg/sentry/socket/netfilter/BUILD +++ b/pkg/sentry/socket/netfilter/BUILD @@ -7,6 +7,8 @@ go_library( srcs = [ "extensions.go", "netfilter.go", + "owner_matcher.go", + "targets.go", "tcp_matcher.go", "udp_matcher.go", ], @@ -21,7 +23,6 @@ go_library( "//pkg/syserr", "//pkg/tcpip", "//pkg/tcpip/header", - "//pkg/tcpip/iptables", "//pkg/tcpip/stack", "//pkg/usermem", ], diff --git a/pkg/sentry/socket/netfilter/extensions.go b/pkg/sentry/socket/netfilter/extensions.go index b4b244abf..0336a32d8 100644 --- a/pkg/sentry/socket/netfilter/extensions.go +++ b/pkg/sentry/socket/netfilter/extensions.go @@ -19,7 +19,7 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/binary" - "gvisor.dev/gvisor/pkg/tcpip/iptables" + "gvisor.dev/gvisor/pkg/tcpip/stack" "gvisor.dev/gvisor/pkg/usermem" ) @@ -37,12 +37,12 @@ type matchMaker interface { // name is the matcher name as stored in the xt_entry_match struct. name() string - // marshal converts from an iptables.Matcher to an ABI struct. - marshal(matcher iptables.Matcher) []byte + // marshal converts from an stack.Matcher to an ABI struct. + marshal(matcher stack.Matcher) []byte // unmarshal converts from the ABI matcher struct to an - // iptables.Matcher. - unmarshal(buf []byte, filter iptables.IPHeaderFilter) (iptables.Matcher, error) + // stack.Matcher. + unmarshal(buf []byte, filter stack.IPHeaderFilter) (stack.Matcher, error) } // matchMakers maps the name of supported matchers to the matchMaker that @@ -58,7 +58,7 @@ func registerMatchMaker(mm matchMaker) { matchMakers[mm.name()] = mm } -func marshalMatcher(matcher iptables.Matcher) []byte { +func marshalMatcher(matcher stack.Matcher) []byte { matchMaker, ok := matchMakers[matcher.Name()] if !ok { panic(fmt.Sprintf("Unknown matcher of type %T.", matcher)) @@ -86,7 +86,7 @@ func marshalEntryMatch(name string, data []byte) []byte { return append(buf, make([]byte, size-len(buf))...) } -func unmarshalMatcher(match linux.XTEntryMatch, filter iptables.IPHeaderFilter, buf []byte) (iptables.Matcher, error) { +func unmarshalMatcher(match linux.XTEntryMatch, filter stack.IPHeaderFilter, buf []byte) (stack.Matcher, error) { matchMaker, ok := matchMakers[match.Name.String()] if !ok { return nil, fmt.Errorf("unsupported matcher with name %q", match.Name.String()) diff --git a/pkg/sentry/socket/netfilter/netfilter.go b/pkg/sentry/socket/netfilter/netfilter.go index 3fc80e0de..66015e2bc 100644 --- a/pkg/sentry/socket/netfilter/netfilter.go +++ b/pkg/sentry/socket/netfilter/netfilter.go @@ -17,6 +17,7 @@ package netfilter import ( + "bytes" "errors" "fmt" @@ -26,7 +27,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/syserr" "gvisor.dev/gvisor/pkg/tcpip" - "gvisor.dev/gvisor/pkg/tcpip/iptables" + "gvisor.dev/gvisor/pkg/tcpip/header" "gvisor.dev/gvisor/pkg/tcpip/stack" "gvisor.dev/gvisor/pkg/usermem" ) @@ -35,6 +36,11 @@ import ( // shouldn't be reached - an error has occurred if we fall through to one. const errorTargetName = "ERROR" +// redirectTargetName is used to mark targets as redirect targets. Redirect +// targets should be reached for only NAT and Mangle tables. These targets will +// change the destination port/destination IP for packets. +const redirectTargetName = "REDIRECT" + // Metadata is used to verify that we are correctly serializing and // deserializing iptables into structs consumable by the iptables tool. We save // a metadata struct when the tables are written, and when they are read out we @@ -48,9 +54,23 @@ type metadata struct { Size uint32 } +// enableLogging controls whether to log the (de)serialization of netfilter +// structs between userspace and netstack. These logs are useful when +// developing iptables, but can pollute sentry logs otherwise. +const enableLogging = false + +// emptyFilter is for comparison with a rule's filters to determine whether it +// is also empty. It is immutable. +var emptyFilter = stack.IPHeaderFilter{ + Dst: "\x00\x00\x00\x00", + DstMask: "\x00\x00\x00\x00", + Src: "\x00\x00\x00\x00", + SrcMask: "\x00\x00\x00\x00", +} + // nflog logs messages related to the writing and reading of iptables. func nflog(format string, args ...interface{}) { - if log.IsLogging(log.Debug) { + if enableLogging && log.IsLogging(log.Debug) { log.Debugf("netfilter: "+format, args...) } } @@ -123,39 +143,35 @@ func GetEntries(t *kernel.Task, stack *stack.Stack, outPtr usermem.Addr, outLen return entries, nil } -func findTable(stack *stack.Stack, tablename linux.TableName) (iptables.Table, error) { - ipt := stack.IPTables() - table, ok := ipt.Tables[tablename.String()] +func findTable(stk *stack.Stack, tablename linux.TableName) (stack.Table, error) { + table, ok := stk.IPTables().GetTable(tablename.String()) if !ok { - return iptables.Table{}, fmt.Errorf("couldn't find table %q", tablename) + return stack.Table{}, fmt.Errorf("couldn't find table %q", tablename) } return table, nil } -// FillDefaultIPTables sets stack's IPTables to the default tables and -// populates them with metadata. -func FillDefaultIPTables(stack *stack.Stack) { - ipt := iptables.DefaultTables() - - // In order to fill in the metadata, we have to translate ipt from its - // netstack format to Linux's giant-binary-blob format. - for name, table := range ipt.Tables { - _, metadata, err := convertNetstackToBinary(name, table) - if err != nil { - panic(fmt.Errorf("Unable to set default IP tables: %v", err)) +// FillIPTablesMetadata populates stack's IPTables with metadata. +func FillIPTablesMetadata(stk *stack.Stack) { + stk.IPTables().ModifyTables(func(tables map[string]stack.Table) { + // In order to fill in the metadata, we have to translate ipt from its + // netstack format to Linux's giant-binary-blob format. + for name, table := range tables { + _, metadata, err := convertNetstackToBinary(name, table) + if err != nil { + panic(fmt.Errorf("Unable to set default IP tables: %v", err)) + } + table.SetMetadata(metadata) + tables[name] = table } - table.SetMetadata(metadata) - ipt.Tables[name] = table - } - - stack.SetIPTables(ipt) + }) } // convertNetstackToBinary converts the iptables as stored in netstack to the // format expected by the iptables tool. Linux stores each table as a binary // blob that can only be traversed by parsing a bit, reading some offsets, // jumping to those offsets, parsing again, etc. -func convertNetstackToBinary(tablename string, table iptables.Table) (linux.KernelIPTGetEntries, metadata, error) { +func convertNetstackToBinary(tablename string, table stack.Table) (linux.KernelIPTGetEntries, metadata, error) { // Return values. var entries linux.KernelIPTGetEntries var meta metadata @@ -194,6 +210,21 @@ func convertNetstackToBinary(tablename string, table iptables.Table) (linux.Kern TargetOffset: linux.SizeOfIPTEntry, }, } + copy(entry.IPTEntry.IP.Dst[:], rule.Filter.Dst) + copy(entry.IPTEntry.IP.DstMask[:], rule.Filter.DstMask) + copy(entry.IPTEntry.IP.Src[:], rule.Filter.Src) + copy(entry.IPTEntry.IP.SrcMask[:], rule.Filter.SrcMask) + copy(entry.IPTEntry.IP.OutputInterface[:], rule.Filter.OutputInterface) + copy(entry.IPTEntry.IP.OutputInterfaceMask[:], rule.Filter.OutputInterfaceMask) + if rule.Filter.DstInvert { + entry.IPTEntry.IP.InverseFlags |= linux.IPT_INV_DSTIP + } + if rule.Filter.SrcInvert { + entry.IPTEntry.IP.InverseFlags |= linux.IPT_INV_SRCIP + } + if rule.Filter.OutputInterfaceInvert { + entry.IPTEntry.IP.InverseFlags |= linux.IPT_INV_VIA_OUT + } for _, matcher := range rule.Matchers { // Serialize the matcher and add it to the @@ -228,25 +259,29 @@ func convertNetstackToBinary(tablename string, table iptables.Table) (linux.Kern return entries, meta, nil } -func marshalTarget(target iptables.Target) []byte { +func marshalTarget(target stack.Target) []byte { switch tg := target.(type) { - case iptables.AcceptTarget: - return marshalStandardTarget(iptables.RuleAccept) - case iptables.DropTarget: - return marshalStandardTarget(iptables.RuleDrop) - case iptables.ErrorTarget: + case stack.AcceptTarget: + return marshalStandardTarget(stack.RuleAccept) + case stack.DropTarget: + return marshalStandardTarget(stack.RuleDrop) + case stack.ErrorTarget: return marshalErrorTarget(errorTargetName) - case iptables.UserChainTarget: + case stack.UserChainTarget: return marshalErrorTarget(tg.Name) - case iptables.ReturnTarget: - return marshalStandardTarget(iptables.RuleReturn) + case stack.ReturnTarget: + return marshalStandardTarget(stack.RuleReturn) + case stack.RedirectTarget: + return marshalRedirectTarget(tg) + case JumpTarget: + return marshalJumpTarget(tg) default: panic(fmt.Errorf("unknown target of type %T", target)) } } -func marshalStandardTarget(verdict iptables.RuleVerdict) []byte { - nflog("convert to binary: marshalling standard target with size %d", linux.SizeOfXTStandardTarget) +func marshalStandardTarget(verdict stack.RuleVerdict) []byte { + nflog("convert to binary: marshalling standard target") // The target's name will be the empty string. target := linux.XTStandardTarget{ @@ -274,15 +309,55 @@ func marshalErrorTarget(errorName string) []byte { return binary.Marshal(ret, usermem.ByteOrder, target) } +func marshalRedirectTarget(rt stack.RedirectTarget) []byte { + // This is a redirect target named redirect + target := linux.XTRedirectTarget{ + Target: linux.XTEntryTarget{ + TargetSize: linux.SizeOfXTRedirectTarget, + }, + } + copy(target.Target.Name[:], redirectTargetName) + + ret := make([]byte, 0, linux.SizeOfXTRedirectTarget) + target.NfRange.RangeSize = 1 + if rt.RangeProtoSpecified { + target.NfRange.RangeIPV4.Flags |= linux.NF_NAT_RANGE_PROTO_SPECIFIED + } + // 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) +} + +func marshalJumpTarget(jt JumpTarget) []byte { + nflog("convert to binary: marshalling jump target") + + // The target's name will be the empty string. + target := linux.XTStandardTarget{ + Target: linux.XTEntryTarget{ + TargetSize: linux.SizeOfXTStandardTarget, + }, + // Verdict is overloaded by the ABI. When positive, it holds + // the jump offset from the start of the table. + Verdict: int32(jt.Offset), + } + + ret := make([]byte, 0, linux.SizeOfXTStandardTarget) + return binary.Marshal(ret, usermem.ByteOrder, target) +} + // translateFromStandardVerdict translates verdicts the same way as the iptables // tool. -func translateFromStandardVerdict(verdict iptables.RuleVerdict) int32 { +func translateFromStandardVerdict(verdict stack.RuleVerdict) int32 { switch verdict { - case iptables.RuleAccept: + case stack.RuleAccept: return -linux.NF_ACCEPT - 1 - case iptables.RuleDrop: + case stack.RuleDrop: return -linux.NF_DROP - 1 - case iptables.RuleReturn: + case stack.RuleReturn: return linux.NF_RETURN default: // TODO(gvisor.dev/issue/170): Support Jump. @@ -291,18 +366,18 @@ func translateFromStandardVerdict(verdict iptables.RuleVerdict) int32 { } // translateToStandardTarget translates from the value in a -// linux.XTStandardTarget to an iptables.Verdict. -func translateToStandardTarget(val int32) (iptables.Target, error) { +// linux.XTStandardTarget to an stack.Verdict. +func translateToStandardTarget(val int32) (stack.Target, error) { // TODO(gvisor.dev/issue/170): Support other verdicts. switch val { case -linux.NF_ACCEPT - 1: - return iptables.AcceptTarget{}, nil + return stack.AcceptTarget{}, nil case -linux.NF_DROP - 1: - return iptables.DropTarget{}, nil + return stack.DropTarget{}, nil case -linux.NF_QUEUE - 1: return nil, errors.New("unsupported iptables verdict QUEUE") case linux.NF_RETURN: - return iptables.ReturnTarget{}, nil + return stack.ReturnTarget{}, nil default: return nil, fmt.Errorf("unknown iptables verdict %d", val) } @@ -310,7 +385,7 @@ func translateToStandardTarget(val int32) (iptables.Target, error) { // SetEntries sets iptables rules for a single table. See // net/ipv4/netfilter/ip_tables.c:translate_table for reference. -func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { +func SetEntries(stk *stack.Stack, optVal []byte) *syserr.Error { // Get the basic rules data (struct ipt_replace). if len(optVal) < linux.SizeOfIPTReplace { nflog("optVal has insufficient size for replace %d", len(optVal)) @@ -322,10 +397,12 @@ func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { binary.Unmarshal(replaceBuf, usermem.ByteOrder, &replace) // TODO(gvisor.dev/issue/170): Support other tables. - var table iptables.Table + var table stack.Table switch replace.Name.String() { - case iptables.TablenameFilter: - table = iptables.EmptyFilterTable() + case stack.TablenameFilter: + table = stack.EmptyFilterTable() + case stack.TablenameNat: + table = stack.EmptyNatTable() default: nflog("we don't yet support writing to the %q table (gvisor.dev/issue/170)", replace.Name.String()) return syserr.ErrInvalidArgument @@ -335,7 +412,8 @@ func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { // Convert input into a list of rules and their offsets. var offset uint32 - var offsets []uint32 + // offsets maps rule byte offsets to their position in table.Rules. + offsets := map[uint32]int{} for entryIdx := uint32(0); entryIdx < replace.NumEntries; entryIdx++ { nflog("set entries: processing entry at offset %d", offset) @@ -384,23 +462,24 @@ func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { nflog("entry doesn't have enough room for its target (only %d bytes remain)", len(optVal)) return syserr.ErrInvalidArgument } - target, err := parseTarget(optVal[:targetSize]) + target, err := parseTarget(filter, optVal[:targetSize]) if err != nil { nflog("failed to parse target: %v", err) return syserr.ErrInvalidArgument } optVal = optVal[targetSize:] - table.Rules = append(table.Rules, iptables.Rule{ + table.Rules = append(table.Rules, stack.Rule{ Filter: filter, Target: target, Matchers: matchers, }) - offsets = append(offsets, offset) + offsets[offset] = int(entryIdx) offset += uint32(entry.NextOffset) if initialOptValLen-len(optVal) != int(entry.NextOffset) { nflog("entry NextOffset is %d, but entry took up %d bytes", entry.NextOffset, initialOptValLen-len(optVal)) + return syserr.ErrInvalidArgument } } @@ -409,23 +488,23 @@ func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { for hook, _ := range replace.HookEntry { if table.ValidHooks()&(1<<hook) != 0 { hk := hookFromLinux(hook) - for ruleIdx, offset := range offsets { + for offset, ruleIdx := range offsets { if offset == replace.HookEntry[hook] { table.BuiltinChains[hk] = ruleIdx } if offset == replace.Underflow[hook] { if !validUnderflow(table.Rules[ruleIdx]) { - nflog("underflow for hook %d isn't an unconditional ACCEPT or DROP.") + nflog("underflow for hook %d isn't an unconditional ACCEPT or DROP", ruleIdx) return syserr.ErrInvalidArgument } table.Underflows[hk] = ruleIdx } } - if ruleIdx := table.BuiltinChains[hk]; ruleIdx == iptables.HookUnset { + if ruleIdx := table.BuiltinChains[hk]; ruleIdx == stack.HookUnset { nflog("hook %v is unset.", hk) return syserr.ErrInvalidArgument } - if ruleIdx := table.Underflows[hk]; ruleIdx == iptables.HookUnset { + if ruleIdx := table.Underflows[hk]; ruleIdx == stack.HookUnset { nflog("underflow %v is unset.", hk) return syserr.ErrInvalidArgument } @@ -434,7 +513,7 @@ func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { // Add the user chains. for ruleIdx, rule := range table.Rules { - target, ok := rule.Target.(iptables.UserChainTarget) + target, ok := rule.Target.(stack.UserChainTarget) if !ok { continue } @@ -444,22 +523,41 @@ func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { // - There's some other rule after it. // - There are no matchers. if ruleIdx == len(table.Rules)-1 { - nflog("user chain must have a rule or default policy.") + nflog("user chain must have a rule or default policy") return syserr.ErrInvalidArgument } if len(table.Rules[ruleIdx].Matchers) != 0 { - nflog("user chain's first node must have no matcheres.") + nflog("user chain's first node must have no matchers") return syserr.ErrInvalidArgument } table.UserChains[target.Name] = ruleIdx + 1 } + // Set each jump to point to the appropriate rule. Right now they hold byte + // offsets. + for ruleIdx, rule := range table.Rules { + jump, ok := rule.Target.(JumpTarget) + if !ok { + continue + } + + // Find the rule corresponding to the jump rule offset. + jumpTo, ok := offsets[jump.Offset] + if !ok { + nflog("failed to find a rule to jump to") + return syserr.ErrInvalidArgument + } + jump.RuleNum = jumpTo + rule.Target = jump + table.Rules[ruleIdx] = rule + } + // TODO(gvisor.dev/issue/170): Support other chains. - // Since we only support modifying the INPUT chain right now, make sure - // all other chains point to ACCEPT rules. + // Since we only support modifying the INPUT, PREROUTING and OUTPUT chain right now, + // make sure all other chains point to ACCEPT rules. for hook, ruleIdx := range table.BuiltinChains { - if hook != iptables.Input { - if _, ok := table.Rules[ruleIdx].Target.(iptables.AcceptTarget); !ok { + if hook == stack.Forward || hook == stack.Postrouting { + if !isUnconditionalAccept(table.Rules[ruleIdx]) { nflog("hook %d is unsupported.", hook) return syserr.ErrInvalidArgument } @@ -471,24 +569,22 @@ func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { // - There are no chains without an unconditional final rule. // - There are no chains without an unconditional underflow rule. - ipt := stack.IPTables() table.SetMetadata(metadata{ HookEntry: replace.HookEntry, Underflow: replace.Underflow, NumEntries: replace.NumEntries, Size: replace.Size, }) - ipt.Tables[replace.Name.String()] = table - stack.SetIPTables(ipt) + stk.IPTables().ReplaceTable(replace.Name.String(), table) return nil } // parseMatchers parses 0 or more matchers from optVal. optVal should contain // only the matchers. -func parseMatchers(filter iptables.IPHeaderFilter, optVal []byte) ([]iptables.Matcher, error) { +func parseMatchers(filter stack.IPHeaderFilter, optVal []byte) ([]stack.Matcher, error) { nflog("set entries: parsing matchers of size %d", len(optVal)) - var matchers []iptables.Matcher + var matchers []stack.Matcher for len(optVal) > 0 { nflog("set entries: optVal has len %d", len(optVal)) @@ -530,7 +626,7 @@ func parseMatchers(filter iptables.IPHeaderFilter, optVal []byte) ([]iptables.Ma // parseTarget parses a target from optVal. optVal should contain only the // target. -func parseTarget(optVal []byte) (iptables.Target, error) { +func parseTarget(filter stack.IPHeaderFilter, optVal []byte) (stack.Target, error) { nflog("set entries: parsing target of size %d", len(optVal)) if len(optVal) < linux.SizeOfXTEntryTarget { return nil, fmt.Errorf("optVal has insufficient size for entry target %d", len(optVal)) @@ -548,7 +644,12 @@ func parseTarget(optVal []byte) (iptables.Target, error) { buf = optVal[:linux.SizeOfXTStandardTarget] binary.Unmarshal(buf, usermem.ByteOrder, &standardTarget) - return translateToStandardTarget(standardTarget.Verdict) + if standardTarget.Verdict < 0 { + // A Verdict < 0 indicates a non-jump verdict. + return translateToStandardTarget(standardTarget.Verdict) + } + // A verdict >= 0 indicates a jump. + return JumpTarget{Offset: uint32(standardTarget.Verdict)}, nil case errorTargetName: // Error target. @@ -569,67 +670,155 @@ func parseTarget(optVal []byte) (iptables.Target, error) { switch name := errorTarget.Name.String(); name { case errorTargetName: nflog("set entries: error target") - return iptables.ErrorTarget{}, nil + return stack.ErrorTarget{}, nil default: // User defined chain. nflog("set entries: user-defined target %q", name) - return iptables.UserChainTarget{Name: name}, nil + return stack.UserChainTarget{Name: name}, nil + } + + case redirectTargetName: + // Redirect target. + if len(optVal) < linux.SizeOfXTRedirectTarget { + return nil, fmt.Errorf("netfilter.SetEntries: optVal has insufficient size for redirect target %d", len(optVal)) + } + + if filter.Protocol != header.TCPProtocolNumber && filter.Protocol != header.UDPProtocolNumber { + return nil, fmt.Errorf("netfilter.SetEntries: invalid argument") + } + + var redirectTarget linux.XTRedirectTarget + buf = optVal[:linux.SizeOfXTRedirectTarget] + binary.Unmarshal(buf, usermem.ByteOrder, &redirectTarget) + + // Copy linux.XTRedirectTarget to stack.RedirectTarget. + var target stack.RedirectTarget + nfRange := redirectTarget.NfRange + + // RangeSize should be 1. + if nfRange.RangeSize != 1 { + return nil, fmt.Errorf("netfilter.SetEntries: invalid argument") + } + + // TODO(gvisor.dev/issue/170): Check if the flags are valid. + // Also check if we need to map ports or IP. + // For now, redirect target only supports destination port change. + // Port range and IP range are not supported yet. + if nfRange.RangeIPV4.Flags&linux.NF_NAT_RANGE_PROTO_SPECIFIED == 0 { + return nil, fmt.Errorf("netfilter.SetEntries: invalid argument") } + target.RangeProtoSpecified = true + + target.MinIP = tcpip.Address(nfRange.RangeIPV4.MinIP[:]) + target.MaxIP = tcpip.Address(nfRange.RangeIPV4.MaxIP[:]) + + // TODO(gvisor.dev/issue/170): Port range is not supported yet. + if nfRange.RangeIPV4.MinPort != nfRange.RangeIPV4.MaxPort { + return nil, fmt.Errorf("netfilter.SetEntries: invalid argument") + } + + // Convert port from big endian to little endian. + port := make([]byte, 2) + binary.BigEndian.PutUint16(port, nfRange.RangeIPV4.MinPort) + target.MinPort = binary.LittleEndian.Uint16(port) + + binary.BigEndian.PutUint16(port, nfRange.RangeIPV4.MaxPort) + target.MaxPort = binary.LittleEndian.Uint16(port) + return target, nil } // Unknown target. return nil, fmt.Errorf("unknown target %q doesn't exist or isn't supported yet.", target.Name.String()) } -func filterFromIPTIP(iptip linux.IPTIP) (iptables.IPHeaderFilter, error) { +func filterFromIPTIP(iptip linux.IPTIP) (stack.IPHeaderFilter, error) { if containsUnsupportedFields(iptip) { - return iptables.IPHeaderFilter{}, fmt.Errorf("unsupported fields in struct iptip: %+v", iptip) + return stack.IPHeaderFilter{}, fmt.Errorf("unsupported fields in struct iptip: %+v", iptip) + } + if len(iptip.Dst) != header.IPv4AddressSize || len(iptip.DstMask) != header.IPv4AddressSize { + return stack.IPHeaderFilter{}, fmt.Errorf("incorrect length of destination (%d) and/or destination mask (%d) fields", len(iptip.Dst), len(iptip.DstMask)) + } + if len(iptip.Src) != header.IPv4AddressSize || len(iptip.SrcMask) != header.IPv4AddressSize { + return stack.IPHeaderFilter{}, fmt.Errorf("incorrect length of source (%d) and/or source mask (%d) fields", len(iptip.Src), len(iptip.SrcMask)) + } + + n := bytes.IndexByte([]byte(iptip.OutputInterface[:]), 0) + if n == -1 { + n = len(iptip.OutputInterface) } - return iptables.IPHeaderFilter{ - Protocol: tcpip.TransportProtocolNumber(iptip.Protocol), + ifname := string(iptip.OutputInterface[:n]) + + n = bytes.IndexByte([]byte(iptip.OutputInterfaceMask[:]), 0) + if n == -1 { + n = len(iptip.OutputInterfaceMask) + } + ifnameMask := string(iptip.OutputInterfaceMask[:n]) + + return stack.IPHeaderFilter{ + Protocol: tcpip.TransportProtocolNumber(iptip.Protocol), + Dst: tcpip.Address(iptip.Dst[:]), + DstMask: tcpip.Address(iptip.DstMask[:]), + DstInvert: iptip.InverseFlags&linux.IPT_INV_DSTIP != 0, + Src: tcpip.Address(iptip.Src[:]), + SrcMask: tcpip.Address(iptip.SrcMask[:]), + SrcInvert: iptip.InverseFlags&linux.IPT_INV_SRCIP != 0, + OutputInterface: ifname, + OutputInterfaceMask: ifnameMask, + OutputInterfaceInvert: iptip.InverseFlags&linux.IPT_INV_VIA_OUT != 0, }, nil } func containsUnsupportedFields(iptip linux.IPTIP) bool { - // Currently we check that everything except protocol is zeroed. - var emptyInetAddr = linux.InetAddr{} + // The following features are supported: + // - Protocol + // - Dst and DstMask + // - Src and SrcMask + // - The inverse destination IP check flag + // - OutputInterface, OutputInterfaceMask and its inverse. var emptyInterface = [linux.IFNAMSIZ]byte{} - return iptip.Dst != emptyInetAddr || - iptip.Src != emptyInetAddr || - iptip.SrcMask != emptyInetAddr || - iptip.DstMask != emptyInetAddr || - iptip.InputInterface != emptyInterface || - iptip.OutputInterface != emptyInterface || + // Disable any supported inverse flags. + inverseMask := uint8(linux.IPT_INV_DSTIP) | uint8(linux.IPT_INV_SRCIP) | uint8(linux.IPT_INV_VIA_OUT) + return iptip.InputInterface != emptyInterface || iptip.InputInterfaceMask != emptyInterface || - iptip.OutputInterfaceMask != emptyInterface || iptip.Flags != 0 || - iptip.InverseFlags != 0 + iptip.InverseFlags&^inverseMask != 0 } -func validUnderflow(rule iptables.Rule) bool { +func validUnderflow(rule stack.Rule) bool { if len(rule.Matchers) != 0 { return false } + if rule.Filter != emptyFilter { + return false + } switch rule.Target.(type) { - case iptables.AcceptTarget, iptables.DropTarget: + case stack.AcceptTarget, stack.DropTarget: return true default: return false } } -func hookFromLinux(hook int) iptables.Hook { +func isUnconditionalAccept(rule stack.Rule) bool { + if !validUnderflow(rule) { + return false + } + _, ok := rule.Target.(stack.AcceptTarget) + return ok +} + +func hookFromLinux(hook int) stack.Hook { switch hook { case linux.NF_INET_PRE_ROUTING: - return iptables.Prerouting + return stack.Prerouting case linux.NF_INET_LOCAL_IN: - return iptables.Input + return stack.Input case linux.NF_INET_FORWARD: - return iptables.Forward + return stack.Forward case linux.NF_INET_LOCAL_OUT: - return iptables.Output + return stack.Output case linux.NF_INET_POST_ROUTING: - return iptables.Postrouting + return stack.Postrouting } panic(fmt.Sprintf("Unknown hook %d does not correspond to a builtin chain", hook)) } diff --git a/pkg/sentry/socket/netfilter/owner_matcher.go b/pkg/sentry/socket/netfilter/owner_matcher.go new file mode 100644 index 000000000..1b4e0ad79 --- /dev/null +++ b/pkg/sentry/socket/netfilter/owner_matcher.go @@ -0,0 +1,149 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package netfilter + +import ( + "fmt" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/binary" + "gvisor.dev/gvisor/pkg/tcpip/stack" + "gvisor.dev/gvisor/pkg/usermem" +) + +const matcherNameOwner = "owner" + +func init() { + registerMatchMaker(ownerMarshaler{}) +} + +// ownerMarshaler implements matchMaker for owner matching. +type ownerMarshaler struct{} + +// name implements matchMaker.name. +func (ownerMarshaler) name() string { + return matcherNameOwner +} + +// marshal implements matchMaker.marshal. +func (ownerMarshaler) marshal(mr stack.Matcher) []byte { + matcher := mr.(*OwnerMatcher) + iptOwnerInfo := linux.IPTOwnerInfo{ + UID: matcher.uid, + GID: matcher.gid, + } + + // Support for UID and GID match. + if matcher.matchUID { + iptOwnerInfo.Match = linux.XT_OWNER_UID + if matcher.invertUID { + iptOwnerInfo.Invert = linux.XT_OWNER_UID + } + } + if matcher.matchGID { + iptOwnerInfo.Match |= linux.XT_OWNER_GID + if matcher.invertGID { + iptOwnerInfo.Invert |= linux.XT_OWNER_GID + } + } + + buf := make([]byte, 0, linux.SizeOfIPTOwnerInfo) + return marshalEntryMatch(matcherNameOwner, binary.Marshal(buf, usermem.ByteOrder, iptOwnerInfo)) +} + +// unmarshal implements matchMaker.unmarshal. +func (ownerMarshaler) unmarshal(buf []byte, filter stack.IPHeaderFilter) (stack.Matcher, error) { + if len(buf) < linux.SizeOfIPTOwnerInfo { + return nil, fmt.Errorf("buf has insufficient size for owner match: %d", len(buf)) + } + + // For alignment reasons, the match's total size may + // exceed what's strictly necessary to hold matchData. + var matchData linux.IPTOwnerInfo + binary.Unmarshal(buf[:linux.SizeOfIPTOwnerInfo], usermem.ByteOrder, &matchData) + nflog("parseMatchers: parsed IPTOwnerInfo: %+v", matchData) + + var owner OwnerMatcher + owner.uid = matchData.UID + owner.gid = matchData.GID + + // Check flags. + if matchData.Match&linux.XT_OWNER_UID != 0 { + owner.matchUID = true + if matchData.Invert&linux.XT_OWNER_UID != 0 { + owner.invertUID = true + } + } + if matchData.Match&linux.XT_OWNER_GID != 0 { + owner.matchGID = true + if matchData.Invert&linux.XT_OWNER_GID != 0 { + owner.invertGID = true + } + } + + return &owner, nil +} + +type OwnerMatcher struct { + uid uint32 + gid uint32 + matchUID bool + matchGID bool + invertUID bool + invertGID bool +} + +// Name implements Matcher.Name. +func (*OwnerMatcher) Name() string { + return matcherNameOwner +} + +// Match implements Matcher.Match. +func (om *OwnerMatcher) Match(hook stack.Hook, pkt *stack.PacketBuffer, interfaceName string) (bool, bool) { + // Support only for OUTPUT chain. + // TODO(gvisor.dev/issue/170): Need to support for POSTROUTING chain also. + if hook != stack.Output { + return false, true + } + + // If the packet owner is not set, drop the packet. + if pkt.Owner == nil { + return false, true + } + + var matches bool + // Check for UID match. + if om.matchUID { + if pkt.Owner.UID() == om.uid { + matches = true + } + if matches == om.invertUID { + return false, false + } + } + + // Check for GID match. + if om.matchGID { + matches = false + if pkt.Owner.GID() == om.gid { + matches = true + } + if matches == om.invertGID { + return false, false + } + } + + return true, false +} diff --git a/pkg/sentry/socket/netfilter/targets.go b/pkg/sentry/socket/netfilter/targets.go new file mode 100644 index 000000000..84abe8d29 --- /dev/null +++ b/pkg/sentry/socket/netfilter/targets.go @@ -0,0 +1,35 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package netfilter + +import ( + "gvisor.dev/gvisor/pkg/tcpip" + "gvisor.dev/gvisor/pkg/tcpip/stack" +) + +// JumpTarget implements stack.Target. +type JumpTarget struct { + // Offset is the byte offset of the rule to jump to. It is used for + // marshaling and unmarshaling. + Offset uint32 + + // RuleNum is the rule to jump to. + RuleNum int +} + +// Action implements stack.Target.Action. +func (jt JumpTarget) Action(*stack.PacketBuffer, *stack.ConnTrackTable, stack.Hook, *stack.GSO, *stack.Route, tcpip.Address) (stack.RuleVerdict, int) { + return stack.RuleJump, jt.RuleNum +} diff --git a/pkg/sentry/socket/netfilter/tcp_matcher.go b/pkg/sentry/socket/netfilter/tcp_matcher.go index f9945e214..4f98ee2d5 100644 --- a/pkg/sentry/socket/netfilter/tcp_matcher.go +++ b/pkg/sentry/socket/netfilter/tcp_matcher.go @@ -19,9 +19,8 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/binary" - "gvisor.dev/gvisor/pkg/tcpip" "gvisor.dev/gvisor/pkg/tcpip/header" - "gvisor.dev/gvisor/pkg/tcpip/iptables" + "gvisor.dev/gvisor/pkg/tcpip/stack" "gvisor.dev/gvisor/pkg/usermem" ) @@ -40,7 +39,7 @@ func (tcpMarshaler) name() string { } // marshal implements matchMaker.marshal. -func (tcpMarshaler) marshal(mr iptables.Matcher) []byte { +func (tcpMarshaler) marshal(mr stack.Matcher) []byte { matcher := mr.(*TCPMatcher) xttcp := linux.XTTCP{ SourcePortStart: matcher.sourcePortStart, @@ -53,7 +52,7 @@ func (tcpMarshaler) marshal(mr iptables.Matcher) []byte { } // unmarshal implements matchMaker.unmarshal. -func (tcpMarshaler) unmarshal(buf []byte, filter iptables.IPHeaderFilter) (iptables.Matcher, error) { +func (tcpMarshaler) unmarshal(buf []byte, filter stack.IPHeaderFilter) (stack.Matcher, error) { if len(buf) < linux.SizeOfXTTCP { return nil, fmt.Errorf("buf has insufficient size for TCP match: %d", len(buf)) } @@ -97,7 +96,7 @@ func (*TCPMatcher) Name() string { } // Match implements Matcher.Match. -func (tm *TCPMatcher) Match(hook iptables.Hook, pkt tcpip.PacketBuffer, interfaceName string) (bool, bool) { +func (tm *TCPMatcher) Match(hook stack.Hook, pkt *stack.PacketBuffer, interfaceName string) (bool, bool) { netHeader := header.IPv4(pkt.NetworkHeader) if netHeader.TransportProtocol() != header.TCPProtocolNumber { @@ -112,22 +111,10 @@ func (tm *TCPMatcher) Match(hook iptables.Hook, pkt tcpip.PacketBuffer, interfac return false, false } - // Now we need the transport header. However, this may not have been set - // yet. - // TODO(gvisor.dev/issue/170): Parsing the transport header should - // ultimately be moved into the iptables.Check codepath as matchers are - // added. - var tcpHeader header.TCP - if pkt.TransportHeader != nil { - tcpHeader = header.TCP(pkt.TransportHeader) - } else { - // The TCP header hasn't been parsed yet. We have to do it here. - if len(pkt.Data.First()) < header.TCPMinimumSize { - // There's no valid TCP header here, so we hotdrop the - // packet. - return false, true - } - tcpHeader = header.TCP(pkt.Data.First()) + tcpHeader := header.TCP(pkt.TransportHeader) + if len(tcpHeader) < header.TCPMinimumSize { + // There's no valid TCP header here, so we drop the packet immediately. + return false, true } // Check whether the source and destination ports are within the diff --git a/pkg/sentry/socket/netfilter/udp_matcher.go b/pkg/sentry/socket/netfilter/udp_matcher.go index 86aa11696..3f20fc891 100644 --- a/pkg/sentry/socket/netfilter/udp_matcher.go +++ b/pkg/sentry/socket/netfilter/udp_matcher.go @@ -19,9 +19,8 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/binary" - "gvisor.dev/gvisor/pkg/tcpip" "gvisor.dev/gvisor/pkg/tcpip/header" - "gvisor.dev/gvisor/pkg/tcpip/iptables" + "gvisor.dev/gvisor/pkg/tcpip/stack" "gvisor.dev/gvisor/pkg/usermem" ) @@ -40,7 +39,7 @@ func (udpMarshaler) name() string { } // marshal implements matchMaker.marshal. -func (udpMarshaler) marshal(mr iptables.Matcher) []byte { +func (udpMarshaler) marshal(mr stack.Matcher) []byte { matcher := mr.(*UDPMatcher) xtudp := linux.XTUDP{ SourcePortStart: matcher.sourcePortStart, @@ -53,7 +52,7 @@ func (udpMarshaler) marshal(mr iptables.Matcher) []byte { } // unmarshal implements matchMaker.unmarshal. -func (udpMarshaler) unmarshal(buf []byte, filter iptables.IPHeaderFilter) (iptables.Matcher, error) { +func (udpMarshaler) unmarshal(buf []byte, filter stack.IPHeaderFilter) (stack.Matcher, error) { if len(buf) < linux.SizeOfXTUDP { return nil, fmt.Errorf("buf has insufficient size for UDP match: %d", len(buf)) } @@ -94,11 +93,11 @@ func (*UDPMatcher) Name() string { } // Match implements Matcher.Match. -func (um *UDPMatcher) Match(hook iptables.Hook, pkt tcpip.PacketBuffer, interfaceName string) (bool, bool) { +func (um *UDPMatcher) Match(hook stack.Hook, pkt *stack.PacketBuffer, interfaceName string) (bool, bool) { netHeader := header.IPv4(pkt.NetworkHeader) // TODO(gvisor.dev/issue/170): Proto checks should ultimately be moved - // into the iptables.Check codepath as matchers are added. + // into the stack.Check codepath as matchers are added. if netHeader.TransportProtocol() != header.UDPProtocolNumber { return false, false } @@ -111,22 +110,10 @@ func (um *UDPMatcher) Match(hook iptables.Hook, pkt tcpip.PacketBuffer, interfac return false, false } - // Now we need the transport header. However, this may not have been set - // yet. - // TODO(gvisor.dev/issue/170): Parsing the transport header should - // ultimately be moved into the iptables.Check codepath as matchers are - // added. - var udpHeader header.UDP - if pkt.TransportHeader != nil { - udpHeader = header.UDP(pkt.TransportHeader) - } else { - // The UDP header hasn't been parsed yet. We have to do it here. - if len(pkt.Data.First()) < header.UDPMinimumSize { - // There's no valid UDP header here, so we hotdrop the - // packet. - return false, true - } - udpHeader = header.UDP(pkt.Data.First()) + udpHeader := header.UDP(pkt.TransportHeader) + if len(udpHeader) < header.UDPMinimumSize { + // There's no valid UDP header here, so we drop the packet immediately. + return false, true } // Check whether the source and destination ports are within the diff --git a/pkg/sentry/socket/netlink/BUILD b/pkg/sentry/socket/netlink/BUILD index 1911cd9b8..7212d8644 100644 --- a/pkg/sentry/socket/netlink/BUILD +++ b/pkg/sentry/socket/netlink/BUILD @@ -7,7 +7,9 @@ go_library( srcs = [ "message.go", "provider.go", + "provider_vfs2.go", "socket.go", + "socket_vfs2.go", ], visibility = ["//pkg/sentry:internal"], deps = [ @@ -18,6 +20,7 @@ go_library( "//pkg/sentry/device", "//pkg/sentry/fs", "//pkg/sentry/fs/fsutil", + "//pkg/sentry/fsimpl/sockfs", "//pkg/sentry/kernel", "//pkg/sentry/kernel/auth", "//pkg/sentry/kernel/time", @@ -25,6 +28,7 @@ go_library( "//pkg/sentry/socket/netlink/port", "//pkg/sentry/socket/unix", "//pkg/sentry/socket/unix/transport", + "//pkg/sentry/vfs", "//pkg/sync", "//pkg/syserr", "//pkg/syserror", diff --git a/pkg/sentry/socket/netlink/provider.go b/pkg/sentry/socket/netlink/provider.go index b0dc70e5c..0d45e5053 100644 --- a/pkg/sentry/socket/netlink/provider.go +++ b/pkg/sentry/socket/netlink/provider.go @@ -67,6 +67,8 @@ func RegisterProvider(protocol int, provider Provider) { protocols[protocol] = provider } +// LINT.IfChange + // socketProvider implements socket.Provider. type socketProvider struct { } @@ -105,7 +107,10 @@ func (*socketProvider) Pair(*kernel.Task, linux.SockType, int) (*fs.File, *fs.Fi return nil, nil, syserr.ErrNotSupported } +// LINT.ThenChange(./provider_vfs2.go) + // init registers the socket provider. func init() { socket.RegisterProvider(linux.AF_NETLINK, &socketProvider{}) + socket.RegisterProviderVFS2(linux.AF_NETLINK, &socketProviderVFS2{}) } diff --git a/pkg/sentry/socket/netlink/provider_vfs2.go b/pkg/sentry/socket/netlink/provider_vfs2.go new file mode 100644 index 000000000..bb205be0d --- /dev/null +++ b/pkg/sentry/socket/netlink/provider_vfs2.go @@ -0,0 +1,69 @@ +// Copyright 2018 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package netlink + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/sockfs" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserr" +) + +// socketProviderVFS2 implements socket.Provider. +type socketProviderVFS2 struct { +} + +// Socket implements socket.Provider.Socket. +func (*socketProviderVFS2) Socket(t *kernel.Task, stype linux.SockType, protocol int) (*vfs.FileDescription, *syserr.Error) { + // Netlink sockets must be specified as datagram or raw, but they + // behave the same regardless of type. + if stype != linux.SOCK_DGRAM && stype != linux.SOCK_RAW { + return nil, syserr.ErrSocketNotSupported + } + + provider, ok := protocols[protocol] + if !ok { + return nil, syserr.ErrProtocolNotSupported + } + + p, err := provider(t) + if err != nil { + return nil, err + } + + s, err := NewVFS2(t, stype, p) + if err != nil { + return nil, err + } + + vfsfd := &s.vfsfd + mnt := t.Kernel().SocketMount() + d := sockfs.NewDentry(t.Credentials(), mnt) + if err := vfsfd.Init(s, linux.O_RDWR, mnt, d, &vfs.FileDescriptionOptions{ + DenyPRead: true, + DenyPWrite: true, + UseDentryMetadata: true, + }); err != nil { + return nil, syserr.FromError(err) + } + return vfsfd, nil +} + +// Pair implements socket.Provider.Pair by returning an error. +func (*socketProviderVFS2) Pair(*kernel.Task, linux.SockType, int) (*vfs.FileDescription, *vfs.FileDescription, *syserr.Error) { + // Netlink sockets never supports creating socket pairs. + return nil, nil, syserr.ErrNotSupported +} diff --git a/pkg/sentry/socket/netlink/socket.go b/pkg/sentry/socket/netlink/socket.go index 2ca02567d..81f34c5a2 100644 --- a/pkg/sentry/socket/netlink/socket.go +++ b/pkg/sentry/socket/netlink/socket.go @@ -58,6 +58,8 @@ var errNoFilter = syserr.New("no filter attached", linux.ENOENT) // netlinkSocketDevice is the netlink socket virtual device. var netlinkSocketDevice = device.NewAnonDevice() +// LINT.IfChange + // Socket is the base socket type for netlink sockets. // // This implementation only supports userspace sending and receiving messages @@ -74,6 +76,14 @@ type Socket struct { fsutil.FileNoSplice `state:"nosave"` fsutil.FileNoopFlush `state:"nosave"` fsutil.FileUseInodeUnstableAttr `state:"nosave"` + + socketOpsCommon +} + +// socketOpsCommon contains the socket operations common to VFS1 and VFS2. +// +// +stateify savable +type socketOpsCommon struct { socket.SendReceiveTimeout // ports provides netlink port allocation. @@ -140,17 +150,19 @@ func NewSocket(t *kernel.Task, skType linux.SockType, protocol Protocol) (*Socke } return &Socket{ - ports: t.Kernel().NetlinkPorts(), - protocol: protocol, - skType: skType, - ep: ep, - connection: connection, - sendBufferSize: defaultSendBufferSize, + socketOpsCommon: socketOpsCommon{ + ports: t.Kernel().NetlinkPorts(), + protocol: protocol, + skType: skType, + ep: ep, + connection: connection, + sendBufferSize: defaultSendBufferSize, + }, }, nil } // Release implements fs.FileOperations.Release. -func (s *Socket) Release() { +func (s *socketOpsCommon) Release() { s.connection.Release() s.ep.Close() @@ -160,7 +172,7 @@ func (s *Socket) Release() { } // Readiness implements waiter.Waitable.Readiness. -func (s *Socket) Readiness(mask waiter.EventMask) waiter.EventMask { +func (s *socketOpsCommon) Readiness(mask waiter.EventMask) waiter.EventMask { // ep holds messages to be read and thus handles EventIn readiness. ready := s.ep.Readiness(mask) @@ -174,18 +186,18 @@ func (s *Socket) Readiness(mask waiter.EventMask) waiter.EventMask { } // EventRegister implements waiter.Waitable.EventRegister. -func (s *Socket) EventRegister(e *waiter.Entry, mask waiter.EventMask) { +func (s *socketOpsCommon) EventRegister(e *waiter.Entry, mask waiter.EventMask) { s.ep.EventRegister(e, mask) // Writable readiness never changes, so no registration is needed. } // EventUnregister implements waiter.Waitable.EventUnregister. -func (s *Socket) EventUnregister(e *waiter.Entry) { +func (s *socketOpsCommon) EventUnregister(e *waiter.Entry) { s.ep.EventUnregister(e) } // Passcred implements transport.Credentialer.Passcred. -func (s *Socket) Passcred() bool { +func (s *socketOpsCommon) Passcred() bool { s.mu.Lock() passcred := s.passcred s.mu.Unlock() @@ -193,7 +205,7 @@ func (s *Socket) Passcred() bool { } // ConnectedPasscred implements transport.Credentialer.ConnectedPasscred. -func (s *Socket) ConnectedPasscred() bool { +func (s *socketOpsCommon) ConnectedPasscred() bool { // This socket is connected to the kernel, which doesn't need creds. // // This is arbitrary, as ConnectedPasscred on this type has no callers. @@ -227,7 +239,7 @@ func ExtractSockAddr(b []byte) (*linux.SockAddrNetlink, *syserr.Error) { // port of 0 defaults to the ThreadGroup ID. // // Preconditions: mu is held. -func (s *Socket) bindPort(t *kernel.Task, port int32) *syserr.Error { +func (s *socketOpsCommon) bindPort(t *kernel.Task, port int32) *syserr.Error { if s.bound { // Re-binding is only allowed if the port doesn't change. if port != s.portID { @@ -251,7 +263,7 @@ func (s *Socket) bindPort(t *kernel.Task, port int32) *syserr.Error { } // Bind implements socket.Socket.Bind. -func (s *Socket) Bind(t *kernel.Task, sockaddr []byte) *syserr.Error { +func (s *socketOpsCommon) Bind(t *kernel.Task, sockaddr []byte) *syserr.Error { a, err := ExtractSockAddr(sockaddr) if err != nil { return err @@ -269,7 +281,7 @@ func (s *Socket) Bind(t *kernel.Task, sockaddr []byte) *syserr.Error { } // Connect implements socket.Socket.Connect. -func (s *Socket) Connect(t *kernel.Task, sockaddr []byte, blocking bool) *syserr.Error { +func (s *socketOpsCommon) Connect(t *kernel.Task, sockaddr []byte, blocking bool) *syserr.Error { a, err := ExtractSockAddr(sockaddr) if err != nil { return err @@ -300,25 +312,25 @@ func (s *Socket) Connect(t *kernel.Task, sockaddr []byte, blocking bool) *syserr } // Accept implements socket.Socket.Accept. -func (s *Socket) Accept(t *kernel.Task, peerRequested bool, flags int, blocking bool) (int32, linux.SockAddr, uint32, *syserr.Error) { +func (s *socketOpsCommon) Accept(t *kernel.Task, peerRequested bool, flags int, blocking bool) (int32, linux.SockAddr, uint32, *syserr.Error) { // Netlink sockets never support accept. return 0, nil, 0, syserr.ErrNotSupported } // Listen implements socket.Socket.Listen. -func (s *Socket) Listen(t *kernel.Task, backlog int) *syserr.Error { +func (s *socketOpsCommon) Listen(t *kernel.Task, backlog int) *syserr.Error { // Netlink sockets never support listen. return syserr.ErrNotSupported } // Shutdown implements socket.Socket.Shutdown. -func (s *Socket) Shutdown(t *kernel.Task, how int) *syserr.Error { +func (s *socketOpsCommon) Shutdown(t *kernel.Task, how int) *syserr.Error { // Netlink sockets never support shutdown. return syserr.ErrNotSupported } // GetSockOpt implements socket.Socket.GetSockOpt. -func (s *Socket) GetSockOpt(t *kernel.Task, level int, name int, outPtr usermem.Addr, outLen int) (interface{}, *syserr.Error) { +func (s *socketOpsCommon) GetSockOpt(t *kernel.Task, level int, name int, outPtr usermem.Addr, outLen int) (interface{}, *syserr.Error) { switch level { case linux.SOL_SOCKET: switch name { @@ -369,7 +381,7 @@ func (s *Socket) GetSockOpt(t *kernel.Task, level int, name int, outPtr usermem. } // SetSockOpt implements socket.Socket.SetSockOpt. -func (s *Socket) SetSockOpt(t *kernel.Task, level int, name int, opt []byte) *syserr.Error { +func (s *socketOpsCommon) SetSockOpt(t *kernel.Task, level int, name int, opt []byte) *syserr.Error { switch level { case linux.SOL_SOCKET: switch name { @@ -466,7 +478,7 @@ func (s *Socket) SetSockOpt(t *kernel.Task, level int, name int, opt []byte) *sy } // GetSockName implements socket.Socket.GetSockName. -func (s *Socket) GetSockName(t *kernel.Task) (linux.SockAddr, uint32, *syserr.Error) { +func (s *socketOpsCommon) GetSockName(t *kernel.Task) (linux.SockAddr, uint32, *syserr.Error) { s.mu.Lock() defer s.mu.Unlock() @@ -478,7 +490,7 @@ func (s *Socket) GetSockName(t *kernel.Task) (linux.SockAddr, uint32, *syserr.Er } // GetPeerName implements socket.Socket.GetPeerName. -func (s *Socket) GetPeerName(t *kernel.Task) (linux.SockAddr, uint32, *syserr.Error) { +func (s *socketOpsCommon) GetPeerName(t *kernel.Task) (linux.SockAddr, uint32, *syserr.Error) { sa := &linux.SockAddrNetlink{ Family: linux.AF_NETLINK, // TODO(b/68878065): Support non-kernel peers. For now the peer @@ -489,7 +501,7 @@ func (s *Socket) GetPeerName(t *kernel.Task) (linux.SockAddr, uint32, *syserr.Er } // RecvMsg implements socket.Socket.RecvMsg. -func (s *Socket) RecvMsg(t *kernel.Task, dst usermem.IOSequence, flags int, haveDeadline bool, deadline ktime.Time, senderRequested bool, controlDataLen uint64) (int, int, linux.SockAddr, uint32, socket.ControlMessages, *syserr.Error) { +func (s *socketOpsCommon) RecvMsg(t *kernel.Task, dst usermem.IOSequence, flags int, haveDeadline bool, deadline ktime.Time, senderRequested bool, controlDataLen uint64) (int, int, linux.SockAddr, uint32, socket.ControlMessages, *syserr.Error) { from := &linux.SockAddrNetlink{ Family: linux.AF_NETLINK, PortID: 0, @@ -590,7 +602,7 @@ func (kernelSCM) Credentials(*kernel.Task) (kernel.ThreadID, auth.UID, auth.GID) var kernelCreds = &kernelSCM{} // sendResponse sends the response messages in ms back to userspace. -func (s *Socket) sendResponse(ctx context.Context, ms *MessageSet) *syserr.Error { +func (s *socketOpsCommon) sendResponse(ctx context.Context, ms *MessageSet) *syserr.Error { // Linux combines multiple netlink messages into a single datagram. bufs := make([][]byte, 0, len(ms.Messages)) for _, m := range ms.Messages { @@ -666,7 +678,7 @@ func dumpAckMesage(hdr linux.NetlinkMessageHeader, ms *MessageSet) { // processMessages handles each message in buf, passing it to the protocol // handler for final handling. -func (s *Socket) processMessages(ctx context.Context, buf []byte) *syserr.Error { +func (s *socketOpsCommon) processMessages(ctx context.Context, buf []byte) *syserr.Error { for len(buf) > 0 { msg, rest, ok := ParseMessage(buf) if !ok { @@ -698,7 +710,7 @@ func (s *Socket) processMessages(ctx context.Context, buf []byte) *syserr.Error } // sendMsg is the core of message send, used for SendMsg and Write. -func (s *Socket) sendMsg(ctx context.Context, src usermem.IOSequence, to []byte, flags int, controlMessages socket.ControlMessages) (int, *syserr.Error) { +func (s *socketOpsCommon) sendMsg(ctx context.Context, src usermem.IOSequence, to []byte, flags int, controlMessages socket.ControlMessages) (int, *syserr.Error) { dstPort := int32(0) if len(to) != 0 { @@ -745,7 +757,7 @@ func (s *Socket) sendMsg(ctx context.Context, src usermem.IOSequence, to []byte, } // SendMsg implements socket.Socket.SendMsg. -func (s *Socket) SendMsg(t *kernel.Task, src usermem.IOSequence, to []byte, flags int, haveDeadline bool, deadline ktime.Time, controlMessages socket.ControlMessages) (int, *syserr.Error) { +func (s *socketOpsCommon) SendMsg(t *kernel.Task, src usermem.IOSequence, to []byte, flags int, haveDeadline bool, deadline ktime.Time, controlMessages socket.ControlMessages) (int, *syserr.Error) { return s.sendMsg(t, src, to, flags, controlMessages) } @@ -756,11 +768,13 @@ func (s *Socket) Write(ctx context.Context, _ *fs.File, src usermem.IOSequence, } // State implements socket.Socket.State. -func (s *Socket) State() uint32 { +func (s *socketOpsCommon) State() uint32 { return s.ep.State() } // Type implements socket.Socket.Type. -func (s *Socket) Type() (family int, skType linux.SockType, protocol int) { +func (s *socketOpsCommon) Type() (family int, skType linux.SockType, protocol int) { return linux.AF_NETLINK, s.skType, s.protocol.Protocol() } + +// LINT.ThenChange(./socket_vfs2.go) diff --git a/pkg/sentry/socket/netlink/socket_vfs2.go b/pkg/sentry/socket/netlink/socket_vfs2.go new file mode 100644 index 000000000..b854bf990 --- /dev/null +++ b/pkg/sentry/socket/netlink/socket_vfs2.go @@ -0,0 +1,138 @@ +// Copyright 2018 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package netlink + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/socket" + "gvisor.dev/gvisor/pkg/sentry/socket/unix" + "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserr" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/tcpip" + "gvisor.dev/gvisor/pkg/usermem" + "gvisor.dev/gvisor/pkg/waiter" +) + +// SocketVFS2 is the base VFS2 socket type for netlink sockets. +// +// This implementation only supports userspace sending and receiving messages +// to/from the kernel. +// +// SocketVFS2 implements socket.SocketVFS2 and transport.Credentialer. +type SocketVFS2 struct { + vfsfd vfs.FileDescription + vfs.FileDescriptionDefaultImpl + vfs.DentryMetadataFileDescriptionImpl + + socketOpsCommon +} + +var _ socket.SocketVFS2 = (*SocketVFS2)(nil) +var _ transport.Credentialer = (*SocketVFS2)(nil) + +// NewVFS2 creates a new SocketVFS2. +func NewVFS2(t *kernel.Task, skType linux.SockType, protocol Protocol) (*SocketVFS2, *syserr.Error) { + // Datagram endpoint used to buffer kernel -> user messages. + ep := transport.NewConnectionless(t) + + // Bind the endpoint for good measure so we can connect to it. The + // bound address will never be exposed. + if err := ep.Bind(tcpip.FullAddress{Addr: "dummy"}, nil); err != nil { + ep.Close() + return nil, err + } + + // Create a connection from which the kernel can write messages. + connection, err := ep.(transport.BoundEndpoint).UnidirectionalConnect(t) + if err != nil { + ep.Close() + return nil, err + } + + return &SocketVFS2{ + socketOpsCommon: socketOpsCommon{ + ports: t.Kernel().NetlinkPorts(), + protocol: protocol, + skType: skType, + ep: ep, + connection: connection, + sendBufferSize: defaultSendBufferSize, + }, + }, nil +} + +// Readiness implements waiter.Waitable.Readiness. +func (s *SocketVFS2) Readiness(mask waiter.EventMask) waiter.EventMask { + return s.socketOpsCommon.Readiness(mask) +} + +// EventRegister implements waiter.Waitable.EventRegister. +func (s *SocketVFS2) EventRegister(e *waiter.Entry, mask waiter.EventMask) { + s.socketOpsCommon.EventRegister(e, mask) +} + +// EventUnregister implements waiter.Waitable.EventUnregister. +func (s *SocketVFS2) EventUnregister(e *waiter.Entry) { + s.socketOpsCommon.EventUnregister(e) +} + +// Ioctl implements vfs.FileDescriptionImpl. +func (*SocketVFS2) Ioctl(context.Context, usermem.IO, arch.SyscallArguments) (uintptr, error) { + // TODO(b/68878065): no ioctls supported. + return 0, syserror.ENOTTY +} + +// PRead implements vfs.FileDescriptionImpl. +func (s *SocketVFS2) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { + return 0, syserror.ESPIPE +} + +// Read implements vfs.FileDescriptionImpl. +func (s *SocketVFS2) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { + // All flags other than RWF_NOWAIT should be ignored. + // TODO(gvisor.dev/issue/2601): Support RWF_NOWAIT. + if opts.Flags != 0 { + return 0, syserror.EOPNOTSUPP + } + + if dst.NumBytes() == 0 { + return 0, nil + } + return dst.CopyOutFrom(ctx, &unix.EndpointReader{ + Endpoint: s.ep, + }) +} + +// PWrite implements vfs.FileDescriptionImpl. +func (s *SocketVFS2) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) { + return 0, syserror.ESPIPE +} + +// Write implements vfs.FileDescriptionImpl. +func (s *SocketVFS2) Write(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) { + // All flags other than RWF_NOWAIT should be ignored. + // TODO(gvisor.dev/issue/2601): Support RWF_NOWAIT. + if opts.Flags != 0 { + return 0, syserror.EOPNOTSUPP + } + + n, err := s.sendMsg(ctx, src, nil, 0, socket.ControlMessages{}) + return int64(n), err.ToError() +} diff --git a/pkg/sentry/socket/netstack/BUILD b/pkg/sentry/socket/netstack/BUILD index ab01cb4fa..8f0f5466e 100644 --- a/pkg/sentry/socket/netstack/BUILD +++ b/pkg/sentry/socket/netstack/BUILD @@ -7,7 +7,9 @@ go_library( srcs = [ "device.go", "netstack.go", + "netstack_vfs2.go", "provider.go", + "provider_vfs2.go", "save_restore.go", "stack.go", ], @@ -16,6 +18,7 @@ go_library( ], deps = [ "//pkg/abi/linux", + "//pkg/amutex", "//pkg/binary", "//pkg/context", "//pkg/log", @@ -25,6 +28,7 @@ go_library( "//pkg/sentry/device", "//pkg/sentry/fs", "//pkg/sentry/fs/fsutil", + "//pkg/sentry/fsimpl/sockfs", "//pkg/sentry/inet", "//pkg/sentry/kernel", "//pkg/sentry/kernel/auth", @@ -32,13 +36,13 @@ go_library( "//pkg/sentry/socket", "//pkg/sentry/socket/netfilter", "//pkg/sentry/unimpl", + "//pkg/sentry/vfs", "//pkg/sync", "//pkg/syserr", "//pkg/syserror", "//pkg/tcpip", "//pkg/tcpip/buffer", "//pkg/tcpip/header", - "//pkg/tcpip/iptables", "//pkg/tcpip/network/ipv4", "//pkg/tcpip/network/ipv6", "//pkg/tcpip/stack", @@ -46,5 +50,6 @@ go_library( "//pkg/tcpip/transport/udp", "//pkg/usermem", "//pkg/waiter", + "@org_golang_x_sys//unix:go_default_library", ], ) diff --git a/pkg/sentry/socket/netstack/netstack.go b/pkg/sentry/socket/netstack/netstack.go index 9757fbfba..e1e0c5931 100644 --- a/pkg/sentry/socket/netstack/netstack.go +++ b/pkg/sentry/socket/netstack/netstack.go @@ -29,10 +29,13 @@ import ( "io" "math" "reflect" + "sync/atomic" "syscall" "time" + "golang.org/x/sys/unix" "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/amutex" "gvisor.dev/gvisor/pkg/binary" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/log" @@ -62,7 +65,13 @@ import ( func mustCreateMetric(name, description string) *tcpip.StatCounter { var cm tcpip.StatCounter - metric.MustRegisterCustomUint64Metric(name, false /* sync */, description, cm.Value) + metric.MustRegisterCustomUint64Metric(name, true /* cumulative */, false /* sync */, description, cm.Value) + return &cm +} + +func mustCreateGauge(name, description string) *tcpip.StatCounter { + var cm tcpip.StatCounter + metric.MustRegisterCustomUint64Metric(name, false /* cumulative */, false /* sync */, description, cm.Value) return &cm } @@ -150,10 +159,10 @@ var Metrics = tcpip.Stats{ TCP: tcpip.TCPStats{ ActiveConnectionOpenings: mustCreateMetric("/netstack/tcp/active_connection_openings", "Number of connections opened successfully via Connect."), PassiveConnectionOpenings: mustCreateMetric("/netstack/tcp/passive_connection_openings", "Number of connections opened successfully via Listen."), - CurrentEstablished: mustCreateMetric("/netstack/tcp/current_established", "Number of connections in ESTABLISHED state now."), - CurrentConnected: mustCreateMetric("/netstack/tcp/current_open", "Number of connections that are in connected state."), + CurrentEstablished: mustCreateGauge("/netstack/tcp/current_established", "Number of connections in ESTABLISHED state now."), + CurrentConnected: mustCreateGauge("/netstack/tcp/current_open", "Number of connections that are in connected state."), EstablishedResets: mustCreateMetric("/netstack/tcp/established_resets", "Number of times TCP connections have made a direct transition to the CLOSED state from either the ESTABLISHED state or the CLOSE-WAIT state"), - EstablishedClosed: mustCreateMetric("/netstack/tcp/established_closed", "number of times established TCP connections made a transition to CLOSED state."), + EstablishedClosed: mustCreateMetric("/netstack/tcp/established_closed", "Number of times established TCP connections made a transition to CLOSED state."), EstablishedTimedout: mustCreateMetric("/netstack/tcp/established_timedout", "Number of times an established connection was reset because of keep-alive time out."), ListenOverflowSynDrop: mustCreateMetric("/netstack/tcp/listen_overflow_syn_drop", "Number of times the listen queue overflowed and a SYN was dropped."), ListenOverflowAckDrop: mustCreateMetric("/netstack/tcp/listen_overflow_ack_drop", "Number of times the listen queue overflowed and the final ACK in the handshake was dropped."), @@ -245,6 +254,8 @@ type commonEndpoint interface { GetSockOptInt(opt tcpip.SockOptInt) (int, *tcpip.Error) } +// LINT.IfChange + // SocketOperations encapsulates all the state needed to represent a network stack // endpoint in the kernel context. // @@ -256,6 +267,14 @@ type SocketOperations struct { fsutil.FileNoFsync `state:"nosave"` fsutil.FileNoMMap `state:"nosave"` fsutil.FileUseInodeUnstableAttr `state:"nosave"` + + socketOpsCommon +} + +// socketOpsCommon contains the socket operations common to VFS1 and VFS2. +// +// +stateify savable +type socketOpsCommon struct { socket.SendReceiveTimeout *waiter.Queue @@ -264,6 +283,12 @@ type SocketOperations struct { skType linux.SockType protocol int + // readViewHasData is 1 iff readView has data to be read, 0 otherwise. + // Must be accessed using atomic operations. It must only be written + // with readMu held but can be read without holding readMu. The latter + // is required to avoid deadlocks in epoll Readiness checks. + readViewHasData uint32 + // readMu protects access to the below fields. readMu sync.Mutex `state:"nosave"` // readView contains the remaining payload from the last packet. @@ -293,7 +318,7 @@ type SocketOperations struct { // New creates a new endpoint socket. func New(t *kernel.Task, family int, skType linux.SockType, protocol int, queue *waiter.Queue, endpoint tcpip.Endpoint) (*fs.File, *syserr.Error) { if skType == linux.SOCK_STREAM { - if err := endpoint.SetSockOptInt(tcpip.DelayOption, 1); err != nil { + if err := endpoint.SetSockOptBool(tcpip.DelayOption, true); err != nil { return nil, syserr.TranslateNetstackError(err) } } @@ -301,11 +326,13 @@ func New(t *kernel.Task, family int, skType linux.SockType, protocol int, queue dirent := socket.NewDirent(t, netstackDevice) defer dirent.DecRef() return fs.NewFile(t, dirent, fs.FileFlags{Read: true, Write: true, NonSeekable: true}, &SocketOperations{ - Queue: queue, - family: family, - Endpoint: endpoint, - skType: skType, - protocol: protocol, + socketOpsCommon: socketOpsCommon{ + Queue: queue, + family: family, + Endpoint: endpoint, + skType: skType, + protocol: protocol, + }, }), nil } @@ -404,33 +431,36 @@ func AddressAndFamily(addr []byte) (tcpip.FullAddress, uint16, *syserr.Error) { } } -func (s *SocketOperations) isPacketBased() bool { +func (s *socketOpsCommon) isPacketBased() bool { return s.skType == linux.SOCK_DGRAM || s.skType == linux.SOCK_SEQPACKET || s.skType == linux.SOCK_RDM || s.skType == linux.SOCK_RAW } // fetchReadView updates the readView field of the socket if it's currently // empty. It assumes that the socket is locked. -func (s *SocketOperations) fetchReadView() *syserr.Error { +// +// Precondition: s.readMu must be held. +func (s *socketOpsCommon) fetchReadView() *syserr.Error { if len(s.readView) > 0 { return nil } - s.readView = nil s.sender = tcpip.FullAddress{} v, cms, err := s.Endpoint.Read(&s.sender) if err != nil { + atomic.StoreUint32(&s.readViewHasData, 0) return syserr.TranslateNetstackError(err) } s.readView = v s.readCM = cms + atomic.StoreUint32(&s.readViewHasData, 1) return nil } // Release implements fs.FileOperations.Release. -func (s *SocketOperations) Release() { +func (s *socketOpsCommon) Release() { s.Endpoint.Close() } @@ -525,11 +555,9 @@ func (s *SocketOperations) Write(ctx context.Context, _ *fs.File, src usermem.IO } if resCh != nil { - t := ctx.(*kernel.Task) - if err := t.Block(resCh); err != nil { - return 0, syserr.FromError(err).ToError() + if err := amutex.Block(ctx, resCh); err != nil { + return 0, err } - n, _, err = s.Endpoint.Write(f, tcpip.WriteOptions{}) } @@ -598,11 +626,9 @@ func (s *SocketOperations) ReadFrom(ctx context.Context, _ *fs.File, r io.Reader } if resCh != nil { - t := ctx.(*kernel.Task) - if err := t.Block(resCh); err != nil { - return 0, syserr.FromError(err).ToError() + if err := amutex.Block(ctx, resCh); err != nil { + return 0, err } - n, _, err = s.Endpoint.Write(f, tcpip.WriteOptions{ Atomic: true, // See above. }) @@ -617,23 +643,21 @@ func (s *SocketOperations) ReadFrom(ctx context.Context, _ *fs.File, r io.Reader } // Readiness returns a mask of ready events for socket s. -func (s *SocketOperations) Readiness(mask waiter.EventMask) waiter.EventMask { +func (s *socketOpsCommon) Readiness(mask waiter.EventMask) waiter.EventMask { r := s.Endpoint.Readiness(mask) // Check our cached value iff the caller asked for readability and the // endpoint itself is currently not readable. if (mask & ^r & waiter.EventIn) != 0 { - s.readMu.Lock() - if len(s.readView) > 0 { + if atomic.LoadUint32(&s.readViewHasData) == 1 { r |= waiter.EventIn } - s.readMu.Unlock() } return r } -func (s *SocketOperations) checkFamily(family uint16, exact bool) *syserr.Error { +func (s *socketOpsCommon) checkFamily(family uint16, exact bool) *syserr.Error { if family == uint16(s.family) { return nil } @@ -655,8 +679,8 @@ func (s *SocketOperations) checkFamily(family uint16, exact bool) *syserr.Error // This is a hack to work around the fact that both IPv4 and IPv6 ANY are // represented by the empty string. // -// TODO(gvisor.dev/issues/1556): remove this function. -func (s *SocketOperations) mapFamily(addr tcpip.FullAddress, family uint16) tcpip.FullAddress { +// TODO(gvisor.dev/issue/1556): remove this function. +func (s *socketOpsCommon) mapFamily(addr tcpip.FullAddress, family uint16) tcpip.FullAddress { if len(addr.Addr) == 0 && s.family == linux.AF_INET6 && family == linux.AF_INET { addr.Addr = "\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\xff\xff\x00\x00\x00\x00" } @@ -665,7 +689,7 @@ func (s *SocketOperations) mapFamily(addr tcpip.FullAddress, family uint16) tcpi // Connect implements the linux syscall connect(2) for sockets backed by // tpcip.Endpoint. -func (s *SocketOperations) Connect(t *kernel.Task, sockaddr []byte, blocking bool) *syserr.Error { +func (s *socketOpsCommon) Connect(t *kernel.Task, sockaddr []byte, blocking bool) *syserr.Error { addr, family, err := AddressAndFamily(sockaddr) if err != nil { return err @@ -696,6 +720,14 @@ func (s *SocketOperations) Connect(t *kernel.Task, sockaddr []byte, blocking boo defer s.EventUnregister(&e) if err := s.Endpoint.Connect(addr); err != tcpip.ErrConnectStarted && err != tcpip.ErrAlreadyConnecting { + if (s.family == unix.AF_INET || s.family == unix.AF_INET6) && s.skType == linux.SOCK_STREAM { + // TCP unlike UDP returns EADDRNOTAVAIL when it can't + // find an available local ephemeral port. + if err == tcpip.ErrNoPortAvailable { + return syserr.ErrAddressNotAvailable + } + } + return syserr.TranslateNetstackError(err) } @@ -711,15 +743,45 @@ func (s *SocketOperations) Connect(t *kernel.Task, sockaddr []byte, blocking boo // Bind implements the linux syscall bind(2) for sockets backed by // tcpip.Endpoint. -func (s *SocketOperations) Bind(t *kernel.Task, sockaddr []byte) *syserr.Error { - addr, family, err := AddressAndFamily(sockaddr) - if err != nil { - return err +func (s *socketOpsCommon) Bind(t *kernel.Task, sockaddr []byte) *syserr.Error { + if len(sockaddr) < 2 { + return syserr.ErrInvalidArgument } - if err := s.checkFamily(family, true /* exact */); err != nil { - return err + + family := usermem.ByteOrder.Uint16(sockaddr) + var addr tcpip.FullAddress + + // Bind for AF_PACKET requires only family, protocol and ifindex. + // In function AddressAndFamily, we check the address length which is + // not needed for AF_PACKET bind. + if family == linux.AF_PACKET { + var a linux.SockAddrLink + if len(sockaddr) < sockAddrLinkSize { + return syserr.ErrInvalidArgument + } + binary.Unmarshal(sockaddr[:sockAddrLinkSize], usermem.ByteOrder, &a) + + if a.Protocol != uint16(s.protocol) { + return syserr.ErrInvalidArgument + } + + addr = tcpip.FullAddress{ + NIC: tcpip.NICID(a.InterfaceIndex), + Addr: tcpip.Address(a.HardwareAddr[:header.EthernetAddressSize]), + } + } else { + var err *syserr.Error + addr, family, err = AddressAndFamily(sockaddr) + if err != nil { + return err + } + + if err = s.checkFamily(family, true /* exact */); err != nil { + return err + } + + addr = s.mapFamily(addr, family) } - addr = s.mapFamily(addr, family) // Issue the bind request to the endpoint. return syserr.TranslateNetstackError(s.Endpoint.Bind(addr)) @@ -727,13 +789,13 @@ func (s *SocketOperations) Bind(t *kernel.Task, sockaddr []byte) *syserr.Error { // Listen implements the linux syscall listen(2) for sockets backed by // tcpip.Endpoint. -func (s *SocketOperations) Listen(t *kernel.Task, backlog int) *syserr.Error { +func (s *socketOpsCommon) Listen(t *kernel.Task, backlog int) *syserr.Error { return syserr.TranslateNetstackError(s.Endpoint.Listen(backlog)) } // blockingAccept implements a blocking version of accept(2), that is, if no // connections are ready to be accept, it will block until one becomes ready. -func (s *SocketOperations) blockingAccept(t *kernel.Task) (tcpip.Endpoint, *waiter.Queue, *syserr.Error) { +func (s *socketOpsCommon) blockingAccept(t *kernel.Task) (tcpip.Endpoint, *waiter.Queue, *syserr.Error) { // Register for notifications. e, ch := waiter.NewChannelEntry(nil) s.EventRegister(&e, waiter.EventIn) @@ -819,7 +881,7 @@ func ConvertShutdown(how int) (tcpip.ShutdownFlags, *syserr.Error) { // Shutdown implements the linux syscall shutdown(2) for sockets backed by // tcpip.Endpoint. -func (s *SocketOperations) Shutdown(t *kernel.Task, how int) *syserr.Error { +func (s *socketOpsCommon) Shutdown(t *kernel.Task, how int) *syserr.Error { f, err := ConvertShutdown(how) if err != nil { return err @@ -902,7 +964,7 @@ func (s *SocketOperations) GetSockOpt(t *kernel.Task, level, name int, outPtr us // GetSockOpt can be used to implement the linux syscall getsockopt(2) for // sockets backed by a commonEndpoint. -func GetSockOpt(t *kernel.Task, s socket.Socket, ep commonEndpoint, family int, skType linux.SockType, level, name, outLen int) (interface{}, *syserr.Error) { +func GetSockOpt(t *kernel.Task, s socket.SocketOps, ep commonEndpoint, family int, skType linux.SockType, level, name, outLen int) (interface{}, *syserr.Error) { switch level { case linux.SOL_SOCKET: return getSockOptSocket(t, s, ep, family, skType, name, outLen) @@ -927,8 +989,15 @@ func GetSockOpt(t *kernel.Task, s socket.Socket, ep commonEndpoint, family int, return nil, syserr.ErrProtocolNotAvailable } +func boolToInt32(v bool) int32 { + if v { + return 1 + } + return 0 +} + // getSockOptSocket implements GetSockOpt when level is SOL_SOCKET. -func getSockOptSocket(t *kernel.Task, s socket.Socket, ep commonEndpoint, family int, skType linux.SockType, name, outLen int) (interface{}, *syserr.Error) { +func getSockOptSocket(t *kernel.Task, s socket.SocketOps, ep commonEndpoint, family int, skType linux.SockType, name, outLen int) (interface{}, *syserr.Error) { // TODO(b/124056281): Stop rejecting short optLen values in getsockopt. switch name { case linux.SO_ERROR: @@ -960,12 +1029,11 @@ func getSockOptSocket(t *kernel.Task, s socket.Socket, ep commonEndpoint, family return nil, syserr.ErrInvalidArgument } - var v tcpip.PasscredOption - if err := ep.GetSockOpt(&v); err != nil { + v, err := ep.GetSockOptBool(tcpip.PasscredOption) + if err != nil { return nil, syserr.TranslateNetstackError(err) } - - return int32(v), nil + return boolToInt32(v), nil case linux.SO_SNDBUF: if outLen < sizeOfInt32 { @@ -1004,24 +1072,22 @@ func getSockOptSocket(t *kernel.Task, s socket.Socket, ep commonEndpoint, family return nil, syserr.ErrInvalidArgument } - var v tcpip.ReuseAddressOption - if err := ep.GetSockOpt(&v); err != nil { + v, err := ep.GetSockOptBool(tcpip.ReuseAddressOption) + if err != nil { return nil, syserr.TranslateNetstackError(err) } - - return int32(v), nil + return boolToInt32(v), nil case linux.SO_REUSEPORT: if outLen < sizeOfInt32 { return nil, syserr.ErrInvalidArgument } - var v tcpip.ReusePortOption - if err := ep.GetSockOpt(&v); err != nil { + v, err := ep.GetSockOptBool(tcpip.ReusePortOption) + if err != nil { return nil, syserr.TranslateNetstackError(err) } - - return int32(v), nil + return boolToInt32(v), nil case linux.SO_BINDTODEVICE: var v tcpip.BindToDeviceOption @@ -1051,24 +1117,22 @@ func getSockOptSocket(t *kernel.Task, s socket.Socket, ep commonEndpoint, family return nil, syserr.ErrInvalidArgument } - var v tcpip.BroadcastOption - if err := ep.GetSockOpt(&v); err != nil { + v, err := ep.GetSockOptBool(tcpip.BroadcastOption) + if err != nil { return nil, syserr.TranslateNetstackError(err) } - - return int32(v), nil + return boolToInt32(v), nil case linux.SO_KEEPALIVE: if outLen < sizeOfInt32 { return nil, syserr.ErrInvalidArgument } - var v tcpip.KeepaliveEnabledOption - if err := ep.GetSockOpt(&v); err != nil { + v, err := ep.GetSockOptBool(tcpip.KeepaliveEnabledOption) + if err != nil { return nil, syserr.TranslateNetstackError(err) } - - return int32(v), nil + return boolToInt32(v), nil case linux.SO_LINGER: if outLen < linux.SizeOfLinger { @@ -1118,47 +1182,41 @@ func getSockOptTCP(t *kernel.Task, ep commonEndpoint, name, outLen int) (interfa return nil, syserr.ErrInvalidArgument } - v, err := ep.GetSockOptInt(tcpip.DelayOption) + v, err := ep.GetSockOptBool(tcpip.DelayOption) if err != nil { return nil, syserr.TranslateNetstackError(err) } - - if v == 0 { - return int32(1), nil - } - return int32(0), nil + return boolToInt32(!v), nil case linux.TCP_CORK: if outLen < sizeOfInt32 { return nil, syserr.ErrInvalidArgument } - var v tcpip.CorkOption - if err := ep.GetSockOpt(&v); err != nil { + v, err := ep.GetSockOptBool(tcpip.CorkOption) + if err != nil { return nil, syserr.TranslateNetstackError(err) } - - return int32(v), nil + return boolToInt32(v), nil case linux.TCP_QUICKACK: if outLen < sizeOfInt32 { return nil, syserr.ErrInvalidArgument } - var v tcpip.QuickAckOption - if err := ep.GetSockOpt(&v); err != nil { + v, err := ep.GetSockOptBool(tcpip.QuickAckOption) + if err != nil { return nil, syserr.TranslateNetstackError(err) } - - return int32(v), nil + return boolToInt32(v), nil case linux.TCP_MAXSEG: if outLen < sizeOfInt32 { return nil, syserr.ErrInvalidArgument } - var v tcpip.MaxSegOption - if err := ep.GetSockOpt(&v); err != nil { + v, err := ep.GetSockOptInt(tcpip.MaxSegOption) + if err != nil { return nil, syserr.TranslateNetstackError(err) } @@ -1272,6 +1330,29 @@ func getSockOptTCP(t *kernel.Task, ep commonEndpoint, name, outLen int) (interfa return int32(time.Duration(v) / time.Second), nil + case linux.TCP_SYNCNT: + if outLen < sizeOfInt32 { + return nil, syserr.ErrInvalidArgument + } + + v, err := ep.GetSockOptInt(tcpip.TCPSynCountOption) + if err != nil { + return nil, syserr.TranslateNetstackError(err) + } + + return int32(v), nil + + case linux.TCP_WINDOW_CLAMP: + if outLen < sizeOfInt32 { + return nil, syserr.ErrInvalidArgument + } + + v, err := ep.GetSockOptInt(tcpip.TCPWindowClampOption) + if err != nil { + return nil, syserr.TranslateNetstackError(err) + } + + return int32(v), nil default: emitUnimplementedEventTCP(t, name) } @@ -1290,11 +1371,7 @@ func getSockOptIPv6(t *kernel.Task, ep commonEndpoint, name, outLen int) (interf if err != nil { return nil, syserr.TranslateNetstackError(err) } - var o int32 - if v { - o = 1 - } - return o, nil + return boolToInt32(v), nil case linux.IPV6_PATHMTU: t.Kernel().EmitUnimplementedEvent(t) @@ -1304,8 +1381,8 @@ func getSockOptIPv6(t *kernel.Task, ep commonEndpoint, name, outLen int) (interf if outLen == 0 { return make([]byte, 0), nil } - var v tcpip.IPv6TrafficClassOption - if err := ep.GetSockOpt(&v); err != nil { + v, err := ep.GetSockOptInt(tcpip.IPv6TrafficClassOption) + if err != nil { return nil, syserr.TranslateNetstackError(err) } @@ -1318,6 +1395,17 @@ func getSockOptIPv6(t *kernel.Task, ep commonEndpoint, name, outLen int) (interf } return ib, nil + case linux.IPV6_RECVTCLASS: + if outLen < sizeOfInt32 { + return nil, syserr.ErrInvalidArgument + } + + v, err := ep.GetSockOptBool(tcpip.ReceiveTClassOption) + if err != nil { + return nil, syserr.TranslateNetstackError(err) + } + return boolToInt32(v), nil + default: emitUnimplementedEventIPv6(t, name) } @@ -1332,8 +1420,8 @@ func getSockOptIP(t *kernel.Task, ep commonEndpoint, name, outLen int, family in return nil, syserr.ErrInvalidArgument } - var v tcpip.TTLOption - if err := ep.GetSockOpt(&v); err != nil { + v, err := ep.GetSockOptInt(tcpip.TTLOption) + if err != nil { return nil, syserr.TranslateNetstackError(err) } @@ -1349,8 +1437,8 @@ func getSockOptIP(t *kernel.Task, ep commonEndpoint, name, outLen int, family in return nil, syserr.ErrInvalidArgument } - var v tcpip.MulticastTTLOption - if err := ep.GetSockOpt(&v); err != nil { + v, err := ep.GetSockOptInt(tcpip.MulticastTTLOption) + if err != nil { return nil, syserr.TranslateNetstackError(err) } @@ -1375,23 +1463,19 @@ func getSockOptIP(t *kernel.Task, ep commonEndpoint, name, outLen int, family in return nil, syserr.ErrInvalidArgument } - var v tcpip.MulticastLoopOption - if err := ep.GetSockOpt(&v); err != nil { + v, err := ep.GetSockOptBool(tcpip.MulticastLoopOption) + if err != nil { return nil, syserr.TranslateNetstackError(err) } - - if v { - return int32(1), nil - } - return int32(0), nil + return boolToInt32(v), nil case linux.IP_TOS: // Length handling for parity with Linux. if outLen == 0 { return []byte(nil), nil } - var v tcpip.IPv4TOSOption - if err := ep.GetSockOpt(&v); err != nil { + v, err := ep.GetSockOptInt(tcpip.IPv4TOSOption) + if err != nil { return nil, syserr.TranslateNetstackError(err) } if outLen < sizeOfInt32 { @@ -1408,11 +1492,7 @@ func getSockOptIP(t *kernel.Task, ep commonEndpoint, name, outLen int, family in if err != nil { return nil, syserr.TranslateNetstackError(err) } - var o int32 - if v { - o = 1 - } - return o, nil + return boolToInt32(v), nil case linux.IP_PKTINFO: if outLen < sizeOfInt32 { @@ -1423,11 +1503,7 @@ func getSockOptIP(t *kernel.Task, ep commonEndpoint, name, outLen int, family in if err != nil { return nil, syserr.TranslateNetstackError(err) } - var o int32 - if v { - o = 1 - } - return o, nil + return boolToInt32(v), nil default: emitUnimplementedEventIP(t, name) @@ -1487,7 +1563,7 @@ func (s *SocketOperations) SetSockOpt(t *kernel.Task, level int, name int, optVa // SetSockOpt can be used to implement the linux syscall setsockopt(2) for // sockets backed by a commonEndpoint. -func SetSockOpt(t *kernel.Task, s socket.Socket, ep commonEndpoint, level int, name int, optVal []byte) *syserr.Error { +func SetSockOpt(t *kernel.Task, s socket.SocketOps, ep commonEndpoint, level int, name int, optVal []byte) *syserr.Error { switch level { case linux.SOL_SOCKET: return setSockOptSocket(t, s, ep, name, optVal) @@ -1514,7 +1590,7 @@ func SetSockOpt(t *kernel.Task, s socket.Socket, ep commonEndpoint, level int, n } // setSockOptSocket implements SetSockOpt when level is SOL_SOCKET. -func setSockOptSocket(t *kernel.Task, s socket.Socket, ep commonEndpoint, name int, optVal []byte) *syserr.Error { +func setSockOptSocket(t *kernel.Task, s socket.SocketOps, ep commonEndpoint, name int, optVal []byte) *syserr.Error { switch name { case linux.SO_SNDBUF: if len(optVal) < sizeOfInt32 { @@ -1538,7 +1614,7 @@ func setSockOptSocket(t *kernel.Task, s socket.Socket, ep commonEndpoint, name i } v := usermem.ByteOrder.Uint32(optVal) - return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.ReuseAddressOption(v))) + return syserr.TranslateNetstackError(ep.SetSockOptBool(tcpip.ReuseAddressOption, v != 0)) case linux.SO_REUSEPORT: if len(optVal) < sizeOfInt32 { @@ -1546,7 +1622,7 @@ func setSockOptSocket(t *kernel.Task, s socket.Socket, ep commonEndpoint, name i } v := usermem.ByteOrder.Uint32(optVal) - return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.ReusePortOption(v))) + return syserr.TranslateNetstackError(ep.SetSockOptBool(tcpip.ReusePortOption, v != 0)) case linux.SO_BINDTODEVICE: n := bytes.IndexByte(optVal, 0) @@ -1574,7 +1650,7 @@ func setSockOptSocket(t *kernel.Task, s socket.Socket, ep commonEndpoint, name i } v := usermem.ByteOrder.Uint32(optVal) - return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.BroadcastOption(v))) + return syserr.TranslateNetstackError(ep.SetSockOptBool(tcpip.BroadcastOption, v != 0)) case linux.SO_PASSCRED: if len(optVal) < sizeOfInt32 { @@ -1582,7 +1658,7 @@ func setSockOptSocket(t *kernel.Task, s socket.Socket, ep commonEndpoint, name i } v := usermem.ByteOrder.Uint32(optVal) - return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.PasscredOption(v))) + return syserr.TranslateNetstackError(ep.SetSockOptBool(tcpip.PasscredOption, v != 0)) case linux.SO_KEEPALIVE: if len(optVal) < sizeOfInt32 { @@ -1590,7 +1666,7 @@ func setSockOptSocket(t *kernel.Task, s socket.Socket, ep commonEndpoint, name i } v := usermem.ByteOrder.Uint32(optVal) - return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.KeepaliveEnabledOption(v))) + return syserr.TranslateNetstackError(ep.SetSockOptBool(tcpip.KeepaliveEnabledOption, v != 0)) case linux.SO_SNDTIMEO: if len(optVal) < linux.SizeOfTimeval { @@ -1662,11 +1738,7 @@ func setSockOptTCP(t *kernel.Task, ep commonEndpoint, name int, optVal []byte) * } v := usermem.ByteOrder.Uint32(optVal) - var o int - if v == 0 { - o = 1 - } - return syserr.TranslateNetstackError(ep.SetSockOptInt(tcpip.DelayOption, o)) + return syserr.TranslateNetstackError(ep.SetSockOptBool(tcpip.DelayOption, v == 0)) case linux.TCP_CORK: if len(optVal) < sizeOfInt32 { @@ -1674,7 +1746,7 @@ func setSockOptTCP(t *kernel.Task, ep commonEndpoint, name int, optVal []byte) * } v := usermem.ByteOrder.Uint32(optVal) - return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.CorkOption(v))) + return syserr.TranslateNetstackError(ep.SetSockOptBool(tcpip.CorkOption, v != 0)) case linux.TCP_QUICKACK: if len(optVal) < sizeOfInt32 { @@ -1682,7 +1754,7 @@ func setSockOptTCP(t *kernel.Task, ep commonEndpoint, name int, optVal []byte) * } v := usermem.ByteOrder.Uint32(optVal) - return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.QuickAckOption(v))) + return syserr.TranslateNetstackError(ep.SetSockOptBool(tcpip.QuickAckOption, v != 0)) case linux.TCP_MAXSEG: if len(optVal) < sizeOfInt32 { @@ -1690,7 +1762,7 @@ func setSockOptTCP(t *kernel.Task, ep commonEndpoint, name int, optVal []byte) * } v := usermem.ByteOrder.Uint32(optVal) - return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.MaxSegOption(v))) + return syserr.TranslateNetstackError(ep.SetSockOptInt(tcpip.MaxSegOption, int(v))) case linux.TCP_KEEPIDLE: if len(optVal) < sizeOfInt32 { @@ -1750,6 +1822,22 @@ func setSockOptTCP(t *kernel.Task, ep commonEndpoint, name int, optVal []byte) * } return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.TCPDeferAcceptOption(time.Second * time.Duration(v)))) + case linux.TCP_SYNCNT: + if len(optVal) < sizeOfInt32 { + return syserr.ErrInvalidArgument + } + v := usermem.ByteOrder.Uint32(optVal) + + return syserr.TranslateNetstackError(ep.SetSockOptInt(tcpip.TCPSynCountOption, int(v))) + + case linux.TCP_WINDOW_CLAMP: + if len(optVal) < sizeOfInt32 { + return syserr.ErrInvalidArgument + } + v := usermem.ByteOrder.Uint32(optVal) + + return syserr.TranslateNetstackError(ep.SetSockOptInt(tcpip.TCPWindowClampOption, int(v))) + case linux.TCP_REPAIR_OPTIONS: t.Kernel().EmitUnimplementedEvent(t) @@ -1801,7 +1889,15 @@ func setSockOptIPv6(t *kernel.Task, ep commonEndpoint, name int, optVal []byte) if v == -1 { v = 0 } - return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.IPv6TrafficClassOption(v))) + return syserr.TranslateNetstackError(ep.SetSockOptInt(tcpip.IPv6TrafficClassOption, int(v))) + + case linux.IPV6_RECVTCLASS: + v, err := parseIntOrChar(optVal) + if err != nil { + return err + } + + return syserr.TranslateNetstackError(ep.SetSockOptBool(tcpip.ReceiveTClassOption, v != 0)) default: emitUnimplementedEventIPv6(t, name) @@ -1878,7 +1974,7 @@ func setSockOptIP(t *kernel.Task, ep commonEndpoint, name int, optVal []byte) *s if v < 0 || v > 255 { return syserr.ErrInvalidArgument } - return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.MulticastTTLOption(v))) + return syserr.TranslateNetstackError(ep.SetSockOptInt(tcpip.MulticastTTLOption, int(v))) case linux.IP_ADD_MEMBERSHIP: req, err := copyInMulticastRequest(optVal, false /* allowAddr */) @@ -1925,9 +2021,7 @@ func setSockOptIP(t *kernel.Task, ep commonEndpoint, name int, optVal []byte) *s return err } - return syserr.TranslateNetstackError(ep.SetSockOpt( - tcpip.MulticastLoopOption(v != 0), - )) + return syserr.TranslateNetstackError(ep.SetSockOptBool(tcpip.MulticastLoopOption, v != 0)) case linux.MCAST_JOIN_GROUP: // FIXME(b/124219304): Implement MCAST_JOIN_GROUP. @@ -1946,7 +2040,7 @@ func setSockOptIP(t *kernel.Task, ep commonEndpoint, name int, optVal []byte) *s } else if v < 1 || v > 255 { return syserr.ErrInvalidArgument } - return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.TTLOption(v))) + return syserr.TranslateNetstackError(ep.SetSockOptInt(tcpip.TTLOption, int(v))) case linux.IP_TOS: if len(optVal) == 0 { @@ -1956,7 +2050,7 @@ func setSockOptIP(t *kernel.Task, ep commonEndpoint, name int, optVal []byte) *s if err != nil { return err } - return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.IPv4TOSOption(v))) + return syserr.TranslateNetstackError(ep.SetSockOptInt(tcpip.IPv4TOSOption, int(v))) case linux.IP_RECVTOS: v, err := parseIntOrChar(optVal) @@ -2086,7 +2180,6 @@ func emitUnimplementedEventIPv6(t *kernel.Task, name int) { linux.IPV6_RECVPATHMTU, linux.IPV6_RECVPKTINFO, linux.IPV6_RECVRTHDR, - linux.IPV6_RECVTCLASS, linux.IPV6_RTHDR, linux.IPV6_RTHDRDSTOPTS, linux.IPV6_TCLASS, @@ -2222,7 +2315,7 @@ func ConvertAddress(family int, addr tcpip.FullAddress) (linux.SockAddr, uint32) // GetSockName implements the linux syscall getsockname(2) for sockets backed by // tcpip.Endpoint. -func (s *SocketOperations) GetSockName(t *kernel.Task) (linux.SockAddr, uint32, *syserr.Error) { +func (s *socketOpsCommon) GetSockName(t *kernel.Task) (linux.SockAddr, uint32, *syserr.Error) { addr, err := s.Endpoint.GetLocalAddress() if err != nil { return nil, 0, syserr.TranslateNetstackError(err) @@ -2234,7 +2327,7 @@ func (s *SocketOperations) GetSockName(t *kernel.Task) (linux.SockAddr, uint32, // GetPeerName implements the linux syscall getpeername(2) for sockets backed by // tcpip.Endpoint. -func (s *SocketOperations) GetPeerName(t *kernel.Task) (linux.SockAddr, uint32, *syserr.Error) { +func (s *socketOpsCommon) GetPeerName(t *kernel.Task) (linux.SockAddr, uint32, *syserr.Error) { addr, err := s.Endpoint.GetRemoteAddress() if err != nil { return nil, 0, syserr.TranslateNetstackError(err) @@ -2249,7 +2342,7 @@ func (s *SocketOperations) GetPeerName(t *kernel.Task) (linux.SockAddr, uint32, // caller. // // Precondition: s.readMu must be locked. -func (s *SocketOperations) coalescingRead(ctx context.Context, dst usermem.IOSequence, discard bool) (int, *syserr.Error) { +func (s *socketOpsCommon) coalescingRead(ctx context.Context, dst usermem.IOSequence, discard bool) (int, *syserr.Error) { var err *syserr.Error var copied int @@ -2281,6 +2374,10 @@ func (s *SocketOperations) coalescingRead(ctx context.Context, dst usermem.IOSeq } copied += n s.readView.TrimFront(n) + if len(s.readView) == 0 { + atomic.StoreUint32(&s.readViewHasData, 0) + } + dst = dst.DropFirst(n) if e != nil { err = syserr.FromError(e) @@ -2297,7 +2394,7 @@ func (s *SocketOperations) coalescingRead(ctx context.Context, dst usermem.IOSeq return 0, err } -func (s *SocketOperations) fillCmsgInq(cmsg *socket.ControlMessages) { +func (s *socketOpsCommon) fillCmsgInq(cmsg *socket.ControlMessages) { if !s.sockOptInq { return } @@ -2312,7 +2409,7 @@ func (s *SocketOperations) fillCmsgInq(cmsg *socket.ControlMessages) { // nonBlockingRead issues a non-blocking read. // // TODO(b/78348848): Support timestamps for stream sockets. -func (s *SocketOperations) nonBlockingRead(ctx context.Context, dst usermem.IOSequence, peek, trunc, senderRequested bool) (int, int, linux.SockAddr, uint32, socket.ControlMessages, *syserr.Error) { +func (s *socketOpsCommon) nonBlockingRead(ctx context.Context, dst usermem.IOSequence, peek, trunc, senderRequested bool) (int, int, linux.SockAddr, uint32, socket.ControlMessages, *syserr.Error) { isPacket := s.isPacketBased() // Fast path for regular reads from stream (e.g., TCP) endpoints. Note @@ -2327,9 +2424,9 @@ func (s *SocketOperations) nonBlockingRead(ctx context.Context, dst usermem.IOSe // caller-supplied buffer. s.readMu.Lock() n, err := s.coalescingRead(ctx, dst, trunc) - s.readMu.Unlock() cmsg := s.controlMessages() s.fillCmsgInq(&cmsg) + s.readMu.Unlock() return n, 0, nil, 0, cmsg, err } @@ -2403,6 +2500,10 @@ func (s *SocketOperations) nonBlockingRead(ctx context.Context, dst usermem.IOSe s.readView.TrimFront(int(n)) } + if len(s.readView) == 0 { + atomic.StoreUint32(&s.readViewHasData, 0) + } + var flags int if msgLen > int(n) { flags |= linux.MSG_TRUNC @@ -2417,13 +2518,15 @@ func (s *SocketOperations) nonBlockingRead(ctx context.Context, dst usermem.IOSe return n, flags, addr, addrLen, cmsg, syserr.FromError(err) } -func (s *SocketOperations) controlMessages() socket.ControlMessages { +func (s *socketOpsCommon) controlMessages() socket.ControlMessages { return socket.ControlMessages{ IP: tcpip.ControlMessages{ HasTimestamp: s.readCM.HasTimestamp && s.sockOptTimestamp, Timestamp: s.readCM.Timestamp, HasTOS: s.readCM.HasTOS, TOS: s.readCM.TOS, + HasTClass: s.readCM.HasTClass, + TClass: s.readCM.TClass, HasIPPacketInfo: s.readCM.HasIPPacketInfo, PacketInfo: s.readCM.PacketInfo, }, @@ -2434,7 +2537,7 @@ func (s *SocketOperations) controlMessages() socket.ControlMessages { // successfully writing packet data out to userspace. // // Precondition: s.readMu must be locked. -func (s *SocketOperations) updateTimestamp() { +func (s *socketOpsCommon) updateTimestamp() { // Save the SIOCGSTAMP timestamp only if SO_TIMESTAMP is disabled. if !s.sockOptTimestamp { s.timestampValid = true @@ -2444,7 +2547,7 @@ func (s *SocketOperations) updateTimestamp() { // RecvMsg implements the linux syscall recvmsg(2) for sockets backed by // tcpip.Endpoint. -func (s *SocketOperations) RecvMsg(t *kernel.Task, dst usermem.IOSequence, flags int, haveDeadline bool, deadline ktime.Time, senderRequested bool, controlDataLen uint64) (n int, msgFlags int, senderAddr linux.SockAddr, senderAddrLen uint32, controlMessages socket.ControlMessages, err *syserr.Error) { +func (s *socketOpsCommon) RecvMsg(t *kernel.Task, dst usermem.IOSequence, flags int, haveDeadline bool, deadline ktime.Time, senderRequested bool, controlDataLen uint64) (n int, msgFlags int, senderAddr linux.SockAddr, senderAddrLen uint32, controlMessages socket.ControlMessages, err *syserr.Error) { trunc := flags&linux.MSG_TRUNC != 0 peek := flags&linux.MSG_PEEK != 0 dontWait := flags&linux.MSG_DONTWAIT != 0 @@ -2512,7 +2615,7 @@ func (s *SocketOperations) RecvMsg(t *kernel.Task, dst usermem.IOSequence, flags // SendMsg implements the linux syscall sendmsg(2) for sockets backed by // tcpip.Endpoint. -func (s *SocketOperations) SendMsg(t *kernel.Task, src usermem.IOSequence, to []byte, flags int, haveDeadline bool, deadline ktime.Time, controlMessages socket.ControlMessages) (int, *syserr.Error) { +func (s *socketOpsCommon) SendMsg(t *kernel.Task, src usermem.IOSequence, to []byte, flags int, haveDeadline bool, deadline ktime.Time, controlMessages socket.ControlMessages) (int, *syserr.Error) { // Reject Unix control messages. if !controlMessages.Unix.Empty() { return 0, syserr.ErrInvalidArgument @@ -2588,6 +2691,10 @@ func (s *SocketOperations) SendMsg(t *kernel.Task, src usermem.IOSequence, to [] // Ioctl implements fs.FileOperations.Ioctl. func (s *SocketOperations) Ioctl(ctx context.Context, _ *fs.File, io usermem.IO, args arch.SyscallArguments) (uintptr, error) { + return s.socketOpsCommon.ioctl(ctx, io, args) +} + +func (s *socketOpsCommon) ioctl(ctx context.Context, io usermem.IO, args arch.SyscallArguments) (uintptr, error) { // SIOCGSTAMP is implemented by netstack rather than all commonEndpoint // sockets. // TODO(b/78348848): Add a commonEndpoint method to support SIOCGSTAMP. @@ -2612,13 +2719,15 @@ func (s *SocketOperations) Ioctl(ctx context.Context, _ *fs.File, io usermem.IO, } // Add bytes removed from the endpoint but not yet sent to the caller. + s.readMu.Lock() v += len(s.readView) + s.readMu.Unlock() if v > math.MaxInt32 { v = math.MaxInt32 } - // Copy result to user-space. + // Copy result to userspace. _, err := usermem.CopyObjectOut(ctx, io, args[2].Pointer(), int32(v), usermem.IOOpts{ AddressSpaceActive: true, }) @@ -2687,7 +2796,7 @@ func Ioctl(ctx context.Context, ep commonEndpoint, io usermem.IO, args arch.Sysc if v > math.MaxInt32 { v = math.MaxInt32 } - // Copy result to user-space. + // Copy result to userspace. _, err := usermem.CopyObjectOut(ctx, io, args[2].Pointer(), int32(v), usermem.IOOpts{ AddressSpaceActive: true, }) @@ -2703,7 +2812,7 @@ func Ioctl(ctx context.Context, ep commonEndpoint, io usermem.IO, args arch.Sysc v = math.MaxInt32 } - // Copy result to user-space. + // Copy result to userspace. _, err := usermem.CopyObjectOut(ctx, io, args[2].Pointer(), int32(v), usermem.IOOpts{ AddressSpaceActive: true, }) @@ -2925,7 +3034,7 @@ func nicStateFlagsToLinux(f stack.NICStateFlags) uint32 { // State implements socket.Socket.State. State translates the internal state // returned by netstack to values defined by Linux. -func (s *SocketOperations) State() uint32 { +func (s *socketOpsCommon) State() uint32 { if s.family != linux.AF_INET && s.family != linux.AF_INET6 { // States not implemented for this socket's family. return 0 @@ -2985,6 +3094,8 @@ func (s *SocketOperations) State() uint32 { } // Type implements socket.Socket.Type. -func (s *SocketOperations) Type() (family int, skType linux.SockType, protocol int) { +func (s *socketOpsCommon) Type() (family int, skType linux.SockType, protocol int) { return s.family, s.skType, s.protocol } + +// LINT.ThenChange(./netstack_vfs2.go) diff --git a/pkg/sentry/socket/netstack/netstack_vfs2.go b/pkg/sentry/socket/netstack/netstack_vfs2.go new file mode 100644 index 000000000..fcd8013c0 --- /dev/null +++ b/pkg/sentry/socket/netstack/netstack_vfs2.go @@ -0,0 +1,317 @@ +// Copyright 2018 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package netstack + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/amutex" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/sockfs" + "gvisor.dev/gvisor/pkg/sentry/inet" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/socket" + "gvisor.dev/gvisor/pkg/sentry/socket/netfilter" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserr" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/tcpip" + "gvisor.dev/gvisor/pkg/usermem" + "gvisor.dev/gvisor/pkg/waiter" +) + +// SocketVFS2 encapsulates all the state needed to represent a network stack +// endpoint in the kernel context. +type SocketVFS2 struct { + vfsfd vfs.FileDescription + vfs.FileDescriptionDefaultImpl + vfs.DentryMetadataFileDescriptionImpl + + socketOpsCommon +} + +var _ = socket.SocketVFS2(&SocketVFS2{}) + +// NewVFS2 creates a new endpoint socket. +func NewVFS2(t *kernel.Task, family int, skType linux.SockType, protocol int, queue *waiter.Queue, endpoint tcpip.Endpoint) (*vfs.FileDescription, *syserr.Error) { + if skType == linux.SOCK_STREAM { + if err := endpoint.SetSockOptBool(tcpip.DelayOption, true); err != nil { + return nil, syserr.TranslateNetstackError(err) + } + } + + mnt := t.Kernel().SocketMount() + d := sockfs.NewDentry(t.Credentials(), mnt) + + s := &SocketVFS2{ + socketOpsCommon: socketOpsCommon{ + Queue: queue, + family: family, + Endpoint: endpoint, + skType: skType, + protocol: protocol, + }, + } + vfsfd := &s.vfsfd + if err := vfsfd.Init(s, linux.O_RDWR, mnt, d, &vfs.FileDescriptionOptions{ + DenyPRead: true, + DenyPWrite: true, + UseDentryMetadata: true, + }); err != nil { + return nil, syserr.FromError(err) + } + return vfsfd, nil +} + +// Readiness implements waiter.Waitable.Readiness. +func (s *SocketVFS2) Readiness(mask waiter.EventMask) waiter.EventMask { + return s.socketOpsCommon.Readiness(mask) +} + +// EventRegister implements waiter.Waitable.EventRegister. +func (s *SocketVFS2) EventRegister(e *waiter.Entry, mask waiter.EventMask) { + s.socketOpsCommon.EventRegister(e, mask) +} + +// EventUnregister implements waiter.Waitable.EventUnregister. +func (s *SocketVFS2) EventUnregister(e *waiter.Entry) { + s.socketOpsCommon.EventUnregister(e) +} + +// Read implements vfs.FileDescriptionImpl. +func (s *SocketVFS2) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { + // All flags other than RWF_NOWAIT should be ignored. + // TODO(gvisor.dev/issue/2601): Support RWF_NOWAIT. + if opts.Flags != 0 { + return 0, syserror.EOPNOTSUPP + } + + if dst.NumBytes() == 0 { + return 0, nil + } + n, _, _, _, _, err := s.nonBlockingRead(ctx, dst, false, false, false) + if err == syserr.ErrWouldBlock { + return int64(n), syserror.ErrWouldBlock + } + if err != nil { + return 0, err.ToError() + } + return int64(n), nil +} + +// Write implements vfs.FileDescriptionImpl. +func (s *SocketVFS2) Write(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) { + // All flags other than RWF_NOWAIT should be ignored. + // TODO(gvisor.dev/issue/2601): Support RWF_NOWAIT. + if opts.Flags != 0 { + return 0, syserror.EOPNOTSUPP + } + + f := &ioSequencePayload{ctx: ctx, src: src} + n, resCh, err := s.Endpoint.Write(f, tcpip.WriteOptions{}) + if err == tcpip.ErrWouldBlock { + return 0, syserror.ErrWouldBlock + } + + if resCh != nil { + if err := amutex.Block(ctx, resCh); err != nil { + return 0, err + } + n, _, err = s.Endpoint.Write(f, tcpip.WriteOptions{}) + } + + if err != nil { + return 0, syserr.TranslateNetstackError(err).ToError() + } + + if int64(n) < src.NumBytes() { + return int64(n), syserror.ErrWouldBlock + } + + return int64(n), nil +} + +// Accept implements the linux syscall accept(2) for sockets backed by +// tcpip.Endpoint. +func (s *SocketVFS2) Accept(t *kernel.Task, peerRequested bool, flags int, blocking bool) (int32, linux.SockAddr, uint32, *syserr.Error) { + // Issue the accept request to get the new endpoint. + ep, wq, terr := s.Endpoint.Accept() + if terr != nil { + if terr != tcpip.ErrWouldBlock || !blocking { + return 0, nil, 0, syserr.TranslateNetstackError(terr) + } + + var err *syserr.Error + ep, wq, err = s.blockingAccept(t) + if err != nil { + return 0, nil, 0, err + } + } + + ns, err := NewVFS2(t, s.family, s.skType, s.protocol, wq, ep) + if err != nil { + return 0, nil, 0, err + } + defer ns.DecRef() + + if err := ns.SetStatusFlags(t, t.Credentials(), uint32(flags&linux.SOCK_NONBLOCK)); err != nil { + return 0, nil, 0, syserr.FromError(err) + } + + var addr linux.SockAddr + var addrLen uint32 + if peerRequested { + // Get address of the peer and write it to peer slice. + var err *syserr.Error + addr, addrLen, err = ns.Impl().(*SocketVFS2).GetPeerName(t) + if err != nil { + return 0, nil, 0, err + } + } + + fd, e := t.NewFDFromVFS2(0, ns, kernel.FDFlags{ + CloseOnExec: flags&linux.SOCK_CLOEXEC != 0, + }) + + t.Kernel().RecordSocketVFS2(ns) + + return fd, addr, addrLen, syserr.FromError(e) +} + +// Ioctl implements vfs.FileDescriptionImpl. +func (s *SocketVFS2) Ioctl(ctx context.Context, uio usermem.IO, args arch.SyscallArguments) (uintptr, error) { + return s.socketOpsCommon.ioctl(ctx, uio, args) +} + +// GetSockOpt implements the linux syscall getsockopt(2) for sockets backed by +// tcpip.Endpoint. +func (s *SocketVFS2) GetSockOpt(t *kernel.Task, level, name int, outPtr usermem.Addr, outLen int) (interface{}, *syserr.Error) { + // TODO(b/78348848): Unlike other socket options, SO_TIMESTAMP is + // implemented specifically for netstack.SocketVFS2 rather than + // commonEndpoint. commonEndpoint should be extended to support socket + // options where the implementation is not shared, as unix sockets need + // their own support for SO_TIMESTAMP. + if level == linux.SOL_SOCKET && name == linux.SO_TIMESTAMP { + if outLen < sizeOfInt32 { + return nil, syserr.ErrInvalidArgument + } + val := int32(0) + s.readMu.Lock() + defer s.readMu.Unlock() + if s.sockOptTimestamp { + val = 1 + } + return val, nil + } + if level == linux.SOL_TCP && name == linux.TCP_INQ { + if outLen < sizeOfInt32 { + return nil, syserr.ErrInvalidArgument + } + val := int32(0) + s.readMu.Lock() + defer s.readMu.Unlock() + if s.sockOptInq { + val = 1 + } + return val, nil + } + + if s.skType == linux.SOCK_RAW && level == linux.IPPROTO_IP { + switch name { + case linux.IPT_SO_GET_INFO: + if outLen < linux.SizeOfIPTGetinfo { + return nil, syserr.ErrInvalidArgument + } + + stack := inet.StackFromContext(t) + if stack == nil { + return nil, syserr.ErrNoDevice + } + info, err := netfilter.GetInfo(t, stack.(*Stack).Stack, outPtr) + if err != nil { + return nil, err + } + return info, nil + + case linux.IPT_SO_GET_ENTRIES: + if outLen < linux.SizeOfIPTGetEntries { + return nil, syserr.ErrInvalidArgument + } + + stack := inet.StackFromContext(t) + if stack == nil { + return nil, syserr.ErrNoDevice + } + entries, err := netfilter.GetEntries(t, stack.(*Stack).Stack, outPtr, outLen) + if err != nil { + return nil, err + } + return entries, nil + + } + } + + return GetSockOpt(t, s, s.Endpoint, s.family, s.skType, level, name, outLen) +} + +// SetSockOpt implements the linux syscall setsockopt(2) for sockets backed by +// tcpip.Endpoint. +func (s *SocketVFS2) SetSockOpt(t *kernel.Task, level int, name int, optVal []byte) *syserr.Error { + // TODO(b/78348848): Unlike other socket options, SO_TIMESTAMP is + // implemented specifically for netstack.SocketVFS2 rather than + // commonEndpoint. commonEndpoint should be extended to support socket + // options where the implementation is not shared, as unix sockets need + // their own support for SO_TIMESTAMP. + if level == linux.SOL_SOCKET && name == linux.SO_TIMESTAMP { + if len(optVal) < sizeOfInt32 { + return syserr.ErrInvalidArgument + } + s.readMu.Lock() + defer s.readMu.Unlock() + s.sockOptTimestamp = usermem.ByteOrder.Uint32(optVal) != 0 + return nil + } + if level == linux.SOL_TCP && name == linux.TCP_INQ { + if len(optVal) < sizeOfInt32 { + return syserr.ErrInvalidArgument + } + s.readMu.Lock() + defer s.readMu.Unlock() + s.sockOptInq = usermem.ByteOrder.Uint32(optVal) != 0 + return nil + } + + if s.skType == linux.SOCK_RAW && level == linux.IPPROTO_IP { + switch name { + case linux.IPT_SO_SET_REPLACE: + if len(optVal) < linux.SizeOfIPTReplace { + return syserr.ErrInvalidArgument + } + + stack := inet.StackFromContext(t) + if stack == nil { + return syserr.ErrNoDevice + } + // Stack must be a netstack stack. + return netfilter.SetEntries(stack.(*Stack).Stack, optVal) + + case linux.IPT_SO_SET_ADD_COUNTERS: + // TODO(gvisor.dev/issue/170): Counter support. + return nil + } + } + + return SetSockOpt(t, s, s.Endpoint, level, name, optVal) +} diff --git a/pkg/sentry/socket/netstack/provider.go b/pkg/sentry/socket/netstack/provider.go index 5afff2564..ead3b2b79 100644 --- a/pkg/sentry/socket/netstack/provider.go +++ b/pkg/sentry/socket/netstack/provider.go @@ -33,6 +33,8 @@ import ( "gvisor.dev/gvisor/pkg/waiter" ) +// LINT.IfChange + // provider is an inet socket provider. type provider struct { family int @@ -62,10 +64,6 @@ func getTransportProtocol(ctx context.Context, stype linux.SockType, protocol in } case linux.SOCK_RAW: - // TODO(b/142504697): "In order to create a raw socket, a - // process must have the CAP_NET_RAW capability in the user - // namespace that governs its network namespace." - raw(7) - // Raw sockets require CAP_NET_RAW. creds := auth.CredentialsFromContext(ctx) if !creds.HasCapability(linux.CAP_NET_RAW) { @@ -75,6 +73,8 @@ func getTransportProtocol(ctx context.Context, stype linux.SockType, protocol in switch protocol { case syscall.IPPROTO_ICMP: return header.ICMPv4ProtocolNumber, true, nil + case syscall.IPPROTO_ICMPV6: + return header.ICMPv6ProtocolNumber, true, nil case syscall.IPPROTO_UDP: return header.UDPProtocolNumber, true, nil case syscall.IPPROTO_TCP: @@ -124,6 +124,12 @@ func (p *provider) Socket(t *kernel.Task, stype linux.SockType, protocol int) (* ep, e = eps.Stack.NewRawEndpoint(transProto, p.netProto, wq, associated) } else { ep, e = eps.Stack.NewEndpoint(transProto, p.netProto, wq) + + // Assign task to PacketOwner interface to get the UID and GID for + // iptables owner matching. + if e == nil { + ep.SetOwner(t) + } } if e != nil { return nil, syserr.TranslateNetstackError(e) @@ -133,10 +139,6 @@ func (p *provider) Socket(t *kernel.Task, stype linux.SockType, protocol int) (* } func packetSocket(t *kernel.Task, epStack *Stack, stype linux.SockType, protocol int) (*fs.File, *syserr.Error) { - // TODO(b/142504697): "In order to create a packet socket, a process - // must have the CAP_NET_RAW capability in the user namespace that - // governs its network namespace." - packet(7) - // Packet sockets require CAP_NET_RAW. creds := auth.CredentialsFromContext(t) if !creds.HasCapability(linux.CAP_NET_RAW) { @@ -167,6 +169,8 @@ func packetSocket(t *kernel.Task, epStack *Stack, stype linux.SockType, protocol return New(t, linux.AF_PACKET, stype, protocol, wq, ep) } +// LINT.ThenChange(./provider_vfs2.go) + // Pair just returns nil sockets (not supported). func (*provider) Pair(*kernel.Task, linux.SockType, int) (*fs.File, *fs.File, *syserr.Error) { return nil, nil, nil diff --git a/pkg/sentry/socket/netstack/provider_vfs2.go b/pkg/sentry/socket/netstack/provider_vfs2.go new file mode 100644 index 000000000..2a01143f6 --- /dev/null +++ b/pkg/sentry/socket/netstack/provider_vfs2.go @@ -0,0 +1,141 @@ +// 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 netstack + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/socket" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserr" + "gvisor.dev/gvisor/pkg/tcpip" + "gvisor.dev/gvisor/pkg/tcpip/network/ipv4" + "gvisor.dev/gvisor/pkg/tcpip/network/ipv6" + "gvisor.dev/gvisor/pkg/waiter" +) + +// providerVFS2 is an inet socket provider. +type providerVFS2 struct { + family int + netProto tcpip.NetworkProtocolNumber +} + +// Socket creates a new socket object for the AF_INET, AF_INET6, or AF_PACKET +// family. +func (p *providerVFS2) Socket(t *kernel.Task, stype linux.SockType, protocol int) (*vfs.FileDescription, *syserr.Error) { + // Fail right away if we don't have a stack. + stack := t.NetworkContext() + if stack == nil { + // Don't propagate an error here. Instead, allow the socket + // code to continue searching for another provider. + return nil, nil + } + eps, ok := stack.(*Stack) + if !ok { + return nil, nil + } + + // Packet sockets are handled separately, since they are neither INET + // nor INET6 specific. + if p.family == linux.AF_PACKET { + return packetSocketVFS2(t, eps, stype, protocol) + } + + // Figure out the transport protocol. + transProto, associated, err := getTransportProtocol(t, stype, protocol) + if err != nil { + return nil, err + } + + // Create the endpoint. + var ep tcpip.Endpoint + var e *tcpip.Error + wq := &waiter.Queue{} + if stype == linux.SOCK_RAW { + ep, e = eps.Stack.NewRawEndpoint(transProto, p.netProto, wq, associated) + } else { + ep, e = eps.Stack.NewEndpoint(transProto, p.netProto, wq) + + // Assign task to PacketOwner interface to get the UID and GID for + // iptables owner matching. + if e == nil { + ep.SetOwner(t) + } + } + if e != nil { + return nil, syserr.TranslateNetstackError(e) + } + + return NewVFS2(t, p.family, stype, int(transProto), wq, ep) +} + +func packetSocketVFS2(t *kernel.Task, epStack *Stack, stype linux.SockType, protocol int) (*vfs.FileDescription, *syserr.Error) { + // Packet sockets require CAP_NET_RAW. + creds := auth.CredentialsFromContext(t) + if !creds.HasCapability(linux.CAP_NET_RAW) { + return nil, syserr.ErrNotPermitted + } + + // "cooked" packets don't contain link layer information. + var cooked bool + switch stype { + case linux.SOCK_DGRAM: + cooked = true + case linux.SOCK_RAW: + cooked = false + default: + return nil, syserr.ErrProtocolNotSupported + } + + // protocol is passed in network byte order, but netstack wants it in + // host order. + netProto := tcpip.NetworkProtocolNumber(ntohs(uint16(protocol))) + + wq := &waiter.Queue{} + ep, err := epStack.Stack.NewPacketEndpoint(cooked, netProto, wq) + if err != nil { + return nil, syserr.TranslateNetstackError(err) + } + + return NewVFS2(t, linux.AF_PACKET, stype, protocol, wq, ep) +} + +// Pair just returns nil sockets (not supported). +func (*providerVFS2) Pair(*kernel.Task, linux.SockType, int) (*vfs.FileDescription, *vfs.FileDescription, *syserr.Error) { + return nil, nil, nil +} + +// init registers socket providers for AF_INET, AF_INET6, and AF_PACKET. +func init() { + // Providers backed by netstack. + p := []providerVFS2{ + { + family: linux.AF_INET, + netProto: ipv4.ProtocolNumber, + }, + { + family: linux.AF_INET6, + netProto: ipv6.ProtocolNumber, + }, + { + family: linux.AF_PACKET, + }, + } + + for i := range p { + socket.RegisterProviderVFS2(p[i].family, &p[i]) + } +} diff --git a/pkg/sentry/socket/netstack/stack.go b/pkg/sentry/socket/netstack/stack.go index 0692482e9..9b44c2b89 100644 --- a/pkg/sentry/socket/netstack/stack.go +++ b/pkg/sentry/socket/netstack/stack.go @@ -23,7 +23,6 @@ import ( "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/tcpip" "gvisor.dev/gvisor/pkg/tcpip/header" - "gvisor.dev/gvisor/pkg/tcpip/iptables" "gvisor.dev/gvisor/pkg/tcpip/network/ipv4" "gvisor.dev/gvisor/pkg/tcpip/network/ipv6" "gvisor.dev/gvisor/pkg/tcpip/stack" @@ -200,36 +199,66 @@ func (s *Stack) SetTCPSACKEnabled(enabled bool) error { // Statistics implements inet.Stack.Statistics. func (s *Stack) Statistics(stat interface{}, arg string) error { switch stats := stat.(type) { + case *inet.StatDev: + for _, ni := range s.Stack.NICInfo() { + if ni.Name != arg { + continue + } + // TODO(gvisor.dev/issue/2103) Support stubbed stats. + *stats = inet.StatDev{ + // Receive section. + ni.Stats.Rx.Bytes.Value(), // bytes. + ni.Stats.Rx.Packets.Value(), // packets. + 0, // errs. + 0, // drop. + 0, // fifo. + 0, // frame. + 0, // compressed. + 0, // multicast. + // Transmit section. + ni.Stats.Tx.Bytes.Value(), // bytes. + ni.Stats.Tx.Packets.Value(), // packets. + 0, // errs. + 0, // drop. + 0, // fifo. + 0, // colls. + 0, // carrier. + 0, // compressed. + } + break + } case *inet.StatSNMPIP: ip := Metrics.IP + // TODO(gvisor.dev/issue/969) Support stubbed stats. *stats = inet.StatSNMPIP{ - 0, // TODO(gvisor.dev/issue/969): Support Ip/Forwarding. - 0, // TODO(gvisor.dev/issue/969): Support Ip/DefaultTTL. + 0, // Ip/Forwarding. + 0, // Ip/DefaultTTL. ip.PacketsReceived.Value(), // InReceives. - 0, // TODO(gvisor.dev/issue/969): Support Ip/InHdrErrors. + 0, // Ip/InHdrErrors. ip.InvalidDestinationAddressesReceived.Value(), // InAddrErrors. - 0, // TODO(gvisor.dev/issue/969): Support Ip/ForwDatagrams. - 0, // TODO(gvisor.dev/issue/969): Support Ip/InUnknownProtos. - 0, // TODO(gvisor.dev/issue/969): Support Ip/InDiscards. + 0, // Ip/ForwDatagrams. + 0, // Ip/InUnknownProtos. + 0, // Ip/InDiscards. ip.PacketsDelivered.Value(), // InDelivers. ip.PacketsSent.Value(), // OutRequests. ip.OutgoingPacketErrors.Value(), // OutDiscards. - 0, // TODO(gvisor.dev/issue/969): Support Ip/OutNoRoutes. - 0, // TODO(gvisor.dev/issue/969): Support Ip/ReasmTimeout. - 0, // TODO(gvisor.dev/issue/969): Support Ip/ReasmReqds. - 0, // TODO(gvisor.dev/issue/969): Support Ip/ReasmOKs. - 0, // TODO(gvisor.dev/issue/969): Support Ip/ReasmFails. - 0, // TODO(gvisor.dev/issue/969): Support Ip/FragOKs. - 0, // TODO(gvisor.dev/issue/969): Support Ip/FragFails. - 0, // TODO(gvisor.dev/issue/969): Support Ip/FragCreates. + 0, // Ip/OutNoRoutes. + 0, // Support Ip/ReasmTimeout. + 0, // Support Ip/ReasmReqds. + 0, // Support Ip/ReasmOKs. + 0, // Support Ip/ReasmFails. + 0, // Support Ip/FragOKs. + 0, // Support Ip/FragFails. + 0, // Support Ip/FragCreates. } case *inet.StatSNMPICMP: in := Metrics.ICMP.V4PacketsReceived.ICMPv4PacketStats out := Metrics.ICMP.V4PacketsSent.ICMPv4PacketStats + // TODO(gvisor.dev/issue/969) Support stubbed stats. *stats = inet.StatSNMPICMP{ - 0, // TODO(gvisor.dev/issue/969): Support Icmp/InMsgs. + 0, // Icmp/InMsgs. Metrics.ICMP.V4PacketsSent.Dropped.Value(), // InErrors. - 0, // TODO(gvisor.dev/issue/969): Support Icmp/InCsumErrors. + 0, // Icmp/InCsumErrors. in.DstUnreachable.Value(), // InDestUnreachs. in.TimeExceeded.Value(), // InTimeExcds. in.ParamProblem.Value(), // InParmProbs. @@ -241,7 +270,7 @@ func (s *Stack) Statistics(stat interface{}, arg string) error { in.TimestampReply.Value(), // InTimestampReps. in.InfoRequest.Value(), // InAddrMasks. in.InfoReply.Value(), // InAddrMaskReps. - 0, // TODO(gvisor.dev/issue/969): Support Icmp/OutMsgs. + 0, // Icmp/OutMsgs. Metrics.ICMP.V4PacketsReceived.Invalid.Value(), // OutErrors. out.DstUnreachable.Value(), // OutDestUnreachs. out.TimeExceeded.Value(), // OutTimeExcds. @@ -277,15 +306,16 @@ func (s *Stack) Statistics(stat interface{}, arg string) error { } case *inet.StatSNMPUDP: udp := Metrics.UDP + // TODO(gvisor.dev/issue/969) Support stubbed stats. *stats = inet.StatSNMPUDP{ udp.PacketsReceived.Value(), // InDatagrams. udp.UnknownPortErrors.Value(), // NoPorts. - 0, // TODO(gvisor.dev/issue/969): Support Udp/InErrors. + 0, // Udp/InErrors. udp.PacketsSent.Value(), // OutDatagrams. udp.ReceiveBufferErrors.Value(), // RcvbufErrors. - 0, // TODO(gvisor.dev/issue/969): Support Udp/SndbufErrors. - 0, // TODO(gvisor.dev/issue/969): Support Udp/InCsumErrors. - 0, // TODO(gvisor.dev/issue/969): Support Udp/IgnoredMulti. + 0, // Udp/SndbufErrors. + 0, // Udp/InCsumErrors. + 0, // Udp/IgnoredMulti. } default: return syserr.ErrEndpointOperation.ToError() @@ -332,14 +362,13 @@ func (s *Stack) RouteTable() []inet.Route { } // IPTables returns the stack's iptables. -func (s *Stack) IPTables() (iptables.IPTables, error) { +func (s *Stack) IPTables() (*stack.IPTables, error) { return s.Stack.IPTables(), nil } -// FillDefaultIPTables sets the stack's iptables to the default tables, which -// allow and do not modify all traffic. -func (s *Stack) FillDefaultIPTables() { - netfilter.FillDefaultIPTables(s.Stack) +// FillIPTablesMetadata populates stack's IPTables with metadata. +func (s *Stack) FillIPTablesMetadata() { + netfilter.FillIPTablesMetadata(s.Stack) } // Resume implements inet.Stack.Resume. diff --git a/pkg/sentry/socket/socket.go b/pkg/sentry/socket/socket.go index 50d9744e6..6580bd6e9 100644 --- a/pkg/sentry/socket/socket.go +++ b/pkg/sentry/socket/socket.go @@ -31,6 +31,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/kernel" ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" + "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserr" "gvisor.dev/gvisor/pkg/tcpip" "gvisor.dev/gvisor/pkg/usermem" @@ -48,11 +49,25 @@ func (c *ControlMessages) Release() { c.Unix.Release() } -// Socket is the interface containing socket syscalls used by the syscall layer -// to redirect them to the appropriate implementation. +// Socket is an interface combining fs.FileOperations and SocketOps, +// representing a VFS1 socket file. type Socket interface { fs.FileOperations + SocketOps +} + +// SocketVFS2 is an interface combining vfs.FileDescription and SocketOps, +// representing a VFS2 socket file. +type SocketVFS2 interface { + vfs.FileDescriptionImpl + SocketOps +} +// SocketOps is the interface containing socket syscalls used by the syscall +// layer to redirect them to the appropriate implementation. +// +// It is implemented by both Socket and SocketVFS2. +type SocketOps interface { // Connect implements the connect(2) linux syscall. Connect(t *kernel.Task, sockaddr []byte, blocking bool) *syserr.Error @@ -153,6 +168,8 @@ var families = make(map[int][]Provider) // RegisterProvider registers the provider of a given address family so that // sockets of that type can be created via socket() and/or socketpair() // syscalls. +// +// This should only be called during the initialization of the address family. func RegisterProvider(family int, provider Provider) { families[family] = append(families[family], provider) } @@ -216,6 +233,74 @@ func NewDirent(ctx context.Context, d *device.Device) *fs.Dirent { return fs.NewDirent(ctx, inode, fmt.Sprintf("socket:[%d]", ino)) } +// ProviderVFS2 is the vfs2 interface implemented by providers of sockets for +// specific address families (e.g., AF_INET). +type ProviderVFS2 interface { + // Socket creates a new socket. + // + // If a nil Socket _and_ a nil error is returned, it means that the + // protocol is not supported. A non-nil error should only be returned + // if the protocol is supported, but an error occurs during creation. + Socket(t *kernel.Task, stype linux.SockType, protocol int) (*vfs.FileDescription, *syserr.Error) + + // Pair creates a pair of connected sockets. + // + // See Socket for error information. + Pair(t *kernel.Task, stype linux.SockType, protocol int) (*vfs.FileDescription, *vfs.FileDescription, *syserr.Error) +} + +// familiesVFS2 holds a map of all known address families and their providers. +var familiesVFS2 = make(map[int][]ProviderVFS2) + +// RegisterProviderVFS2 registers the provider of a given address family so that +// sockets of that type can be created via socket() and/or socketpair() +// syscalls. +// +// This should only be called during the initialization of the address family. +func RegisterProviderVFS2(family int, provider ProviderVFS2) { + familiesVFS2[family] = append(familiesVFS2[family], provider) +} + +// NewVFS2 creates a new socket with the given family, type and protocol. +func NewVFS2(t *kernel.Task, family int, stype linux.SockType, protocol int) (*vfs.FileDescription, *syserr.Error) { + for _, p := range familiesVFS2[family] { + s, err := p.Socket(t, stype, protocol) + if err != nil { + return nil, err + } + if s != nil { + t.Kernel().RecordSocketVFS2(s) + return s, nil + } + } + + return nil, syserr.ErrAddressFamilyNotSupported +} + +// PairVFS2 creates a new connected socket pair with the given family, type and +// protocol. +func PairVFS2(t *kernel.Task, family int, stype linux.SockType, protocol int) (*vfs.FileDescription, *vfs.FileDescription, *syserr.Error) { + providers, ok := familiesVFS2[family] + if !ok { + return nil, nil, syserr.ErrAddressFamilyNotSupported + } + + for _, p := range providers { + s1, s2, err := p.Pair(t, stype, protocol) + if err != nil { + return nil, nil, err + } + if s1 != nil && s2 != nil { + k := t.Kernel() + k.RecordSocketVFS2(s1) + k.RecordSocketVFS2(s2) + return s1, s2, nil + } + } + + return nil, nil, syserr.ErrSocketNotSupported +} + // SendReceiveTimeout stores timeouts for send and receive calls. // // It is meant to be embedded into Socket implementations to help satisfy the diff --git a/pkg/sentry/socket/unix/BUILD b/pkg/sentry/socket/unix/BUILD index 08743deba..de2cc4bdf 100644 --- a/pkg/sentry/socket/unix/BUILD +++ b/pkg/sentry/socket/unix/BUILD @@ -8,23 +8,27 @@ go_library( "device.go", "io.go", "unix.go", + "unix_vfs2.go", ], visibility = ["//pkg/sentry:internal"], deps = [ "//pkg/abi/linux", "//pkg/context", + "//pkg/fspath", "//pkg/refs", "//pkg/safemem", "//pkg/sentry/arch", "//pkg/sentry/device", "//pkg/sentry/fs", "//pkg/sentry/fs/fsutil", + "//pkg/sentry/fsimpl/sockfs", "//pkg/sentry/kernel", "//pkg/sentry/kernel/time", "//pkg/sentry/socket", "//pkg/sentry/socket/control", "//pkg/sentry/socket/netstack", "//pkg/sentry/socket/unix/transport", + "//pkg/sentry/vfs", "//pkg/syserr", "//pkg/syserror", "//pkg/tcpip", diff --git a/pkg/sentry/socket/unix/transport/BUILD b/pkg/sentry/socket/unix/transport/BUILD index 74bcd6300..c708b6030 100644 --- a/pkg/sentry/socket/unix/transport/BUILD +++ b/pkg/sentry/socket/unix/transport/BUILD @@ -30,6 +30,7 @@ go_library( "//pkg/abi/linux", "//pkg/context", "//pkg/ilist", + "//pkg/log", "//pkg/refs", "//pkg/sync", "//pkg/syserr", diff --git a/pkg/sentry/socket/unix/transport/connectioned.go b/pkg/sentry/socket/unix/transport/connectioned.go index ce5b94ee7..09c6d3b27 100644 --- a/pkg/sentry/socket/unix/transport/connectioned.go +++ b/pkg/sentry/socket/unix/transport/connectioned.go @@ -252,7 +252,7 @@ func (e *connectionedEndpoint) Close() { // BidirectionalConnect implements BoundEndpoint.BidirectionalConnect. func (e *connectionedEndpoint) BidirectionalConnect(ctx context.Context, ce ConnectingEndpoint, returnConnect func(Receiver, ConnectedEndpoint)) *syserr.Error { if ce.Type() != e.stype { - return syserr.ErrConnectionRefused + return syserr.ErrWrongProtocolForSocket } // Check if ce is e to avoid a deadlock. diff --git a/pkg/sentry/socket/unix/transport/unix.go b/pkg/sentry/socket/unix/transport/unix.go index 2ef654235..2f1b127df 100644 --- a/pkg/sentry/socket/unix/transport/unix.go +++ b/pkg/sentry/socket/unix/transport/unix.go @@ -20,6 +20,7 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/log" "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/syserr" "gvisor.dev/gvisor/pkg/tcpip" @@ -838,24 +839,43 @@ func (e *baseEndpoint) SendMsg(ctx context.Context, data [][]byte, c ControlMess // SetSockOpt sets a socket option. Currently not supported. func (e *baseEndpoint) SetSockOpt(opt interface{}) *tcpip.Error { - switch v := opt.(type) { - case tcpip.PasscredOption: - e.setPasscred(v != 0) - return nil - } return nil } func (e *baseEndpoint) SetSockOptBool(opt tcpip.SockOptBool, v bool) *tcpip.Error { + switch opt { + case tcpip.BroadcastOption: + case tcpip.PasscredOption: + e.setPasscred(v) + case tcpip.ReuseAddressOption: + default: + log.Warningf("Unsupported socket option: %d", opt) + } return nil } func (e *baseEndpoint) SetSockOptInt(opt tcpip.SockOptInt, v int) *tcpip.Error { + switch opt { + case tcpip.SendBufferSizeOption: + case tcpip.ReceiveBufferSizeOption: + default: + log.Warningf("Unsupported socket option: %d", opt) + } return nil } func (e *baseEndpoint) GetSockOptBool(opt tcpip.SockOptBool) (bool, *tcpip.Error) { - return false, tcpip.ErrUnknownProtocolOption + switch opt { + case tcpip.KeepaliveEnabledOption: + return false, nil + + case tcpip.PasscredOption: + return e.Passcred(), nil + + default: + log.Warningf("Unsupported socket option: %d", opt) + return false, tcpip.ErrUnknownProtocolOption + } } func (e *baseEndpoint) GetSockOptInt(opt tcpip.SockOptInt) (int, *tcpip.Error) { @@ -914,29 +934,19 @@ func (e *baseEndpoint) GetSockOptInt(opt tcpip.SockOptInt) (int, *tcpip.Error) { return int(v), nil default: + log.Warningf("Unsupported socket option: %d", opt) return -1, tcpip.ErrUnknownProtocolOption } } // GetSockOpt implements tcpip.Endpoint.GetSockOpt. func (e *baseEndpoint) GetSockOpt(opt interface{}) *tcpip.Error { - switch o := opt.(type) { + switch opt.(type) { case tcpip.ErrorOption: return nil - case *tcpip.PasscredOption: - if e.Passcred() { - *o = tcpip.PasscredOption(1) - } else { - *o = tcpip.PasscredOption(0) - } - return nil - - case *tcpip.KeepaliveEnabledOption: - *o = 0 - return nil - default: + log.Warningf("Unsupported socket option: %T", opt) return tcpip.ErrUnknownProtocolOption } } diff --git a/pkg/sentry/socket/unix/unix.go b/pkg/sentry/socket/unix/unix.go index 4d30aa714..c4c9db81b 100644 --- a/pkg/sentry/socket/unix/unix.go +++ b/pkg/sentry/socket/unix/unix.go @@ -23,6 +23,7 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/fspath" "gvisor.dev/gvisor/pkg/refs" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/fs" @@ -33,6 +34,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/socket/control" "gvisor.dev/gvisor/pkg/sentry/socket/netstack" "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" + "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserr" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/tcpip" @@ -52,11 +54,8 @@ type SocketOperations struct { fsutil.FileNoSplice `state:"nosave"` fsutil.FileNoopFlush `state:"nosave"` fsutil.FileUseInodeUnstableAttr `state:"nosave"` - refs.AtomicRefCount - socket.SendReceiveTimeout - ep transport.Endpoint - stype linux.SockType + socketOpsCommon } // New creates a new unix socket. @@ -75,29 +74,42 @@ func NewWithDirent(ctx context.Context, d *fs.Dirent, ep transport.Endpoint, sty } s := SocketOperations{ - ep: ep, - stype: stype, + socketOpsCommon: socketOpsCommon{ + ep: ep, + stype: stype, + }, } s.EnableLeakCheck("unix.SocketOperations") return fs.NewFile(ctx, d, flags, &s) } +// socketOpsCommon contains the socket operations common to VFS1 and VFS2. +// +// +stateify savable +type socketOpsCommon struct { + refs.AtomicRefCount + socket.SendReceiveTimeout + + ep transport.Endpoint + stype linux.SockType +} + // DecRef implements RefCounter.DecRef. -func (s *SocketOperations) DecRef() { +func (s *socketOpsCommon) DecRef() { s.DecRefWithDestructor(func() { s.ep.Close() }) } // Release implemements fs.FileOperations.Release. -func (s *SocketOperations) Release() { +func (s *socketOpsCommon) Release() { // Release only decrements a reference on s because s may be referenced in // the abstract socket namespace. s.DecRef() } -func (s *SocketOperations) isPacket() bool { +func (s *socketOpsCommon) isPacket() bool { switch s.stype { case linux.SOCK_DGRAM, linux.SOCK_SEQPACKET: return true @@ -110,7 +122,7 @@ func (s *SocketOperations) isPacket() bool { } // Endpoint extracts the transport.Endpoint. -func (s *SocketOperations) Endpoint() transport.Endpoint { +func (s *socketOpsCommon) Endpoint() transport.Endpoint { return s.ep } @@ -143,7 +155,7 @@ func extractPath(sockaddr []byte) (string, *syserr.Error) { // GetPeerName implements the linux syscall getpeername(2) for sockets backed by // a transport.Endpoint. -func (s *SocketOperations) GetPeerName(t *kernel.Task) (linux.SockAddr, uint32, *syserr.Error) { +func (s *socketOpsCommon) GetPeerName(t *kernel.Task) (linux.SockAddr, uint32, *syserr.Error) { addr, err := s.ep.GetRemoteAddress() if err != nil { return nil, 0, syserr.TranslateNetstackError(err) @@ -155,7 +167,7 @@ func (s *SocketOperations) GetPeerName(t *kernel.Task) (linux.SockAddr, uint32, // GetSockName implements the linux syscall getsockname(2) for sockets backed by // a transport.Endpoint. -func (s *SocketOperations) GetSockName(t *kernel.Task) (linux.SockAddr, uint32, *syserr.Error) { +func (s *socketOpsCommon) GetSockName(t *kernel.Task) (linux.SockAddr, uint32, *syserr.Error) { addr, err := s.ep.GetLocalAddress() if err != nil { return nil, 0, syserr.TranslateNetstackError(err) @@ -178,7 +190,7 @@ func (s *SocketOperations) GetSockOpt(t *kernel.Task, level, name int, outPtr us // Listen implements the linux syscall listen(2) for sockets backed by // a transport.Endpoint. -func (s *SocketOperations) Listen(t *kernel.Task, backlog int) *syserr.Error { +func (s *socketOpsCommon) Listen(t *kernel.Task, backlog int) *syserr.Error { return s.ep.Listen(backlog) } @@ -310,6 +322,11 @@ func (s *SocketOperations) Bind(t *kernel.Task, sockaddr []byte) *syserr.Error { } // Create the socket. + // + // Note that the file permissions here are not set correctly (see + // gvisor.dev/issue/2324). There is no convenient way to get permissions + // on the socket referred to by s, so we will leave this discrepancy + // unresolved until VFS2 replaces this code. childDir, err := d.Bind(t, t.FSContext().RootDirectory(), name, bep, fs.FilePermissions{User: fs.PermMask{Read: true}}) if err != nil { return syserr.ErrPortInUse @@ -345,6 +362,31 @@ func extractEndpoint(t *kernel.Task, sockaddr []byte) (transport.BoundEndpoint, return ep, nil } + if kernel.VFS2Enabled { + p := fspath.Parse(path) + root := t.FSContext().RootDirectoryVFS2() + start := root + relPath := !p.Absolute + if relPath { + start = t.FSContext().WorkingDirectoryVFS2() + } + pop := vfs.PathOperation{ + Root: root, + Start: start, + Path: p, + FollowFinalSymlink: true, + } + ep, e := t.Kernel().VFS().BoundEndpointAt(t, t.Credentials(), &pop, &vfs.BoundEndpointOptions{path}) + root.DecRef() + if relPath { + start.DecRef() + } + if e != nil { + return nil, syserr.FromError(e) + } + return ep, nil + } + // Find the node in the filesystem. root := t.FSContext().RootDirectory() cwd := t.FSContext().WorkingDirectory() @@ -363,12 +405,11 @@ func extractEndpoint(t *kernel.Task, sockaddr []byte) (transport.BoundEndpoint, // No socket! return nil, syserr.ErrConnectionRefused } - return ep, nil } // Connect implements the linux syscall connect(2) for unix sockets. -func (s *SocketOperations) Connect(t *kernel.Task, sockaddr []byte, blocking bool) *syserr.Error { +func (s *socketOpsCommon) Connect(t *kernel.Task, sockaddr []byte, blocking bool) *syserr.Error { ep, err := extractEndpoint(t, sockaddr) if err != nil { return err @@ -376,10 +417,21 @@ func (s *SocketOperations) Connect(t *kernel.Task, sockaddr []byte, blocking boo defer ep.Release() // Connect the server endpoint. - return s.ep.Connect(t, ep) + err = s.ep.Connect(t, ep) + + if err == syserr.ErrWrongProtocolForSocket { + // Linux for abstract sockets returns ErrConnectionRefused + // instead of ErrWrongProtocolForSocket. + path, _ := extractPath(sockaddr) + if len(path) > 0 && path[0] == 0 { + err = syserr.ErrConnectionRefused + } + } + + return err } -// Writev implements fs.FileOperations.Write. +// Write implements fs.FileOperations.Write. func (s *SocketOperations) Write(ctx context.Context, _ *fs.File, src usermem.IOSequence, _ int64) (int64, error) { t := kernel.TaskFromContext(ctx) ctrl := control.New(t, s.ep, nil) @@ -399,7 +451,7 @@ func (s *SocketOperations) Write(ctx context.Context, _ *fs.File, src usermem.IO // SendMsg implements the linux syscall sendmsg(2) for unix sockets backed by // a transport.Endpoint. -func (s *SocketOperations) SendMsg(t *kernel.Task, src usermem.IOSequence, to []byte, flags int, haveDeadline bool, deadline ktime.Time, controlMessages socket.ControlMessages) (int, *syserr.Error) { +func (s *socketOpsCommon) SendMsg(t *kernel.Task, src usermem.IOSequence, to []byte, flags int, haveDeadline bool, deadline ktime.Time, controlMessages socket.ControlMessages) (int, *syserr.Error) { w := EndpointWriter{ Ctx: t, Endpoint: s.ep, @@ -453,27 +505,27 @@ func (s *SocketOperations) SendMsg(t *kernel.Task, src usermem.IOSequence, to [] } // Passcred implements transport.Credentialer.Passcred. -func (s *SocketOperations) Passcred() bool { +func (s *socketOpsCommon) Passcred() bool { return s.ep.Passcred() } // ConnectedPasscred implements transport.Credentialer.ConnectedPasscred. -func (s *SocketOperations) ConnectedPasscred() bool { +func (s *socketOpsCommon) ConnectedPasscred() bool { return s.ep.ConnectedPasscred() } // Readiness implements waiter.Waitable.Readiness. -func (s *SocketOperations) Readiness(mask waiter.EventMask) waiter.EventMask { +func (s *socketOpsCommon) Readiness(mask waiter.EventMask) waiter.EventMask { return s.ep.Readiness(mask) } // EventRegister implements waiter.Waitable.EventRegister. -func (s *SocketOperations) EventRegister(e *waiter.Entry, mask waiter.EventMask) { +func (s *socketOpsCommon) EventRegister(e *waiter.Entry, mask waiter.EventMask) { s.ep.EventRegister(e, mask) } // EventUnregister implements waiter.Waitable.EventUnregister. -func (s *SocketOperations) EventUnregister(e *waiter.Entry) { +func (s *socketOpsCommon) EventUnregister(e *waiter.Entry) { s.ep.EventUnregister(e) } @@ -485,7 +537,7 @@ func (s *SocketOperations) SetSockOpt(t *kernel.Task, level int, name int, optVa // Shutdown implements the linux syscall shutdown(2) for sockets backed by // a transport.Endpoint. -func (s *SocketOperations) Shutdown(t *kernel.Task, how int) *syserr.Error { +func (s *socketOpsCommon) Shutdown(t *kernel.Task, how int) *syserr.Error { f, err := netstack.ConvertShutdown(how) if err != nil { return err @@ -511,7 +563,7 @@ func (s *SocketOperations) Read(ctx context.Context, _ *fs.File, dst usermem.IOS // RecvMsg implements the linux syscall recvmsg(2) for sockets backed by // a transport.Endpoint. -func (s *SocketOperations) RecvMsg(t *kernel.Task, dst usermem.IOSequence, flags int, haveDeadline bool, deadline ktime.Time, senderRequested bool, controlDataLen uint64) (n int, msgFlags int, senderAddr linux.SockAddr, senderAddrLen uint32, controlMessages socket.ControlMessages, err *syserr.Error) { +func (s *socketOpsCommon) RecvMsg(t *kernel.Task, dst usermem.IOSequence, flags int, haveDeadline bool, deadline ktime.Time, senderRequested bool, controlDataLen uint64) (n int, msgFlags int, senderAddr linux.SockAddr, senderAddrLen uint32, controlMessages socket.ControlMessages, err *syserr.Error) { trunc := flags&linux.MSG_TRUNC != 0 peek := flags&linux.MSG_PEEK != 0 dontWait := flags&linux.MSG_DONTWAIT != 0 @@ -648,12 +700,12 @@ func (s *SocketOperations) RecvMsg(t *kernel.Task, dst usermem.IOSequence, flags } // State implements socket.Socket.State. -func (s *SocketOperations) State() uint32 { +func (s *socketOpsCommon) State() uint32 { return s.ep.State() } // Type implements socket.Socket.Type. -func (s *SocketOperations) Type() (family int, skType linux.SockType, protocol int) { +func (s *socketOpsCommon) Type() (family int, skType linux.SockType, protocol int) { // Unix domain sockets always have a protocol of 0. return linux.AF_UNIX, s.stype, 0 } @@ -706,4 +758,5 @@ func (*provider) Pair(t *kernel.Task, stype linux.SockType, protocol int) (*fs.F func init() { socket.RegisterProvider(linux.AF_UNIX, &provider{}) + socket.RegisterProviderVFS2(linux.AF_UNIX, &providerVFS2{}) } diff --git a/pkg/sentry/socket/unix/unix_vfs2.go b/pkg/sentry/socket/unix/unix_vfs2.go new file mode 100644 index 000000000..45e109361 --- /dev/null +++ b/pkg/sentry/socket/unix/unix_vfs2.go @@ -0,0 +1,358 @@ +// Copyright 2018 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package unix + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/sockfs" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/socket" + "gvisor.dev/gvisor/pkg/sentry/socket/control" + "gvisor.dev/gvisor/pkg/sentry/socket/netstack" + "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserr" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/tcpip" + "gvisor.dev/gvisor/pkg/usermem" + "gvisor.dev/gvisor/pkg/waiter" +) + +// SocketVFS2 implements socket.SocketVFS2 (and by extension, +// vfs.FileDescriptionImpl) for Unix sockets. +type SocketVFS2 struct { + vfsfd vfs.FileDescription + vfs.FileDescriptionDefaultImpl + vfs.DentryMetadataFileDescriptionImpl + + socketOpsCommon +} + +var _ = socket.SocketVFS2(&SocketVFS2{}) + +// NewSockfsFile creates a new socket file in the global sockfs mount and +// returns a corresponding file description. +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) + + fd, err := NewFileDescription(ep, stype, linux.O_RDWR, mnt, d) + if err != nil { + return nil, syserr.FromError(err) + } + return fd, nil +} + +// NewFileDescription creates and returns a socket file description +// corresponding to the given mount and dentry. +func NewFileDescription(ep transport.Endpoint, stype linux.SockType, flags uint32, mnt *vfs.Mount, d *vfs.Dentry) (*vfs.FileDescription, error) { + // You can create AF_UNIX, SOCK_RAW sockets. They're the same as + // SOCK_DGRAM and don't require CAP_NET_RAW. + if stype == linux.SOCK_RAW { + stype = linux.SOCK_DGRAM + } + + sock := &SocketVFS2{ + socketOpsCommon: socketOpsCommon{ + ep: ep, + stype: stype, + }, + } + vfsfd := &sock.vfsfd + if err := vfsfd.Init(sock, flags, mnt, d, &vfs.FileDescriptionOptions{ + DenyPRead: true, + DenyPWrite: true, + UseDentryMetadata: true, + }); err != nil { + return nil, err + } + return vfsfd, nil +} + +// GetSockOpt implements the linux syscall getsockopt(2) for sockets backed by +// a transport.Endpoint. +func (s *SocketVFS2) GetSockOpt(t *kernel.Task, level int, name int, outPtr usermem.Addr, outLen int) (interface{}, *syserr.Error) { + return netstack.GetSockOpt(t, s, s.ep, linux.AF_UNIX, s.ep.Type(), level, name, outLen) +} + +// blockingAccept implements a blocking version of accept(2), that is, if no +// connections are ready to be accept, it will block until one becomes ready. +func (s *SocketVFS2) blockingAccept(t *kernel.Task) (transport.Endpoint, *syserr.Error) { + // Register for notifications. + e, ch := waiter.NewChannelEntry(nil) + s.socketOpsCommon.EventRegister(&e, waiter.EventIn) + defer s.socketOpsCommon.EventUnregister(&e) + + // Try to accept the connection; if it fails, then wait until we get a + // notification. + for { + if ep, err := s.ep.Accept(); err != syserr.ErrWouldBlock { + return ep, err + } + + if err := t.Block(ch); err != nil { + return nil, syserr.FromError(err) + } + } +} + +// Accept implements the linux syscall accept(2) for sockets backed by +// a transport.Endpoint. +func (s *SocketVFS2) Accept(t *kernel.Task, peerRequested bool, flags int, blocking bool) (int32, linux.SockAddr, uint32, *syserr.Error) { + // Issue the accept request to get the new endpoint. + ep, err := s.ep.Accept() + if err != nil { + if err != syserr.ErrWouldBlock || !blocking { + return 0, nil, 0, err + } + + var err *syserr.Error + ep, err = s.blockingAccept(t) + if err != nil { + return 0, nil, 0, err + } + } + + ns, err := NewSockfsFile(t, ep, s.stype) + if err != nil { + return 0, nil, 0, err + } + defer ns.DecRef() + + if flags&linux.SOCK_NONBLOCK != 0 { + ns.SetStatusFlags(t, t.Credentials(), linux.SOCK_NONBLOCK) + } + + var addr linux.SockAddr + var addrLen uint32 + if peerRequested { + // Get address of the peer. + var err *syserr.Error + addr, addrLen, err = ns.Impl().(*SocketVFS2).GetPeerName(t) + if err != nil { + return 0, nil, 0, err + } + } + + fd, e := t.NewFDFromVFS2(0, ns, kernel.FDFlags{ + CloseOnExec: flags&linux.SOCK_CLOEXEC != 0, + }) + if e != nil { + return 0, nil, 0, syserr.FromError(e) + } + + t.Kernel().RecordSocketVFS2(ns) + return fd, addr, addrLen, nil +} + +// Bind implements the linux syscall bind(2) for unix sockets. +func (s *SocketVFS2) Bind(t *kernel.Task, sockaddr []byte) *syserr.Error { + p, e := extractPath(sockaddr) + if e != nil { + return e + } + + bep, ok := s.ep.(transport.BoundEndpoint) + if !ok { + // This socket can't be bound. + return syserr.ErrInvalidArgument + } + + return s.ep.Bind(tcpip.FullAddress{Addr: tcpip.Address(p)}, func() *syserr.Error { + // Is it abstract? + if p[0] == 0 { + if t.IsNetworkNamespaced() { + return syserr.ErrInvalidEndpointState + } + if err := t.AbstractSockets().Bind(p[1:], bep, s); err != nil { + // syserr.ErrPortInUse corresponds to EADDRINUSE. + return syserr.ErrPortInUse + } + } else { + path := fspath.Parse(p) + root := t.FSContext().RootDirectoryVFS2() + defer root.DecRef() + start := root + relPath := !path.Absolute + if relPath { + start = t.FSContext().WorkingDirectoryVFS2() + defer start.DecRef() + } + pop := vfs.PathOperation{ + Root: root, + Start: start, + Path: path, + } + stat, err := s.vfsfd.Stat(t, vfs.StatOptions{Mask: linux.STATX_MODE}) + if err != nil { + return syserr.FromError(err) + } + err = t.Kernel().VFS().MknodAt(t, t.Credentials(), &pop, &vfs.MknodOptions{ + // File permissions correspond to net/unix/af_unix.c:unix_bind. + Mode: linux.FileMode(linux.S_IFSOCK | uint(stat.Mode)&^t.FSContext().Umask()), + Endpoint: bep, + }) + if err == syserror.EEXIST { + return syserr.ErrAddressInUse + } + return syserr.FromError(err) + } + + return nil + }) +} + +// Ioctl implements vfs.FileDescriptionImpl. +func (s *SocketVFS2) Ioctl(ctx context.Context, uio usermem.IO, args arch.SyscallArguments) (uintptr, error) { + return netstack.Ioctl(ctx, s.ep, uio, args) +} + +// PRead implements vfs.FileDescriptionImpl. +func (s *SocketVFS2) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { + return 0, syserror.ESPIPE +} + +// Read implements vfs.FileDescriptionImpl. +func (s *SocketVFS2) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { + // All flags other than RWF_NOWAIT should be ignored. + // TODO(gvisor.dev/issue/2601): Support RWF_NOWAIT. + if opts.Flags != 0 { + return 0, syserror.EOPNOTSUPP + } + + if dst.NumBytes() == 0 { + return 0, nil + } + return dst.CopyOutFrom(ctx, &EndpointReader{ + Ctx: ctx, + Endpoint: s.ep, + NumRights: 0, + Peek: false, + From: nil, + }) +} + +// PWrite implements vfs.FileDescriptionImpl. +func (s *SocketVFS2) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) { + return 0, syserror.ESPIPE +} + +// Write implements vfs.FileDescriptionImpl. +func (s *SocketVFS2) Write(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) { + // All flags other than RWF_NOWAIT should be ignored. + // TODO(gvisor.dev/issue/2601): Support RWF_NOWAIT. + if opts.Flags != 0 { + return 0, syserror.EOPNOTSUPP + } + + t := kernel.TaskFromContext(ctx) + ctrl := control.New(t, s.ep, nil) + + if src.NumBytes() == 0 { + nInt, err := s.ep.SendMsg(ctx, [][]byte{}, ctrl, nil) + return int64(nInt), err.ToError() + } + + return src.CopyInTo(ctx, &EndpointWriter{ + Ctx: ctx, + Endpoint: s.ep, + Control: ctrl, + To: nil, + }) +} + +// Readiness implements waiter.Waitable.Readiness. +func (s *SocketVFS2) Readiness(mask waiter.EventMask) waiter.EventMask { + return s.socketOpsCommon.Readiness(mask) +} + +// EventRegister implements waiter.Waitable.EventRegister. +func (s *SocketVFS2) EventRegister(e *waiter.Entry, mask waiter.EventMask) { + s.socketOpsCommon.EventRegister(e, mask) +} + +// EventUnregister implements waiter.Waitable.EventUnregister. +func (s *SocketVFS2) EventUnregister(e *waiter.Entry) { + s.socketOpsCommon.EventUnregister(e) +} + +// SetSockOpt implements the linux syscall setsockopt(2) for sockets backed by +// a transport.Endpoint. +func (s *SocketVFS2) SetSockOpt(t *kernel.Task, level int, name int, optVal []byte) *syserr.Error { + return netstack.SetSockOpt(t, s, s.ep, level, name, optVal) +} + +// providerVFS2 is a unix domain socket provider for VFS2. +type providerVFS2 struct{} + +func (*providerVFS2) Socket(t *kernel.Task, stype linux.SockType, protocol int) (*vfs.FileDescription, *syserr.Error) { + // Check arguments. + if protocol != 0 && protocol != linux.AF_UNIX /* PF_UNIX */ { + return nil, syserr.ErrProtocolNotSupported + } + + // Create the endpoint and socket. + var ep transport.Endpoint + switch stype { + case linux.SOCK_DGRAM, linux.SOCK_RAW: + ep = transport.NewConnectionless(t) + case linux.SOCK_SEQPACKET, linux.SOCK_STREAM: + ep = transport.NewConnectioned(t, stype, t.Kernel()) + default: + return nil, syserr.ErrInvalidArgument + } + + f, err := NewSockfsFile(t, ep, stype) + if err != nil { + ep.Close() + return nil, err + } + return f, nil +} + +// Pair creates a new pair of AF_UNIX connected sockets. +func (*providerVFS2) Pair(t *kernel.Task, stype linux.SockType, protocol int) (*vfs.FileDescription, *vfs.FileDescription, *syserr.Error) { + // Check arguments. + if protocol != 0 && protocol != linux.AF_UNIX /* PF_UNIX */ { + return nil, nil, syserr.ErrProtocolNotSupported + } + + switch stype { + case linux.SOCK_STREAM, linux.SOCK_DGRAM, linux.SOCK_SEQPACKET, linux.SOCK_RAW: + // Ok + default: + return nil, nil, syserr.ErrInvalidArgument + } + + // Create the endpoints and sockets. + ep1, ep2 := transport.NewPair(t, stype, t.Kernel()) + s1, err := NewSockfsFile(t, ep1, stype) + if err != nil { + ep1.Close() + ep2.Close() + return nil, nil, err + } + s2, err := NewSockfsFile(t, ep2, stype) + if err != nil { + s1.DecRef() + ep2.Close() + return nil, nil, err + } + + return s1, s2, nil +} diff --git a/pkg/sentry/strace/BUILD b/pkg/sentry/strace/BUILD index 2f39a6f2b..88d5db9fc 100644 --- a/pkg/sentry/strace/BUILD +++ b/pkg/sentry/strace/BUILD @@ -7,6 +7,7 @@ go_library( srcs = [ "capability.go", "clone.go", + "epoll.go", "futex.go", "linux64_amd64.go", "linux64_arm64.go", diff --git a/pkg/sentry/strace/epoll.go b/pkg/sentry/strace/epoll.go new file mode 100644 index 000000000..a6e48b836 --- /dev/null +++ b/pkg/sentry/strace/epoll.go @@ -0,0 +1,89 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package strace + +import ( + "fmt" + "strings" + + "gvisor.dev/gvisor/pkg/abi" + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/usermem" +) + +func epollEvent(t *kernel.Task, eventAddr usermem.Addr) string { + var e linux.EpollEvent + if _, err := t.CopyIn(eventAddr, &e); err != nil { + return fmt.Sprintf("%#x {error reading event: %v}", eventAddr, err) + } + var sb strings.Builder + fmt.Fprintf(&sb, "%#x ", eventAddr) + writeEpollEvent(&sb, e) + return sb.String() +} + +func epollEvents(t *kernel.Task, eventsAddr usermem.Addr, numEvents, maxBytes uint64) string { + var sb strings.Builder + fmt.Fprintf(&sb, "%#x {", eventsAddr) + addr := eventsAddr + for i := uint64(0); i < numEvents; i++ { + var e linux.EpollEvent + if _, err := t.CopyIn(addr, &e); err != nil { + fmt.Fprintf(&sb, "{error reading event at %#x: %v}", addr, err) + continue + } + writeEpollEvent(&sb, e) + if uint64(sb.Len()) >= maxBytes { + sb.WriteString("...") + break + } + if _, ok := addr.AddLength(uint64(linux.SizeOfEpollEvent)); !ok { + fmt.Fprintf(&sb, "{error reading event at %#x: EFAULT}", addr) + continue + } + } + sb.WriteString("}") + return sb.String() +} + +func writeEpollEvent(sb *strings.Builder, e linux.EpollEvent) { + events := epollEventEvents.Parse(uint64(e.Events)) + fmt.Fprintf(sb, "{events=%s data=[%#x, %#x]}", events, e.Data[0], e.Data[1]) +} + +var epollCtlOps = abi.ValueSet{ + linux.EPOLL_CTL_ADD: "EPOLL_CTL_ADD", + linux.EPOLL_CTL_DEL: "EPOLL_CTL_DEL", + linux.EPOLL_CTL_MOD: "EPOLL_CTL_MOD", +} + +var epollEventEvents = abi.FlagSet{ + {Flag: linux.EPOLLIN, Name: "EPOLLIN"}, + {Flag: linux.EPOLLPRI, Name: "EPOLLPRI"}, + {Flag: linux.EPOLLOUT, Name: "EPOLLOUT"}, + {Flag: linux.EPOLLERR, Name: "EPOLLERR"}, + {Flag: linux.EPOLLHUP, Name: "EPULLHUP"}, + {Flag: linux.EPOLLRDNORM, Name: "EPOLLRDNORM"}, + {Flag: linux.EPOLLRDBAND, Name: "EPOLLRDBAND"}, + {Flag: linux.EPOLLWRNORM, Name: "EPOLLWRNORM"}, + {Flag: linux.EPOLLWRBAND, Name: "EPOLLWRBAND"}, + {Flag: linux.EPOLLMSG, Name: "EPOLLMSG"}, + {Flag: linux.EPOLLRDHUP, Name: "EPOLLRDHUP"}, + {Flag: linux.EPOLLEXCLUSIVE, Name: "EPOLLEXCLUSIVE"}, + {Flag: linux.EPOLLWAKEUP, Name: "EPOLLWAKEUP"}, + {Flag: linux.EPOLLONESHOT, Name: "EPOLLONESHOT"}, + {Flag: linux.EPOLLET, Name: "EPOLLET"}, +} diff --git a/pkg/sentry/strace/linux64_amd64.go b/pkg/sentry/strace/linux64_amd64.go index a4de545e9..71b92eaee 100644 --- a/pkg/sentry/strace/linux64_amd64.go +++ b/pkg/sentry/strace/linux64_amd64.go @@ -256,8 +256,8 @@ var linuxAMD64 = SyscallMap{ 229: makeSyscallInfo("clock_getres", Hex, PostTimespec), 230: makeSyscallInfo("clock_nanosleep", Hex, Hex, Timespec, PostTimespec), 231: makeSyscallInfo("exit_group", Hex), - 232: makeSyscallInfo("epoll_wait", Hex, Hex, Hex, Hex), - 233: makeSyscallInfo("epoll_ctl", Hex, Hex, FD, Hex), + 232: makeSyscallInfo("epoll_wait", FD, EpollEvents, Hex, Hex), + 233: makeSyscallInfo("epoll_ctl", FD, EpollCtlOp, FD, EpollEvent), 234: makeSyscallInfo("tgkill", Hex, Hex, Signal), 235: makeSyscallInfo("utimes", Path, Timeval), // 236: vserver (not implemented in the Linux kernel) @@ -305,7 +305,7 @@ var linuxAMD64 = SyscallMap{ 278: makeSyscallInfo("vmsplice", FD, Hex, Hex, Hex), 279: makeSyscallInfo("move_pages", Hex, Hex, Hex, Hex, Hex, Hex), 280: makeSyscallInfo("utimensat", FD, Path, UTimeTimespec, Hex), - 281: makeSyscallInfo("epoll_pwait", Hex, Hex, Hex, Hex, SigSet, Hex), + 281: makeSyscallInfo("epoll_pwait", FD, EpollEvents, Hex, Hex, SigSet, Hex), 282: makeSyscallInfo("signalfd", Hex, Hex, Hex), 283: makeSyscallInfo("timerfd_create", Hex, Hex), 284: makeSyscallInfo("eventfd", Hex), diff --git a/pkg/sentry/strace/linux64_arm64.go b/pkg/sentry/strace/linux64_arm64.go index 8bc38545f..bd7361a52 100644 --- a/pkg/sentry/strace/linux64_arm64.go +++ b/pkg/sentry/strace/linux64_arm64.go @@ -45,8 +45,8 @@ var linuxARM64 = SyscallMap{ 18: makeSyscallInfo("lookup_dcookie", Hex, Hex, Hex), 19: makeSyscallInfo("eventfd2", Hex, Hex), 20: makeSyscallInfo("epoll_create1", Hex), - 21: makeSyscallInfo("epoll_ctl", Hex, Hex, FD, Hex), - 22: makeSyscallInfo("epoll_pwait", Hex, Hex, Hex, Hex, SigSet, Hex), + 21: makeSyscallInfo("epoll_ctl", FD, EpollCtlOp, FD, EpollEvent), + 22: makeSyscallInfo("epoll_pwait", FD, EpollEvents, Hex, Hex, SigSet, Hex), 23: makeSyscallInfo("dup", FD), 24: makeSyscallInfo("dup3", FD, FD, Hex), 25: makeSyscallInfo("fcntl", FD, Hex, Hex), diff --git a/pkg/sentry/strace/socket.go b/pkg/sentry/strace/socket.go index 51e6d81b2..c0512de89 100644 --- a/pkg/sentry/strace/socket.go +++ b/pkg/sentry/strace/socket.go @@ -632,4 +632,13 @@ var sockOptNames = map[uint64]abi.ValueSet{ linux.MCAST_MSFILTER: "MCAST_MSFILTER", linux.IPV6_ADDRFORM: "IPV6_ADDRFORM", }, + linux.SOL_NETLINK: { + linux.NETLINK_BROADCAST_ERROR: "NETLINK_BROADCAST_ERROR", + linux.NETLINK_CAP_ACK: "NETLINK_CAP_ACK", + linux.NETLINK_DUMP_STRICT_CHK: "NETLINK_DUMP_STRICT_CHK", + linux.NETLINK_EXT_ACK: "NETLINK_EXT_ACK", + linux.NETLINK_LIST_MEMBERSHIPS: "NETLINK_LIST_MEMBERSHIPS", + linux.NETLINK_NO_ENOBUFS: "NETLINK_NO_ENOBUFS", + linux.NETLINK_PKTINFO: "NETLINK_PKTINFO", + }, } diff --git a/pkg/sentry/strace/strace.go b/pkg/sentry/strace/strace.go index a796b2396..68ca537c8 100644 --- a/pkg/sentry/strace/strace.go +++ b/pkg/sentry/strace/strace.go @@ -141,6 +141,10 @@ func path(t *kernel.Task, addr usermem.Addr) string { } func fd(t *kernel.Task, fd int32) string { + if kernel.VFS2Enabled { + return fdVFS2(t, fd) + } + root := t.FSContext().RootDirectory() if root != nil { defer root.DecRef() @@ -169,6 +173,30 @@ func fd(t *kernel.Task, fd int32) string { return fmt.Sprintf("%#x %s", fd, name) } +func fdVFS2(t *kernel.Task, fd int32) string { + root := t.FSContext().RootDirectoryVFS2() + defer root.DecRef() + + vfsObj := root.Mount().Filesystem().VirtualFilesystem() + if fd == linux.AT_FDCWD { + wd := t.FSContext().WorkingDirectoryVFS2() + defer wd.DecRef() + + name, _ := vfsObj.PathnameWithDeleted(t, root, wd) + return fmt.Sprintf("AT_FDCWD %s", name) + } + + file := t.GetFileVFS2(fd) + if file == nil { + // Cast FD to uint64 to avoid printing negative hex. + return fmt.Sprintf("%#x (bad FD)", uint64(fd)) + } + defer file.DecRef() + + name, _ := vfsObj.PathnameWithDeleted(t, root, file.VirtualDentry()) + return fmt.Sprintf("%#x %s", fd, name) +} + func fdpair(t *kernel.Task, addr usermem.Addr) string { var fds [2]int32 _, err := t.CopyIn(addr, &fds) @@ -453,6 +481,12 @@ func (i *SyscallInfo) pre(t *kernel.Task, args arch.SyscallArguments, maximumBlo output = append(output, capData(t, args[arg-1].Pointer(), args[arg].Pointer())) case PollFDs: output = append(output, pollFDs(t, args[arg].Pointer(), uint(args[arg+1].Uint()), false)) + case EpollCtlOp: + output = append(output, epollCtlOps.Parse(uint64(args[arg].Int()))) + case EpollEvent: + output = append(output, epollEvent(t, args[arg].Pointer())) + case EpollEvents: + output = append(output, epollEvents(t, args[arg].Pointer(), 0 /* numEvents */, uint64(maximumBlobSize))) case SelectFDSet: output = append(output, fdSet(t, int(args[0].Int()), args[arg].Pointer())) case Oct: @@ -521,6 +555,8 @@ func (i *SyscallInfo) post(t *kernel.Task, args arch.SyscallArguments, rval uint output[arg] = capData(t, args[arg-1].Pointer(), args[arg].Pointer()) case PollFDs: output[arg] = pollFDs(t, args[arg].Pointer(), uint(args[arg+1].Uint()), true) + case EpollEvents: + output[arg] = epollEvents(t, args[arg].Pointer(), uint64(rval), uint64(maximumBlobSize)) case GetSockOptVal: output[arg] = getSockOptVal(t, args[arg-2].Uint64() /* level */, args[arg-1].Uint64() /* optName */, args[arg].Pointer() /* optVal */, args[arg+1].Pointer() /* optLen */, maximumBlobSize, rval) case SetSockOptVal: @@ -683,7 +719,7 @@ func (s SyscallMap) SyscallEnter(t *kernel.Task, sysno uintptr, args arch.Syscal // SyscallExit implements kernel.Stracer.SyscallExit. It logs the syscall // exit trace. func (s SyscallMap) SyscallExit(context interface{}, t *kernel.Task, sysno, rval uintptr, err error) { - errno := t.ExtractErrno(err, int(sysno)) + errno := kernel.ExtractErrno(err, int(sysno)) c := context.(*syscallContext) elapsed := time.Since(c.start) @@ -742,9 +778,6 @@ func (s SyscallMap) Name(sysno uintptr) string { // // N.B. This is not in an init function because we can't be sure all syscall // tables are registered with the kernel when init runs. -// -// TODO(gvisor.dev/issue/155): remove kernel package dependencies from this -// package and have the kernel package self-initialize all syscall tables. func Initialize() { for _, table := range kernel.SyscallTables() { // Is this known? diff --git a/pkg/sentry/strace/syscalls.go b/pkg/sentry/strace/syscalls.go index 446d1e0f6..7e69b9279 100644 --- a/pkg/sentry/strace/syscalls.go +++ b/pkg/sentry/strace/syscalls.go @@ -228,6 +228,16 @@ const ( // SockOptLevel is the optname argument in getsockopt(2) and // setsockopt(2). SockOptName + + // EpollCtlOp is the op argument to epoll_ctl(2). + EpollCtlOp + + // EpollEvent is the event argument in epoll_ctl(2). + EpollEvent + + // EpollEvents is an array of struct epoll_event. It is the events + // argument in epoll_wait(2)/epoll_pwait(2). + EpollEvents ) // defaultFormat is the syscall argument format to use if the actual format is diff --git a/pkg/sentry/syscalls/epoll.go b/pkg/sentry/syscalls/epoll.go index 87dcad18b..d9fb808c0 100644 --- a/pkg/sentry/syscalls/epoll.go +++ b/pkg/sentry/syscalls/epoll.go @@ -17,6 +17,7 @@ package syscalls import ( "time" + "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/kernel/epoll" ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" @@ -118,7 +119,7 @@ func RemoveEpoll(t *kernel.Task, epfd int32, fd int32) error { } // WaitEpoll implements the epoll_wait(2) linux syscall. -func WaitEpoll(t *kernel.Task, fd int32, max int, timeout int) ([]epoll.Event, error) { +func WaitEpoll(t *kernel.Task, fd int32, max int, timeout int) ([]linux.EpollEvent, error) { // Get epoll from the file descriptor. epollfile := t.GetFile(fd) if epollfile == nil { diff --git a/pkg/sentry/syscalls/linux/BUILD b/pkg/sentry/syscalls/linux/BUILD index be16ee686..217fcfef2 100644 --- a/pkg/sentry/syscalls/linux/BUILD +++ b/pkg/sentry/syscalls/linux/BUILD @@ -8,8 +8,6 @@ go_library( "error.go", "flags.go", "linux64.go", - "linux64_amd64.go", - "linux64_arm64.go", "sigset.go", "sys_aio.go", "sys_capability.go", @@ -51,7 +49,8 @@ go_library( "sys_time.go", "sys_timer.go", "sys_timerfd.go", - "sys_tls.go", + "sys_tls_amd64.go", + "sys_tls_arm64.go", "sys_utsname.go", "sys_write.go", "sys_xattr.go", @@ -74,6 +73,7 @@ go_library( "//pkg/sentry/fs/lock", "//pkg/sentry/fs/timerfd", "//pkg/sentry/fs/tmpfs", + "//pkg/sentry/fsbridge", "//pkg/sentry/kernel", "//pkg/sentry/kernel/auth", "//pkg/sentry/kernel/epoll", diff --git a/pkg/sentry/syscalls/linux/linux64.go b/pkg/sentry/syscalls/linux/linux64.go index 68589a377..ea4f9b1a7 100644 --- a/pkg/sentry/syscalls/linux/linux64.go +++ b/pkg/sentry/syscalls/linux/linux64.go @@ -15,6 +15,16 @@ // Package linux provides syscall tables for amd64 Linux. package linux +import ( + "gvisor.dev/gvisor/pkg/abi" + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/syscalls" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + const ( // LinuxSysname is the OS name advertised by gVisor. LinuxSysname = "Linux" @@ -25,3 +35,702 @@ const ( // LinuxVersion is the version info advertised by gVisor. LinuxVersion = "#1 SMP Sun Jan 10 15:06:54 PST 2016" ) + +// AMD64 is a table of Linux amd64 syscall API with the corresponding syscall +// numbers from Linux 4.4. +var AMD64 = &kernel.SyscallTable{ + OS: abi.Linux, + Arch: arch.AMD64, + Version: kernel.Version{ + // Version 4.4 is chosen as a stable, longterm version of Linux, which + // guides the interface provided by this syscall table. The build + // version is that for a clean build with default kernel config, at 5 + // minutes after v4.4 was tagged. + Sysname: LinuxSysname, + Release: LinuxRelease, + Version: LinuxVersion, + }, + AuditNumber: linux.AUDIT_ARCH_X86_64, + Table: map[uintptr]kernel.Syscall{ + 0: syscalls.Supported("read", Read), + 1: syscalls.Supported("write", Write), + 2: syscalls.PartiallySupported("open", Open, "Options O_DIRECT, O_NOATIME, O_PATH, O_TMPFILE, O_SYNC are not supported.", nil), + 3: syscalls.Supported("close", Close), + 4: syscalls.Supported("stat", Stat), + 5: syscalls.Supported("fstat", Fstat), + 6: syscalls.Supported("lstat", Lstat), + 7: syscalls.Supported("poll", Poll), + 8: syscalls.Supported("lseek", Lseek), + 9: syscalls.PartiallySupported("mmap", Mmap, "Generally supported with exceptions. Options MAP_FIXED_NOREPLACE, MAP_SHARED_VALIDATE, MAP_SYNC MAP_GROWSDOWN, MAP_HUGETLB are not supported.", nil), + 10: syscalls.Supported("mprotect", Mprotect), + 11: syscalls.Supported("munmap", Munmap), + 12: syscalls.Supported("brk", Brk), + 13: syscalls.Supported("rt_sigaction", RtSigaction), + 14: syscalls.Supported("rt_sigprocmask", RtSigprocmask), + 15: syscalls.Supported("rt_sigreturn", RtSigreturn), + 16: syscalls.PartiallySupported("ioctl", Ioctl, "Only a few ioctls are implemented for backing devices and file systems.", nil), + 17: syscalls.Supported("pread64", Pread64), + 18: syscalls.Supported("pwrite64", Pwrite64), + 19: syscalls.Supported("readv", Readv), + 20: syscalls.Supported("writev", Writev), + 21: syscalls.Supported("access", Access), + 22: syscalls.Supported("pipe", Pipe), + 23: syscalls.Supported("select", Select), + 24: syscalls.Supported("sched_yield", SchedYield), + 25: syscalls.Supported("mremap", Mremap), + 26: syscalls.PartiallySupported("msync", Msync, "Full data flush is not guaranteed at this time.", nil), + 27: syscalls.PartiallySupported("mincore", Mincore, "Stub implementation. The sandbox does not have access to this information. Reports all mapped pages are resident.", nil), + 28: syscalls.PartiallySupported("madvise", Madvise, "Options MADV_DONTNEED, MADV_DONTFORK are supported. Other advice is ignored.", nil), + 29: syscalls.PartiallySupported("shmget", Shmget, "Option SHM_HUGETLB is not supported.", nil), + 30: syscalls.PartiallySupported("shmat", Shmat, "Option SHM_RND is not supported.", nil), + 31: syscalls.PartiallySupported("shmctl", Shmctl, "Options SHM_LOCK, SHM_UNLOCK are not supported.", nil), + 32: syscalls.Supported("dup", Dup), + 33: syscalls.Supported("dup2", Dup2), + 34: syscalls.Supported("pause", Pause), + 35: syscalls.Supported("nanosleep", Nanosleep), + 36: syscalls.Supported("getitimer", Getitimer), + 37: syscalls.Supported("alarm", Alarm), + 38: syscalls.Supported("setitimer", Setitimer), + 39: syscalls.Supported("getpid", Getpid), + 40: syscalls.Supported("sendfile", Sendfile), + 41: syscalls.PartiallySupported("socket", Socket, "Limited support for AF_NETLINK, NETLINK_ROUTE sockets. Limited support for SOCK_RAW.", nil), + 42: syscalls.Supported("connect", Connect), + 43: syscalls.Supported("accept", Accept), + 44: syscalls.Supported("sendto", SendTo), + 45: syscalls.Supported("recvfrom", RecvFrom), + 46: syscalls.Supported("sendmsg", SendMsg), + 47: syscalls.PartiallySupported("recvmsg", RecvMsg, "Not all flags and control messages are supported.", nil), + 48: syscalls.PartiallySupported("shutdown", Shutdown, "Not all flags and control messages are supported.", nil), + 49: syscalls.PartiallySupported("bind", Bind, "Autobind for abstract Unix sockets is not supported.", nil), + 50: syscalls.Supported("listen", Listen), + 51: syscalls.Supported("getsockname", GetSockName), + 52: syscalls.Supported("getpeername", GetPeerName), + 53: syscalls.Supported("socketpair", SocketPair), + 54: syscalls.PartiallySupported("setsockopt", SetSockOpt, "Not all socket options are supported.", nil), + 55: syscalls.PartiallySupported("getsockopt", GetSockOpt, "Not all socket options are supported.", nil), + 56: syscalls.PartiallySupported("clone", Clone, "Mount namespace (CLONE_NEWNS) not supported. Options CLONE_PARENT, CLONE_SYSVSEM not supported.", nil), + 57: syscalls.Supported("fork", Fork), + 58: syscalls.Supported("vfork", Vfork), + 59: syscalls.Supported("execve", Execve), + 60: syscalls.Supported("exit", Exit), + 61: syscalls.Supported("wait4", Wait4), + 62: syscalls.Supported("kill", Kill), + 63: syscalls.Supported("uname", Uname), + 64: syscalls.Supported("semget", Semget), + 65: syscalls.PartiallySupported("semop", Semop, "Option SEM_UNDO not supported.", nil), + 66: syscalls.PartiallySupported("semctl", Semctl, "Options IPC_INFO, SEM_INFO, IPC_STAT, SEM_STAT, SEM_STAT_ANY, GETNCNT, GETZCNT not supported.", nil), + 67: syscalls.Supported("shmdt", Shmdt), + 68: syscalls.ErrorWithEvent("msgget", syserror.ENOSYS, "", []string{"gvisor.dev/issue/135"}), // TODO(b/29354921) + 69: syscalls.ErrorWithEvent("msgsnd", syserror.ENOSYS, "", []string{"gvisor.dev/issue/135"}), // TODO(b/29354921) + 70: syscalls.ErrorWithEvent("msgrcv", syserror.ENOSYS, "", []string{"gvisor.dev/issue/135"}), // TODO(b/29354921) + 71: syscalls.ErrorWithEvent("msgctl", syserror.ENOSYS, "", []string{"gvisor.dev/issue/135"}), // TODO(b/29354921) + 72: syscalls.PartiallySupported("fcntl", Fcntl, "Not all options are supported.", nil), + 73: syscalls.PartiallySupported("flock", Flock, "Locks are held within the sandbox only.", nil), + 74: syscalls.PartiallySupported("fsync", Fsync, "Full data flush is not guaranteed at this time.", nil), + 75: syscalls.PartiallySupported("fdatasync", Fdatasync, "Full data flush is not guaranteed at this time.", nil), + 76: syscalls.Supported("truncate", Truncate), + 77: syscalls.Supported("ftruncate", Ftruncate), + 78: syscalls.Supported("getdents", Getdents), + 79: syscalls.Supported("getcwd", Getcwd), + 80: syscalls.Supported("chdir", Chdir), + 81: syscalls.Supported("fchdir", Fchdir), + 82: syscalls.Supported("rename", Rename), + 83: syscalls.Supported("mkdir", Mkdir), + 84: syscalls.Supported("rmdir", Rmdir), + 85: syscalls.Supported("creat", Creat), + 86: syscalls.Supported("link", Link), + 87: syscalls.Supported("unlink", Unlink), + 88: syscalls.Supported("symlink", Symlink), + 89: syscalls.Supported("readlink", Readlink), + 90: syscalls.Supported("chmod", Chmod), + 91: syscalls.PartiallySupported("fchmod", Fchmod, "Options S_ISUID and S_ISGID not supported.", nil), + 92: syscalls.Supported("chown", Chown), + 93: syscalls.Supported("fchown", Fchown), + 94: syscalls.Supported("lchown", Lchown), + 95: syscalls.Supported("umask", Umask), + 96: syscalls.Supported("gettimeofday", Gettimeofday), + 97: syscalls.Supported("getrlimit", Getrlimit), + 98: syscalls.PartiallySupported("getrusage", Getrusage, "Fields ru_maxrss, ru_minflt, ru_majflt, ru_inblock, ru_oublock are not supported. Fields ru_utime and ru_stime have low precision.", nil), + 99: syscalls.PartiallySupported("sysinfo", Sysinfo, "Fields loads, sharedram, bufferram, totalswap, freeswap, totalhigh, freehigh not supported.", nil), + 100: syscalls.Supported("times", Times), + 101: syscalls.PartiallySupported("ptrace", Ptrace, "Options PTRACE_PEEKSIGINFO, PTRACE_SECCOMP_GET_FILTER not supported.", nil), + 102: syscalls.Supported("getuid", Getuid), + 103: syscalls.PartiallySupported("syslog", Syslog, "Outputs a dummy message for security reasons.", nil), + 104: syscalls.Supported("getgid", Getgid), + 105: syscalls.Supported("setuid", Setuid), + 106: syscalls.Supported("setgid", Setgid), + 107: syscalls.Supported("geteuid", Geteuid), + 108: syscalls.Supported("getegid", Getegid), + 109: syscalls.Supported("setpgid", Setpgid), + 110: syscalls.Supported("getppid", Getppid), + 111: syscalls.Supported("getpgrp", Getpgrp), + 112: syscalls.Supported("setsid", Setsid), + 113: syscalls.Supported("setreuid", Setreuid), + 114: syscalls.Supported("setregid", Setregid), + 115: syscalls.Supported("getgroups", Getgroups), + 116: syscalls.Supported("setgroups", Setgroups), + 117: syscalls.Supported("setresuid", Setresuid), + 118: syscalls.Supported("getresuid", Getresuid), + 119: syscalls.Supported("setresgid", Setresgid), + 120: syscalls.Supported("getresgid", Getresgid), + 121: syscalls.Supported("getpgid", Getpgid), + 122: syscalls.ErrorWithEvent("setfsuid", syserror.ENOSYS, "", []string{"gvisor.dev/issue/260"}), // TODO(b/112851702) + 123: syscalls.ErrorWithEvent("setfsgid", syserror.ENOSYS, "", []string{"gvisor.dev/issue/260"}), // TODO(b/112851702) + 124: syscalls.Supported("getsid", Getsid), + 125: syscalls.Supported("capget", Capget), + 126: syscalls.Supported("capset", Capset), + 127: syscalls.Supported("rt_sigpending", RtSigpending), + 128: syscalls.Supported("rt_sigtimedwait", RtSigtimedwait), + 129: syscalls.Supported("rt_sigqueueinfo", RtSigqueueinfo), + 130: syscalls.Supported("rt_sigsuspend", RtSigsuspend), + 131: syscalls.Supported("sigaltstack", Sigaltstack), + 132: syscalls.Supported("utime", Utime), + 133: syscalls.PartiallySupported("mknod", Mknod, "Device creation is not generally supported. Only regular file and FIFO creation are supported.", nil), + 134: syscalls.Error("uselib", syserror.ENOSYS, "Obsolete", nil), + 135: syscalls.ErrorWithEvent("personality", syserror.EINVAL, "Unable to change personality.", nil), + 136: syscalls.ErrorWithEvent("ustat", syserror.ENOSYS, "Needs filesystem support.", nil), + 137: syscalls.PartiallySupported("statfs", Statfs, "Depends on the backing file system implementation.", nil), + 138: syscalls.PartiallySupported("fstatfs", Fstatfs, "Depends on the backing file system implementation.", nil), + 139: syscalls.ErrorWithEvent("sysfs", syserror.ENOSYS, "", []string{"gvisor.dev/issue/165"}), + 140: syscalls.PartiallySupported("getpriority", Getpriority, "Stub implementation.", nil), + 141: syscalls.PartiallySupported("setpriority", Setpriority, "Stub implementation.", nil), + 142: syscalls.CapError("sched_setparam", linux.CAP_SYS_NICE, "", nil), + 143: syscalls.PartiallySupported("sched_getparam", SchedGetparam, "Stub implementation.", nil), + 144: syscalls.PartiallySupported("sched_setscheduler", SchedSetscheduler, "Stub implementation.", nil), + 145: syscalls.PartiallySupported("sched_getscheduler", SchedGetscheduler, "Stub implementation.", nil), + 146: syscalls.PartiallySupported("sched_get_priority_max", SchedGetPriorityMax, "Stub implementation.", nil), + 147: syscalls.PartiallySupported("sched_get_priority_min", SchedGetPriorityMin, "Stub implementation.", nil), + 148: syscalls.ErrorWithEvent("sched_rr_get_interval", syserror.EPERM, "", nil), + 149: syscalls.PartiallySupported("mlock", Mlock, "Stub implementation. The sandbox lacks appropriate permissions.", nil), + 150: syscalls.PartiallySupported("munlock", Munlock, "Stub implementation. The sandbox lacks appropriate permissions.", nil), + 151: syscalls.PartiallySupported("mlockall", Mlockall, "Stub implementation. The sandbox lacks appropriate permissions.", nil), + 152: syscalls.PartiallySupported("munlockall", Munlockall, "Stub implementation. The sandbox lacks appropriate permissions.", nil), + 153: syscalls.CapError("vhangup", linux.CAP_SYS_TTY_CONFIG, "", nil), + 154: syscalls.Error("modify_ldt", syserror.EPERM, "", nil), + 155: syscalls.Error("pivot_root", syserror.EPERM, "", nil), + 156: syscalls.Error("sysctl", syserror.EPERM, "Deprecated. Use /proc/sys instead.", nil), + 157: syscalls.PartiallySupported("prctl", Prctl, "Not all options are supported.", nil), + 158: syscalls.PartiallySupported("arch_prctl", ArchPrctl, "Options ARCH_GET_GS, ARCH_SET_GS not supported.", nil), + 159: syscalls.CapError("adjtimex", linux.CAP_SYS_TIME, "", nil), + 160: syscalls.PartiallySupported("setrlimit", Setrlimit, "Not all rlimits are enforced.", nil), + 161: syscalls.Supported("chroot", Chroot), + 162: syscalls.PartiallySupported("sync", Sync, "Full data flush is not guaranteed at this time.", nil), + 163: syscalls.CapError("acct", linux.CAP_SYS_PACCT, "", nil), + 164: syscalls.CapError("settimeofday", linux.CAP_SYS_TIME, "", nil), + 165: syscalls.PartiallySupported("mount", Mount, "Not all options or file systems are supported.", nil), + 166: syscalls.PartiallySupported("umount2", Umount2, "Not all options or file systems are supported.", nil), + 167: syscalls.CapError("swapon", linux.CAP_SYS_ADMIN, "", nil), + 168: syscalls.CapError("swapoff", linux.CAP_SYS_ADMIN, "", nil), + 169: syscalls.CapError("reboot", linux.CAP_SYS_BOOT, "", nil), + 170: syscalls.Supported("sethostname", Sethostname), + 171: syscalls.Supported("setdomainname", Setdomainname), + 172: syscalls.CapError("iopl", linux.CAP_SYS_RAWIO, "", nil), + 173: syscalls.CapError("ioperm", linux.CAP_SYS_RAWIO, "", nil), + 174: syscalls.CapError("create_module", linux.CAP_SYS_MODULE, "", nil), + 175: syscalls.CapError("init_module", linux.CAP_SYS_MODULE, "", nil), + 176: syscalls.CapError("delete_module", linux.CAP_SYS_MODULE, "", nil), + 177: syscalls.Error("get_kernel_syms", syserror.ENOSYS, "Not supported in Linux > 2.6.", nil), + 178: syscalls.Error("query_module", syserror.ENOSYS, "Not supported in Linux > 2.6.", nil), + 179: syscalls.CapError("quotactl", linux.CAP_SYS_ADMIN, "", nil), // requires cap_sys_admin for most operations + 180: syscalls.Error("nfsservctl", syserror.ENOSYS, "Removed after Linux 3.1.", nil), + 181: syscalls.Error("getpmsg", syserror.ENOSYS, "Not implemented in Linux.", nil), + 182: syscalls.Error("putpmsg", syserror.ENOSYS, "Not implemented in Linux.", nil), + 183: syscalls.Error("afs_syscall", syserror.ENOSYS, "Not implemented in Linux.", nil), + 184: syscalls.Error("tuxcall", syserror.ENOSYS, "Not implemented in Linux.", nil), + 185: syscalls.Error("security", syserror.ENOSYS, "Not implemented in Linux.", nil), + 186: syscalls.Supported("gettid", Gettid), + 187: syscalls.Supported("readahead", Readahead), + 188: syscalls.PartiallySupported("setxattr", SetXattr, "Only supported for tmpfs.", nil), + 189: syscalls.PartiallySupported("lsetxattr", LSetXattr, "Only supported for tmpfs.", nil), + 190: syscalls.PartiallySupported("fsetxattr", FSetXattr, "Only supported for tmpfs.", nil), + 191: syscalls.PartiallySupported("getxattr", GetXattr, "Only supported for tmpfs.", nil), + 192: syscalls.PartiallySupported("lgetxattr", LGetXattr, "Only supported for tmpfs.", nil), + 193: syscalls.PartiallySupported("fgetxattr", FGetXattr, "Only supported for tmpfs.", nil), + 194: syscalls.PartiallySupported("listxattr", ListXattr, "Only supported for tmpfs", nil), + 195: syscalls.PartiallySupported("llistxattr", LListXattr, "Only supported for tmpfs", nil), + 196: syscalls.PartiallySupported("flistxattr", FListXattr, "Only supported for tmpfs", nil), + 197: syscalls.PartiallySupported("removexattr", RemoveXattr, "Only supported for tmpfs", nil), + 198: syscalls.PartiallySupported("lremovexattr", LRemoveXattr, "Only supported for tmpfs", nil), + 199: syscalls.PartiallySupported("fremovexattr", FRemoveXattr, "Only supported for tmpfs", nil), + 200: syscalls.Supported("tkill", Tkill), + 201: syscalls.Supported("time", Time), + 202: syscalls.PartiallySupported("futex", Futex, "Robust futexes not supported.", nil), + 203: syscalls.PartiallySupported("sched_setaffinity", SchedSetaffinity, "Stub implementation.", nil), + 204: syscalls.PartiallySupported("sched_getaffinity", SchedGetaffinity, "Stub implementation.", nil), + 205: syscalls.Error("set_thread_area", syserror.ENOSYS, "Expected to return ENOSYS on 64-bit", nil), + 206: syscalls.PartiallySupported("io_setup", IoSetup, "Generally supported with exceptions. User ring optimizations are not implemented.", []string{"gvisor.dev/issue/204"}), + 207: syscalls.PartiallySupported("io_destroy", IoDestroy, "Generally supported with exceptions. User ring optimizations are not implemented.", []string{"gvisor.dev/issue/204"}), + 208: syscalls.PartiallySupported("io_getevents", IoGetevents, "Generally supported with exceptions. User ring optimizations are not implemented.", []string{"gvisor.dev/issue/204"}), + 209: syscalls.PartiallySupported("io_submit", IoSubmit, "Generally supported with exceptions. User ring optimizations are not implemented.", []string{"gvisor.dev/issue/204"}), + 210: syscalls.PartiallySupported("io_cancel", IoCancel, "Generally supported with exceptions. User ring optimizations are not implemented.", []string{"gvisor.dev/issue/204"}), + 211: syscalls.Error("get_thread_area", syserror.ENOSYS, "Expected to return ENOSYS on 64-bit", nil), + 212: syscalls.CapError("lookup_dcookie", linux.CAP_SYS_ADMIN, "", nil), + 213: syscalls.Supported("epoll_create", EpollCreate), + 214: syscalls.ErrorWithEvent("epoll_ctl_old", syserror.ENOSYS, "Deprecated.", nil), + 215: syscalls.ErrorWithEvent("epoll_wait_old", syserror.ENOSYS, "Deprecated.", nil), + 216: syscalls.ErrorWithEvent("remap_file_pages", syserror.ENOSYS, "Deprecated since Linux 3.16.", nil), + 217: syscalls.Supported("getdents64", Getdents64), + 218: syscalls.Supported("set_tid_address", SetTidAddress), + 219: syscalls.Supported("restart_syscall", RestartSyscall), + 220: syscalls.ErrorWithEvent("semtimedop", syserror.ENOSYS, "", []string{"gvisor.dev/issue/137"}), + 221: syscalls.PartiallySupported("fadvise64", Fadvise64, "Not all options are supported.", nil), + 222: syscalls.Supported("timer_create", TimerCreate), + 223: syscalls.Supported("timer_settime", TimerSettime), + 224: syscalls.Supported("timer_gettime", TimerGettime), + 225: syscalls.Supported("timer_getoverrun", TimerGetoverrun), + 226: syscalls.Supported("timer_delete", TimerDelete), + 227: syscalls.Supported("clock_settime", ClockSettime), + 228: syscalls.Supported("clock_gettime", ClockGettime), + 229: syscalls.Supported("clock_getres", ClockGetres), + 230: syscalls.Supported("clock_nanosleep", ClockNanosleep), + 231: syscalls.Supported("exit_group", ExitGroup), + 232: syscalls.Supported("epoll_wait", EpollWait), + 233: syscalls.Supported("epoll_ctl", EpollCtl), + 234: syscalls.Supported("tgkill", Tgkill), + 235: syscalls.Supported("utimes", Utimes), + 236: syscalls.Error("vserver", syserror.ENOSYS, "Not implemented by Linux", nil), + 237: syscalls.PartiallySupported("mbind", Mbind, "Stub implementation. Only a single NUMA node is advertised, and mempolicy is ignored accordingly, but mbind() will succeed and has effects reflected by get_mempolicy.", []string{"gvisor.dev/issue/262"}), + 238: syscalls.PartiallySupported("set_mempolicy", SetMempolicy, "Stub implementation.", nil), + 239: syscalls.PartiallySupported("get_mempolicy", GetMempolicy, "Stub implementation.", nil), + 240: syscalls.ErrorWithEvent("mq_open", syserror.ENOSYS, "", []string{"gvisor.dev/issue/136"}), // TODO(b/29354921) + 241: syscalls.ErrorWithEvent("mq_unlink", syserror.ENOSYS, "", []string{"gvisor.dev/issue/136"}), // TODO(b/29354921) + 242: syscalls.ErrorWithEvent("mq_timedsend", syserror.ENOSYS, "", []string{"gvisor.dev/issue/136"}), // TODO(b/29354921) + 243: syscalls.ErrorWithEvent("mq_timedreceive", syserror.ENOSYS, "", []string{"gvisor.dev/issue/136"}), // TODO(b/29354921) + 244: syscalls.ErrorWithEvent("mq_notify", syserror.ENOSYS, "", []string{"gvisor.dev/issue/136"}), // TODO(b/29354921) + 245: syscalls.ErrorWithEvent("mq_getsetattr", syserror.ENOSYS, "", []string{"gvisor.dev/issue/136"}), // TODO(b/29354921) + 246: syscalls.CapError("kexec_load", linux.CAP_SYS_BOOT, "", nil), + 247: syscalls.Supported("waitid", Waitid), + 248: syscalls.Error("add_key", syserror.EACCES, "Not available to user.", nil), + 249: syscalls.Error("request_key", syserror.EACCES, "Not available to user.", nil), + 250: syscalls.Error("keyctl", syserror.EACCES, "Not available to user.", nil), + 251: syscalls.CapError("ioprio_set", linux.CAP_SYS_ADMIN, "", nil), // requires cap_sys_nice or cap_sys_admin (depending) + 252: syscalls.CapError("ioprio_get", linux.CAP_SYS_ADMIN, "", nil), // requires cap_sys_nice or cap_sys_admin (depending) + 253: syscalls.PartiallySupported("inotify_init", InotifyInit, "inotify events are only available inside the sandbox.", nil), + 254: syscalls.PartiallySupported("inotify_add_watch", InotifyAddWatch, "inotify events are only available inside the sandbox.", nil), + 255: syscalls.PartiallySupported("inotify_rm_watch", InotifyRmWatch, "inotify events are only available inside the sandbox.", nil), + 256: syscalls.CapError("migrate_pages", linux.CAP_SYS_NICE, "", nil), + 257: syscalls.Supported("openat", Openat), + 258: syscalls.Supported("mkdirat", Mkdirat), + 259: syscalls.Supported("mknodat", Mknodat), + 260: syscalls.Supported("fchownat", Fchownat), + 261: syscalls.Supported("futimesat", Futimesat), + 262: syscalls.Supported("fstatat", Fstatat), + 263: syscalls.Supported("unlinkat", Unlinkat), + 264: syscalls.Supported("renameat", Renameat), + 265: syscalls.Supported("linkat", Linkat), + 266: syscalls.Supported("symlinkat", Symlinkat), + 267: syscalls.Supported("readlinkat", Readlinkat), + 268: syscalls.Supported("fchmodat", Fchmodat), + 269: syscalls.Supported("faccessat", Faccessat), + 270: syscalls.Supported("pselect", Pselect), + 271: syscalls.Supported("ppoll", Ppoll), + 272: syscalls.PartiallySupported("unshare", Unshare, "Mount, cgroup namespaces not supported. Network namespaces supported but must be empty.", nil), + 273: syscalls.Error("set_robust_list", syserror.ENOSYS, "Obsolete.", nil), + 274: syscalls.Error("get_robust_list", syserror.ENOSYS, "Obsolete.", nil), + 275: syscalls.Supported("splice", Splice), + 276: syscalls.Supported("tee", Tee), + 277: syscalls.PartiallySupported("sync_file_range", SyncFileRange, "Full data flush is not guaranteed at this time.", nil), + 278: syscalls.ErrorWithEvent("vmsplice", syserror.ENOSYS, "", []string{"gvisor.dev/issue/138"}), // TODO(b/29354098) + 279: syscalls.CapError("move_pages", linux.CAP_SYS_NICE, "", nil), // requires cap_sys_nice (mostly) + 280: syscalls.Supported("utimensat", Utimensat), + 281: syscalls.Supported("epoll_pwait", EpollPwait), + 282: syscalls.PartiallySupported("signalfd", Signalfd, "Semantics are slightly different.", []string{"gvisor.dev/issue/139"}), + 283: syscalls.Supported("timerfd_create", TimerfdCreate), + 284: syscalls.Supported("eventfd", Eventfd), + 285: syscalls.PartiallySupported("fallocate", Fallocate, "Not all options are supported.", nil), + 286: syscalls.Supported("timerfd_settime", TimerfdSettime), + 287: syscalls.Supported("timerfd_gettime", TimerfdGettime), + 288: syscalls.Supported("accept4", Accept4), + 289: syscalls.PartiallySupported("signalfd4", Signalfd4, "Semantics are slightly different.", []string{"gvisor.dev/issue/139"}), + 290: syscalls.Supported("eventfd2", Eventfd2), + 291: syscalls.Supported("epoll_create1", EpollCreate1), + 292: syscalls.Supported("dup3", Dup3), + 293: syscalls.Supported("pipe2", Pipe2), + 294: syscalls.Supported("inotify_init1", InotifyInit1), + 295: syscalls.Supported("preadv", Preadv), + 296: syscalls.Supported("pwritev", Pwritev), + 297: syscalls.Supported("rt_tgsigqueueinfo", RtTgsigqueueinfo), + 298: syscalls.ErrorWithEvent("perf_event_open", syserror.ENODEV, "No support for perf counters", nil), + 299: syscalls.PartiallySupported("recvmmsg", RecvMMsg, "Not all flags and control messages are supported.", nil), + 300: syscalls.ErrorWithEvent("fanotify_init", syserror.ENOSYS, "Needs CONFIG_FANOTIFY", nil), + 301: syscalls.ErrorWithEvent("fanotify_mark", syserror.ENOSYS, "Needs CONFIG_FANOTIFY", nil), + 302: syscalls.Supported("prlimit64", Prlimit64), + 303: syscalls.Error("name_to_handle_at", syserror.EOPNOTSUPP, "Not supported by gVisor filesystems", nil), + 304: syscalls.Error("open_by_handle_at", syserror.EOPNOTSUPP, "Not supported by gVisor filesystems", nil), + 305: syscalls.CapError("clock_adjtime", linux.CAP_SYS_TIME, "", nil), + 306: syscalls.PartiallySupported("syncfs", Syncfs, "Depends on backing file system.", nil), + 307: syscalls.PartiallySupported("sendmmsg", SendMMsg, "Not all flags and control messages are supported.", nil), + 308: syscalls.ErrorWithEvent("setns", syserror.EOPNOTSUPP, "Needs filesystem support", []string{"gvisor.dev/issue/140"}), // TODO(b/29354995) + 309: syscalls.Supported("getcpu", Getcpu), + 310: syscalls.ErrorWithEvent("process_vm_readv", syserror.ENOSYS, "", []string{"gvisor.dev/issue/158"}), + 311: syscalls.ErrorWithEvent("process_vm_writev", syserror.ENOSYS, "", []string{"gvisor.dev/issue/158"}), + 312: syscalls.CapError("kcmp", linux.CAP_SYS_PTRACE, "", nil), + 313: syscalls.CapError("finit_module", linux.CAP_SYS_MODULE, "", nil), + 314: syscalls.ErrorWithEvent("sched_setattr", syserror.ENOSYS, "gVisor does not implement a scheduler.", []string{"gvisor.dev/issue/264"}), // TODO(b/118902272) + 315: syscalls.ErrorWithEvent("sched_getattr", syserror.ENOSYS, "gVisor does not implement a scheduler.", []string{"gvisor.dev/issue/264"}), // TODO(b/118902272) + 316: syscalls.ErrorWithEvent("renameat2", syserror.ENOSYS, "", []string{"gvisor.dev/issue/263"}), // TODO(b/118902772) + 317: syscalls.Supported("seccomp", Seccomp), + 318: syscalls.Supported("getrandom", GetRandom), + 319: syscalls.Supported("memfd_create", MemfdCreate), + 320: syscalls.CapError("kexec_file_load", linux.CAP_SYS_BOOT, "", nil), + 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) + 325: syscalls.PartiallySupported("mlock2", Mlock2, "Stub implementation. The sandbox lacks appropriate permissions.", nil), + + // Syscalls implemented after 325 are "backports" from versions + // of Linux after 4.4. + 326: syscalls.ErrorWithEvent("copy_file_range", syserror.ENOSYS, "", nil), + 327: syscalls.Supported("preadv2", Preadv2), + 328: syscalls.PartiallySupported("pwritev2", Pwritev2, "Flag RWF_HIPRI is not supported.", nil), + 329: syscalls.ErrorWithEvent("pkey_mprotect", syserror.ENOSYS, "", nil), + 330: syscalls.ErrorWithEvent("pkey_alloc", syserror.ENOSYS, "", nil), + 331: syscalls.ErrorWithEvent("pkey_free", syserror.ENOSYS, "", nil), + 332: syscalls.Supported("statx", Statx), + 333: syscalls.ErrorWithEvent("io_pgetevents", syserror.ENOSYS, "", nil), + 334: syscalls.PartiallySupported("rseq", RSeq, "Not supported on all platforms.", nil), + + // Linux skips ahead to syscall 424 to sync numbers between arches. + 424: syscalls.ErrorWithEvent("pidfd_send_signal", syserror.ENOSYS, "", nil), + 425: syscalls.ErrorWithEvent("io_uring_setup", syserror.ENOSYS, "", nil), + 426: syscalls.ErrorWithEvent("io_uring_enter", syserror.ENOSYS, "", nil), + 427: syscalls.ErrorWithEvent("io_uring_register", syserror.ENOSYS, "", nil), + 428: syscalls.ErrorWithEvent("open_tree", syserror.ENOSYS, "", nil), + 429: syscalls.ErrorWithEvent("move_mount", syserror.ENOSYS, "", nil), + 430: syscalls.ErrorWithEvent("fsopen", syserror.ENOSYS, "", nil), + 431: syscalls.ErrorWithEvent("fsconfig", syserror.ENOSYS, "", nil), + 432: syscalls.ErrorWithEvent("fsmount", syserror.ENOSYS, "", nil), + 433: syscalls.ErrorWithEvent("fspick", syserror.ENOSYS, "", nil), + 434: syscalls.ErrorWithEvent("pidfd_open", syserror.ENOSYS, "", nil), + 435: syscalls.ErrorWithEvent("clone3", syserror.ENOSYS, "", nil), + }, + Emulate: map[usermem.Addr]uintptr{ + 0xffffffffff600000: 96, // vsyscall gettimeofday(2) + 0xffffffffff600400: 201, // vsyscall time(2) + 0xffffffffff600800: 309, // vsyscall getcpu(2) + }, + Missing: func(t *kernel.Task, sysno uintptr, args arch.SyscallArguments) (uintptr, error) { + t.Kernel().EmitUnimplementedEvent(t) + return 0, syserror.ENOSYS + }, +} + +// ARM64 is a table of Linux arm64 syscall API with the corresponding syscall +// numbers from Linux 4.4. +var ARM64 = &kernel.SyscallTable{ + OS: abi.Linux, + Arch: arch.ARM64, + Version: kernel.Version{ + Sysname: LinuxSysname, + Release: LinuxRelease, + Version: LinuxVersion, + }, + AuditNumber: linux.AUDIT_ARCH_AARCH64, + Table: map[uintptr]kernel.Syscall{ + 0: syscalls.PartiallySupported("io_setup", IoSetup, "Generally supported with exceptions. User ring optimizations are not implemented.", []string{"gvisor.dev/issue/204"}), + 1: syscalls.PartiallySupported("io_destroy", IoDestroy, "Generally supported with exceptions. User ring optimizations are not implemented.", []string{"gvisor.dev/issue/204"}), + 2: syscalls.PartiallySupported("io_submit", IoSubmit, "Generally supported with exceptions. User ring optimizations are not implemented.", []string{"gvisor.dev/issue/204"}), + 3: syscalls.PartiallySupported("io_cancel", IoCancel, "Generally supported with exceptions. User ring optimizations are not implemented.", []string{"gvisor.dev/issue/204"}), + 4: syscalls.PartiallySupported("io_getevents", IoGetevents, "Generally supported with exceptions. User ring optimizations are not implemented.", []string{"gvisor.dev/issue/204"}), + 5: syscalls.PartiallySupported("setxattr", SetXattr, "Only supported for tmpfs.", nil), + 6: syscalls.PartiallySupported("lsetxattr", LSetXattr, "Only supported for tmpfs.", nil), + 7: syscalls.PartiallySupported("fsetxattr", FSetXattr, "Only supported for tmpfs.", nil), + 8: syscalls.PartiallySupported("getxattr", GetXattr, "Only supported for tmpfs.", nil), + 9: syscalls.PartiallySupported("lgetxattr", LGetXattr, "Only supported for tmpfs.", nil), + 10: syscalls.PartiallySupported("fgetxattr", FGetXattr, "Only supported for tmpfs.", nil), + 11: syscalls.PartiallySupported("listxattr", ListXattr, "Only supported for tmpfs", nil), + 12: syscalls.PartiallySupported("llistxattr", LListXattr, "Only supported for tmpfs", nil), + 13: syscalls.PartiallySupported("flistxattr", FListXattr, "Only supported for tmpfs", nil), + 14: syscalls.PartiallySupported("removexattr", RemoveXattr, "Only supported for tmpfs", nil), + 15: syscalls.PartiallySupported("lremovexattr", LRemoveXattr, "Only supported for tmpfs", nil), + 16: syscalls.PartiallySupported("fremovexattr", FRemoveXattr, "Only supported for tmpfs", nil), + 17: syscalls.Supported("getcwd", Getcwd), + 18: syscalls.CapError("lookup_dcookie", linux.CAP_SYS_ADMIN, "", nil), + 19: syscalls.Supported("eventfd2", Eventfd2), + 20: syscalls.Supported("epoll_create1", EpollCreate1), + 21: syscalls.Supported("epoll_ctl", EpollCtl), + 22: syscalls.Supported("epoll_pwait", EpollPwait), + 23: syscalls.Supported("dup", Dup), + 24: syscalls.Supported("dup3", Dup3), + 25: syscalls.PartiallySupported("fcntl", Fcntl, "Not all options are supported.", nil), + 26: syscalls.Supported("inotify_init1", InotifyInit1), + 27: syscalls.PartiallySupported("inotify_add_watch", InotifyAddWatch, "inotify events are only available inside the sandbox.", nil), + 28: syscalls.PartiallySupported("inotify_rm_watch", InotifyRmWatch, "inotify events are only available inside the sandbox.", nil), + 29: syscalls.PartiallySupported("ioctl", Ioctl, "Only a few ioctls are implemented for backing devices and file systems.", nil), + 30: syscalls.CapError("ioprio_set", linux.CAP_SYS_ADMIN, "", nil), // requires cap_sys_nice or cap_sys_admin (depending) + 31: syscalls.CapError("ioprio_get", linux.CAP_SYS_ADMIN, "", nil), // requires cap_sys_nice or cap_sys_admin (depending) + 32: syscalls.PartiallySupported("flock", Flock, "Locks are held within the sandbox only.", nil), + 33: syscalls.Supported("mknodat", Mknodat), + 34: syscalls.Supported("mkdirat", Mkdirat), + 35: syscalls.Supported("unlinkat", Unlinkat), + 36: syscalls.Supported("symlinkat", Symlinkat), + 37: syscalls.Supported("linkat", Linkat), + 38: syscalls.Supported("renameat", Renameat), + 39: syscalls.PartiallySupported("umount2", Umount2, "Not all options or file systems are supported.", nil), + 40: syscalls.PartiallySupported("mount", Mount, "Not all options or file systems are supported.", nil), + 41: syscalls.Error("pivot_root", syserror.EPERM, "", nil), + 42: syscalls.Error("nfsservctl", syserror.ENOSYS, "Removed after Linux 3.1.", nil), + 43: syscalls.PartiallySupported("statfs", Statfs, "Depends on the backing file system implementation.", nil), + 44: syscalls.PartiallySupported("fstatfs", Fstatfs, "Depends on the backing file system implementation.", nil), + 45: syscalls.Supported("truncate", Truncate), + 46: syscalls.Supported("ftruncate", Ftruncate), + 47: syscalls.PartiallySupported("fallocate", Fallocate, "Not all options are supported.", nil), + 48: syscalls.Supported("faccessat", Faccessat), + 49: syscalls.Supported("chdir", Chdir), + 50: syscalls.Supported("fchdir", Fchdir), + 51: syscalls.Supported("chroot", Chroot), + 52: syscalls.PartiallySupported("fchmod", Fchmod, "Options S_ISUID and S_ISGID not supported.", nil), + 53: syscalls.Supported("fchmodat", Fchmodat), + 54: syscalls.Supported("fchownat", Fchownat), + 55: syscalls.Supported("fchown", Fchown), + 56: syscalls.Supported("openat", Openat), + 57: syscalls.Supported("close", Close), + 58: syscalls.CapError("vhangup", linux.CAP_SYS_TTY_CONFIG, "", nil), + 59: syscalls.Supported("pipe2", Pipe2), + 60: syscalls.CapError("quotactl", linux.CAP_SYS_ADMIN, "", nil), // requires cap_sys_admin for most operations + 61: syscalls.Supported("getdents64", Getdents64), + 62: syscalls.Supported("lseek", Lseek), + 63: syscalls.Supported("read", Read), + 64: syscalls.Supported("write", Write), + 65: syscalls.Supported("readv", Readv), + 66: syscalls.Supported("writev", Writev), + 67: syscalls.Supported("pread64", Pread64), + 68: syscalls.Supported("pwrite64", Pwrite64), + 69: syscalls.Supported("preadv", Preadv), + 70: syscalls.Supported("pwritev", Pwritev), + 71: syscalls.Supported("sendfile", Sendfile), + 72: syscalls.Supported("pselect", Pselect), + 73: syscalls.Supported("ppoll", Ppoll), + 74: syscalls.PartiallySupported("signalfd4", Signalfd4, "Semantics are slightly different.", []string{"gvisor.dev/issue/139"}), + 75: syscalls.ErrorWithEvent("vmsplice", syserror.ENOSYS, "", []string{"gvisor.dev/issue/138"}), // TODO(b/29354098) + 76: syscalls.PartiallySupported("splice", Splice, "Stub implementation.", []string{"gvisor.dev/issue/138"}), // TODO(b/29354098) + 77: syscalls.Supported("tee", Tee), + 78: syscalls.Supported("readlinkat", Readlinkat), + 79: syscalls.Supported("fstatat", Fstatat), + 80: syscalls.Supported("fstat", Fstat), + 81: syscalls.PartiallySupported("sync", Sync, "Full data flush is not guaranteed at this time.", nil), + 82: syscalls.PartiallySupported("fsync", Fsync, "Full data flush is not guaranteed at this time.", nil), + 83: syscalls.PartiallySupported("fdatasync", Fdatasync, "Full data flush is not guaranteed at this time.", nil), + 84: syscalls.PartiallySupported("sync_file_range", SyncFileRange, "Full data flush is not guaranteed at this time.", nil), + 85: syscalls.Supported("timerfd_create", TimerfdCreate), + 86: syscalls.Supported("timerfd_settime", TimerfdSettime), + 87: syscalls.Supported("timerfd_gettime", TimerfdGettime), + 88: syscalls.Supported("utimensat", Utimensat), + 89: syscalls.CapError("acct", linux.CAP_SYS_PACCT, "", nil), + 90: syscalls.Supported("capget", Capget), + 91: syscalls.Supported("capset", Capset), + 92: syscalls.ErrorWithEvent("personality", syserror.EINVAL, "Unable to change personality.", nil), + 93: syscalls.Supported("exit", Exit), + 94: syscalls.Supported("exit_group", ExitGroup), + 95: syscalls.Supported("waitid", Waitid), + 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), + 101: syscalls.Supported("nanosleep", Nanosleep), + 102: syscalls.Supported("getitimer", Getitimer), + 103: syscalls.Supported("setitimer", Setitimer), + 104: syscalls.CapError("kexec_load", linux.CAP_SYS_BOOT, "", nil), + 105: syscalls.CapError("init_module", linux.CAP_SYS_MODULE, "", nil), + 106: syscalls.CapError("delete_module", linux.CAP_SYS_MODULE, "", nil), + 107: syscalls.Supported("timer_create", TimerCreate), + 108: syscalls.Supported("timer_gettime", TimerGettime), + 109: syscalls.Supported("timer_getoverrun", TimerGetoverrun), + 110: syscalls.Supported("timer_settime", TimerSettime), + 111: syscalls.Supported("timer_delete", TimerDelete), + 112: syscalls.Supported("clock_settime", ClockSettime), + 113: syscalls.Supported("clock_gettime", ClockGettime), + 114: syscalls.Supported("clock_getres", ClockGetres), + 115: syscalls.Supported("clock_nanosleep", ClockNanosleep), + 116: syscalls.PartiallySupported("syslog", Syslog, "Outputs a dummy message for security reasons.", nil), + 117: syscalls.PartiallySupported("ptrace", Ptrace, "Options PTRACE_PEEKSIGINFO, PTRACE_SECCOMP_GET_FILTER not supported.", nil), + 118: syscalls.CapError("sched_setparam", linux.CAP_SYS_NICE, "", nil), + 119: syscalls.PartiallySupported("sched_setscheduler", SchedSetscheduler, "Stub implementation.", nil), + 120: syscalls.PartiallySupported("sched_getscheduler", SchedGetscheduler, "Stub implementation.", nil), + 121: syscalls.PartiallySupported("sched_getparam", SchedGetparam, "Stub implementation.", nil), + 122: syscalls.PartiallySupported("sched_setaffinity", SchedSetaffinity, "Stub implementation.", nil), + 123: syscalls.PartiallySupported("sched_getaffinity", SchedGetaffinity, "Stub implementation.", nil), + 124: syscalls.Supported("sched_yield", SchedYield), + 125: syscalls.PartiallySupported("sched_get_priority_max", SchedGetPriorityMax, "Stub implementation.", nil), + 126: syscalls.PartiallySupported("sched_get_priority_min", SchedGetPriorityMin, "Stub implementation.", nil), + 127: syscalls.ErrorWithEvent("sched_rr_get_interval", syserror.EPERM, "", nil), + 128: syscalls.Supported("restart_syscall", RestartSyscall), + 129: syscalls.Supported("kill", Kill), + 130: syscalls.Supported("tkill", Tkill), + 131: syscalls.Supported("tgkill", Tgkill), + 132: syscalls.Supported("sigaltstack", Sigaltstack), + 133: syscalls.Supported("rt_sigsuspend", RtSigsuspend), + 134: syscalls.Supported("rt_sigaction", RtSigaction), + 135: syscalls.Supported("rt_sigprocmask", RtSigprocmask), + 136: syscalls.Supported("rt_sigpending", RtSigpending), + 137: syscalls.Supported("rt_sigtimedwait", RtSigtimedwait), + 138: syscalls.Supported("rt_sigqueueinfo", RtSigqueueinfo), + 139: syscalls.Supported("rt_sigreturn", RtSigreturn), + 140: syscalls.PartiallySupported("setpriority", Setpriority, "Stub implementation.", nil), + 141: syscalls.PartiallySupported("getpriority", Getpriority, "Stub implementation.", nil), + 142: syscalls.CapError("reboot", linux.CAP_SYS_BOOT, "", nil), + 143: syscalls.Supported("setregid", Setregid), + 144: syscalls.Supported("setgid", Setgid), + 145: syscalls.Supported("setreuid", Setreuid), + 146: syscalls.Supported("setuid", Setuid), + 147: syscalls.Supported("setresuid", Setresuid), + 148: syscalls.Supported("getresuid", Getresuid), + 149: syscalls.Supported("setresgid", Setresgid), + 150: syscalls.Supported("getresgid", Getresgid), + 151: syscalls.ErrorWithEvent("setfsuid", syserror.ENOSYS, "", []string{"gvisor.dev/issue/260"}), // TODO(b/112851702) + 152: syscalls.ErrorWithEvent("setfsgid", syserror.ENOSYS, "", []string{"gvisor.dev/issue/260"}), // TODO(b/112851702) + 153: syscalls.Supported("times", Times), + 154: syscalls.Supported("setpgid", Setpgid), + 155: syscalls.Supported("getpgid", Getpgid), + 156: syscalls.Supported("getsid", Getsid), + 157: syscalls.Supported("setsid", Setsid), + 158: syscalls.Supported("getgroups", Getgroups), + 159: syscalls.Supported("setgroups", Setgroups), + 160: syscalls.Supported("uname", Uname), + 161: syscalls.Supported("sethostname", Sethostname), + 162: syscalls.Supported("setdomainname", Setdomainname), + 163: syscalls.Supported("getrlimit", Getrlimit), + 164: syscalls.PartiallySupported("setrlimit", Setrlimit, "Not all rlimits are enforced.", nil), + 165: syscalls.PartiallySupported("getrusage", Getrusage, "Fields ru_maxrss, ru_minflt, ru_majflt, ru_inblock, ru_oublock are not supported. Fields ru_utime and ru_stime have low precision.", nil), + 166: syscalls.Supported("umask", Umask), + 167: syscalls.PartiallySupported("prctl", Prctl, "Not all options are supported.", nil), + 168: syscalls.Supported("getcpu", Getcpu), + 169: syscalls.Supported("gettimeofday", Gettimeofday), + 170: syscalls.CapError("settimeofday", linux.CAP_SYS_TIME, "", nil), + 171: syscalls.CapError("adjtimex", linux.CAP_SYS_TIME, "", nil), + 172: syscalls.Supported("getpid", Getpid), + 173: syscalls.Supported("getppid", Getppid), + 174: syscalls.Supported("getuid", Getuid), + 175: syscalls.Supported("geteuid", Geteuid), + 176: syscalls.Supported("getgid", Getgid), + 177: syscalls.Supported("getegid", Getegid), + 178: syscalls.Supported("gettid", Gettid), + 179: syscalls.PartiallySupported("sysinfo", Sysinfo, "Fields loads, sharedram, bufferram, totalswap, freeswap, totalhigh, freehigh not supported.", nil), + 180: syscalls.ErrorWithEvent("mq_open", syserror.ENOSYS, "", []string{"gvisor.dev/issue/136"}), // TODO(b/29354921) + 181: syscalls.ErrorWithEvent("mq_unlink", syserror.ENOSYS, "", []string{"gvisor.dev/issue/136"}), // TODO(b/29354921) + 182: syscalls.ErrorWithEvent("mq_timedsend", syserror.ENOSYS, "", []string{"gvisor.dev/issue/136"}), // TODO(b/29354921) + 183: syscalls.ErrorWithEvent("mq_timedreceive", syserror.ENOSYS, "", []string{"gvisor.dev/issue/136"}), // TODO(b/29354921) + 184: syscalls.ErrorWithEvent("mq_notify", syserror.ENOSYS, "", []string{"gvisor.dev/issue/136"}), // TODO(b/29354921) + 185: syscalls.ErrorWithEvent("mq_getsetattr", syserror.ENOSYS, "", []string{"gvisor.dev/issue/136"}), // TODO(b/29354921) + 186: syscalls.ErrorWithEvent("msgget", syserror.ENOSYS, "", []string{"gvisor.dev/issue/135"}), // TODO(b/29354921) + 187: syscalls.ErrorWithEvent("msgctl", syserror.ENOSYS, "", []string{"gvisor.dev/issue/135"}), // TODO(b/29354921) + 188: syscalls.ErrorWithEvent("msgrcv", syserror.ENOSYS, "", []string{"gvisor.dev/issue/135"}), // TODO(b/29354921) + 189: syscalls.ErrorWithEvent("msgsnd", syserror.ENOSYS, "", []string{"gvisor.dev/issue/135"}), // TODO(b/29354921) + 190: syscalls.Supported("semget", Semget), + 191: syscalls.PartiallySupported("semctl", Semctl, "Options IPC_INFO, SEM_INFO, IPC_STAT, SEM_STAT, SEM_STAT_ANY, GETNCNT, GETZCNT not supported.", nil), + 192: syscalls.ErrorWithEvent("semtimedop", syserror.ENOSYS, "", []string{"gvisor.dev/issue/137"}), + 193: syscalls.PartiallySupported("semop", Semop, "Option SEM_UNDO not supported.", nil), + 194: syscalls.PartiallySupported("shmget", Shmget, "Option SHM_HUGETLB is not supported.", nil), + 195: syscalls.PartiallySupported("shmctl", Shmctl, "Options SHM_LOCK, SHM_UNLOCK are not supported.", nil), + 196: syscalls.PartiallySupported("shmat", Shmat, "Option SHM_RND is not supported.", nil), + 197: syscalls.Supported("shmdt", Shmdt), + 198: syscalls.PartiallySupported("socket", Socket, "Limited support for AF_NETLINK, NETLINK_ROUTE sockets. Limited support for SOCK_RAW.", nil), + 199: syscalls.Supported("socketpair", SocketPair), + 200: syscalls.PartiallySupported("bind", Bind, "Autobind for abstract Unix sockets is not supported.", nil), + 201: syscalls.Supported("listen", Listen), + 202: syscalls.Supported("accept", Accept), + 203: syscalls.Supported("connect", Connect), + 204: syscalls.Supported("getsockname", GetSockName), + 205: syscalls.Supported("getpeername", GetPeerName), + 206: syscalls.Supported("sendto", SendTo), + 207: syscalls.Supported("recvfrom", RecvFrom), + 208: syscalls.PartiallySupported("setsockopt", SetSockOpt, "Not all socket options are supported.", nil), + 209: syscalls.PartiallySupported("getsockopt", GetSockOpt, "Not all socket options are supported.", nil), + 210: syscalls.PartiallySupported("shutdown", Shutdown, "Not all flags and control messages are supported.", nil), + 211: syscalls.Supported("sendmsg", SendMsg), + 212: syscalls.PartiallySupported("recvmsg", RecvMsg, "Not all flags and control messages are supported.", nil), + 213: syscalls.Supported("readahead", Readahead), + 214: syscalls.Supported("brk", Brk), + 215: syscalls.Supported("munmap", Munmap), + 216: syscalls.Supported("mremap", Mremap), + 217: syscalls.Error("add_key", syserror.EACCES, "Not available to user.", nil), + 218: syscalls.Error("request_key", syserror.EACCES, "Not available to user.", nil), + 219: syscalls.Error("keyctl", syserror.EACCES, "Not available to user.", nil), + 220: syscalls.PartiallySupported("clone", Clone, "Mount namespace (CLONE_NEWNS) not supported. Options CLONE_PARENT, CLONE_SYSVSEM not supported.", nil), + 221: syscalls.Supported("execve", Execve), + 222: syscalls.PartiallySupported("mmap", Mmap, "Generally supported with exceptions. Options MAP_FIXED_NOREPLACE, MAP_SHARED_VALIDATE, MAP_SYNC MAP_GROWSDOWN, MAP_HUGETLB are not supported.", nil), + 223: syscalls.PartiallySupported("fadvise64", Fadvise64, "Not all options are supported.", nil), + 224: syscalls.CapError("swapon", linux.CAP_SYS_ADMIN, "", nil), + 225: syscalls.CapError("swapoff", linux.CAP_SYS_ADMIN, "", nil), + 226: syscalls.Supported("mprotect", Mprotect), + 227: syscalls.PartiallySupported("msync", Msync, "Full data flush is not guaranteed at this time.", nil), + 228: syscalls.PartiallySupported("mlock", Mlock, "Stub implementation. The sandbox lacks appropriate permissions.", nil), + 229: syscalls.PartiallySupported("munlock", Munlock, "Stub implementation. The sandbox lacks appropriate permissions.", nil), + 230: syscalls.PartiallySupported("mlockall", Mlockall, "Stub implementation. The sandbox lacks appropriate permissions.", nil), + 231: syscalls.PartiallySupported("munlockall", Munlockall, "Stub implementation. The sandbox lacks appropriate permissions.", nil), + 232: syscalls.PartiallySupported("mincore", Mincore, "Stub implementation. The sandbox does not have access to this information. Reports all mapped pages are resident.", nil), + 233: syscalls.PartiallySupported("madvise", Madvise, "Options MADV_DONTNEED, MADV_DONTFORK are supported. Other advice is ignored.", nil), + 234: syscalls.ErrorWithEvent("remap_file_pages", syserror.ENOSYS, "Deprecated since Linux 3.16.", nil), + 235: syscalls.PartiallySupported("mbind", Mbind, "Stub implementation. Only a single NUMA node is advertised, and mempolicy is ignored accordingly, but mbind() will succeed and has effects reflected by get_mempolicy.", []string{"gvisor.dev/issue/262"}), + 236: syscalls.PartiallySupported("get_mempolicy", GetMempolicy, "Stub implementation.", nil), + 237: syscalls.PartiallySupported("set_mempolicy", SetMempolicy, "Stub implementation.", nil), + 238: syscalls.CapError("migrate_pages", linux.CAP_SYS_NICE, "", nil), + 239: syscalls.CapError("move_pages", linux.CAP_SYS_NICE, "", nil), // requires cap_sys_nice (mostly) + 240: syscalls.Supported("rt_tgsigqueueinfo", RtTgsigqueueinfo), + 241: syscalls.ErrorWithEvent("perf_event_open", syserror.ENODEV, "No support for perf counters", nil), + 242: syscalls.Supported("accept4", Accept4), + 243: syscalls.PartiallySupported("recvmmsg", RecvMMsg, "Not all flags and control messages are supported.", nil), + 260: syscalls.Supported("wait4", Wait4), + 261: syscalls.Supported("prlimit64", Prlimit64), + 262: syscalls.ErrorWithEvent("fanotify_init", syserror.ENOSYS, "Needs CONFIG_FANOTIFY", nil), + 263: syscalls.ErrorWithEvent("fanotify_mark", syserror.ENOSYS, "Needs CONFIG_FANOTIFY", nil), + 264: syscalls.Error("name_to_handle_at", syserror.EOPNOTSUPP, "Not supported by gVisor filesystems", nil), + 265: syscalls.Error("open_by_handle_at", syserror.EOPNOTSUPP, "Not supported by gVisor filesystems", nil), + 266: syscalls.CapError("clock_adjtime", linux.CAP_SYS_TIME, "", nil), + 267: syscalls.PartiallySupported("syncfs", Syncfs, "Depends on backing file system.", nil), + 268: syscalls.ErrorWithEvent("setns", syserror.EOPNOTSUPP, "Needs filesystem support", []string{"gvisor.dev/issue/140"}), // TODO(b/29354995) + 269: syscalls.PartiallySupported("sendmmsg", SendMMsg, "Not all flags and control messages are supported.", nil), + 270: syscalls.ErrorWithEvent("process_vm_readv", syserror.ENOSYS, "", []string{"gvisor.dev/issue/158"}), + 271: syscalls.ErrorWithEvent("process_vm_writev", syserror.ENOSYS, "", []string{"gvisor.dev/issue/158"}), + 272: syscalls.CapError("kcmp", linux.CAP_SYS_PTRACE, "", nil), + 273: syscalls.CapError("finit_module", linux.CAP_SYS_MODULE, "", nil), + 274: syscalls.ErrorWithEvent("sched_setattr", syserror.ENOSYS, "gVisor does not implement a scheduler.", []string{"gvisor.dev/issue/264"}), // TODO(b/118902272) + 275: syscalls.ErrorWithEvent("sched_getattr", syserror.ENOSYS, "gVisor does not implement a scheduler.", []string{"gvisor.dev/issue/264"}), // TODO(b/118902272) + 276: syscalls.ErrorWithEvent("renameat2", syserror.ENOSYS, "", []string{"gvisor.dev/issue/263"}), // TODO(b/118902772) + 277: syscalls.Supported("seccomp", Seccomp), + 278: syscalls.Supported("getrandom", GetRandom), + 279: syscalls.Supported("memfd_create", MemfdCreate), + 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) + 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. + 285: syscalls.ErrorWithEvent("copy_file_range", syserror.ENOSYS, "", nil), + 286: syscalls.Supported("preadv2", Preadv2), + 287: syscalls.PartiallySupported("pwritev2", Pwritev2, "Flag RWF_HIPRI is not supported.", nil), + 288: syscalls.ErrorWithEvent("pkey_mprotect", syserror.ENOSYS, "", nil), + 289: syscalls.ErrorWithEvent("pkey_alloc", syserror.ENOSYS, "", nil), + 290: syscalls.ErrorWithEvent("pkey_free", syserror.ENOSYS, "", nil), + 291: syscalls.Supported("statx", Statx), + 292: syscalls.ErrorWithEvent("io_pgetevents", syserror.ENOSYS, "", nil), + 293: syscalls.PartiallySupported("rseq", RSeq, "Not supported on all platforms.", nil), + + // Linux skips ahead to syscall 424 to sync numbers between arches. + 424: syscalls.ErrorWithEvent("pidfd_send_signal", syserror.ENOSYS, "", nil), + 425: syscalls.ErrorWithEvent("io_uring_setup", syserror.ENOSYS, "", nil), + 426: syscalls.ErrorWithEvent("io_uring_enter", syserror.ENOSYS, "", nil), + 427: syscalls.ErrorWithEvent("io_uring_register", syserror.ENOSYS, "", nil), + 428: syscalls.ErrorWithEvent("open_tree", syserror.ENOSYS, "", nil), + 429: syscalls.ErrorWithEvent("move_mount", syserror.ENOSYS, "", nil), + 430: syscalls.ErrorWithEvent("fsopen", syserror.ENOSYS, "", nil), + 431: syscalls.ErrorWithEvent("fsconfig", syserror.ENOSYS, "", nil), + 432: syscalls.ErrorWithEvent("fsmount", syserror.ENOSYS, "", nil), + 433: syscalls.ErrorWithEvent("fspick", syserror.ENOSYS, "", nil), + 434: syscalls.ErrorWithEvent("pidfd_open", syserror.ENOSYS, "", nil), + 435: syscalls.ErrorWithEvent("clone3", syserror.ENOSYS, "", nil), + }, + Emulate: map[usermem.Addr]uintptr{}, + Missing: func(t *kernel.Task, sysno uintptr, args arch.SyscallArguments) (uintptr, error) { + t.Kernel().EmitUnimplementedEvent(t) + return 0, syserror.ENOSYS + }, +} + +func init() { + kernel.RegisterSyscallTable(AMD64) + kernel.RegisterSyscallTable(ARM64) +} diff --git a/pkg/sentry/syscalls/linux/linux64_amd64.go b/pkg/sentry/syscalls/linux/linux64_amd64.go deleted file mode 100644 index 79066ad2a..000000000 --- a/pkg/sentry/syscalls/linux/linux64_amd64.go +++ /dev/null @@ -1,406 +0,0 @@ -// Copyright 2018 The gVisor Authors. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -package linux - -import ( - "gvisor.dev/gvisor/pkg/abi" - "gvisor.dev/gvisor/pkg/abi/linux" - "gvisor.dev/gvisor/pkg/sentry/arch" - "gvisor.dev/gvisor/pkg/sentry/kernel" - "gvisor.dev/gvisor/pkg/sentry/syscalls" - "gvisor.dev/gvisor/pkg/syserror" - "gvisor.dev/gvisor/pkg/usermem" -) - -// AMD64 is a table of Linux amd64 syscall API with the corresponding syscall -// numbers from Linux 4.4. -var AMD64 = &kernel.SyscallTable{ - OS: abi.Linux, - Arch: arch.AMD64, - Version: kernel.Version{ - // Version 4.4 is chosen as a stable, longterm version of Linux, which - // guides the interface provided by this syscall table. The build - // version is that for a clean build with default kernel config, at 5 - // minutes after v4.4 was tagged. - Sysname: LinuxSysname, - Release: LinuxRelease, - Version: LinuxVersion, - }, - AuditNumber: linux.AUDIT_ARCH_X86_64, - Table: map[uintptr]kernel.Syscall{ - 0: syscalls.Supported("read", Read), - 1: syscalls.Supported("write", Write), - 2: syscalls.PartiallySupported("open", Open, "Options O_DIRECT, O_NOATIME, O_PATH, O_TMPFILE, O_SYNC are not supported.", nil), - 3: syscalls.Supported("close", Close), - 4: syscalls.Supported("stat", Stat), - 5: syscalls.Supported("fstat", Fstat), - 6: syscalls.Supported("lstat", Lstat), - 7: syscalls.Supported("poll", Poll), - 8: syscalls.Supported("lseek", Lseek), - 9: syscalls.PartiallySupported("mmap", Mmap, "Generally supported with exceptions. Options MAP_FIXED_NOREPLACE, MAP_SHARED_VALIDATE, MAP_SYNC MAP_GROWSDOWN, MAP_HUGETLB are not supported.", nil), - 10: syscalls.Supported("mprotect", Mprotect), - 11: syscalls.Supported("munmap", Munmap), - 12: syscalls.Supported("brk", Brk), - 13: syscalls.Supported("rt_sigaction", RtSigaction), - 14: syscalls.Supported("rt_sigprocmask", RtSigprocmask), - 15: syscalls.Supported("rt_sigreturn", RtSigreturn), - 16: syscalls.PartiallySupported("ioctl", Ioctl, "Only a few ioctls are implemented for backing devices and file systems.", nil), - 17: syscalls.Supported("pread64", Pread64), - 18: syscalls.Supported("pwrite64", Pwrite64), - 19: syscalls.Supported("readv", Readv), - 20: syscalls.Supported("writev", Writev), - 21: syscalls.Supported("access", Access), - 22: syscalls.Supported("pipe", Pipe), - 23: syscalls.Supported("select", Select), - 24: syscalls.Supported("sched_yield", SchedYield), - 25: syscalls.Supported("mremap", Mremap), - 26: syscalls.PartiallySupported("msync", Msync, "Full data flush is not guaranteed at this time.", nil), - 27: syscalls.PartiallySupported("mincore", Mincore, "Stub implementation. The sandbox does not have access to this information. Reports all mapped pages are resident.", nil), - 28: syscalls.PartiallySupported("madvise", Madvise, "Options MADV_DONTNEED, MADV_DONTFORK are supported. Other advice is ignored.", nil), - 29: syscalls.PartiallySupported("shmget", Shmget, "Option SHM_HUGETLB is not supported.", nil), - 30: syscalls.PartiallySupported("shmat", Shmat, "Option SHM_RND is not supported.", nil), - 31: syscalls.PartiallySupported("shmctl", Shmctl, "Options SHM_LOCK, SHM_UNLOCK are not supported.", nil), - 32: syscalls.Supported("dup", Dup), - 33: syscalls.Supported("dup2", Dup2), - 34: syscalls.Supported("pause", Pause), - 35: syscalls.Supported("nanosleep", Nanosleep), - 36: syscalls.Supported("getitimer", Getitimer), - 37: syscalls.Supported("alarm", Alarm), - 38: syscalls.Supported("setitimer", Setitimer), - 39: syscalls.Supported("getpid", Getpid), - 40: syscalls.Supported("sendfile", Sendfile), - 41: syscalls.PartiallySupported("socket", Socket, "Limited support for AF_NETLINK, NETLINK_ROUTE sockets. Limited support for SOCK_RAW.", nil), - 42: syscalls.Supported("connect", Connect), - 43: syscalls.Supported("accept", Accept), - 44: syscalls.Supported("sendto", SendTo), - 45: syscalls.Supported("recvfrom", RecvFrom), - 46: syscalls.Supported("sendmsg", SendMsg), - 47: syscalls.PartiallySupported("recvmsg", RecvMsg, "Not all flags and control messages are supported.", nil), - 48: syscalls.PartiallySupported("shutdown", Shutdown, "Not all flags and control messages are supported.", nil), - 49: syscalls.PartiallySupported("bind", Bind, "Autobind for abstract Unix sockets is not supported.", nil), - 50: syscalls.Supported("listen", Listen), - 51: syscalls.Supported("getsockname", GetSockName), - 52: syscalls.Supported("getpeername", GetPeerName), - 53: syscalls.Supported("socketpair", SocketPair), - 54: syscalls.PartiallySupported("setsockopt", SetSockOpt, "Not all socket options are supported.", nil), - 55: syscalls.PartiallySupported("getsockopt", GetSockOpt, "Not all socket options are supported.", nil), - 56: syscalls.PartiallySupported("clone", Clone, "Mount namespace (CLONE_NEWNS) not supported. Options CLONE_PARENT, CLONE_SYSVSEM not supported.", nil), - 57: syscalls.Supported("fork", Fork), - 58: syscalls.Supported("vfork", Vfork), - 59: syscalls.Supported("execve", Execve), - 60: syscalls.Supported("exit", Exit), - 61: syscalls.Supported("wait4", Wait4), - 62: syscalls.Supported("kill", Kill), - 63: syscalls.Supported("uname", Uname), - 64: syscalls.Supported("semget", Semget), - 65: syscalls.PartiallySupported("semop", Semop, "Option SEM_UNDO not supported.", nil), - 66: syscalls.PartiallySupported("semctl", Semctl, "Options IPC_INFO, SEM_INFO, IPC_STAT, SEM_STAT, SEM_STAT_ANY, GETNCNT, GETZCNT not supported.", nil), - 67: syscalls.Supported("shmdt", Shmdt), - 68: syscalls.ErrorWithEvent("msgget", syserror.ENOSYS, "", []string{"gvisor.dev/issue/135"}), // TODO(b/29354921) - 69: syscalls.ErrorWithEvent("msgsnd", syserror.ENOSYS, "", []string{"gvisor.dev/issue/135"}), // TODO(b/29354921) - 70: syscalls.ErrorWithEvent("msgrcv", syserror.ENOSYS, "", []string{"gvisor.dev/issue/135"}), // TODO(b/29354921) - 71: syscalls.ErrorWithEvent("msgctl", syserror.ENOSYS, "", []string{"gvisor.dev/issue/135"}), // TODO(b/29354921) - 72: syscalls.PartiallySupported("fcntl", Fcntl, "Not all options are supported.", nil), - 73: syscalls.PartiallySupported("flock", Flock, "Locks are held within the sandbox only.", nil), - 74: syscalls.PartiallySupported("fsync", Fsync, "Full data flush is not guaranteed at this time.", nil), - 75: syscalls.PartiallySupported("fdatasync", Fdatasync, "Full data flush is not guaranteed at this time.", nil), - 76: syscalls.Supported("truncate", Truncate), - 77: syscalls.Supported("ftruncate", Ftruncate), - 78: syscalls.Supported("getdents", Getdents), - 79: syscalls.Supported("getcwd", Getcwd), - 80: syscalls.Supported("chdir", Chdir), - 81: syscalls.Supported("fchdir", Fchdir), - 82: syscalls.Supported("rename", Rename), - 83: syscalls.Supported("mkdir", Mkdir), - 84: syscalls.Supported("rmdir", Rmdir), - 85: syscalls.Supported("creat", Creat), - 86: syscalls.Supported("link", Link), - 87: syscalls.Supported("unlink", Unlink), - 88: syscalls.Supported("symlink", Symlink), - 89: syscalls.Supported("readlink", Readlink), - 90: syscalls.Supported("chmod", Chmod), - 91: syscalls.PartiallySupported("fchmod", Fchmod, "Options S_ISUID and S_ISGID not supported.", nil), - 92: syscalls.Supported("chown", Chown), - 93: syscalls.Supported("fchown", Fchown), - 94: syscalls.Supported("lchown", Lchown), - 95: syscalls.Supported("umask", Umask), - 96: syscalls.Supported("gettimeofday", Gettimeofday), - 97: syscalls.Supported("getrlimit", Getrlimit), - 98: syscalls.PartiallySupported("getrusage", Getrusage, "Fields ru_maxrss, ru_minflt, ru_majflt, ru_inblock, ru_oublock are not supported. Fields ru_utime and ru_stime have low precision.", nil), - 99: syscalls.PartiallySupported("sysinfo", Sysinfo, "Fields loads, sharedram, bufferram, totalswap, freeswap, totalhigh, freehigh not supported.", nil), - 100: syscalls.Supported("times", Times), - 101: syscalls.PartiallySupported("ptrace", Ptrace, "Options PTRACE_PEEKSIGINFO, PTRACE_SECCOMP_GET_FILTER not supported.", nil), - 102: syscalls.Supported("getuid", Getuid), - 103: syscalls.PartiallySupported("syslog", Syslog, "Outputs a dummy message for security reasons.", nil), - 104: syscalls.Supported("getgid", Getgid), - 105: syscalls.Supported("setuid", Setuid), - 106: syscalls.Supported("setgid", Setgid), - 107: syscalls.Supported("geteuid", Geteuid), - 108: syscalls.Supported("getegid", Getegid), - 109: syscalls.Supported("setpgid", Setpgid), - 110: syscalls.Supported("getppid", Getppid), - 111: syscalls.Supported("getpgrp", Getpgrp), - 112: syscalls.Supported("setsid", Setsid), - 113: syscalls.Supported("setreuid", Setreuid), - 114: syscalls.Supported("setregid", Setregid), - 115: syscalls.Supported("getgroups", Getgroups), - 116: syscalls.Supported("setgroups", Setgroups), - 117: syscalls.Supported("setresuid", Setresuid), - 118: syscalls.Supported("getresuid", Getresuid), - 119: syscalls.Supported("setresgid", Setresgid), - 120: syscalls.Supported("getresgid", Getresgid), - 121: syscalls.Supported("getpgid", Getpgid), - 122: syscalls.ErrorWithEvent("setfsuid", syserror.ENOSYS, "", []string{"gvisor.dev/issue/260"}), // TODO(b/112851702) - 123: syscalls.ErrorWithEvent("setfsgid", syserror.ENOSYS, "", []string{"gvisor.dev/issue/260"}), // TODO(b/112851702) - 124: syscalls.Supported("getsid", Getsid), - 125: syscalls.Supported("capget", Capget), - 126: syscalls.Supported("capset", Capset), - 127: syscalls.Supported("rt_sigpending", RtSigpending), - 128: syscalls.Supported("rt_sigtimedwait", RtSigtimedwait), - 129: syscalls.Supported("rt_sigqueueinfo", RtSigqueueinfo), - 130: syscalls.Supported("rt_sigsuspend", RtSigsuspend), - 131: syscalls.Supported("sigaltstack", Sigaltstack), - 132: syscalls.Supported("utime", Utime), - 133: syscalls.PartiallySupported("mknod", Mknod, "Device creation is not generally supported. Only regular file and FIFO creation are supported.", nil), - 134: syscalls.Error("uselib", syserror.ENOSYS, "Obsolete", nil), - 135: syscalls.ErrorWithEvent("personality", syserror.EINVAL, "Unable to change personality.", nil), - 136: syscalls.ErrorWithEvent("ustat", syserror.ENOSYS, "Needs filesystem support.", nil), - 137: syscalls.PartiallySupported("statfs", Statfs, "Depends on the backing file system implementation.", nil), - 138: syscalls.PartiallySupported("fstatfs", Fstatfs, "Depends on the backing file system implementation.", nil), - 139: syscalls.ErrorWithEvent("sysfs", syserror.ENOSYS, "", []string{"gvisor.dev/issue/165"}), - 140: syscalls.PartiallySupported("getpriority", Getpriority, "Stub implementation.", nil), - 141: syscalls.PartiallySupported("setpriority", Setpriority, "Stub implementation.", nil), - 142: syscalls.CapError("sched_setparam", linux.CAP_SYS_NICE, "", nil), - 143: syscalls.PartiallySupported("sched_getparam", SchedGetparam, "Stub implementation.", nil), - 144: syscalls.PartiallySupported("sched_setscheduler", SchedSetscheduler, "Stub implementation.", nil), - 145: syscalls.PartiallySupported("sched_getscheduler", SchedGetscheduler, "Stub implementation.", nil), - 146: syscalls.PartiallySupported("sched_get_priority_max", SchedGetPriorityMax, "Stub implementation.", nil), - 147: syscalls.PartiallySupported("sched_get_priority_min", SchedGetPriorityMin, "Stub implementation.", nil), - 148: syscalls.ErrorWithEvent("sched_rr_get_interval", syserror.EPERM, "", nil), - 149: syscalls.PartiallySupported("mlock", Mlock, "Stub implementation. The sandbox lacks appropriate permissions.", nil), - 150: syscalls.PartiallySupported("munlock", Munlock, "Stub implementation. The sandbox lacks appropriate permissions.", nil), - 151: syscalls.PartiallySupported("mlockall", Mlockall, "Stub implementation. The sandbox lacks appropriate permissions.", nil), - 152: syscalls.PartiallySupported("munlockall", Munlockall, "Stub implementation. The sandbox lacks appropriate permissions.", nil), - 153: syscalls.CapError("vhangup", linux.CAP_SYS_TTY_CONFIG, "", nil), - 154: syscalls.Error("modify_ldt", syserror.EPERM, "", nil), - 155: syscalls.Error("pivot_root", syserror.EPERM, "", nil), - 156: syscalls.Error("sysctl", syserror.EPERM, "Deprecated. Use /proc/sys instead.", nil), - 157: syscalls.PartiallySupported("prctl", Prctl, "Not all options are supported.", nil), - 158: syscalls.PartiallySupported("arch_prctl", ArchPrctl, "Options ARCH_GET_GS, ARCH_SET_GS not supported.", nil), - 159: syscalls.CapError("adjtimex", linux.CAP_SYS_TIME, "", nil), - 160: syscalls.PartiallySupported("setrlimit", Setrlimit, "Not all rlimits are enforced.", nil), - 161: syscalls.Supported("chroot", Chroot), - 162: syscalls.PartiallySupported("sync", Sync, "Full data flush is not guaranteed at this time.", nil), - 163: syscalls.CapError("acct", linux.CAP_SYS_PACCT, "", nil), - 164: syscalls.CapError("settimeofday", linux.CAP_SYS_TIME, "", nil), - 165: syscalls.PartiallySupported("mount", Mount, "Not all options or file systems are supported.", nil), - 166: syscalls.PartiallySupported("umount2", Umount2, "Not all options or file systems are supported.", nil), - 167: syscalls.CapError("swapon", linux.CAP_SYS_ADMIN, "", nil), - 168: syscalls.CapError("swapoff", linux.CAP_SYS_ADMIN, "", nil), - 169: syscalls.CapError("reboot", linux.CAP_SYS_BOOT, "", nil), - 170: syscalls.Supported("sethostname", Sethostname), - 171: syscalls.Supported("setdomainname", Setdomainname), - 172: syscalls.CapError("iopl", linux.CAP_SYS_RAWIO, "", nil), - 173: syscalls.CapError("ioperm", linux.CAP_SYS_RAWIO, "", nil), - 174: syscalls.CapError("create_module", linux.CAP_SYS_MODULE, "", nil), - 175: syscalls.CapError("init_module", linux.CAP_SYS_MODULE, "", nil), - 176: syscalls.CapError("delete_module", linux.CAP_SYS_MODULE, "", nil), - 177: syscalls.Error("get_kernel_syms", syserror.ENOSYS, "Not supported in Linux > 2.6.", nil), - 178: syscalls.Error("query_module", syserror.ENOSYS, "Not supported in Linux > 2.6.", nil), - 179: syscalls.CapError("quotactl", linux.CAP_SYS_ADMIN, "", nil), // requires cap_sys_admin for most operations - 180: syscalls.Error("nfsservctl", syserror.ENOSYS, "Removed after Linux 3.1.", nil), - 181: syscalls.Error("getpmsg", syserror.ENOSYS, "Not implemented in Linux.", nil), - 182: syscalls.Error("putpmsg", syserror.ENOSYS, "Not implemented in Linux.", nil), - 183: syscalls.Error("afs_syscall", syserror.ENOSYS, "Not implemented in Linux.", nil), - 184: syscalls.Error("tuxcall", syserror.ENOSYS, "Not implemented in Linux.", nil), - 185: syscalls.Error("security", syserror.ENOSYS, "Not implemented in Linux.", nil), - 186: syscalls.Supported("gettid", Gettid), - 187: syscalls.Supported("readahead", Readahead), - 188: syscalls.PartiallySupported("setxattr", SetXattr, "Only supported for tmpfs.", nil), - 189: syscalls.PartiallySupported("lsetxattr", LSetXattr, "Only supported for tmpfs.", nil), - 190: syscalls.PartiallySupported("fsetxattr", FSetXattr, "Only supported for tmpfs.", nil), - 191: syscalls.PartiallySupported("getxattr", GetXattr, "Only supported for tmpfs.", nil), - 192: syscalls.PartiallySupported("lgetxattr", LGetXattr, "Only supported for tmpfs.", nil), - 193: syscalls.PartiallySupported("fgetxattr", FGetXattr, "Only supported for tmpfs.", nil), - 194: syscalls.PartiallySupported("listxattr", ListXattr, "Only supported for tmpfs", nil), - 195: syscalls.PartiallySupported("llistxattr", LListXattr, "Only supported for tmpfs", nil), - 196: syscalls.PartiallySupported("flistxattr", FListXattr, "Only supported for tmpfs", nil), - 197: syscalls.PartiallySupported("removexattr", RemoveXattr, "Only supported for tmpfs", nil), - 198: syscalls.PartiallySupported("lremovexattr", LRemoveXattr, "Only supported for tmpfs", nil), - 199: syscalls.PartiallySupported("fremovexattr", FRemoveXattr, "Only supported for tmpfs", nil), - 200: syscalls.Supported("tkill", Tkill), - 201: syscalls.Supported("time", Time), - 202: syscalls.PartiallySupported("futex", Futex, "Robust futexes not supported.", nil), - 203: syscalls.PartiallySupported("sched_setaffinity", SchedSetaffinity, "Stub implementation.", nil), - 204: syscalls.PartiallySupported("sched_getaffinity", SchedGetaffinity, "Stub implementation.", nil), - 205: syscalls.Error("set_thread_area", syserror.ENOSYS, "Expected to return ENOSYS on 64-bit", nil), - 206: syscalls.PartiallySupported("io_setup", IoSetup, "Generally supported with exceptions. User ring optimizations are not implemented.", []string{"gvisor.dev/issue/204"}), - 207: syscalls.PartiallySupported("io_destroy", IoDestroy, "Generally supported with exceptions. User ring optimizations are not implemented.", []string{"gvisor.dev/issue/204"}), - 208: syscalls.PartiallySupported("io_getevents", IoGetevents, "Generally supported with exceptions. User ring optimizations are not implemented.", []string{"gvisor.dev/issue/204"}), - 209: syscalls.PartiallySupported("io_submit", IoSubmit, "Generally supported with exceptions. User ring optimizations are not implemented.", []string{"gvisor.dev/issue/204"}), - 210: syscalls.PartiallySupported("io_cancel", IoCancel, "Generally supported with exceptions. User ring optimizations are not implemented.", []string{"gvisor.dev/issue/204"}), - 211: syscalls.Error("get_thread_area", syserror.ENOSYS, "Expected to return ENOSYS on 64-bit", nil), - 212: syscalls.CapError("lookup_dcookie", linux.CAP_SYS_ADMIN, "", nil), - 213: syscalls.Supported("epoll_create", EpollCreate), - 214: syscalls.ErrorWithEvent("epoll_ctl_old", syserror.ENOSYS, "Deprecated.", nil), - 215: syscalls.ErrorWithEvent("epoll_wait_old", syserror.ENOSYS, "Deprecated.", nil), - 216: syscalls.ErrorWithEvent("remap_file_pages", syserror.ENOSYS, "Deprecated since Linux 3.16.", nil), - 217: syscalls.Supported("getdents64", Getdents64), - 218: syscalls.Supported("set_tid_address", SetTidAddress), - 219: syscalls.Supported("restart_syscall", RestartSyscall), - 220: syscalls.ErrorWithEvent("semtimedop", syserror.ENOSYS, "", []string{"gvisor.dev/issue/137"}), - 221: syscalls.PartiallySupported("fadvise64", Fadvise64, "Not all options are supported.", nil), - 222: syscalls.Supported("timer_create", TimerCreate), - 223: syscalls.Supported("timer_settime", TimerSettime), - 224: syscalls.Supported("timer_gettime", TimerGettime), - 225: syscalls.Supported("timer_getoverrun", TimerGetoverrun), - 226: syscalls.Supported("timer_delete", TimerDelete), - 227: syscalls.Supported("clock_settime", ClockSettime), - 228: syscalls.Supported("clock_gettime", ClockGettime), - 229: syscalls.Supported("clock_getres", ClockGetres), - 230: syscalls.Supported("clock_nanosleep", ClockNanosleep), - 231: syscalls.Supported("exit_group", ExitGroup), - 232: syscalls.Supported("epoll_wait", EpollWait), - 233: syscalls.Supported("epoll_ctl", EpollCtl), - 234: syscalls.Supported("tgkill", Tgkill), - 235: syscalls.Supported("utimes", Utimes), - 236: syscalls.Error("vserver", syserror.ENOSYS, "Not implemented by Linux", nil), - 237: syscalls.PartiallySupported("mbind", Mbind, "Stub implementation. Only a single NUMA node is advertised, and mempolicy is ignored accordingly, but mbind() will succeed and has effects reflected by get_mempolicy.", []string{"gvisor.dev/issue/262"}), - 238: syscalls.PartiallySupported("set_mempolicy", SetMempolicy, "Stub implementation.", nil), - 239: syscalls.PartiallySupported("get_mempolicy", GetMempolicy, "Stub implementation.", nil), - 240: syscalls.ErrorWithEvent("mq_open", syserror.ENOSYS, "", []string{"gvisor.dev/issue/136"}), // TODO(b/29354921) - 241: syscalls.ErrorWithEvent("mq_unlink", syserror.ENOSYS, "", []string{"gvisor.dev/issue/136"}), // TODO(b/29354921) - 242: syscalls.ErrorWithEvent("mq_timedsend", syserror.ENOSYS, "", []string{"gvisor.dev/issue/136"}), // TODO(b/29354921) - 243: syscalls.ErrorWithEvent("mq_timedreceive", syserror.ENOSYS, "", []string{"gvisor.dev/issue/136"}), // TODO(b/29354921) - 244: syscalls.ErrorWithEvent("mq_notify", syserror.ENOSYS, "", []string{"gvisor.dev/issue/136"}), // TODO(b/29354921) - 245: syscalls.ErrorWithEvent("mq_getsetattr", syserror.ENOSYS, "", []string{"gvisor.dev/issue/136"}), // TODO(b/29354921) - 246: syscalls.CapError("kexec_load", linux.CAP_SYS_BOOT, "", nil), - 247: syscalls.Supported("waitid", Waitid), - 248: syscalls.Error("add_key", syserror.EACCES, "Not available to user.", nil), - 249: syscalls.Error("request_key", syserror.EACCES, "Not available to user.", nil), - 250: syscalls.Error("keyctl", syserror.EACCES, "Not available to user.", nil), - 251: syscalls.CapError("ioprio_set", linux.CAP_SYS_ADMIN, "", nil), // requires cap_sys_nice or cap_sys_admin (depending) - 252: syscalls.CapError("ioprio_get", linux.CAP_SYS_ADMIN, "", nil), // requires cap_sys_nice or cap_sys_admin (depending) - 253: syscalls.PartiallySupported("inotify_init", InotifyInit, "inotify events are only available inside the sandbox.", nil), - 254: syscalls.PartiallySupported("inotify_add_watch", InotifyAddWatch, "inotify events are only available inside the sandbox.", nil), - 255: syscalls.PartiallySupported("inotify_rm_watch", InotifyRmWatch, "inotify events are only available inside the sandbox.", nil), - 256: syscalls.CapError("migrate_pages", linux.CAP_SYS_NICE, "", nil), - 257: syscalls.Supported("openat", Openat), - 258: syscalls.Supported("mkdirat", Mkdirat), - 259: syscalls.Supported("mknodat", Mknodat), - 260: syscalls.Supported("fchownat", Fchownat), - 261: syscalls.Supported("futimesat", Futimesat), - 262: syscalls.Supported("fstatat", Fstatat), - 263: syscalls.Supported("unlinkat", Unlinkat), - 264: syscalls.Supported("renameat", Renameat), - 265: syscalls.Supported("linkat", Linkat), - 266: syscalls.Supported("symlinkat", Symlinkat), - 267: syscalls.Supported("readlinkat", Readlinkat), - 268: syscalls.Supported("fchmodat", Fchmodat), - 269: syscalls.Supported("faccessat", Faccessat), - 270: syscalls.Supported("pselect", Pselect), - 271: syscalls.Supported("ppoll", Ppoll), - 272: syscalls.PartiallySupported("unshare", Unshare, "Mount, cgroup namespaces not supported. Network namespaces supported but must be empty.", nil), - 273: syscalls.Error("set_robust_list", syserror.ENOSYS, "Obsolete.", nil), - 274: syscalls.Error("get_robust_list", syserror.ENOSYS, "Obsolete.", nil), - 275: syscalls.Supported("splice", Splice), - 276: syscalls.Supported("tee", Tee), - 277: syscalls.PartiallySupported("sync_file_range", SyncFileRange, "Full data flush is not guaranteed at this time.", nil), - 278: syscalls.ErrorWithEvent("vmsplice", syserror.ENOSYS, "", []string{"gvisor.dev/issue/138"}), // TODO(b/29354098) - 279: syscalls.CapError("move_pages", linux.CAP_SYS_NICE, "", nil), // requires cap_sys_nice (mostly) - 280: syscalls.Supported("utimensat", Utimensat), - 281: syscalls.Supported("epoll_pwait", EpollPwait), - 282: syscalls.PartiallySupported("signalfd", Signalfd, "Semantics are slightly different.", []string{"gvisor.dev/issue/139"}), - 283: syscalls.Supported("timerfd_create", TimerfdCreate), - 284: syscalls.Supported("eventfd", Eventfd), - 285: syscalls.PartiallySupported("fallocate", Fallocate, "Not all options are supported.", nil), - 286: syscalls.Supported("timerfd_settime", TimerfdSettime), - 287: syscalls.Supported("timerfd_gettime", TimerfdGettime), - 288: syscalls.Supported("accept4", Accept4), - 289: syscalls.PartiallySupported("signalfd4", Signalfd4, "Semantics are slightly different.", []string{"gvisor.dev/issue/139"}), - 290: syscalls.Supported("eventfd2", Eventfd2), - 291: syscalls.Supported("epoll_create1", EpollCreate1), - 292: syscalls.Supported("dup3", Dup3), - 293: syscalls.Supported("pipe2", Pipe2), - 294: syscalls.Supported("inotify_init1", InotifyInit1), - 295: syscalls.Supported("preadv", Preadv), - 296: syscalls.Supported("pwritev", Pwritev), - 297: syscalls.Supported("rt_tgsigqueueinfo", RtTgsigqueueinfo), - 298: syscalls.ErrorWithEvent("perf_event_open", syserror.ENODEV, "No support for perf counters", nil), - 299: syscalls.PartiallySupported("recvmmsg", RecvMMsg, "Not all flags and control messages are supported.", nil), - 300: syscalls.ErrorWithEvent("fanotify_init", syserror.ENOSYS, "Needs CONFIG_FANOTIFY", nil), - 301: syscalls.ErrorWithEvent("fanotify_mark", syserror.ENOSYS, "Needs CONFIG_FANOTIFY", nil), - 302: syscalls.Supported("prlimit64", Prlimit64), - 303: syscalls.Error("name_to_handle_at", syserror.EOPNOTSUPP, "Not supported by gVisor filesystems", nil), - 304: syscalls.Error("open_by_handle_at", syserror.EOPNOTSUPP, "Not supported by gVisor filesystems", nil), - 305: syscalls.CapError("clock_adjtime", linux.CAP_SYS_TIME, "", nil), - 306: syscalls.PartiallySupported("syncfs", Syncfs, "Depends on backing file system.", nil), - 307: syscalls.PartiallySupported("sendmmsg", SendMMsg, "Not all flags and control messages are supported.", nil), - 308: syscalls.ErrorWithEvent("setns", syserror.EOPNOTSUPP, "Needs filesystem support", []string{"gvisor.dev/issue/140"}), // TODO(b/29354995) - 309: syscalls.Supported("getcpu", Getcpu), - 310: syscalls.ErrorWithEvent("process_vm_readv", syserror.ENOSYS, "", []string{"gvisor.dev/issue/158"}), - 311: syscalls.ErrorWithEvent("process_vm_writev", syserror.ENOSYS, "", []string{"gvisor.dev/issue/158"}), - 312: syscalls.CapError("kcmp", linux.CAP_SYS_PTRACE, "", nil), - 313: syscalls.CapError("finit_module", linux.CAP_SYS_MODULE, "", nil), - 314: syscalls.ErrorWithEvent("sched_setattr", syserror.ENOSYS, "gVisor does not implement a scheduler.", []string{"gvisor.dev/issue/264"}), // TODO(b/118902272) - 315: syscalls.ErrorWithEvent("sched_getattr", syserror.ENOSYS, "gVisor does not implement a scheduler.", []string{"gvisor.dev/issue/264"}), // TODO(b/118902272) - 316: syscalls.ErrorWithEvent("renameat2", syserror.ENOSYS, "", []string{"gvisor.dev/issue/263"}), // TODO(b/118902772) - 317: syscalls.Supported("seccomp", Seccomp), - 318: syscalls.Supported("getrandom", GetRandom), - 319: syscalls.Supported("memfd_create", MemfdCreate), - 320: syscalls.CapError("kexec_file_load", linux.CAP_SYS_BOOT, "", nil), - 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) - 325: syscalls.PartiallySupported("mlock2", Mlock2, "Stub implementation. The sandbox lacks appropriate permissions.", nil), - - // Syscalls implemented after 325 are "backports" from versions - // of Linux after 4.4. - 326: syscalls.ErrorWithEvent("copy_file_range", syserror.ENOSYS, "", nil), - 327: syscalls.Supported("preadv2", Preadv2), - 328: syscalls.PartiallySupported("pwritev2", Pwritev2, "Flag RWF_HIPRI is not supported.", nil), - 329: syscalls.ErrorWithEvent("pkey_mprotect", syserror.ENOSYS, "", nil), - 330: syscalls.ErrorWithEvent("pkey_alloc", syserror.ENOSYS, "", nil), - 331: syscalls.ErrorWithEvent("pkey_free", syserror.ENOSYS, "", nil), - 332: syscalls.Supported("statx", Statx), - 333: syscalls.ErrorWithEvent("io_pgetevents", syserror.ENOSYS, "", nil), - 334: syscalls.PartiallySupported("rseq", RSeq, "Not supported on all platforms.", nil), - - // Linux skips ahead to syscall 424 to sync numbers between arches. - 424: syscalls.ErrorWithEvent("pidfd_send_signal", syserror.ENOSYS, "", nil), - 425: syscalls.ErrorWithEvent("io_uring_setup", syserror.ENOSYS, "", nil), - 426: syscalls.ErrorWithEvent("io_uring_enter", syserror.ENOSYS, "", nil), - 427: syscalls.ErrorWithEvent("io_uring_register", syserror.ENOSYS, "", nil), - 428: syscalls.ErrorWithEvent("open_tree", syserror.ENOSYS, "", nil), - 429: syscalls.ErrorWithEvent("move_mount", syserror.ENOSYS, "", nil), - 430: syscalls.ErrorWithEvent("fsopen", syserror.ENOSYS, "", nil), - 431: syscalls.ErrorWithEvent("fsconfig", syserror.ENOSYS, "", nil), - 432: syscalls.ErrorWithEvent("fsmount", syserror.ENOSYS, "", nil), - 433: syscalls.ErrorWithEvent("fspick", syserror.ENOSYS, "", nil), - 434: syscalls.ErrorWithEvent("pidfd_open", syserror.ENOSYS, "", nil), - 435: syscalls.ErrorWithEvent("clone3", syserror.ENOSYS, "", nil), - }, - - Emulate: map[usermem.Addr]uintptr{ - 0xffffffffff600000: 96, // vsyscall gettimeofday(2) - 0xffffffffff600400: 201, // vsyscall time(2) - 0xffffffffff600800: 309, // vsyscall getcpu(2) - }, - Missing: func(t *kernel.Task, sysno uintptr, args arch.SyscallArguments) (uintptr, error) { - t.Kernel().EmitUnimplementedEvent(t) - return 0, syserror.ENOSYS - }, -} diff --git a/pkg/sentry/syscalls/linux/linux64_arm64.go b/pkg/sentry/syscalls/linux/linux64_arm64.go deleted file mode 100644 index 7421619de..000000000 --- a/pkg/sentry/syscalls/linux/linux64_arm64.go +++ /dev/null @@ -1,340 +0,0 @@ -// Copyright 2019 The gVisor Authors. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -package linux - -import ( - "gvisor.dev/gvisor/pkg/abi" - "gvisor.dev/gvisor/pkg/abi/linux" - "gvisor.dev/gvisor/pkg/sentry/arch" - "gvisor.dev/gvisor/pkg/sentry/kernel" - "gvisor.dev/gvisor/pkg/sentry/syscalls" - "gvisor.dev/gvisor/pkg/syserror" - "gvisor.dev/gvisor/pkg/usermem" -) - -// ARM64 is a table of Linux arm64 syscall API with the corresponding syscall -// numbers from Linux 4.4. -var ARM64 = &kernel.SyscallTable{ - OS: abi.Linux, - Arch: arch.ARM64, - Version: kernel.Version{ - Sysname: LinuxSysname, - Release: LinuxRelease, - Version: LinuxVersion, - }, - AuditNumber: linux.AUDIT_ARCH_AARCH64, - Table: map[uintptr]kernel.Syscall{ - 0: syscalls.PartiallySupported("io_setup", IoSetup, "Generally supported with exceptions. User ring optimizations are not implemented.", []string{"gvisor.dev/issue/204"}), - 1: syscalls.PartiallySupported("io_destroy", IoDestroy, "Generally supported with exceptions. User ring optimizations are not implemented.", []string{"gvisor.dev/issue/204"}), - 2: syscalls.PartiallySupported("io_submit", IoSubmit, "Generally supported with exceptions. User ring optimizations are not implemented.", []string{"gvisor.dev/issue/204"}), - 3: syscalls.PartiallySupported("io_cancel", IoCancel, "Generally supported with exceptions. User ring optimizations are not implemented.", []string{"gvisor.dev/issue/204"}), - 4: syscalls.PartiallySupported("io_getevents", IoGetevents, "Generally supported with exceptions. User ring optimizations are not implemented.", []string{"gvisor.dev/issue/204"}), - 5: syscalls.PartiallySupported("setxattr", SetXattr, "Only supported for tmpfs.", nil), - 6: syscalls.PartiallySupported("lsetxattr", LSetXattr, "Only supported for tmpfs.", nil), - 7: syscalls.PartiallySupported("fsetxattr", FSetXattr, "Only supported for tmpfs.", nil), - 8: syscalls.PartiallySupported("getxattr", GetXattr, "Only supported for tmpfs.", nil), - 9: syscalls.PartiallySupported("lgetxattr", LGetXattr, "Only supported for tmpfs.", nil), - 10: syscalls.PartiallySupported("fgetxattr", FGetXattr, "Only supported for tmpfs.", nil), - 11: syscalls.PartiallySupported("listxattr", ListXattr, "Only supported for tmpfs", nil), - 12: syscalls.PartiallySupported("llistxattr", LListXattr, "Only supported for tmpfs", nil), - 13: syscalls.PartiallySupported("flistxattr", FListXattr, "Only supported for tmpfs", nil), - 14: syscalls.PartiallySupported("removexattr", RemoveXattr, "Only supported for tmpfs", nil), - 15: syscalls.PartiallySupported("lremovexattr", LRemoveXattr, "Only supported for tmpfs", nil), - 16: syscalls.PartiallySupported("fremovexattr", FRemoveXattr, "Only supported for tmpfs", nil), - 17: syscalls.Supported("getcwd", Getcwd), - 18: syscalls.CapError("lookup_dcookie", linux.CAP_SYS_ADMIN, "", nil), - 19: syscalls.Supported("eventfd2", Eventfd2), - 20: syscalls.Supported("epoll_create1", EpollCreate1), - 21: syscalls.Supported("epoll_ctl", EpollCtl), - 22: syscalls.Supported("epoll_pwait", EpollPwait), - 23: syscalls.Supported("dup", Dup), - 24: syscalls.Supported("dup3", Dup3), - 25: syscalls.PartiallySupported("fcntl", Fcntl, "Not all options are supported.", nil), - 26: syscalls.Supported("inotify_init1", InotifyInit1), - 27: syscalls.PartiallySupported("inotify_add_watch", InotifyAddWatch, "inotify events are only available inside the sandbox.", nil), - 28: syscalls.PartiallySupported("inotify_rm_watch", InotifyRmWatch, "inotify events are only available inside the sandbox.", nil), - 29: syscalls.PartiallySupported("ioctl", Ioctl, "Only a few ioctls are implemented for backing devices and file systems.", nil), - 30: syscalls.CapError("ioprio_set", linux.CAP_SYS_ADMIN, "", nil), // requires cap_sys_nice or cap_sys_admin (depending) - 31: syscalls.CapError("ioprio_get", linux.CAP_SYS_ADMIN, "", nil), // requires cap_sys_nice or cap_sys_admin (depending) - 32: syscalls.PartiallySupported("flock", Flock, "Locks are held within the sandbox only.", nil), - 33: syscalls.Supported("mknodat", Mknodat), - 34: syscalls.Supported("mkdirat", Mkdirat), - 35: syscalls.Supported("unlinkat", Unlinkat), - 36: syscalls.Supported("symlinkat", Symlinkat), - 37: syscalls.Supported("linkat", Linkat), - 38: syscalls.Supported("renameat", Renameat), - 39: syscalls.PartiallySupported("umount2", Umount2, "Not all options or file systems are supported.", nil), - 40: syscalls.PartiallySupported("mount", Mount, "Not all options or file systems are supported.", nil), - 41: syscalls.Error("pivot_root", syserror.EPERM, "", nil), - 42: syscalls.Error("nfsservctl", syserror.ENOSYS, "Removed after Linux 3.1.", nil), - 43: syscalls.PartiallySupported("statfs", Statfs, "Depends on the backing file system implementation.", nil), - 44: syscalls.PartiallySupported("fstatfs", Fstatfs, "Depends on the backing file system implementation.", nil), - 45: syscalls.Supported("truncate", Truncate), - 46: syscalls.Supported("ftruncate", Ftruncate), - 47: syscalls.PartiallySupported("fallocate", Fallocate, "Not all options are supported.", nil), - 48: syscalls.Supported("faccessat", Faccessat), - 49: syscalls.Supported("chdir", Chdir), - 50: syscalls.Supported("fchdir", Fchdir), - 51: syscalls.Supported("chroot", Chroot), - 52: syscalls.PartiallySupported("fchmod", Fchmod, "Options S_ISUID and S_ISGID not supported.", nil), - 53: syscalls.Supported("fchmodat", Fchmodat), - 54: syscalls.Supported("fchownat", Fchownat), - 55: syscalls.Supported("fchown", Fchown), - 56: syscalls.Supported("openat", Openat), - 57: syscalls.Supported("close", Close), - 58: syscalls.CapError("vhangup", linux.CAP_SYS_TTY_CONFIG, "", nil), - 59: syscalls.Supported("pipe2", Pipe2), - 60: syscalls.CapError("quotactl", linux.CAP_SYS_ADMIN, "", nil), // requires cap_sys_admin for most operations - 61: syscalls.Supported("getdents64", Getdents64), - 62: syscalls.Supported("lseek", Lseek), - 63: syscalls.Supported("read", Read), - 64: syscalls.Supported("write", Write), - 65: syscalls.Supported("readv", Readv), - 66: syscalls.Supported("writev", Writev), - 67: syscalls.Supported("pread64", Pread64), - 68: syscalls.Supported("pwrite64", Pwrite64), - 69: syscalls.Supported("preadv", Preadv), - 70: syscalls.Supported("pwritev", Pwritev), - 71: syscalls.Supported("sendfile", Sendfile), - 72: syscalls.Supported("pselect", Pselect), - 73: syscalls.Supported("ppoll", Ppoll), - 74: syscalls.PartiallySupported("signalfd4", Signalfd4, "Semantics are slightly different.", []string{"gvisor.dev/issue/139"}), - 75: syscalls.ErrorWithEvent("vmsplice", syserror.ENOSYS, "", []string{"gvisor.dev/issue/138"}), // TODO(b/29354098) - 76: syscalls.PartiallySupported("splice", Splice, "Stub implementation.", []string{"gvisor.dev/issue/138"}), // TODO(b/29354098) - 77: syscalls.Supported("tee", Tee), - 78: syscalls.Supported("readlinkat", Readlinkat), - 79: syscalls.Supported("fstatat", Fstatat), - 80: syscalls.Supported("fstat", Fstat), - 81: syscalls.PartiallySupported("sync", Sync, "Full data flush is not guaranteed at this time.", nil), - 82: syscalls.PartiallySupported("fsync", Fsync, "Full data flush is not guaranteed at this time.", nil), - 83: syscalls.PartiallySupported("fdatasync", Fdatasync, "Full data flush is not guaranteed at this time.", nil), - 84: syscalls.PartiallySupported("sync_file_range", SyncFileRange, "Full data flush is not guaranteed at this time.", nil), - 85: syscalls.Supported("timerfd_create", TimerfdCreate), - 86: syscalls.Supported("timerfd_settime", TimerfdSettime), - 87: syscalls.Supported("timerfd_gettime", TimerfdGettime), - 88: syscalls.Supported("utimensat", Utimensat), - 89: syscalls.CapError("acct", linux.CAP_SYS_PACCT, "", nil), - 90: syscalls.Supported("capget", Capget), - 91: syscalls.Supported("capset", Capset), - 92: syscalls.ErrorWithEvent("personality", syserror.EINVAL, "Unable to change personality.", nil), - 93: syscalls.Supported("exit", Exit), - 94: syscalls.Supported("exit_group", ExitGroup), - 95: syscalls.Supported("waitid", Waitid), - 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), - 101: syscalls.Supported("nanosleep", Nanosleep), - 102: syscalls.Supported("getitimer", Getitimer), - 103: syscalls.Supported("setitimer", Setitimer), - 104: syscalls.CapError("kexec_load", linux.CAP_SYS_BOOT, "", nil), - 105: syscalls.CapError("init_module", linux.CAP_SYS_MODULE, "", nil), - 106: syscalls.CapError("delete_module", linux.CAP_SYS_MODULE, "", nil), - 107: syscalls.Supported("timer_create", TimerCreate), - 108: syscalls.Supported("timer_gettime", TimerGettime), - 109: syscalls.Supported("timer_getoverrun", TimerGetoverrun), - 110: syscalls.Supported("timer_settime", TimerSettime), - 111: syscalls.Supported("timer_delete", TimerDelete), - 112: syscalls.Supported("clock_settime", ClockSettime), - 113: syscalls.Supported("clock_gettime", ClockGettime), - 114: syscalls.Supported("clock_getres", ClockGetres), - 115: syscalls.Supported("clock_nanosleep", ClockNanosleep), - 116: syscalls.PartiallySupported("syslog", Syslog, "Outputs a dummy message for security reasons.", nil), - 117: syscalls.PartiallySupported("ptrace", Ptrace, "Options PTRACE_PEEKSIGINFO, PTRACE_SECCOMP_GET_FILTER not supported.", nil), - 118: syscalls.CapError("sched_setparam", linux.CAP_SYS_NICE, "", nil), - 119: syscalls.PartiallySupported("sched_setscheduler", SchedSetscheduler, "Stub implementation.", nil), - 120: syscalls.PartiallySupported("sched_getscheduler", SchedGetscheduler, "Stub implementation.", nil), - 121: syscalls.PartiallySupported("sched_getparam", SchedGetparam, "Stub implementation.", nil), - 122: syscalls.PartiallySupported("sched_setaffinity", SchedSetaffinity, "Stub implementation.", nil), - 123: syscalls.PartiallySupported("sched_getaffinity", SchedGetaffinity, "Stub implementation.", nil), - 124: syscalls.Supported("sched_yield", SchedYield), - 125: syscalls.PartiallySupported("sched_get_priority_max", SchedGetPriorityMax, "Stub implementation.", nil), - 126: syscalls.PartiallySupported("sched_get_priority_min", SchedGetPriorityMin, "Stub implementation.", nil), - 127: syscalls.ErrorWithEvent("sched_rr_get_interval", syserror.EPERM, "", nil), - 128: syscalls.Supported("restart_syscall", RestartSyscall), - 129: syscalls.Supported("kill", Kill), - 130: syscalls.Supported("tkill", Tkill), - 131: syscalls.Supported("tgkill", Tgkill), - 132: syscalls.Supported("sigaltstack", Sigaltstack), - 133: syscalls.Supported("rt_sigsuspend", RtSigsuspend), - 134: syscalls.Supported("rt_sigaction", RtSigaction), - 135: syscalls.Supported("rt_sigprocmask", RtSigprocmask), - 136: syscalls.Supported("rt_sigpending", RtSigpending), - 137: syscalls.Supported("rt_sigtimedwait", RtSigtimedwait), - 138: syscalls.Supported("rt_sigqueueinfo", RtSigqueueinfo), - 139: syscalls.Supported("rt_sigreturn", RtSigreturn), - 140: syscalls.PartiallySupported("setpriority", Setpriority, "Stub implementation.", nil), - 141: syscalls.PartiallySupported("getpriority", Getpriority, "Stub implementation.", nil), - 142: syscalls.CapError("reboot", linux.CAP_SYS_BOOT, "", nil), - 143: syscalls.Supported("setregid", Setregid), - 144: syscalls.Supported("setgid", Setgid), - 145: syscalls.Supported("setreuid", Setreuid), - 146: syscalls.Supported("setuid", Setuid), - 147: syscalls.Supported("setresuid", Setresuid), - 148: syscalls.Supported("getresuid", Getresuid), - 149: syscalls.Supported("setresgid", Setresgid), - 150: syscalls.Supported("getresgid", Getresgid), - 151: syscalls.ErrorWithEvent("setfsuid", syserror.ENOSYS, "", []string{"gvisor.dev/issue/260"}), // TODO(b/112851702) - 152: syscalls.ErrorWithEvent("setfsgid", syserror.ENOSYS, "", []string{"gvisor.dev/issue/260"}), // TODO(b/112851702) - 153: syscalls.Supported("times", Times), - 154: syscalls.Supported("setpgid", Setpgid), - 155: syscalls.Supported("getpgid", Getpgid), - 156: syscalls.Supported("getsid", Getsid), - 157: syscalls.Supported("setsid", Setsid), - 158: syscalls.Supported("getgroups", Getgroups), - 159: syscalls.Supported("setgroups", Setgroups), - 160: syscalls.Supported("uname", Uname), - 161: syscalls.Supported("sethostname", Sethostname), - 162: syscalls.Supported("setdomainname", Setdomainname), - 163: syscalls.Supported("getrlimit", Getrlimit), - 164: syscalls.PartiallySupported("setrlimit", Setrlimit, "Not all rlimits are enforced.", nil), - 165: syscalls.PartiallySupported("getrusage", Getrusage, "Fields ru_maxrss, ru_minflt, ru_majflt, ru_inblock, ru_oublock are not supported. Fields ru_utime and ru_stime have low precision.", nil), - 166: syscalls.Supported("umask", Umask), - 167: syscalls.PartiallySupported("prctl", Prctl, "Not all options are supported.", nil), - 168: syscalls.Supported("getcpu", Getcpu), - 169: syscalls.Supported("gettimeofday", Gettimeofday), - 170: syscalls.CapError("settimeofday", linux.CAP_SYS_TIME, "", nil), - 171: syscalls.CapError("adjtimex", linux.CAP_SYS_TIME, "", nil), - 172: syscalls.Supported("getpid", Getpid), - 173: syscalls.Supported("getppid", Getppid), - 174: syscalls.Supported("getuid", Getuid), - 175: syscalls.Supported("geteuid", Geteuid), - 176: syscalls.Supported("getgid", Getgid), - 177: syscalls.Supported("getegid", Getegid), - 178: syscalls.Supported("gettid", Gettid), - 179: syscalls.PartiallySupported("sysinfo", Sysinfo, "Fields loads, sharedram, bufferram, totalswap, freeswap, totalhigh, freehigh not supported.", nil), - 180: syscalls.ErrorWithEvent("mq_open", syserror.ENOSYS, "", []string{"gvisor.dev/issue/136"}), // TODO(b/29354921) - 181: syscalls.ErrorWithEvent("mq_unlink", syserror.ENOSYS, "", []string{"gvisor.dev/issue/136"}), // TODO(b/29354921) - 182: syscalls.ErrorWithEvent("mq_timedsend", syserror.ENOSYS, "", []string{"gvisor.dev/issue/136"}), // TODO(b/29354921) - 183: syscalls.ErrorWithEvent("mq_timedreceive", syserror.ENOSYS, "", []string{"gvisor.dev/issue/136"}), // TODO(b/29354921) - 184: syscalls.ErrorWithEvent("mq_notify", syserror.ENOSYS, "", []string{"gvisor.dev/issue/136"}), // TODO(b/29354921) - 185: syscalls.ErrorWithEvent("mq_getsetattr", syserror.ENOSYS, "", []string{"gvisor.dev/issue/136"}), // TODO(b/29354921) - 186: syscalls.ErrorWithEvent("msgget", syserror.ENOSYS, "", []string{"gvisor.dev/issue/135"}), // TODO(b/29354921) - 187: syscalls.ErrorWithEvent("msgctl", syserror.ENOSYS, "", []string{"gvisor.dev/issue/135"}), // TODO(b/29354921) - 188: syscalls.ErrorWithEvent("msgrcv", syserror.ENOSYS, "", []string{"gvisor.dev/issue/135"}), // TODO(b/29354921) - 189: syscalls.ErrorWithEvent("msgsnd", syserror.ENOSYS, "", []string{"gvisor.dev/issue/135"}), // TODO(b/29354921) - 190: syscalls.Supported("semget", Semget), - 191: syscalls.PartiallySupported("semctl", Semctl, "Options IPC_INFO, SEM_INFO, IPC_STAT, SEM_STAT, SEM_STAT_ANY, GETNCNT, GETZCNT not supported.", nil), - 192: syscalls.ErrorWithEvent("semtimedop", syserror.ENOSYS, "", []string{"gvisor.dev/issue/137"}), - 193: syscalls.PartiallySupported("semop", Semop, "Option SEM_UNDO not supported.", nil), - 194: syscalls.PartiallySupported("shmget", Shmget, "Option SHM_HUGETLB is not supported.", nil), - 195: syscalls.PartiallySupported("shmctl", Shmctl, "Options SHM_LOCK, SHM_UNLOCK are not supported.", nil), - 196: syscalls.PartiallySupported("shmat", Shmat, "Option SHM_RND is not supported.", nil), - 197: syscalls.Supported("shmdt", Shmdt), - 198: syscalls.PartiallySupported("socket", Socket, "Limited support for AF_NETLINK, NETLINK_ROUTE sockets. Limited support for SOCK_RAW.", nil), - 199: syscalls.Supported("socketpair", SocketPair), - 200: syscalls.PartiallySupported("bind", Bind, "Autobind for abstract Unix sockets is not supported.", nil), - 201: syscalls.Supported("listen", Listen), - 202: syscalls.Supported("accept", Accept), - 203: syscalls.Supported("connect", Connect), - 204: syscalls.Supported("getsockname", GetSockName), - 205: syscalls.Supported("getpeername", GetPeerName), - 206: syscalls.Supported("sendto", SendTo), - 207: syscalls.Supported("recvfrom", RecvFrom), - 208: syscalls.PartiallySupported("setsockopt", SetSockOpt, "Not all socket options are supported.", nil), - 209: syscalls.PartiallySupported("getsockopt", GetSockOpt, "Not all socket options are supported.", nil), - 210: syscalls.PartiallySupported("shutdown", Shutdown, "Not all flags and control messages are supported.", nil), - 211: syscalls.Supported("sendmsg", SendMsg), - 212: syscalls.PartiallySupported("recvmsg", RecvMsg, "Not all flags and control messages are supported.", nil), - 213: syscalls.Supported("readahead", Readahead), - 214: syscalls.Supported("brk", Brk), - 215: syscalls.Supported("munmap", Munmap), - 216: syscalls.Supported("mremap", Mremap), - 217: syscalls.Error("add_key", syserror.EACCES, "Not available to user.", nil), - 218: syscalls.Error("request_key", syserror.EACCES, "Not available to user.", nil), - 219: syscalls.Error("keyctl", syserror.EACCES, "Not available to user.", nil), - 220: syscalls.PartiallySupported("clone", Clone, "Mount namespace (CLONE_NEWNS) not supported. Options CLONE_PARENT, CLONE_SYSVSEM not supported.", nil), - 221: syscalls.Supported("execve", Execve), - 222: syscalls.PartiallySupported("mmap", Mmap, "Generally supported with exceptions. Options MAP_FIXED_NOREPLACE, MAP_SHARED_VALIDATE, MAP_SYNC MAP_GROWSDOWN, MAP_HUGETLB are not supported.", nil), - 223: syscalls.PartiallySupported("fadvise64", Fadvise64, "Not all options are supported.", nil), - 224: syscalls.CapError("swapon", linux.CAP_SYS_ADMIN, "", nil), - 225: syscalls.CapError("swapoff", linux.CAP_SYS_ADMIN, "", nil), - 226: syscalls.Supported("mprotect", Mprotect), - 227: syscalls.PartiallySupported("msync", Msync, "Full data flush is not guaranteed at this time.", nil), - 228: syscalls.PartiallySupported("mlock", Mlock, "Stub implementation. The sandbox lacks appropriate permissions.", nil), - 229: syscalls.PartiallySupported("munlock", Munlock, "Stub implementation. The sandbox lacks appropriate permissions.", nil), - 230: syscalls.PartiallySupported("mlockall", Mlockall, "Stub implementation. The sandbox lacks appropriate permissions.", nil), - 231: syscalls.PartiallySupported("munlockall", Munlockall, "Stub implementation. The sandbox lacks appropriate permissions.", nil), - 232: syscalls.PartiallySupported("mincore", Mincore, "Stub implementation. The sandbox does not have access to this information. Reports all mapped pages are resident.", nil), - 233: syscalls.PartiallySupported("madvise", Madvise, "Options MADV_DONTNEED, MADV_DONTFORK are supported. Other advice is ignored.", nil), - 234: syscalls.ErrorWithEvent("remap_file_pages", syserror.ENOSYS, "Deprecated since Linux 3.16.", nil), - 235: syscalls.PartiallySupported("mbind", Mbind, "Stub implementation. Only a single NUMA node is advertised, and mempolicy is ignored accordingly, but mbind() will succeed and has effects reflected by get_mempolicy.", []string{"gvisor.dev/issue/262"}), - 236: syscalls.PartiallySupported("get_mempolicy", GetMempolicy, "Stub implementation.", nil), - 237: syscalls.PartiallySupported("set_mempolicy", SetMempolicy, "Stub implementation.", nil), - 238: syscalls.CapError("migrate_pages", linux.CAP_SYS_NICE, "", nil), - 239: syscalls.CapError("move_pages", linux.CAP_SYS_NICE, "", nil), // requires cap_sys_nice (mostly) - 240: syscalls.Supported("rt_tgsigqueueinfo", RtTgsigqueueinfo), - 241: syscalls.ErrorWithEvent("perf_event_open", syserror.ENODEV, "No support for perf counters", nil), - 242: syscalls.Supported("accept4", Accept4), - 243: syscalls.PartiallySupported("recvmmsg", RecvMMsg, "Not all flags and control messages are supported.", nil), - 260: syscalls.Supported("wait4", Wait4), - 261: syscalls.Supported("prlimit64", Prlimit64), - 262: syscalls.ErrorWithEvent("fanotify_init", syserror.ENOSYS, "Needs CONFIG_FANOTIFY", nil), - 263: syscalls.ErrorWithEvent("fanotify_mark", syserror.ENOSYS, "Needs CONFIG_FANOTIFY", nil), - 264: syscalls.Error("name_to_handle_at", syserror.EOPNOTSUPP, "Not supported by gVisor filesystems", nil), - 265: syscalls.Error("open_by_handle_at", syserror.EOPNOTSUPP, "Not supported by gVisor filesystems", nil), - 266: syscalls.CapError("clock_adjtime", linux.CAP_SYS_TIME, "", nil), - 267: syscalls.PartiallySupported("syncfs", Syncfs, "Depends on backing file system.", nil), - 268: syscalls.ErrorWithEvent("setns", syserror.EOPNOTSUPP, "Needs filesystem support", []string{"gvisor.dev/issue/140"}), // TODO(b/29354995) - 269: syscalls.PartiallySupported("sendmmsg", SendMMsg, "Not all flags and control messages are supported.", nil), - 270: syscalls.ErrorWithEvent("process_vm_readv", syserror.ENOSYS, "", []string{"gvisor.dev/issue/158"}), - 271: syscalls.ErrorWithEvent("process_vm_writev", syserror.ENOSYS, "", []string{"gvisor.dev/issue/158"}), - 272: syscalls.CapError("kcmp", linux.CAP_SYS_PTRACE, "", nil), - 273: syscalls.CapError("finit_module", linux.CAP_SYS_MODULE, "", nil), - 274: syscalls.ErrorWithEvent("sched_setattr", syserror.ENOSYS, "gVisor does not implement a scheduler.", []string{"gvisor.dev/issue/264"}), // TODO(b/118902272) - 275: syscalls.ErrorWithEvent("sched_getattr", syserror.ENOSYS, "gVisor does not implement a scheduler.", []string{"gvisor.dev/issue/264"}), // TODO(b/118902272) - 276: syscalls.ErrorWithEvent("renameat2", syserror.ENOSYS, "", []string{"gvisor.dev/issue/263"}), // TODO(b/118902772) - 277: syscalls.Supported("seccomp", Seccomp), - 278: syscalls.Supported("getrandom", GetRandom), - 279: syscalls.Supported("memfd_create", MemfdCreate), - 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) - 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. - 285: syscalls.ErrorWithEvent("copy_file_range", syserror.ENOSYS, "", nil), - 286: syscalls.Supported("preadv2", Preadv2), - 287: syscalls.PartiallySupported("pwritev2", Pwritev2, "Flag RWF_HIPRI is not supported.", nil), - 288: syscalls.ErrorWithEvent("pkey_mprotect", syserror.ENOSYS, "", nil), - 289: syscalls.ErrorWithEvent("pkey_alloc", syserror.ENOSYS, "", nil), - 290: syscalls.ErrorWithEvent("pkey_free", syserror.ENOSYS, "", nil), - 291: syscalls.Supported("statx", Statx), - 292: syscalls.ErrorWithEvent("io_pgetevents", syserror.ENOSYS, "", nil), - 293: syscalls.PartiallySupported("rseq", RSeq, "Not supported on all platforms.", nil), - - // Linux skips ahead to syscall 424 to sync numbers between arches. - 424: syscalls.ErrorWithEvent("pidfd_send_signal", syserror.ENOSYS, "", nil), - 425: syscalls.ErrorWithEvent("io_uring_setup", syserror.ENOSYS, "", nil), - 426: syscalls.ErrorWithEvent("io_uring_enter", syserror.ENOSYS, "", nil), - 427: syscalls.ErrorWithEvent("io_uring_register", syserror.ENOSYS, "", nil), - 428: syscalls.ErrorWithEvent("open_tree", syserror.ENOSYS, "", nil), - 429: syscalls.ErrorWithEvent("move_mount", syserror.ENOSYS, "", nil), - 430: syscalls.ErrorWithEvent("fsopen", syserror.ENOSYS, "", nil), - 431: syscalls.ErrorWithEvent("fsconfig", syserror.ENOSYS, "", nil), - 432: syscalls.ErrorWithEvent("fsmount", syserror.ENOSYS, "", nil), - 433: syscalls.ErrorWithEvent("fspick", syserror.ENOSYS, "", nil), - 434: syscalls.ErrorWithEvent("pidfd_open", syserror.ENOSYS, "", nil), - 435: syscalls.ErrorWithEvent("clone3", syserror.ENOSYS, "", nil), - }, - Emulate: map[usermem.Addr]uintptr{}, - - Missing: func(t *kernel.Task, sysno uintptr, args arch.SyscallArguments) (uintptr, error) { - t.Kernel().EmitUnimplementedEvent(t) - return 0, syserror.ENOSYS - }, -} diff --git a/pkg/sentry/syscalls/linux/sigset.go b/pkg/sentry/syscalls/linux/sigset.go index 2ddb2b146..434559b80 100644 --- a/pkg/sentry/syscalls/linux/sigset.go +++ b/pkg/sentry/syscalls/linux/sigset.go @@ -21,9 +21,13 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) -// copyInSigSet copies in a sigset_t, checks its size, and ensures that KILL and +// CopyInSigSet copies in a sigset_t, checks its size, and ensures that KILL and // STOP are clear. -func copyInSigSet(t *kernel.Task, sigSetAddr usermem.Addr, size uint) (linux.SignalSet, error) { +// +// TODO(gvisor.dev/issue/1624): This is only exported because +// syscalls/vfs2/signal.go depends on it. Once vfs1 is deleted and the vfs2 +// syscalls are moved into this package, then they can be unexported. +func CopyInSigSet(t *kernel.Task, sigSetAddr usermem.Addr, size uint) (linux.SignalSet, error) { if size != linux.SignalSetSize { return 0, syserror.EINVAL } diff --git a/pkg/sentry/syscalls/linux/sys_aio.go b/pkg/sentry/syscalls/linux/sys_aio.go index b401978db..d781d6a04 100644 --- a/pkg/sentry/syscalls/linux/sys_aio.go +++ b/pkg/sentry/syscalls/linux/sys_aio.go @@ -114,14 +114,28 @@ func IoSetup(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysca func IoDestroy(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { id := args[0].Uint64() - // Destroy the given context. - if !t.MemoryManager().DestroyAIOContext(t, id) { + ctx := t.MemoryManager().DestroyAIOContext(t, id) + if ctx == nil { // Does not exist. return 0, nil, syserror.EINVAL } - // FIXME(fvoznika): Linux blocks until all AIO to the destroyed context is - // done. - return 0, nil, nil + + // Drain completed requests amd wait for pending requests until there are no + // more. + for { + ctx.Drain() + + ch := ctx.WaitChannel() + if ch == nil { + // No more requests, we're done. + return 0, nil, nil + } + // The task cannot be interrupted during the wait. Equivalent to + // TASK_UNINTERRUPTIBLE in Linux. + t.UninterruptibleSleepStart(true /* deactivate */) + <-ch + t.UninterruptibleSleepFinish(true /* activate */) + } } // IoGetevents implements linux syscall io_getevents(2). @@ -200,13 +214,13 @@ func IoGetevents(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.S func waitForRequest(ctx *mm.AIOContext, t *kernel.Task, haveDeadline bool, deadline ktime.Time) (interface{}, error) { for { if v, ok := ctx.PopRequest(); ok { - // Request was readly available. Just return it. + // Request was readily available. Just return it. return v, nil } // Need to wait for request completion. - done, active := ctx.WaitChannel() - if !active { + done := ctx.WaitChannel() + if done == nil { // Context has been destroyed. return nil, syserror.EINVAL } @@ -248,6 +262,10 @@ func memoryFor(t *kernel.Task, cb *ioCallback) (usermem.IOSequence, error) { } func performCallback(t *kernel.Task, file *fs.File, cbAddr usermem.Addr, cb *ioCallback, ioseq usermem.IOSequence, ctx *mm.AIOContext, eventFile *fs.File) { + if ctx.Dead() { + ctx.CancelPendingRequest() + return + } ev := &ioEvent{ Data: cb.Data, Obj: uint64(cbAddr), @@ -272,7 +290,7 @@ func performCallback(t *kernel.Task, file *fs.File, cbAddr usermem.Addr, cb *ioC // Update the result. if err != nil { err = handleIOError(t, ev.Result != 0 /* partial */, err, nil /* never interrupted */, "aio", file) - ev.Result = -int64(t.ExtractErrno(err, 0)) + ev.Result = -int64(kernel.ExtractErrno(err, 0)) } file.DecRef() diff --git a/pkg/sentry/syscalls/linux/sys_epoll.go b/pkg/sentry/syscalls/linux/sys_epoll.go index 5f11b496c..7f460d30b 100644 --- a/pkg/sentry/syscalls/linux/sys_epoll.go +++ b/pkg/sentry/syscalls/linux/sys_epoll.go @@ -21,10 +21,11 @@ import ( "gvisor.dev/gvisor/pkg/sentry/kernel/epoll" "gvisor.dev/gvisor/pkg/sentry/syscalls" "gvisor.dev/gvisor/pkg/syserror" - "gvisor.dev/gvisor/pkg/usermem" "gvisor.dev/gvisor/pkg/waiter" ) +// LINT.IfChange + // EpollCreate1 implements the epoll_create1(2) linux syscall. func EpollCreate1(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { flags := args[0].Int() @@ -70,7 +71,7 @@ func EpollCtl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysc var data [2]int32 if op != linux.EPOLL_CTL_DEL { var e linux.EpollEvent - if _, err := t.CopyIn(eventAddr, &e); err != nil { + if _, err := e.CopyIn(t, eventAddr); err != nil { return 0, nil, err } @@ -83,8 +84,7 @@ func EpollCtl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysc } mask = waiter.EventMaskFromLinux(e.Events) - data[0] = e.Fd - data[1] = e.Data + data = e.Data } // Perform the requested operations. @@ -104,28 +104,6 @@ func EpollCtl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysc } } -// copyOutEvents copies epoll events from the kernel to user memory. -func copyOutEvents(t *kernel.Task, addr usermem.Addr, e []epoll.Event) error { - const itemLen = 12 - buffLen := len(e) * itemLen - if _, ok := addr.AddLength(uint64(buffLen)); !ok { - return syserror.EFAULT - } - - b := t.CopyScratchBuffer(buffLen) - for i := range e { - usermem.ByteOrder.PutUint32(b[i*itemLen:], e[i].Events) - usermem.ByteOrder.PutUint32(b[i*itemLen+4:], uint32(e[i].Data[0])) - usermem.ByteOrder.PutUint32(b[i*itemLen+8:], uint32(e[i].Data[1])) - } - - if _, err := t.CopyOutBytes(addr, b); err != nil { - return err - } - - return nil -} - // EpollWait implements the epoll_wait(2) linux syscall. func EpollWait(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { epfd := args[0].Int() @@ -139,7 +117,7 @@ func EpollWait(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sys } if len(r) != 0 { - if err := copyOutEvents(t, eventsAddr, r); err != nil { + if _, err := linux.CopyEpollEventSliceOut(t, eventsAddr, r); err != nil { return 0, nil, err } } @@ -153,7 +131,7 @@ func EpollPwait(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sy maskSize := uint(args[5].Uint()) if maskAddr != 0 { - mask, err := copyInSigSet(t, maskAddr, maskSize) + mask, err := CopyInSigSet(t, maskAddr, maskSize) if err != nil { return 0, nil, err } @@ -165,3 +143,5 @@ func EpollPwait(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sy return EpollWait(t, args) } + +// LINT.ThenChange(vfs2/epoll.go) diff --git a/pkg/sentry/syscalls/linux/sys_eventfd.go b/pkg/sentry/syscalls/linux/sys_eventfd.go index 8a34c4e99..ed3413ca6 100644 --- a/pkg/sentry/syscalls/linux/sys_eventfd.go +++ b/pkg/sentry/syscalls/linux/sys_eventfd.go @@ -15,6 +15,7 @@ package linux import ( + "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/kernel" @@ -22,32 +23,24 @@ import ( "gvisor.dev/gvisor/pkg/syserror" ) -const ( - // EFD_SEMAPHORE is a flag used in syscall eventfd(2) and eventfd2(2). Please - // see its man page for more information. - EFD_SEMAPHORE = 1 - EFD_NONBLOCK = 0x800 - EFD_CLOEXEC = 0x80000 -) - // Eventfd2 implements linux syscall eventfd2(2). func Eventfd2(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { initVal := args[0].Int() flags := uint(args[1].Uint()) - allOps := uint(EFD_SEMAPHORE | EFD_NONBLOCK | EFD_CLOEXEC) + allOps := uint(linux.EFD_SEMAPHORE | linux.EFD_NONBLOCK | linux.EFD_CLOEXEC) if flags & ^allOps != 0 { return 0, nil, syserror.EINVAL } - event := eventfd.New(t, uint64(initVal), flags&EFD_SEMAPHORE != 0) + event := eventfd.New(t, uint64(initVal), flags&linux.EFD_SEMAPHORE != 0) event.SetFlags(fs.SettableFileFlags{ - NonBlocking: flags&EFD_NONBLOCK != 0, + NonBlocking: flags&linux.EFD_NONBLOCK != 0, }) defer event.DecRef() fd, err := t.NewFDFrom(0, event, kernel.FDFlags{ - CloseOnExec: flags&EFD_CLOEXEC != 0, + CloseOnExec: flags&linux.EFD_CLOEXEC != 0, }) if err != nil { return 0, nil, err diff --git a/pkg/sentry/syscalls/linux/sys_file.go b/pkg/sentry/syscalls/linux/sys_file.go index 421845ebb..35a98212a 100644 --- a/pkg/sentry/syscalls/linux/sys_file.go +++ b/pkg/sentry/syscalls/linux/sys_file.go @@ -130,6 +130,8 @@ func copyInPath(t *kernel.Task, addr usermem.Addr, allowEmpty bool) (path string return path, dirPath, nil } +// LINT.IfChange + func openAt(t *kernel.Task, dirFD int32, addr usermem.Addr, flags uint) (fd uintptr, err error) { path, dirPath, err := copyInPath(t, addr, false /* allowEmpty */) if err != nil { @@ -512,7 +514,7 @@ func (ac accessContext) Value(key interface{}) interface{} { } } -func accessAt(t *kernel.Task, dirFD int32, addr usermem.Addr, resolve bool, mode uint) error { +func accessAt(t *kernel.Task, dirFD int32, addr usermem.Addr, mode uint) error { const rOK = 4 const wOK = 2 const xOK = 1 @@ -527,7 +529,7 @@ func accessAt(t *kernel.Task, dirFD int32, addr usermem.Addr, resolve bool, mode return syserror.EINVAL } - return fileOpOn(t, dirFD, path, resolve, func(root *fs.Dirent, d *fs.Dirent, _ uint) error { + return fileOpOn(t, dirFD, path, true /* resolve */, func(root *fs.Dirent, d *fs.Dirent, _ uint) error { // access(2) and faccessat(2) check permissions using real // UID/GID, not effective UID/GID. // @@ -562,19 +564,29 @@ func Access(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal addr := args[0].Pointer() mode := args[1].ModeT() - return 0, nil, accessAt(t, linux.AT_FDCWD, addr, true, mode) + return 0, nil, accessAt(t, linux.AT_FDCWD, addr, mode) } // Faccessat implements linux syscall faccessat(2). +// +// Note that the faccessat() system call does not take a flags argument: +// "The raw faccessat() system call takes only the first three arguments. The +// AT_EACCESS and AT_SYMLINK_NOFOLLOW flags are actually implemented within +// the glibc wrapper function for faccessat(). If either of these flags is +// specified, then the wrapper function employs fstatat(2) to determine access +// permissions." - faccessat(2) func Faccessat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { dirFD := args[0].Int() addr := args[1].Pointer() mode := args[2].ModeT() - flags := args[3].Int() - return 0, nil, accessAt(t, dirFD, addr, flags&linux.AT_SYMLINK_NOFOLLOW == 0, mode) + return 0, nil, accessAt(t, dirFD, addr, mode) } +// LINT.ThenChange(vfs2/filesystem.go) + +// LINT.IfChange + // Ioctl implements linux syscall ioctl(2). func Ioctl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { fd := args[0].Int() @@ -650,6 +662,10 @@ func Ioctl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall } } +// LINT.ThenChange(vfs2/ioctl.go) + +// LINT.IfChange + // Getcwd implements the linux syscall getcwd(2). func Getcwd(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { addr := args[0].Pointer() @@ -760,6 +776,10 @@ func Fchdir(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal return 0, nil, nil } +// LINT.ThenChange(vfs2/fscontext.go) + +// LINT.IfChange + // Close implements linux syscall close(2). func Close(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { fd := args[0].Int() @@ -1094,6 +1114,8 @@ func Fcntl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall } } +// LINT.ThenChange(vfs2/fd.go) + const ( _FADV_NORMAL = 0 _FADV_RANDOM = 1 @@ -1141,6 +1163,8 @@ func Fadvise64(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sys return 0, nil, nil } +// LINT.IfChange + func mkdirAt(t *kernel.Task, dirFD int32, addr usermem.Addr, mode linux.FileMode) error { path, _, err := copyInPath(t, addr, false /* allowEmpty */) if err != nil { @@ -1218,7 +1242,7 @@ func rmdirAt(t *kernel.Task, dirFD int32, addr usermem.Addr) error { return syserror.ENOTEMPTY } - if err := fs.MayDelete(t, root, d, name); err != nil { + if err := d.MayDelete(t, root, name); err != nil { return err } @@ -1421,6 +1445,10 @@ func Linkat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal return 0, nil, linkAt(t, oldDirFD, oldAddr, newDirFD, newAddr, resolve, allowEmpty) } +// LINT.ThenChange(vfs2/filesystem.go) + +// LINT.IfChange + func readlinkAt(t *kernel.Task, dirFD int32, addr usermem.Addr, bufAddr usermem.Addr, size uint) (copied uintptr, err error) { path, dirPath, err := copyInPath(t, addr, false /* allowEmpty */) if err != nil { @@ -1480,6 +1508,10 @@ func Readlinkat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sy return n, nil, err } +// LINT.ThenChange(vfs2/stat.go) + +// LINT.IfChange + func unlinkAt(t *kernel.Task, dirFD int32, addr usermem.Addr) error { path, dirPath, err := copyInPath(t, addr, false /* allowEmpty */) if err != nil { @@ -1491,7 +1523,7 @@ func unlinkAt(t *kernel.Task, dirFD int32, addr usermem.Addr) error { return syserror.ENOTDIR } - if err := fs.MayDelete(t, root, d, name); err != nil { + if err := d.MayDelete(t, root, name); err != nil { return err } @@ -1516,6 +1548,10 @@ func Unlinkat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysc return 0, nil, unlinkAt(t, dirFD, addr) } +// LINT.ThenChange(vfs2/filesystem.go) + +// LINT.IfChange + // Truncate implements linux syscall truncate(2). func Truncate(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { addr := args[0].Pointer() @@ -1614,6 +1650,8 @@ func Ftruncate(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sys return 0, nil, nil } +// LINT.ThenChange(vfs2/setstat.go) + // Umask implements linux syscall umask(2). func Umask(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { mask := args[0].ModeT() @@ -1621,6 +1659,8 @@ func Umask(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall return uintptr(mask), nil, nil } +// LINT.IfChange + // Change ownership of a file. // // uid and gid may be -1, in which case they will not be changed. @@ -1987,6 +2027,10 @@ func Futimesat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sys return 0, nil, utimes(t, dirFD, pathnameAddr, ts, true) } +// LINT.ThenChange(vfs2/setstat.go) + +// LINT.IfChange + func renameAt(t *kernel.Task, oldDirFD int32, oldAddr usermem.Addr, newDirFD int32, newAddr usermem.Addr) error { newPath, _, err := copyInPath(t, newAddr, false /* allowEmpty */) if err != nil { @@ -2042,6 +2086,8 @@ func Renameat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysc return 0, nil, renameAt(t, oldDirFD, oldPathAddr, newDirFD, newPathAddr) } +// LINT.ThenChange(vfs2/filesystem.go) + // Fallocate implements linux system call fallocate(2). func Fallocate(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { fd := args[0].Int() diff --git a/pkg/sentry/syscalls/linux/sys_getdents.go b/pkg/sentry/syscalls/linux/sys_getdents.go index f66f4ffde..b126fecc0 100644 --- a/pkg/sentry/syscalls/linux/sys_getdents.go +++ b/pkg/sentry/syscalls/linux/sys_getdents.go @@ -27,6 +27,8 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) +// LINT.IfChange + // Getdents implements linux syscall getdents(2) for 64bit systems. func Getdents(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { fd := args[0].Int() @@ -244,3 +246,5 @@ func (ds *direntSerializer) CopyOut(name string, attr fs.DentAttr) error { func (ds *direntSerializer) Written() int { return ds.written } + +// LINT.ThenChange(vfs2/getdents.go) diff --git a/pkg/sentry/syscalls/linux/sys_lseek.go b/pkg/sentry/syscalls/linux/sys_lseek.go index 297e920c4..3f7691eae 100644 --- a/pkg/sentry/syscalls/linux/sys_lseek.go +++ b/pkg/sentry/syscalls/linux/sys_lseek.go @@ -21,6 +21,8 @@ import ( "gvisor.dev/gvisor/pkg/syserror" ) +// LINT.IfChange + // Lseek implements linux syscall lseek(2). func Lseek(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { fd := args[0].Int() @@ -52,3 +54,5 @@ func Lseek(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall } return uintptr(offset), nil, err } + +// LINT.ThenChange(vfs2/read_write.go) diff --git a/pkg/sentry/syscalls/linux/sys_mempolicy.go b/pkg/sentry/syscalls/linux/sys_mempolicy.go index ac934dc6f..9b4a5c3f1 100644 --- a/pkg/sentry/syscalls/linux/sys_mempolicy.go +++ b/pkg/sentry/syscalls/linux/sys_mempolicy.go @@ -162,10 +162,10 @@ func GetMempolicy(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel. if err != nil { return 0, nil, err } - policy = 0 // maxNodes == 1 + policy = linux.MPOL_DEFAULT // maxNodes == 1 } if mode != 0 { - if _, err := t.CopyOut(mode, policy); err != nil { + if _, err := policy.CopyOut(t, mode); err != nil { return 0, nil, err } } @@ -199,10 +199,10 @@ func GetMempolicy(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel. if policy&^linux.MPOL_MODE_FLAGS != linux.MPOL_INTERLEAVE { return 0, nil, syserror.EINVAL } - policy = 0 // maxNodes == 1 + policy = linux.MPOL_DEFAULT // maxNodes == 1 } if mode != 0 { - if _, err := t.CopyOut(mode, policy); err != nil { + if _, err := policy.CopyOut(t, mode); err != nil { return 0, nil, err } } @@ -216,7 +216,7 @@ func GetMempolicy(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel. // SetMempolicy implements the syscall set_mempolicy(2). func SetMempolicy(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { - modeWithFlags := args[0].Int() + modeWithFlags := linux.NumaPolicy(args[0].Int()) nodemask := args[1].Pointer() maxnode := args[2].Uint() @@ -233,7 +233,7 @@ func SetMempolicy(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel. func Mbind(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { addr := args[0].Pointer() length := args[1].Uint64() - mode := args[2].Int() + mode := linux.NumaPolicy(args[2].Int()) nodemask := args[3].Pointer() maxnode := args[4].Uint() flags := args[5].Uint() @@ -258,9 +258,9 @@ func Mbind(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall return 0, nil, err } -func copyInMempolicyNodemask(t *kernel.Task, modeWithFlags int32, nodemask usermem.Addr, maxnode uint32) (int32, uint64, error) { - flags := modeWithFlags & linux.MPOL_MODE_FLAGS - mode := modeWithFlags &^ linux.MPOL_MODE_FLAGS +func copyInMempolicyNodemask(t *kernel.Task, modeWithFlags linux.NumaPolicy, nodemask usermem.Addr, maxnode uint32) (linux.NumaPolicy, uint64, error) { + flags := linux.NumaPolicy(modeWithFlags & linux.MPOL_MODE_FLAGS) + mode := linux.NumaPolicy(modeWithFlags &^ linux.MPOL_MODE_FLAGS) if flags == linux.MPOL_MODE_FLAGS { // Can't specify both mode flags simultaneously. return 0, 0, syserror.EINVAL diff --git a/pkg/sentry/syscalls/linux/sys_mmap.go b/pkg/sentry/syscalls/linux/sys_mmap.go index 9959f6e61..91694d374 100644 --- a/pkg/sentry/syscalls/linux/sys_mmap.go +++ b/pkg/sentry/syscalls/linux/sys_mmap.go @@ -35,6 +35,8 @@ func Brk(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallCo return uintptr(addr), nil, nil } +// LINT.IfChange + // Mmap implements linux syscall mmap(2). func Mmap(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { prot := args[2].Int() @@ -104,6 +106,8 @@ func Mmap(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallC return uintptr(rv), nil, err } +// LINT.ThenChange(vfs2/mmap.go) + // Munmap implements linux syscall munmap(2). func Munmap(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { return 0, nil, t.MemoryManager().MUnmap(t, args[0].Pointer(), args[1].Uint64()) diff --git a/pkg/sentry/syscalls/linux/sys_pipe.go b/pkg/sentry/syscalls/linux/sys_pipe.go index 798344042..43c510930 100644 --- a/pkg/sentry/syscalls/linux/sys_pipe.go +++ b/pkg/sentry/syscalls/linux/sys_pipe.go @@ -24,6 +24,8 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) +// LINT.IfChange + // pipe2 implements the actual system call with flags. func pipe2(t *kernel.Task, addr usermem.Addr, flags uint) (uintptr, error) { if flags&^(linux.O_NONBLOCK|linux.O_CLOEXEC) != 0 { @@ -45,10 +47,12 @@ func pipe2(t *kernel.Task, addr usermem.Addr, flags uint) (uintptr, error) { } if _, err := t.CopyOut(addr, fds); err != nil { - // The files are not closed in this case, the exact semantics - // of this error case are not well defined, but they could have - // already been observed by user space. - return 0, syserror.EFAULT + for _, fd := range fds { + if file, _ := t.FDTable().Remove(fd); file != nil { + file.DecRef() + } + } + return 0, err } return 0, nil } @@ -69,3 +73,5 @@ func Pipe2(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall n, err := pipe2(t, addr, flags) return n, nil, err } + +// LINT.ThenChange(vfs2/pipe.go) diff --git a/pkg/sentry/syscalls/linux/sys_poll.go b/pkg/sentry/syscalls/linux/sys_poll.go index 4f8762d7d..f0198141c 100644 --- a/pkg/sentry/syscalls/linux/sys_poll.go +++ b/pkg/sentry/syscalls/linux/sys_poll.go @@ -443,7 +443,7 @@ func Ppoll(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall } if maskAddr != 0 { - mask, err := copyInSigSet(t, maskAddr, maskSize) + mask, err := CopyInSigSet(t, maskAddr, maskSize) if err != nil { return 0, nil, err } @@ -525,7 +525,7 @@ func Pselect(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysca } if maskAddr != 0 { - mask, err := copyInSigSet(t, maskAddr, size) + mask, err := CopyInSigSet(t, maskAddr, size) if err != nil { return 0, nil, err } diff --git a/pkg/sentry/syscalls/linux/sys_prctl.go b/pkg/sentry/syscalls/linux/sys_prctl.go index 98db32d77..f92bf8096 100644 --- a/pkg/sentry/syscalls/linux/sys_prctl.go +++ b/pkg/sentry/syscalls/linux/sys_prctl.go @@ -20,6 +20,7 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/mm" @@ -135,7 +136,7 @@ func Prctl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall } // Set the underlying executable. - t.MemoryManager().SetExecutable(file.Dirent) + t.MemoryManager().SetExecutable(fsbridge.NewFSFile(file)) case linux.PR_SET_MM_AUXV, linux.PR_SET_MM_START_CODE, @@ -160,8 +161,8 @@ func Prctl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall if args[1].Int() != 1 || args[2].Int() != 0 || args[3].Int() != 0 || args[4].Int() != 0 { return 0, nil, syserror.EINVAL } - // no_new_privs is assumed to always be set. See - // kernel.Task.updateCredsForExec. + // PR_SET_NO_NEW_PRIVS is assumed to always be set. + // See kernel.Task.updateCredsForExecLocked. return 0, nil, nil case linux.PR_GET_NO_NEW_PRIVS: diff --git a/pkg/sentry/syscalls/linux/sys_read.go b/pkg/sentry/syscalls/linux/sys_read.go index 227692f06..071b4bacc 100644 --- a/pkg/sentry/syscalls/linux/sys_read.go +++ b/pkg/sentry/syscalls/linux/sys_read.go @@ -28,6 +28,8 @@ import ( "gvisor.dev/gvisor/pkg/waiter" ) +// LINT.IfChange + const ( // EventMaskRead contains events that can be triggered on reads. EventMaskRead = waiter.EventIn | waiter.EventHUp | waiter.EventErr @@ -94,8 +96,8 @@ func Readahead(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sys return 0, nil, syserror.EINVAL } - // Check that the offset is legitimate. - if offset < 0 { + // Check that the offset is legitimate and does not overflow. + if offset < 0 || offset+int64(size) < 0 { return 0, nil, syserror.EINVAL } @@ -118,8 +120,8 @@ func Pread64(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysca } defer file.DecRef() - // Check that the offset is legitimate. - if offset < 0 { + // Check that the offset is legitimate and does not overflow. + if offset < 0 || offset+int64(size) < 0 { return 0, nil, syserror.EINVAL } @@ -388,3 +390,5 @@ func preadv(t *kernel.Task, f *fs.File, dst usermem.IOSequence, offset int64) (i return total, err } + +// LINT.ThenChange(vfs2/read_write.go) diff --git a/pkg/sentry/syscalls/linux/sys_rlimit.go b/pkg/sentry/syscalls/linux/sys_rlimit.go index e08c333d6..d5d5b6959 100644 --- a/pkg/sentry/syscalls/linux/sys_rlimit.go +++ b/pkg/sentry/syscalls/linux/sys_rlimit.go @@ -197,7 +197,7 @@ func Prlimit64(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sys // saved set user IDs of the target process must match the real user ID of // the caller and the real, effective, and saved set group IDs of the // target process must match the real group ID of the caller." - if !t.HasCapabilityIn(linux.CAP_SYS_RESOURCE, t.PIDNamespace().UserNamespace()) { + if ot != t && !t.HasCapabilityIn(linux.CAP_SYS_RESOURCE, t.PIDNamespace().UserNamespace()) { cred, tcred := t.Credentials(), ot.Credentials() if cred.RealKUID != tcred.RealKUID || cred.RealKUID != tcred.EffectiveKUID || diff --git a/pkg/sentry/syscalls/linux/sys_signal.go b/pkg/sentry/syscalls/linux/sys_signal.go index 7e1747a0c..d2b0012ae 100644 --- a/pkg/sentry/syscalls/linux/sys_signal.go +++ b/pkg/sentry/syscalls/linux/sys_signal.go @@ -295,7 +295,7 @@ func RtSigprocmask(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel } oldmask := t.SignalMask() if setaddr != 0 { - mask, err := copyInSigSet(t, setaddr, sigsetsize) + mask, err := CopyInSigSet(t, setaddr, sigsetsize) if err != nil { return 0, nil, err } @@ -355,7 +355,7 @@ func Pause(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall func RtSigpending(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { addr := args[0].Pointer() pending := t.PendingSignals() - _, err := t.CopyOut(addr, pending) + _, err := pending.CopyOut(t, addr) return 0, nil, err } @@ -366,7 +366,7 @@ func RtSigtimedwait(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kerne timespec := args[2].Pointer() sigsetsize := args[3].SizeT() - mask, err := copyInSigSet(t, sigset, sigsetsize) + mask, err := CopyInSigSet(t, sigset, sigsetsize) if err != nil { return 0, nil, err } @@ -392,7 +392,7 @@ func RtSigtimedwait(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kerne if siginfo != 0 { si.FixSignalCodeForUser() - if _, err := t.CopyOut(siginfo, si); err != nil { + if _, err := si.CopyOut(t, siginfo); err != nil { return 0, nil, err } } @@ -411,7 +411,7 @@ func RtSigqueueinfo(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kerne // same way), and that the code is in the allowed set. This same logic // appears below in RtSigtgqueueinfo and should be kept in sync. var info arch.SignalInfo - if _, err := t.CopyIn(infoAddr, &info); err != nil { + if _, err := info.CopyIn(t, infoAddr); err != nil { return 0, nil, err } info.Signo = int32(sig) @@ -455,7 +455,7 @@ func RtTgsigqueueinfo(t *kernel.Task, args arch.SyscallArguments) (uintptr, *ker // Copy in the info. See RtSigqueueinfo above. var info arch.SignalInfo - if _, err := t.CopyIn(infoAddr, &info); err != nil { + if _, err := info.CopyIn(t, infoAddr); err != nil { return 0, nil, err } info.Signo = int32(sig) @@ -485,7 +485,7 @@ func RtSigsuspend(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel. // Copy in the signal mask. var mask linux.SignalSet - if _, err := t.CopyIn(sigset, &mask); err != nil { + if _, err := mask.CopyIn(t, sigset); err != nil { return 0, nil, err } mask &^= kernel.UnblockableSignals @@ -518,7 +518,7 @@ func RestartSyscall(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kerne // sharedSignalfd is shared between the two calls. func sharedSignalfd(t *kernel.Task, fd int32, sigset usermem.Addr, sigsetsize uint, flags int32) (uintptr, *kernel.SyscallControl, error) { // Copy in the signal mask. - mask, err := copyInSigSet(t, sigset, sigsetsize) + mask, err := CopyInSigSet(t, sigset, sigsetsize) if err != nil { return 0, nil, err } diff --git a/pkg/sentry/syscalls/linux/sys_socket.go b/pkg/sentry/syscalls/linux/sys_socket.go index 2919228d0..0760af77b 100644 --- a/pkg/sentry/syscalls/linux/sys_socket.go +++ b/pkg/sentry/syscalls/linux/sys_socket.go @@ -31,6 +31,8 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) +// LINT.IfChange + // minListenBacklog is the minimum reasonable backlog for listening sockets. const minListenBacklog = 8 @@ -244,7 +246,11 @@ func SocketPair(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sy // Copy the file descriptors out. if _, err := t.CopyOut(socks, fds); err != nil { - // Note that we don't close files here; see pipe(2) also. + for _, fd := range fds { + if file, _ := t.FDTable().Remove(fd); file != nil { + file.DecRef() + } + } return 0, nil, err } @@ -1128,3 +1134,5 @@ func SendTo(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal n, err := sendTo(t, fd, bufPtr, bufLen, flags, namePtr, nameLen) return n, nil, err } + +// LINT.ThenChange(./vfs2/socket.go) diff --git a/pkg/sentry/syscalls/linux/sys_splice.go b/pkg/sentry/syscalls/linux/sys_splice.go index fd642834b..77c78889d 100644 --- a/pkg/sentry/syscalls/linux/sys_splice.go +++ b/pkg/sentry/syscalls/linux/sys_splice.go @@ -25,10 +25,14 @@ import ( // doSplice implements a blocking splice operation. func doSplice(t *kernel.Task, outFile, inFile *fs.File, opts fs.SpliceOpts, nonBlocking bool) (int64, error) { - if opts.Length < 0 || opts.SrcStart < 0 || opts.DstStart < 0 { + if opts.Length < 0 || opts.SrcStart < 0 || opts.DstStart < 0 || (opts.SrcStart+opts.Length < 0) { return 0, syserror.EINVAL } + if opts.Length > int64(kernel.MAX_RW_COUNT) { + opts.Length = int64(kernel.MAX_RW_COUNT) + } + var ( total int64 n int64 @@ -76,6 +80,12 @@ func doSplice(t *kernel.Task, outFile, inFile *fs.File, opts fs.SpliceOpts, nonB } } + if total > 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.Dirent.InotifyEvent(linux.IN_MODIFY, 0) + } return total, err } diff --git a/pkg/sentry/syscalls/linux/sys_stat.go b/pkg/sentry/syscalls/linux/sys_stat.go index c841abccb..46ebf27a2 100644 --- a/pkg/sentry/syscalls/linux/sys_stat.go +++ b/pkg/sentry/syscalls/linux/sys_stat.go @@ -23,6 +23,8 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) +// LINT.IfChange + // Stat implements linux syscall stat(2). func Stat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { addr := args[0].Pointer() @@ -112,7 +114,9 @@ func stat(t *kernel.Task, d *fs.Dirent, dirPath bool, statAddr usermem.Addr) err if err != nil { return err } - return copyOutStat(t, statAddr, d.Inode.StableAttr, uattr) + s := statFromAttrs(t, d.Inode.StableAttr, uattr) + _, err = s.CopyOut(t, statAddr) + return err } // fstat implements fstat for the given *fs.File. @@ -121,7 +125,9 @@ func fstat(t *kernel.Task, f *fs.File, statAddr usermem.Addr) error { if err != nil { return err } - return copyOutStat(t, statAddr, f.Dirent.Inode.StableAttr, uattr) + s := statFromAttrs(t, f.Dirent.Inode.StableAttr, uattr) + _, err = s.CopyOut(t, statAddr) + return err } // Statx implements linux syscall statx(2). @@ -132,7 +138,10 @@ func Statx(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall mask := args[3].Uint() statxAddr := args[4].Pointer() - if mask&linux.STATX__RESERVED > 0 { + if mask&linux.STATX__RESERVED != 0 { + return 0, nil, syserror.EINVAL + } + if flags&^(linux.AT_SYMLINK_NOFOLLOW|linux.AT_EMPTY_PATH|linux.AT_STATX_SYNC_TYPE) != 0 { return 0, nil, syserror.EINVAL } if flags&linux.AT_STATX_SYNC_TYPE == linux.AT_STATX_SYNC_TYPE { @@ -277,3 +286,5 @@ func statfsImpl(t *kernel.Task, d *fs.Dirent, addr usermem.Addr) error { _, err = t.CopyOut(addr, &statfs) return err } + +// LINT.ThenChange(vfs2/stat.go) diff --git a/pkg/sentry/syscalls/linux/sys_stat_amd64.go b/pkg/sentry/syscalls/linux/sys_stat_amd64.go index 75a567bd4..0a04a6113 100644 --- a/pkg/sentry/syscalls/linux/sys_stat_amd64.go +++ b/pkg/sentry/syscalls/linux/sys_stat_amd64.go @@ -12,64 +12,34 @@ // See the License for the specific language governing permissions and // limitations under the License. -//+build amd64 +// +build amd64 package linux import ( "gvisor.dev/gvisor/pkg/abi/linux" - "gvisor.dev/gvisor/pkg/binary" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/kernel" - "gvisor.dev/gvisor/pkg/usermem" ) -// copyOutStat copies the attributes (sattr, uattr) to the struct stat at -// address dst in t's address space. It encodes the stat struct to bytes -// manually, as stat() is a very common syscall for many applications, and -// t.CopyObjectOut has noticeable performance impact due to its many slice -// allocations and use of reflection. -func copyOutStat(t *kernel.Task, dst usermem.Addr, sattr fs.StableAttr, uattr fs.UnstableAttr) error { - b := t.CopyScratchBuffer(int(linux.SizeOfStat))[:0] - - // Dev (uint64) - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(sattr.DeviceID)) - // Ino (uint64) - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(sattr.InodeID)) - // Nlink (uint64) - b = binary.AppendUint64(b, usermem.ByteOrder, uattr.Links) - // Mode (uint32) - b = binary.AppendUint32(b, usermem.ByteOrder, sattr.Type.LinuxType()|uint32(uattr.Perms.LinuxMode())) - // UID (uint32) - b = binary.AppendUint32(b, usermem.ByteOrder, uint32(uattr.Owner.UID.In(t.UserNamespace()).OrOverflow())) - // GID (uint32) - b = binary.AppendUint32(b, usermem.ByteOrder, uint32(uattr.Owner.GID.In(t.UserNamespace()).OrOverflow())) - // Padding (uint32) - b = binary.AppendUint32(b, usermem.ByteOrder, 0) - // Rdev (uint64) - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(linux.MakeDeviceID(sattr.DeviceFileMajor, sattr.DeviceFileMinor))) - // Size (uint64) - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(uattr.Size)) - // Blksize (uint64) - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(sattr.BlockSize)) - // Blocks (uint64) - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(uattr.Usage/512)) - - // ATime - atime := uattr.AccessTime.Timespec() - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(atime.Sec)) - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(atime.Nsec)) - - // MTime - mtime := uattr.ModificationTime.Timespec() - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(mtime.Sec)) - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(mtime.Nsec)) - - // CTime - ctime := uattr.StatusChangeTime.Timespec() - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(ctime.Sec)) - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(ctime.Nsec)) - - _, err := t.CopyOutBytes(dst, b) - return err +// LINT.IfChange + +func statFromAttrs(t *kernel.Task, sattr fs.StableAttr, uattr fs.UnstableAttr) linux.Stat { + return linux.Stat{ + Dev: sattr.DeviceID, + Ino: sattr.InodeID, + Nlink: uattr.Links, + Mode: sattr.Type.LinuxType() | uint32(uattr.Perms.LinuxMode()), + UID: uint32(uattr.Owner.UID.In(t.UserNamespace()).OrOverflow()), + GID: uint32(uattr.Owner.GID.In(t.UserNamespace()).OrOverflow()), + Rdev: uint64(linux.MakeDeviceID(sattr.DeviceFileMajor, sattr.DeviceFileMinor)), + Size: uattr.Size, + Blksize: sattr.BlockSize, + Blocks: uattr.Usage / 512, + ATime: uattr.AccessTime.Timespec(), + MTime: uattr.ModificationTime.Timespec(), + CTime: uattr.StatusChangeTime.Timespec(), + } } + +// LINT.ThenChange(vfs2/stat_amd64.go) diff --git a/pkg/sentry/syscalls/linux/sys_stat_arm64.go b/pkg/sentry/syscalls/linux/sys_stat_arm64.go index 80c98d05c..5a3b1bfad 100644 --- a/pkg/sentry/syscalls/linux/sys_stat_arm64.go +++ b/pkg/sentry/syscalls/linux/sys_stat_arm64.go @@ -12,66 +12,34 @@ // See the License for the specific language governing permissions and // limitations under the License. -//+build arm64 +// +build arm64 package linux import ( "gvisor.dev/gvisor/pkg/abi/linux" - "gvisor.dev/gvisor/pkg/binary" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/kernel" - "gvisor.dev/gvisor/pkg/usermem" ) -// copyOutStat copies the attributes (sattr, uattr) to the struct stat at -// address dst in t's address space. It encodes the stat struct to bytes -// manually, as stat() is a very common syscall for many applications, and -// t.CopyObjectOut has noticeable performance impact due to its many slice -// allocations and use of reflection. -func copyOutStat(t *kernel.Task, dst usermem.Addr, sattr fs.StableAttr, uattr fs.UnstableAttr) error { - b := t.CopyScratchBuffer(int(linux.SizeOfStat))[:0] - - // Dev (uint64) - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(sattr.DeviceID)) - // Ino (uint64) - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(sattr.InodeID)) - // Mode (uint32) - b = binary.AppendUint32(b, usermem.ByteOrder, sattr.Type.LinuxType()|uint32(uattr.Perms.LinuxMode())) - // Nlink (uint32) - b = binary.AppendUint32(b, usermem.ByteOrder, uint32(uattr.Links)) - // UID (uint32) - b = binary.AppendUint32(b, usermem.ByteOrder, uint32(uattr.Owner.UID.In(t.UserNamespace()).OrOverflow())) - // GID (uint32) - b = binary.AppendUint32(b, usermem.ByteOrder, uint32(uattr.Owner.GID.In(t.UserNamespace()).OrOverflow())) - // Rdev (uint64) - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(linux.MakeDeviceID(sattr.DeviceFileMajor, sattr.DeviceFileMinor))) - // Padding (uint64) - b = binary.AppendUint64(b, usermem.ByteOrder, 0) - // Size (uint64) - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(uattr.Size)) - // Blksize (uint32) - b = binary.AppendUint32(b, usermem.ByteOrder, uint32(sattr.BlockSize)) - // Padding (uint32) - b = binary.AppendUint32(b, usermem.ByteOrder, 0) - // Blocks (uint64) - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(uattr.Usage/512)) - - // ATime - atime := uattr.AccessTime.Timespec() - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(atime.Sec)) - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(atime.Nsec)) - - // MTime - mtime := uattr.ModificationTime.Timespec() - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(mtime.Sec)) - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(mtime.Nsec)) - - // CTime - ctime := uattr.StatusChangeTime.Timespec() - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(ctime.Sec)) - b = binary.AppendUint64(b, usermem.ByteOrder, uint64(ctime.Nsec)) - - _, err := t.CopyOutBytes(dst, b) - return err +// LINT.IfChange + +func statFromAttrs(t *kernel.Task, sattr fs.StableAttr, uattr fs.UnstableAttr) linux.Stat { + return linux.Stat{ + Dev: sattr.DeviceID, + Ino: sattr.InodeID, + Nlink: uint32(uattr.Links), + Mode: sattr.Type.LinuxType() | uint32(uattr.Perms.LinuxMode()), + UID: uint32(uattr.Owner.UID.In(t.UserNamespace()).OrOverflow()), + GID: uint32(uattr.Owner.GID.In(t.UserNamespace()).OrOverflow()), + Rdev: uint64(linux.MakeDeviceID(sattr.DeviceFileMajor, sattr.DeviceFileMinor)), + Size: uattr.Size, + Blksize: int32(sattr.BlockSize), + Blocks: uattr.Usage / 512, + ATime: uattr.AccessTime.Timespec(), + MTime: uattr.ModificationTime.Timespec(), + CTime: uattr.StatusChangeTime.Timespec(), + } } + +// LINT.ThenChange(vfs2/stat_arm64.go) diff --git a/pkg/sentry/syscalls/linux/sys_sync.go b/pkg/sentry/syscalls/linux/sys_sync.go index 3e55235bd..5ad465ae3 100644 --- a/pkg/sentry/syscalls/linux/sys_sync.go +++ b/pkg/sentry/syscalls/linux/sys_sync.go @@ -22,6 +22,8 @@ import ( "gvisor.dev/gvisor/pkg/syserror" ) +// LINT.IfChange + // Sync implements linux system call sync(2). func Sync(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { t.MountNamespace().SyncAll(t) @@ -135,3 +137,5 @@ func SyncFileRange(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel return 0, nil, syserror.ConvertIntr(err, kernel.ERESTARTSYS) } + +// LINT.ThenChange(vfs2/sync.go) diff --git a/pkg/sentry/syscalls/linux/sys_sysinfo.go b/pkg/sentry/syscalls/linux/sys_sysinfo.go index a65b560c8..297de052a 100644 --- a/pkg/sentry/syscalls/linux/sys_sysinfo.go +++ b/pkg/sentry/syscalls/linux/sys_sysinfo.go @@ -29,13 +29,18 @@ func Sysinfo(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysca mf.UpdateUsage() _, totalUsage := usage.MemoryAccounting.Copy() totalSize := usage.TotalMemory(mf.TotalSize(), totalUsage) + memFree := totalSize - totalUsage + if memFree > totalSize { + // Underflow. + memFree = 0 + } // Only a subset of the fields in sysinfo_t make sense to return. si := linux.Sysinfo{ Procs: uint16(len(t.PIDNamespace().Tasks())), Uptime: t.Kernel().MonotonicClock().Now().Seconds(), TotalRAM: totalSize, - FreeRAM: totalSize - totalUsage, + FreeRAM: memFree, Unit: 1, } _, err := t.CopyOut(addr, si) diff --git a/pkg/sentry/syscalls/linux/sys_thread.go b/pkg/sentry/syscalls/linux/sys_thread.go index 0c9e2255d..00915fdde 100644 --- a/pkg/sentry/syscalls/linux/sys_thread.go +++ b/pkg/sentry/syscalls/linux/sys_thread.go @@ -21,6 +21,7 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/kernel/sched" "gvisor.dev/gvisor/pkg/sentry/loader" @@ -119,7 +120,7 @@ func execveat(t *kernel.Task, dirFD int32, pathnameAddr, argvAddr, envvAddr user defer root.DecRef() var wd *fs.Dirent - var executable *fs.File + var executable fsbridge.File var closeOnExec bool if dirFD == linux.AT_FDCWD || path.IsAbs(pathname) { // Even if the pathname is absolute, we may still need the wd @@ -136,7 +137,15 @@ func execveat(t *kernel.Task, dirFD int32, pathnameAddr, argvAddr, envvAddr user closeOnExec = fdFlags.CloseOnExec if atEmptyPath && len(pathname) == 0 { - executable = f + // TODO(gvisor.dev/issue/160): Linux requires only execute permission, + // not read. However, our backing filesystems may prevent us from reading + // the file without read permission. Additionally, a task with a + // non-readable executable has additional constraints on access via + // ptrace and procfs. + if err := f.Dirent.Inode.CheckPermission(t, fs.PermMask{Read: true, Execute: true}); err != nil { + return 0, nil, err + } + executable = fsbridge.NewFSFile(f) } else { wd = f.Dirent wd.IncRef() @@ -152,9 +161,7 @@ func execveat(t *kernel.Task, dirFD int32, pathnameAddr, argvAddr, envvAddr user // Load the new TaskContext. remainingTraversals := uint(linux.MaxSymlinkTraversals) loadArgs := loader.LoadArgs{ - Mounts: t.MountNamespace(), - Root: root, - WorkingDirectory: wd, + Opener: fsbridge.NewFSLookup(t.MountNamespace(), root, wd), RemainingTraversals: &remainingTraversals, ResolveFinal: resolveFinal, Filename: pathname, diff --git a/pkg/sentry/syscalls/linux/sys_tls.go b/pkg/sentry/syscalls/linux/sys_tls_amd64.go index b3eb96a1c..b3eb96a1c 100644 --- a/pkg/sentry/syscalls/linux/sys_tls.go +++ b/pkg/sentry/syscalls/linux/sys_tls_amd64.go diff --git a/pkg/sentry/syscalls/linux/vfs2/linux64_override_amd64.go b/pkg/sentry/syscalls/linux/sys_tls_arm64.go index c134714ee..fb08a356e 100644 --- a/pkg/sentry/syscalls/linux/vfs2/linux64_override_amd64.go +++ b/pkg/sentry/syscalls/linux/sys_tls_arm64.go @@ -1,4 +1,4 @@ -// Copyright 2020 The gVisor Authors. +// Copyright 2018 The gVisor Authors. // // Licensed under the Apache License, Version 2.0 (the "License"); // you may not use this file except in compliance with the License. @@ -12,14 +12,17 @@ // See the License for the specific language governing permissions and // limitations under the License. -package vfs2 +//+build arm64 + +package linux import ( + "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/kernel" - "gvisor.dev/gvisor/pkg/sentry/syscalls" + "gvisor.dev/gvisor/pkg/syserror" ) -// Override syscall table to add syscalls implementations from this package. -func Override(table map[uintptr]kernel.Syscall) { - table[0] = syscalls.Supported("read", Read) +// ArchPrctl is not defined for ARM64. +func ArchPrctl(*kernel.Task, arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + return 0, nil, syserror.ENOSYS } diff --git a/pkg/sentry/syscalls/linux/sys_utsname.go b/pkg/sentry/syscalls/linux/sys_utsname.go index a393e28c1..e9d702e8e 100644 --- a/pkg/sentry/syscalls/linux/sys_utsname.go +++ b/pkg/sentry/syscalls/linux/sys_utsname.go @@ -12,8 +12,6 @@ // See the License for the specific language governing permissions and // limitations under the License. -// +build amd64 arm64 - package linux import ( diff --git a/pkg/sentry/syscalls/linux/sys_write.go b/pkg/sentry/syscalls/linux/sys_write.go index aba892939..6ec0de96e 100644 --- a/pkg/sentry/syscalls/linux/sys_write.go +++ b/pkg/sentry/syscalls/linux/sys_write.go @@ -28,6 +28,8 @@ import ( "gvisor.dev/gvisor/pkg/waiter" ) +// LINT.IfChange + const ( // EventMaskWrite contains events that can be triggered on writes. // @@ -85,8 +87,8 @@ func Pwrite64(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysc } defer file.DecRef() - // Check that the offset is legitimate. - if offset < 0 { + // Check that the offset is legitimate and does not overflow. + if offset < 0 || offset+int64(size) < 0 { return 0, nil, syserror.EINVAL } @@ -358,3 +360,5 @@ func pwritev(t *kernel.Task, f *fs.File, src usermem.IOSequence, offset int64) ( return total, err } + +// LINT.ThenChange(vfs2/read_write.go) diff --git a/pkg/sentry/syscalls/linux/sys_xattr.go b/pkg/sentry/syscalls/linux/sys_xattr.go index 9d8140b8a..c24946160 100644 --- a/pkg/sentry/syscalls/linux/sys_xattr.go +++ b/pkg/sentry/syscalls/linux/sys_xattr.go @@ -25,6 +25,8 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) +// LINT.IfChange + // GetXattr implements linux syscall getxattr(2). func GetXattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { return getXattrFromPath(t, args, true) @@ -205,7 +207,11 @@ func setXattr(t *kernel.Task, d *fs.Dirent, nameAddr, valueAddr usermem.Addr, si return syserror.EOPNOTSUPP } - return d.Inode.SetXattr(t, d, name, value, flags) + if err := d.Inode.SetXattr(t, d, name, value, flags); err != nil { + return err + } + d.InotifyEvent(linux.IN_ATTRIB, 0) + return nil } func copyInXattrName(t *kernel.Task, nameAddr usermem.Addr) (string, error) { @@ -416,5 +422,11 @@ func removeXattr(t *kernel.Task, d *fs.Dirent, nameAddr usermem.Addr) error { return syserror.EOPNOTSUPP } - return d.Inode.RemoveXattr(t, d, name) + if err := d.Inode.RemoveXattr(t, d, name); err != nil { + return err + } + d.InotifyEvent(linux.IN_ATTRIB, 0) + return nil } + +// LINT.ThenChange(vfs2/xattr.go) diff --git a/pkg/sentry/syscalls/linux/vfs2/BUILD b/pkg/sentry/syscalls/linux/vfs2/BUILD index 6b8a00b6e..c0d005247 100644 --- a/pkg/sentry/syscalls/linux/vfs2/BUILD +++ b/pkg/sentry/syscalls/linux/vfs2/BUILD @@ -5,18 +5,64 @@ package(licenses = ["notice"]) go_library( name = "vfs2", srcs = [ - "linux64.go", - "linux64_override_amd64.go", - "linux64_override_arm64.go", - "sys_read.go", + "epoll.go", + "eventfd.go", + "execve.go", + "fd.go", + "filesystem.go", + "fscontext.go", + "getdents.go", + "inotify.go", + "ioctl.go", + "memfd.go", + "mmap.go", + "mount.go", + "path.go", + "pipe.go", + "poll.go", + "read_write.go", + "setstat.go", + "signal.go", + "socket.go", + "splice.go", + "stat.go", + "stat_amd64.go", + "stat_arm64.go", + "sync.go", + "timerfd.go", + "vfs2.go", + "xattr.go", ], + marshal = True, visibility = ["//:sandbox"], deps = [ + "//pkg/abi/linux", + "//pkg/binary", + "//pkg/bits", + "//pkg/fspath", + "//pkg/gohacks", "//pkg/sentry/arch", + "//pkg/sentry/fsbridge", + "//pkg/sentry/fsimpl/eventfd", + "//pkg/sentry/fsimpl/pipefs", + "//pkg/sentry/fsimpl/signalfd", + "//pkg/sentry/fsimpl/timerfd", + "//pkg/sentry/fsimpl/tmpfs", "//pkg/sentry/kernel", + "//pkg/sentry/kernel/auth", + "//pkg/sentry/kernel/pipe", + "//pkg/sentry/kernel/time", + "//pkg/sentry/limits", + "//pkg/sentry/loader", + "//pkg/sentry/memmap", + "//pkg/sentry/socket", + "//pkg/sentry/socket/control", + "//pkg/sentry/socket/unix/transport", "//pkg/sentry/syscalls", "//pkg/sentry/syscalls/linux", "//pkg/sentry/vfs", + "//pkg/sync", + "//pkg/syserr", "//pkg/syserror", "//pkg/usermem", "//pkg/waiter", diff --git a/pkg/sentry/syscalls/linux/vfs2/epoll.go b/pkg/sentry/syscalls/linux/vfs2/epoll.go new file mode 100644 index 000000000..34c90ae3e --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/epoll.go @@ -0,0 +1,228 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "math" + "time" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" + "gvisor.dev/gvisor/pkg/waiter" +) + +var sizeofEpollEvent = (*linux.EpollEvent)(nil).SizeBytes() + +// EpollCreate1 implements Linux syscall epoll_create1(2). +func EpollCreate1(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + flags := args[0].Int() + if flags&^linux.EPOLL_CLOEXEC != 0 { + return 0, nil, syserror.EINVAL + } + + file, err := t.Kernel().VFS().NewEpollInstanceFD() + if err != nil { + return 0, nil, err + } + defer file.DecRef() + + fd, err := t.NewFDFromVFS2(0, file, kernel.FDFlags{ + CloseOnExec: flags&linux.EPOLL_CLOEXEC != 0, + }) + if err != nil { + return 0, nil, err + } + return uintptr(fd), nil, nil +} + +// EpollCreate implements Linux syscall epoll_create(2). +func EpollCreate(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + size := args[0].Int() + + // "Since Linux 2.6.8, the size argument is ignored, but must be greater + // than zero" - epoll_create(2) + if size <= 0 { + return 0, nil, syserror.EINVAL + } + + file, err := t.Kernel().VFS().NewEpollInstanceFD() + if err != nil { + return 0, nil, err + } + defer file.DecRef() + + fd, err := t.NewFDFromVFS2(0, file, kernel.FDFlags{}) + if err != nil { + return 0, nil, err + } + return uintptr(fd), nil, nil +} + +// EpollCtl implements Linux syscall epoll_ctl(2). +func EpollCtl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + epfd := args[0].Int() + op := args[1].Int() + fd := args[2].Int() + eventAddr := args[3].Pointer() + + epfile := t.GetFileVFS2(epfd) + if epfile == nil { + return 0, nil, syserror.EBADF + } + defer epfile.DecRef() + ep, ok := epfile.Impl().(*vfs.EpollInstance) + if !ok { + return 0, nil, syserror.EINVAL + } + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + if epfile == file { + return 0, nil, syserror.EINVAL + } + + var event linux.EpollEvent + switch op { + case linux.EPOLL_CTL_ADD: + if _, err := event.CopyIn(t, eventAddr); err != nil { + return 0, nil, err + } + return 0, nil, ep.AddInterest(file, fd, event) + case linux.EPOLL_CTL_DEL: + return 0, nil, ep.DeleteInterest(file, fd) + case linux.EPOLL_CTL_MOD: + if _, err := event.CopyIn(t, eventAddr); err != nil { + return 0, nil, err + } + return 0, nil, ep.ModifyInterest(file, fd, event) + default: + return 0, nil, syserror.EINVAL + } +} + +// EpollWait implements Linux syscall epoll_wait(2). +func EpollWait(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + epfd := args[0].Int() + eventsAddr := args[1].Pointer() + maxEvents := int(args[2].Int()) + timeout := int(args[3].Int()) + + var _EP_MAX_EVENTS = math.MaxInt32 / sizeofEpollEvent // Linux: fs/eventpoll.c:EP_MAX_EVENTS + if maxEvents <= 0 || maxEvents > _EP_MAX_EVENTS { + return 0, nil, syserror.EINVAL + } + + epfile := t.GetFileVFS2(epfd) + if epfile == nil { + return 0, nil, syserror.EBADF + } + defer epfile.DecRef() + ep, ok := epfile.Impl().(*vfs.EpollInstance) + if !ok { + return 0, nil, syserror.EINVAL + } + + // Use a fixed-size buffer in a loop, instead of make([]linux.EpollEvent, + // maxEvents), so that the buffer can be allocated on the stack. + var ( + events [16]linux.EpollEvent + total int + ch chan struct{} + haveDeadline bool + deadline ktime.Time + ) + for { + batchEvents := len(events) + if batchEvents > maxEvents { + batchEvents = maxEvents + } + n := ep.ReadEvents(events[:batchEvents]) + maxEvents -= n + if n != 0 { + // Copy what we read out. + copiedBytes, err := linux.CopyEpollEventSliceOut(t, eventsAddr, events[:n]) + copiedEvents := copiedBytes / sizeofEpollEvent // rounded down + eventsAddr += usermem.Addr(copiedEvents * sizeofEpollEvent) + total += copiedEvents + if err != nil { + if total != 0 { + return uintptr(total), nil, nil + } + return 0, nil, err + } + // If we've filled the application's event buffer, we're done. + if maxEvents == 0 { + return uintptr(total), nil, nil + } + // Loop if we read a full batch, under the expectation that there + // may be more events to read. + if n == batchEvents { + continue + } + } + // We get here if n != batchEvents. If we read any number of events + // (just now, or in a previous iteration of this loop), or if timeout + // is 0 (such that epoll_wait should be non-blocking), return the + // events we've read so far to the application. + if total != 0 || timeout == 0 { + return uintptr(total), nil, nil + } + // In the first iteration of this loop, register with the epoll + // instance for readability events, but then immediately continue the + // loop since we need to retry ReadEvents() before blocking. In all + // subsequent iterations, block until events are available, the timeout + // expires, or an interrupt arrives. + if ch == nil { + var w waiter.Entry + w, ch = waiter.NewChannelEntry(nil) + epfile.EventRegister(&w, waiter.EventIn) + defer epfile.EventUnregister(&w) + } else { + // Set up the timer if a timeout was specified. + if timeout > 0 && !haveDeadline { + timeoutDur := time.Duration(timeout) * time.Millisecond + deadline = t.Kernel().MonotonicClock().Now().Add(timeoutDur) + haveDeadline = true + } + if err := t.BlockWithDeadline(ch, haveDeadline, deadline); err != nil { + if err == syserror.ETIMEDOUT { + err = nil + } + // total must be 0 since otherwise we would have returned + // above. + return 0, nil, err + } + } + } +} + +// EpollPwait implements Linux syscall epoll_pwait(2). +func EpollPwait(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + maskAddr := args[4].Pointer() + maskSize := uint(args[5].Uint()) + + if err := setTempSignalSet(t, maskAddr, maskSize); err != nil { + return 0, nil, err + } + + return EpollWait(t, args) +} diff --git a/pkg/sentry/syscalls/linux/vfs2/eventfd.go b/pkg/sentry/syscalls/linux/vfs2/eventfd.go new file mode 100644 index 000000000..aff1a2070 --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/eventfd.go @@ -0,0 +1,61 @@ +// Copyright 2018 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/eventfd" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/syserror" +) + +// Eventfd2 implements linux syscall eventfd2(2). +func Eventfd2(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + initVal := uint64(args[0].Uint()) + flags := uint(args[1].Uint()) + allOps := uint(linux.EFD_SEMAPHORE | linux.EFD_NONBLOCK | linux.EFD_CLOEXEC) + + if flags & ^allOps != 0 { + return 0, nil, syserror.EINVAL + } + + vfsObj := t.Kernel().VFS() + fileFlags := uint32(linux.O_RDWR) + if flags&linux.EFD_NONBLOCK != 0 { + fileFlags |= linux.O_NONBLOCK + } + semMode := flags&linux.EFD_SEMAPHORE != 0 + eventfd, err := eventfd.New(vfsObj, initVal, semMode, fileFlags) + if err != nil { + return 0, nil, err + } + defer eventfd.DecRef() + + fd, err := t.NewFDFromVFS2(0, eventfd, kernel.FDFlags{ + CloseOnExec: flags&linux.EFD_CLOEXEC != 0, + }) + if err != nil { + return 0, nil, err + } + + return uintptr(fd), nil, nil +} + +// Eventfd implements linux syscall eventfd(2). +func Eventfd(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + args[1].Value = 0 + return Eventfd2(t, args) +} diff --git a/pkg/sentry/syscalls/linux/vfs2/execve.go b/pkg/sentry/syscalls/linux/vfs2/execve.go new file mode 100644 index 000000000..aef0078a8 --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/execve.go @@ -0,0 +1,137 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/loader" + slinux "gvisor.dev/gvisor/pkg/sentry/syscalls/linux" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +// Execve implements linux syscall execve(2). +func Execve(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + pathnameAddr := args[0].Pointer() + argvAddr := args[1].Pointer() + envvAddr := args[2].Pointer() + return execveat(t, linux.AT_FDCWD, pathnameAddr, argvAddr, envvAddr, 0 /* flags */) +} + +// Execveat implements linux syscall execveat(2). +func Execveat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + dirfd := args[0].Int() + pathnameAddr := args[1].Pointer() + argvAddr := args[2].Pointer() + envvAddr := args[3].Pointer() + flags := args[4].Int() + return execveat(t, dirfd, pathnameAddr, argvAddr, envvAddr, flags) +} + +func execveat(t *kernel.Task, dirfd int32, pathnameAddr, argvAddr, envvAddr usermem.Addr, flags int32) (uintptr, *kernel.SyscallControl, error) { + if flags&^(linux.AT_EMPTY_PATH|linux.AT_SYMLINK_NOFOLLOW) != 0 { + return 0, nil, syserror.EINVAL + } + + pathname, err := t.CopyInString(pathnameAddr, linux.PATH_MAX) + if err != nil { + return 0, nil, err + } + var argv, envv []string + if argvAddr != 0 { + var err error + argv, err = t.CopyInVector(argvAddr, slinux.ExecMaxElemSize, slinux.ExecMaxTotalSize) + if err != nil { + return 0, nil, err + } + } + if envvAddr != 0 { + var err error + envv, err = t.CopyInVector(envvAddr, slinux.ExecMaxElemSize, slinux.ExecMaxTotalSize) + if err != nil { + return 0, nil, err + } + } + + root := t.FSContext().RootDirectoryVFS2() + defer root.DecRef() + var executable fsbridge.File + closeOnExec := false + if path := fspath.Parse(pathname); dirfd != linux.AT_FDCWD && !path.Absolute { + // We must open the executable ourselves since dirfd is used as the + // starting point while resolving path, but the task working directory + // is used as the starting point while resolving interpreters (Linux: + // fs/binfmt_script.c:load_script() => fs/exec.c:open_exec() => + // do_open_execat(fd=AT_FDCWD)), and the loader package is currently + // incapable of handling this correctly. + if !path.HasComponents() && flags&linux.AT_EMPTY_PATH == 0 { + return 0, nil, syserror.ENOENT + } + dirfile, dirfileFlags := t.FDTable().GetVFS2(dirfd) + if dirfile == nil { + return 0, nil, syserror.EBADF + } + start := dirfile.VirtualDentry() + start.IncRef() + dirfile.DecRef() + closeOnExec = dirfileFlags.CloseOnExec + file, err := t.Kernel().VFS().OpenAt(t, t.Credentials(), &vfs.PathOperation{ + Root: root, + Start: start, + Path: path, + FollowFinalSymlink: flags&linux.AT_SYMLINK_NOFOLLOW == 0, + }, &vfs.OpenOptions{ + Flags: linux.O_RDONLY, + FileExec: true, + }) + start.DecRef() + if err != nil { + return 0, nil, err + } + defer file.DecRef() + executable = fsbridge.NewVFSFile(file) + } + + // Load the new TaskContext. + mntns := t.MountNamespaceVFS2() // FIXME(jamieliu): useless refcount change + defer mntns.DecRef() + wd := t.FSContext().WorkingDirectoryVFS2() + defer wd.DecRef() + remainingTraversals := uint(linux.MaxSymlinkTraversals) + loadArgs := loader.LoadArgs{ + Opener: fsbridge.NewVFSLookup(mntns, root, wd), + RemainingTraversals: &remainingTraversals, + ResolveFinal: flags&linux.AT_SYMLINK_NOFOLLOW == 0, + Filename: pathname, + File: executable, + CloseOnExec: closeOnExec, + Argv: argv, + Envv: envv, + Features: t.Arch().FeatureSet(), + } + + tc, se := t.Kernel().LoadTaskImage(t, loadArgs) + if se != nil { + return 0, nil, se.ToError() + } + + ctrl, err := t.Execve(tc) + return 0, ctrl, err +} diff --git a/pkg/sentry/syscalls/linux/vfs2/fd.go b/pkg/sentry/syscalls/linux/vfs2/fd.go new file mode 100644 index 000000000..ca0f7fd1e --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/fd.go @@ -0,0 +1,174 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/tmpfs" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/kernel/pipe" + slinux "gvisor.dev/gvisor/pkg/sentry/syscalls/linux" + "gvisor.dev/gvisor/pkg/syserror" +) + +// Close implements Linux syscall close(2). +func Close(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + + // Note that Remove provides a reference on the file that we may use to + // flush. It is still active until we drop the final reference below + // (and other reference-holding operations complete). + _, file := t.FDTable().Remove(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + err := file.OnClose(t) + return 0, nil, slinux.HandleIOErrorVFS2(t, false /* partial */, err, syserror.EINTR, "close", file) +} + +// Dup implements Linux syscall dup(2). +func Dup(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + newFD, err := t.NewFDFromVFS2(0, file, kernel.FDFlags{}) + if err != nil { + return 0, nil, syserror.EMFILE + } + return uintptr(newFD), nil, nil +} + +// Dup2 implements Linux syscall dup2(2). +func Dup2(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + oldfd := args[0].Int() + newfd := args[1].Int() + + if oldfd == newfd { + // As long as oldfd is valid, dup2() does nothing and returns newfd. + file := t.GetFileVFS2(oldfd) + if file == nil { + return 0, nil, syserror.EBADF + } + file.DecRef() + return uintptr(newfd), nil, nil + } + + return dup3(t, oldfd, newfd, 0) +} + +// Dup3 implements Linux syscall dup3(2). +func Dup3(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + oldfd := args[0].Int() + newfd := args[1].Int() + flags := args[2].Uint() + + if oldfd == newfd { + return 0, nil, syserror.EINVAL + } + + return dup3(t, oldfd, newfd, flags) +} + +func dup3(t *kernel.Task, oldfd, newfd int32, flags uint32) (uintptr, *kernel.SyscallControl, error) { + if flags&^linux.O_CLOEXEC != 0 { + return 0, nil, syserror.EINVAL + } + + file := t.GetFileVFS2(oldfd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + err := t.NewFDAtVFS2(newfd, file, kernel.FDFlags{ + CloseOnExec: flags&linux.O_CLOEXEC != 0, + }) + if err != nil { + return 0, nil, err + } + return uintptr(newfd), nil, nil +} + +// Fcntl implements linux syscall fcntl(2). +func Fcntl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + cmd := args[1].Int() + + file, flags := t.FDTable().GetVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + switch cmd { + case linux.F_DUPFD, linux.F_DUPFD_CLOEXEC: + minfd := args[2].Int() + fd, err := t.NewFDFromVFS2(minfd, file, kernel.FDFlags{ + CloseOnExec: cmd == linux.F_DUPFD_CLOEXEC, + }) + if err != nil { + return 0, nil, err + } + return uintptr(fd), nil, nil + case linux.F_GETFD: + return uintptr(flags.ToLinuxFDFlags()), nil, nil + case linux.F_SETFD: + flags := args[2].Uint() + t.FDTable().SetFlags(fd, kernel.FDFlags{ + CloseOnExec: flags&linux.FD_CLOEXEC != 0, + }) + return 0, nil, nil + case linux.F_GETFL: + return uintptr(file.StatusFlags()), nil, nil + case linux.F_SETFL: + return 0, nil, file.SetStatusFlags(t, t.Credentials(), args[2].Uint()) + case linux.F_SETPIPE_SZ: + pipefile, ok := file.Impl().(*pipe.VFSPipeFD) + if !ok { + return 0, nil, syserror.EBADF + } + n, err := pipefile.SetPipeSize(int64(args[2].Int())) + if err != nil { + return 0, nil, err + } + return uintptr(n), nil, nil + case linux.F_GETPIPE_SZ: + pipefile, ok := file.Impl().(*pipe.VFSPipeFD) + if !ok { + return 0, nil, syserror.EBADF + } + return uintptr(pipefile.PipeSize()), nil, nil + case linux.F_GET_SEALS: + val, err := tmpfs.GetSeals(file) + return uintptr(val), nil, err + case linux.F_ADD_SEALS: + if !file.IsWritable() { + return 0, nil, syserror.EPERM + } + err := tmpfs.AddSeals(file, args[2].Uint()) + return 0, nil, err + default: + // TODO(gvisor.dev/issue/1623): Everything else is not yet supported. + return 0, nil, syserror.EINVAL + } +} diff --git a/pkg/sentry/syscalls/linux/vfs2/filesystem.go b/pkg/sentry/syscalls/linux/vfs2/filesystem.go new file mode 100644 index 000000000..46d3e189c --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/filesystem.go @@ -0,0 +1,326 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +// Link implements Linux syscall link(2). +func Link(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + oldpathAddr := args[0].Pointer() + newpathAddr := args[1].Pointer() + return 0, nil, linkat(t, linux.AT_FDCWD, oldpathAddr, linux.AT_FDCWD, newpathAddr, 0 /* flags */) +} + +// Linkat implements Linux syscall linkat(2). +func Linkat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + olddirfd := args[0].Int() + oldpathAddr := args[1].Pointer() + newdirfd := args[2].Int() + newpathAddr := args[3].Pointer() + flags := args[4].Int() + return 0, nil, linkat(t, olddirfd, oldpathAddr, newdirfd, newpathAddr, flags) +} + +func linkat(t *kernel.Task, olddirfd int32, oldpathAddr usermem.Addr, newdirfd int32, newpathAddr usermem.Addr, flags int32) error { + if flags&^(linux.AT_EMPTY_PATH|linux.AT_SYMLINK_FOLLOW) != 0 { + return syserror.EINVAL + } + if flags&linux.AT_EMPTY_PATH != 0 && !t.HasCapability(linux.CAP_DAC_READ_SEARCH) { + return syserror.ENOENT + } + + oldpath, err := copyInPath(t, oldpathAddr) + if err != nil { + return err + } + oldtpop, err := getTaskPathOperation(t, olddirfd, oldpath, shouldAllowEmptyPath(flags&linux.AT_EMPTY_PATH != 0), shouldFollowFinalSymlink(flags&linux.AT_SYMLINK_FOLLOW != 0)) + if err != nil { + return err + } + defer oldtpop.Release() + + newpath, err := copyInPath(t, newpathAddr) + if err != nil { + return err + } + newtpop, err := getTaskPathOperation(t, newdirfd, newpath, disallowEmptyPath, nofollowFinalSymlink) + if err != nil { + return err + } + defer newtpop.Release() + + return t.Kernel().VFS().LinkAt(t, t.Credentials(), &oldtpop.pop, &newtpop.pop) +} + +// Mkdir implements Linux syscall mkdir(2). +func Mkdir(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + addr := args[0].Pointer() + mode := args[1].ModeT() + return 0, nil, mkdirat(t, linux.AT_FDCWD, addr, mode) +} + +// Mkdirat implements Linux syscall mkdirat(2). +func Mkdirat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + dirfd := args[0].Int() + addr := args[1].Pointer() + mode := args[2].ModeT() + return 0, nil, mkdirat(t, dirfd, addr, mode) +} + +func mkdirat(t *kernel.Task, dirfd int32, addr usermem.Addr, mode uint) error { + path, err := copyInPath(t, addr) + if err != nil { + return err + } + tpop, err := getTaskPathOperation(t, dirfd, path, disallowEmptyPath, nofollowFinalSymlink) + if err != nil { + return err + } + defer tpop.Release() + return t.Kernel().VFS().MkdirAt(t, t.Credentials(), &tpop.pop, &vfs.MkdirOptions{ + Mode: linux.FileMode(mode & (0777 | linux.S_ISVTX) &^ t.FSContext().Umask()), + }) +} + +// Mknod implements Linux syscall mknod(2). +func Mknod(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + addr := args[0].Pointer() + mode := args[1].ModeT() + dev := args[2].Uint() + return 0, nil, mknodat(t, linux.AT_FDCWD, addr, mode, dev) +} + +// Mknodat implements Linux syscall mknodat(2). +func Mknodat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + dirfd := args[0].Int() + addr := args[1].Pointer() + mode := args[2].ModeT() + dev := args[3].Uint() + return 0, nil, mknodat(t, dirfd, addr, mode, dev) +} + +func mknodat(t *kernel.Task, dirfd int32, addr usermem.Addr, mode uint, dev uint32) error { + path, err := copyInPath(t, addr) + if err != nil { + return err + } + tpop, err := getTaskPathOperation(t, dirfd, path, disallowEmptyPath, nofollowFinalSymlink) + if err != nil { + return err + } + defer tpop.Release() + major, minor := linux.DecodeDeviceID(dev) + return t.Kernel().VFS().MknodAt(t, t.Credentials(), &tpop.pop, &vfs.MknodOptions{ + Mode: linux.FileMode(mode &^ t.FSContext().Umask()), + DevMajor: uint32(major), + DevMinor: minor, + }) +} + +// Open implements Linux syscall open(2). +func Open(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + addr := args[0].Pointer() + flags := args[1].Uint() + mode := args[2].ModeT() + return openat(t, linux.AT_FDCWD, addr, flags, mode) +} + +// Openat implements Linux syscall openat(2). +func Openat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + dirfd := args[0].Int() + addr := args[1].Pointer() + flags := args[2].Uint() + mode := args[3].ModeT() + return openat(t, dirfd, addr, flags, mode) +} + +// Creat implements Linux syscall creat(2). +func Creat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + addr := args[0].Pointer() + mode := args[1].ModeT() + return openat(t, linux.AT_FDCWD, addr, linux.O_WRONLY|linux.O_CREAT|linux.O_TRUNC, mode) +} + +func openat(t *kernel.Task, dirfd int32, pathAddr usermem.Addr, flags uint32, mode uint) (uintptr, *kernel.SyscallControl, error) { + path, err := copyInPath(t, pathAddr) + if err != nil { + return 0, nil, err + } + tpop, err := getTaskPathOperation(t, dirfd, path, disallowEmptyPath, shouldFollowFinalSymlink(flags&linux.O_NOFOLLOW == 0)) + if err != nil { + return 0, nil, err + } + defer tpop.Release() + + file, err := t.Kernel().VFS().OpenAt(t, t.Credentials(), &tpop.pop, &vfs.OpenOptions{ + Flags: flags | linux.O_LARGEFILE, + Mode: linux.FileMode(mode & (0777 | linux.S_ISUID | linux.S_ISGID | linux.S_ISVTX) &^ t.FSContext().Umask()), + }) + if err != nil { + return 0, nil, err + } + defer file.DecRef() + + fd, err := t.NewFDFromVFS2(0, file, kernel.FDFlags{ + CloseOnExec: flags&linux.O_CLOEXEC != 0, + }) + return uintptr(fd), nil, err +} + +// Rename implements Linux syscall rename(2). +func Rename(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + oldpathAddr := args[0].Pointer() + newpathAddr := args[1].Pointer() + return 0, nil, renameat(t, linux.AT_FDCWD, oldpathAddr, linux.AT_FDCWD, newpathAddr, 0 /* flags */) +} + +// Renameat implements Linux syscall renameat(2). +func Renameat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + olddirfd := args[0].Int() + oldpathAddr := args[1].Pointer() + newdirfd := args[2].Int() + newpathAddr := args[3].Pointer() + return 0, nil, renameat(t, olddirfd, oldpathAddr, newdirfd, newpathAddr, 0 /* flags */) +} + +// Renameat2 implements Linux syscall renameat2(2). +func Renameat2(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + olddirfd := args[0].Int() + oldpathAddr := args[1].Pointer() + newdirfd := args[2].Int() + newpathAddr := args[3].Pointer() + flags := args[4].Uint() + return 0, nil, renameat(t, olddirfd, oldpathAddr, newdirfd, newpathAddr, flags) +} + +func renameat(t *kernel.Task, olddirfd int32, oldpathAddr usermem.Addr, newdirfd int32, newpathAddr usermem.Addr, flags uint32) error { + oldpath, err := copyInPath(t, oldpathAddr) + if err != nil { + return err + } + // "If oldpath refers to a symbolic link, the link is renamed" - rename(2) + oldtpop, err := getTaskPathOperation(t, olddirfd, oldpath, disallowEmptyPath, nofollowFinalSymlink) + if err != nil { + return err + } + defer oldtpop.Release() + + newpath, err := copyInPath(t, newpathAddr) + if err != nil { + return err + } + newtpop, err := getTaskPathOperation(t, newdirfd, newpath, disallowEmptyPath, nofollowFinalSymlink) + if err != nil { + return err + } + defer newtpop.Release() + + return t.Kernel().VFS().RenameAt(t, t.Credentials(), &oldtpop.pop, &newtpop.pop, &vfs.RenameOptions{ + Flags: flags, + }) +} + +// Rmdir implements Linux syscall rmdir(2). +func Rmdir(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + pathAddr := args[0].Pointer() + return 0, nil, rmdirat(t, linux.AT_FDCWD, pathAddr) +} + +func rmdirat(t *kernel.Task, dirfd int32, pathAddr usermem.Addr) error { + path, err := copyInPath(t, pathAddr) + if err != nil { + return err + } + tpop, err := getTaskPathOperation(t, dirfd, path, disallowEmptyPath, nofollowFinalSymlink) + if err != nil { + return err + } + defer tpop.Release() + return t.Kernel().VFS().RmdirAt(t, t.Credentials(), &tpop.pop) +} + +// Unlink implements Linux syscall unlink(2). +func Unlink(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + pathAddr := args[0].Pointer() + return 0, nil, unlinkat(t, linux.AT_FDCWD, pathAddr) +} + +func unlinkat(t *kernel.Task, dirfd int32, pathAddr usermem.Addr) error { + path, err := copyInPath(t, pathAddr) + if err != nil { + return err + } + tpop, err := getTaskPathOperation(t, dirfd, path, disallowEmptyPath, nofollowFinalSymlink) + if err != nil { + return err + } + defer tpop.Release() + return t.Kernel().VFS().UnlinkAt(t, t.Credentials(), &tpop.pop) +} + +// Unlinkat implements Linux syscall unlinkat(2). +func Unlinkat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + dirfd := args[0].Int() + pathAddr := args[1].Pointer() + flags := args[2].Int() + + if flags&^linux.AT_REMOVEDIR != 0 { + return 0, nil, syserror.EINVAL + } + + if flags&linux.AT_REMOVEDIR != 0 { + return 0, nil, rmdirat(t, dirfd, pathAddr) + } + return 0, nil, unlinkat(t, dirfd, pathAddr) +} + +// Symlink implements Linux syscall symlink(2). +func Symlink(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + targetAddr := args[0].Pointer() + linkpathAddr := args[1].Pointer() + return 0, nil, symlinkat(t, targetAddr, linux.AT_FDCWD, linkpathAddr) +} + +// Symlinkat implements Linux syscall symlinkat(2). +func Symlinkat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + targetAddr := args[0].Pointer() + newdirfd := args[1].Int() + linkpathAddr := args[2].Pointer() + return 0, nil, symlinkat(t, targetAddr, newdirfd, linkpathAddr) +} + +func symlinkat(t *kernel.Task, targetAddr usermem.Addr, newdirfd int32, linkpathAddr usermem.Addr) error { + target, err := t.CopyInString(targetAddr, linux.PATH_MAX) + if err != nil { + return err + } + linkpath, err := copyInPath(t, linkpathAddr) + if err != nil { + return err + } + tpop, err := getTaskPathOperation(t, newdirfd, linkpath, disallowEmptyPath, nofollowFinalSymlink) + if err != nil { + return err + } + defer tpop.Release() + return t.Kernel().VFS().SymlinkAt(t, t.Credentials(), &tpop.pop, target) +} diff --git a/pkg/sentry/syscalls/linux/vfs2/fscontext.go b/pkg/sentry/syscalls/linux/vfs2/fscontext.go new file mode 100644 index 000000000..317409a18 --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/fscontext.go @@ -0,0 +1,131 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" +) + +// Getcwd implements Linux syscall getcwd(2). +func Getcwd(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + addr := args[0].Pointer() + size := args[1].SizeT() + + root := t.FSContext().RootDirectoryVFS2() + wd := t.FSContext().WorkingDirectoryVFS2() + s, err := t.Kernel().VFS().PathnameForGetcwd(t, root, wd) + root.DecRef() + wd.DecRef() + if err != nil { + return 0, nil, err + } + + // Note this is >= because we need a terminator. + if uint(len(s)) >= size { + return 0, nil, syserror.ERANGE + } + + // Construct a byte slice containing a NUL terminator. + buf := t.CopyScratchBuffer(len(s) + 1) + copy(buf, s) + buf[len(buf)-1] = 0 + + // Write the pathname slice. + n, err := t.CopyOutBytes(addr, buf) + if err != nil { + return 0, nil, err + } + return uintptr(n), nil, nil +} + +// Chdir implements Linux syscall chdir(2). +func Chdir(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + addr := args[0].Pointer() + + path, err := copyInPath(t, addr) + if err != nil { + return 0, nil, err + } + tpop, err := getTaskPathOperation(t, linux.AT_FDCWD, path, disallowEmptyPath, followFinalSymlink) + if err != nil { + return 0, nil, err + } + defer tpop.Release() + + vd, err := t.Kernel().VFS().GetDentryAt(t, t.Credentials(), &tpop.pop, &vfs.GetDentryOptions{ + CheckSearchable: true, + }) + if err != nil { + return 0, nil, err + } + t.FSContext().SetWorkingDirectoryVFS2(vd) + vd.DecRef() + return 0, nil, nil +} + +// Fchdir implements Linux syscall fchdir(2). +func Fchdir(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + + tpop, err := getTaskPathOperation(t, fd, fspath.Path{}, allowEmptyPath, nofollowFinalSymlink) + if err != nil { + return 0, nil, err + } + defer tpop.Release() + + vd, err := t.Kernel().VFS().GetDentryAt(t, t.Credentials(), &tpop.pop, &vfs.GetDentryOptions{ + CheckSearchable: true, + }) + if err != nil { + return 0, nil, err + } + t.FSContext().SetWorkingDirectoryVFS2(vd) + vd.DecRef() + return 0, nil, nil +} + +// Chroot implements Linux syscall chroot(2). +func Chroot(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + addr := args[0].Pointer() + + if !t.HasCapability(linux.CAP_SYS_CHROOT) { + return 0, nil, syserror.EPERM + } + + path, err := copyInPath(t, addr) + if err != nil { + return 0, nil, err + } + tpop, err := getTaskPathOperation(t, linux.AT_FDCWD, path, disallowEmptyPath, followFinalSymlink) + if err != nil { + return 0, nil, err + } + defer tpop.Release() + + vd, err := t.Kernel().VFS().GetDentryAt(t, t.Credentials(), &tpop.pop, &vfs.GetDentryOptions{ + CheckSearchable: true, + }) + if err != nil { + return 0, nil, err + } + t.FSContext().SetRootDirectoryVFS2(vd) + vd.DecRef() + return 0, nil, nil +} diff --git a/pkg/sentry/syscalls/linux/vfs2/getdents.go b/pkg/sentry/syscalls/linux/vfs2/getdents.go new file mode 100644 index 000000000..c7c7bf7ce --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/getdents.go @@ -0,0 +1,161 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "fmt" + + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +// Getdents implements Linux syscall getdents(2). +func Getdents(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + return getdents(t, args, false /* isGetdents64 */) +} + +// Getdents64 implements Linux syscall getdents64(2). +func Getdents64(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + return getdents(t, args, true /* isGetdents64 */) +} + +func getdents(t *kernel.Task, args arch.SyscallArguments, isGetdents64 bool) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + addr := args[1].Pointer() + size := int(args[2].Uint()) + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + cb := getGetdentsCallback(t, addr, size, isGetdents64) + err := file.IterDirents(t, cb) + n := size - cb.remaining + putGetdentsCallback(cb) + if n == 0 { + return 0, nil, err + } + return uintptr(n), nil, nil +} + +type getdentsCallback struct { + t *kernel.Task + addr usermem.Addr + remaining int + isGetdents64 bool +} + +var getdentsCallbackPool = sync.Pool{ + New: func() interface{} { + return &getdentsCallback{} + }, +} + +func getGetdentsCallback(t *kernel.Task, addr usermem.Addr, size int, isGetdents64 bool) *getdentsCallback { + cb := getdentsCallbackPool.Get().(*getdentsCallback) + *cb = getdentsCallback{ + t: t, + addr: addr, + remaining: size, + isGetdents64: isGetdents64, + } + return cb +} + +func putGetdentsCallback(cb *getdentsCallback) { + cb.t = nil + getdentsCallbackPool.Put(cb) +} + +// Handle implements vfs.IterDirentsCallback.Handle. +func (cb *getdentsCallback) Handle(dirent vfs.Dirent) error { + var buf []byte + if cb.isGetdents64 { + // struct linux_dirent64 { + // ino64_t d_ino; /* 64-bit inode number */ + // off64_t d_off; /* 64-bit offset to next structure */ + // unsigned short d_reclen; /* Size of this dirent */ + // unsigned char d_type; /* File type */ + // char d_name[]; /* Filename (null-terminated) */ + // }; + size := 8 + 8 + 2 + 1 + 1 + len(dirent.Name) + size = (size + 7) &^ 7 // round up to multiple of 8 + if size > cb.remaining { + return syserror.EINVAL + } + buf = cb.t.CopyScratchBuffer(size) + usermem.ByteOrder.PutUint64(buf[0:8], dirent.Ino) + usermem.ByteOrder.PutUint64(buf[8:16], uint64(dirent.NextOff)) + usermem.ByteOrder.PutUint16(buf[16:18], uint16(size)) + buf[18] = dirent.Type + copy(buf[19:], dirent.Name) + // Zero out all remaining bytes in buf, including the NUL terminator + // after dirent.Name. + bufTail := buf[19+len(dirent.Name):] + for i := range bufTail { + bufTail[i] = 0 + } + } else { + // struct linux_dirent { + // unsigned long d_ino; /* Inode number */ + // unsigned long d_off; /* Offset to next linux_dirent */ + // unsigned short d_reclen; /* Length of this linux_dirent */ + // char d_name[]; /* Filename (null-terminated) */ + // /* length is actually (d_reclen - 2 - + // offsetof(struct linux_dirent, d_name)) */ + // /* + // char pad; // Zero padding byte + // char d_type; // File type (only since Linux + // // 2.6.4); offset is (d_reclen - 1) + // */ + // }; + if cb.t.Arch().Width() != 8 { + panic(fmt.Sprintf("unsupported sizeof(unsigned long): %d", cb.t.Arch().Width())) + } + size := 8 + 8 + 2 + 1 + 1 + len(dirent.Name) + size = (size + 7) &^ 7 // round up to multiple of sizeof(long) + if size > cb.remaining { + return syserror.EINVAL + } + buf = cb.t.CopyScratchBuffer(size) + usermem.ByteOrder.PutUint64(buf[0:8], dirent.Ino) + usermem.ByteOrder.PutUint64(buf[8:16], uint64(dirent.NextOff)) + usermem.ByteOrder.PutUint16(buf[16:18], uint16(size)) + copy(buf[18:], dirent.Name) + // Zero out all remaining bytes in buf, including the NUL terminator + // after dirent.Name and the zero padding byte between the name and + // dirent type. + bufTail := buf[18+len(dirent.Name) : size-1] + for i := range bufTail { + bufTail[i] = 0 + } + buf[size-1] = dirent.Type + } + n, err := cb.t.CopyOutBytes(cb.addr, buf) + if err != nil { + // Don't report partially-written dirents by advancing cb.addr or + // cb.remaining. + return err + } + cb.addr += usermem.Addr(n) + cb.remaining -= n + return nil +} diff --git a/pkg/sentry/syscalls/linux/vfs2/inotify.go b/pkg/sentry/syscalls/linux/vfs2/inotify.go new file mode 100644 index 000000000..7d50b6a16 --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/inotify.go @@ -0,0 +1,134 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" +) + +const allFlags = linux.IN_NONBLOCK | linux.IN_CLOEXEC + +// InotifyInit1 implements the inotify_init1() syscalls. +func InotifyInit1(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + flags := args[0].Int() + if flags&^allFlags != 0 { + return 0, nil, syserror.EINVAL + } + + ino, err := vfs.NewInotifyFD(t, t.Kernel().VFS(), uint32(flags)) + if err != nil { + return 0, nil, err + } + defer ino.DecRef() + + fd, err := t.NewFDFromVFS2(0, ino, kernel.FDFlags{ + CloseOnExec: flags&linux.IN_CLOEXEC != 0, + }) + + if err != nil { + return 0, nil, err + } + + return uintptr(fd), nil, nil +} + +// InotifyInit implements the inotify_init() syscalls. +func InotifyInit(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + args[0].Value = 0 + return InotifyInit1(t, args) +} + +// fdToInotify resolves an fd to an inotify object. If successful, the file will +// have an extra ref and the caller is responsible for releasing the ref. +func fdToInotify(t *kernel.Task, fd int32) (*vfs.Inotify, *vfs.FileDescription, error) { + f := t.GetFileVFS2(fd) + if f == nil { + // Invalid fd. + return nil, nil, syserror.EBADF + } + + ino, ok := f.Impl().(*vfs.Inotify) + if !ok { + // Not an inotify fd. + f.DecRef() + return nil, nil, syserror.EINVAL + } + + return ino, f, nil +} + +// InotifyAddWatch implements the inotify_add_watch() syscall. +func InotifyAddWatch(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + addr := args[1].Pointer() + mask := args[2].Uint() + + // "EINVAL: The given event mask contains no valid events." + // -- inotify_add_watch(2) + if validBits := mask & linux.ALL_INOTIFY_BITS; validBits == 0 { + return 0, nil, syserror.EINVAL + } + + // "IN_DONT_FOLLOW: Don't dereference pathname if it is a symbolic link." + // -- inotify(7) + follow := followFinalSymlink + if mask&linux.IN_DONT_FOLLOW == 0 { + follow = nofollowFinalSymlink + } + + ino, f, err := fdToInotify(t, fd) + if err != nil { + return 0, nil, err + } + defer f.DecRef() + + path, err := copyInPath(t, addr) + if err != nil { + return 0, nil, err + } + if mask&linux.IN_ONLYDIR != 0 { + path.Dir = true + } + tpop, err := getTaskPathOperation(t, linux.AT_FDCWD, path, disallowEmptyPath, follow) + if err != nil { + return 0, nil, err + } + defer tpop.Release() + d, err := t.Kernel().VFS().GetDentryAt(t, t.Credentials(), &tpop.pop, &vfs.GetDentryOptions{}) + if err != nil { + return 0, nil, err + } + defer d.DecRef() + + fd = ino.AddWatch(d.Dentry(), mask) + return uintptr(fd), nil, err +} + +// InotifyRmWatch implements the inotify_rm_watch() syscall. +func InotifyRmWatch(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + wd := args[1].Int() + + ino, f, err := fdToInotify(t, fd) + if err != nil { + return 0, nil, err + } + defer f.DecRef() + return 0, nil, ino.RmWatch(wd) +} diff --git a/pkg/sentry/arch/arch_state_aarch64.go b/pkg/sentry/syscalls/linux/vfs2/ioctl.go index 0136a85ad..5a2418da9 100644 --- a/pkg/sentry/arch/arch_state_aarch64.go +++ b/pkg/sentry/syscalls/linux/vfs2/ioctl.go @@ -12,27 +12,24 @@ // See the License for the specific language governing permissions and // limitations under the License. -// +build arm64 - -package arch +package vfs2 import ( - "syscall" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/syserror" ) -type syscallPtraceRegs struct { - Regs [31]uint64 - Sp uint64 - Pc uint64 - Pstate uint64 -} +// Ioctl implements Linux syscall ioctl(2). +func Ioctl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() -// saveRegs is invoked by stateify. -func (s *State) saveRegs() syscallPtraceRegs { - return syscallPtraceRegs(s.Regs) -} + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() -// loadRegs is invoked by stateify. -func (s *State) loadRegs(r syscallPtraceRegs) { - s.Regs = syscall.PtraceRegs(r) + ret, err := file.Ioctl(t, t.MemoryManager(), args) + return ret, nil, err } diff --git a/pkg/sentry/syscalls/linux/vfs2/memfd.go b/pkg/sentry/syscalls/linux/vfs2/memfd.go new file mode 100644 index 000000000..bbe248d17 --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/memfd.go @@ -0,0 +1,63 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/tmpfs" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/syserror" +) + +const ( + memfdPrefix = "memfd:" + memfdMaxNameLen = linux.NAME_MAX - len(memfdPrefix) + memfdAllFlags = uint32(linux.MFD_CLOEXEC | linux.MFD_ALLOW_SEALING) +) + +// MemfdCreate implements the linux syscall memfd_create(2). +func MemfdCreate(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + addr := args[0].Pointer() + flags := args[1].Uint() + + if flags&^memfdAllFlags != 0 { + // Unknown bits in flags. + return 0, nil, syserror.EINVAL + } + + allowSeals := flags&linux.MFD_ALLOW_SEALING != 0 + cloExec := flags&linux.MFD_CLOEXEC != 0 + + name, err := t.CopyInString(addr, memfdMaxNameLen) + if err != nil { + return 0, nil, err + } + + shmMount := t.Kernel().ShmMount() + file, err := tmpfs.NewMemfd(shmMount, t.Credentials(), allowSeals, memfdPrefix+name) + if err != nil { + return 0, nil, err + } + + fd, err := t.NewFDFromVFS2(0, file, kernel.FDFlags{ + CloseOnExec: cloExec, + }) + if err != nil { + return 0, nil, err + } + + return uintptr(fd), nil, nil +} diff --git a/pkg/sentry/syscalls/linux/vfs2/mmap.go b/pkg/sentry/syscalls/linux/vfs2/mmap.go new file mode 100644 index 000000000..60a43f0a0 --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/mmap.go @@ -0,0 +1,92 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/memmap" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +// Mmap implements Linux syscall mmap(2). +func Mmap(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + prot := args[2].Int() + flags := args[3].Int() + fd := args[4].Int() + fixed := flags&linux.MAP_FIXED != 0 + private := flags&linux.MAP_PRIVATE != 0 + shared := flags&linux.MAP_SHARED != 0 + anon := flags&linux.MAP_ANONYMOUS != 0 + map32bit := flags&linux.MAP_32BIT != 0 + + // Require exactly one of MAP_PRIVATE and MAP_SHARED. + if private == shared { + return 0, nil, syserror.EINVAL + } + + opts := memmap.MMapOpts{ + Length: args[1].Uint64(), + Offset: args[5].Uint64(), + Addr: args[0].Pointer(), + Fixed: fixed, + Unmap: fixed, + Map32Bit: map32bit, + Private: private, + Perms: usermem.AccessType{ + Read: linux.PROT_READ&prot != 0, + Write: linux.PROT_WRITE&prot != 0, + Execute: linux.PROT_EXEC&prot != 0, + }, + MaxPerms: usermem.AnyAccess, + GrowsDown: linux.MAP_GROWSDOWN&flags != 0, + Precommit: linux.MAP_POPULATE&flags != 0, + } + if linux.MAP_LOCKED&flags != 0 { + opts.MLockMode = memmap.MLockEager + } + defer func() { + if opts.MappingIdentity != nil { + opts.MappingIdentity.DecRef() + } + }() + + if !anon { + // Convert the passed FD to a file reference. + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // mmap unconditionally requires that the FD is readable. + if !file.IsReadable() { + return 0, nil, syserror.EACCES + } + // MAP_SHARED requires that the FD be writable for PROT_WRITE. + if shared && !file.IsWritable() { + opts.MaxPerms.Write = false + } + + if err := file.ConfigureMMap(t, &opts); err != nil { + return 0, nil, err + } + } + + rv, err := t.MemoryManager().MMap(t, opts) + return uintptr(rv), nil, err +} diff --git a/pkg/sentry/syscalls/linux/vfs2/mount.go b/pkg/sentry/syscalls/linux/vfs2/mount.go new file mode 100644 index 000000000..adeaa39cc --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/mount.go @@ -0,0 +1,145 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +// Mount implements Linux syscall mount(2). +func Mount(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + sourceAddr := args[0].Pointer() + targetAddr := args[1].Pointer() + typeAddr := args[2].Pointer() + flags := args[3].Uint64() + dataAddr := args[4].Pointer() + + // For null-terminated strings related to mount(2), Linux copies in at most + // a page worth of data. See fs/namespace.c:copy_mount_string(). + fsType, err := t.CopyInString(typeAddr, usermem.PageSize) + if err != nil { + return 0, nil, err + } + source, err := t.CopyInString(sourceAddr, usermem.PageSize) + if err != nil { + return 0, nil, err + } + + targetPath, err := copyInPath(t, targetAddr) + if err != nil { + return 0, nil, err + } + + data := "" + if dataAddr != 0 { + // In Linux, a full page is always copied in regardless of null + // character placement, and the address is passed to each file system. + // Most file systems always treat this data as a string, though, and so + // do all of the ones we implement. + data, err = t.CopyInString(dataAddr, usermem.PageSize) + if err != nil { + return 0, nil, err + } + } + + // Ignore magic value that was required before Linux 2.4. + if flags&linux.MS_MGC_MSK == linux.MS_MGC_VAL { + flags = flags &^ linux.MS_MGC_MSK + } + + // Must have CAP_SYS_ADMIN in the current mount namespace's associated user + // namespace. + creds := t.Credentials() + if !creds.HasCapabilityIn(linux.CAP_SYS_ADMIN, t.MountNamespaceVFS2().Owner) { + return 0, nil, syserror.EPERM + } + + const unsupportedOps = linux.MS_REMOUNT | linux.MS_BIND | + linux.MS_SHARED | linux.MS_PRIVATE | linux.MS_SLAVE | + linux.MS_UNBINDABLE | linux.MS_MOVE + + // Silently allow MS_NOSUID, since we don't implement set-id bits + // anyway. + const unsupportedFlags = linux.MS_NODEV | + linux.MS_NODIRATIME | linux.MS_STRICTATIME + + // Linux just allows passing any flags to mount(2) - it won't fail when + // unknown or unsupported flags are passed. Since we don't implement + // everything, we fail explicitly on flags that are unimplemented. + if flags&(unsupportedOps|unsupportedFlags) != 0 { + return 0, nil, syserror.EINVAL + } + + var opts vfs.MountOptions + if flags&linux.MS_NOATIME == linux.MS_NOATIME { + opts.Flags.NoATime = true + } + if flags&linux.MS_NOEXEC == linux.MS_NOEXEC { + opts.Flags.NoExec = true + } + if flags&linux.MS_RDONLY == linux.MS_RDONLY { + opts.ReadOnly = true + } + opts.GetFilesystemOptions.Data = data + + target, err := getTaskPathOperation(t, linux.AT_FDCWD, targetPath, disallowEmptyPath, nofollowFinalSymlink) + if err != nil { + return 0, nil, err + } + defer target.Release() + + return 0, nil, t.Kernel().VFS().MountAt(t, creds, source, &target.pop, fsType, &opts) +} + +// Umount2 implements Linux syscall umount2(2). +func Umount2(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + addr := args[0].Pointer() + flags := args[1].Int() + + // Must have CAP_SYS_ADMIN in the mount namespace's associated user + // namespace. + // + // Currently, this is always the init task's user namespace. + creds := t.Credentials() + if !creds.HasCapabilityIn(linux.CAP_SYS_ADMIN, t.MountNamespaceVFS2().Owner) { + return 0, nil, syserror.EPERM + } + + const unsupported = linux.MNT_FORCE | linux.MNT_EXPIRE + if flags&unsupported != 0 { + return 0, nil, syserror.EINVAL + } + + path, err := copyInPath(t, addr) + if err != nil { + return 0, nil, err + } + tpop, err := getTaskPathOperation(t, linux.AT_FDCWD, path, disallowEmptyPath, nofollowFinalSymlink) + if err != nil { + return 0, nil, err + } + defer tpop.Release() + + opts := vfs.UmountOptions{ + Flags: uint32(flags), + } + + return 0, nil, t.Kernel().VFS().UmountAt(t, creds, &tpop.pop, &opts) +} diff --git a/pkg/sentry/syscalls/linux/vfs2/path.go b/pkg/sentry/syscalls/linux/vfs2/path.go new file mode 100644 index 000000000..97da6c647 --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/path.go @@ -0,0 +1,94 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +func copyInPath(t *kernel.Task, addr usermem.Addr) (fspath.Path, error) { + pathname, err := t.CopyInString(addr, linux.PATH_MAX) + if err != nil { + return fspath.Path{}, err + } + return fspath.Parse(pathname), nil +} + +type taskPathOperation struct { + pop vfs.PathOperation + haveStartRef bool +} + +func getTaskPathOperation(t *kernel.Task, dirfd int32, path fspath.Path, shouldAllowEmptyPath shouldAllowEmptyPath, shouldFollowFinalSymlink shouldFollowFinalSymlink) (taskPathOperation, error) { + root := t.FSContext().RootDirectoryVFS2() + start := root + haveStartRef := false + if !path.Absolute { + if !path.HasComponents() && !bool(shouldAllowEmptyPath) { + root.DecRef() + return taskPathOperation{}, syserror.ENOENT + } + if dirfd == linux.AT_FDCWD { + start = t.FSContext().WorkingDirectoryVFS2() + haveStartRef = true + } else { + dirfile := t.GetFileVFS2(dirfd) + if dirfile == nil { + root.DecRef() + return taskPathOperation{}, syserror.EBADF + } + start = dirfile.VirtualDentry() + start.IncRef() + haveStartRef = true + dirfile.DecRef() + } + } + return taskPathOperation{ + pop: vfs.PathOperation{ + Root: root, + Start: start, + Path: path, + FollowFinalSymlink: bool(shouldFollowFinalSymlink), + }, + haveStartRef: haveStartRef, + }, nil +} + +func (tpop *taskPathOperation) Release() { + tpop.pop.Root.DecRef() + if tpop.haveStartRef { + tpop.pop.Start.DecRef() + tpop.haveStartRef = false + } +} + +type shouldAllowEmptyPath bool + +const ( + disallowEmptyPath shouldAllowEmptyPath = false + allowEmptyPath shouldAllowEmptyPath = true +) + +type shouldFollowFinalSymlink bool + +const ( + nofollowFinalSymlink shouldFollowFinalSymlink = false + followFinalSymlink shouldFollowFinalSymlink = true +) diff --git a/pkg/sentry/syscalls/linux/vfs2/pipe.go b/pkg/sentry/syscalls/linux/vfs2/pipe.go new file mode 100644 index 000000000..4a01e4209 --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/pipe.go @@ -0,0 +1,63 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/pipefs" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +// Pipe implements Linux syscall pipe(2). +func Pipe(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + addr := args[0].Pointer() + return 0, nil, pipe2(t, addr, 0) +} + +// Pipe2 implements Linux syscall pipe2(2). +func Pipe2(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + addr := args[0].Pointer() + flags := args[1].Int() + return 0, nil, pipe2(t, addr, flags) +} + +func pipe2(t *kernel.Task, addr usermem.Addr, flags int32) error { + if flags&^(linux.O_NONBLOCK|linux.O_CLOEXEC) != 0 { + return syserror.EINVAL + } + r, w := pipefs.NewConnectedPipeFDs(t, t.Kernel().PipeMount(), uint32(flags&linux.O_NONBLOCK)) + defer r.DecRef() + defer w.DecRef() + + fds, err := t.NewFDsVFS2(0, []*vfs.FileDescription{r, w}, kernel.FDFlags{ + CloseOnExec: flags&linux.O_CLOEXEC != 0, + }) + if err != nil { + return err + } + if _, err := t.CopyOut(addr, fds); err != nil { + for _, fd := range fds { + if _, file := t.FDTable().Remove(fd); file != nil { + file.DecRef() + } + } + return err + } + return nil +} diff --git a/pkg/sentry/syscalls/linux/vfs2/poll.go b/pkg/sentry/syscalls/linux/vfs2/poll.go new file mode 100644 index 000000000..ff1b25d7b --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/poll.go @@ -0,0 +1,586 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "fmt" + "time" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" + "gvisor.dev/gvisor/pkg/sentry/limits" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" + "gvisor.dev/gvisor/pkg/waiter" +) + +// fileCap is the maximum allowable files for poll & select. This has no +// equivalent in Linux; it exists in gVisor since allocation failure in Go is +// unrecoverable. +const fileCap = 1024 * 1024 + +// Masks for "readable", "writable", and "exceptional" events as defined by +// select(2). +const ( + // selectReadEvents is analogous to the Linux kernel's + // fs/select.c:POLLIN_SET. + selectReadEvents = linux.POLLIN | linux.POLLHUP | linux.POLLERR + + // selectWriteEvents is analogous to the Linux kernel's + // fs/select.c:POLLOUT_SET. + selectWriteEvents = linux.POLLOUT | linux.POLLERR + + // selectExceptEvents is analogous to the Linux kernel's + // fs/select.c:POLLEX_SET. + selectExceptEvents = linux.POLLPRI +) + +// pollState tracks the associated file description and waiter of a PollFD. +type pollState struct { + file *vfs.FileDescription + waiter waiter.Entry +} + +// initReadiness gets the current ready mask for the file represented by the FD +// stored in pfd.FD. If a channel is passed in, the waiter entry in "state" is +// used to register with the file for event notifications, and a reference to +// the file is stored in "state". +func initReadiness(t *kernel.Task, pfd *linux.PollFD, state *pollState, ch chan struct{}) { + if pfd.FD < 0 { + pfd.REvents = 0 + return + } + + file := t.GetFileVFS2(pfd.FD) + if file == nil { + pfd.REvents = linux.POLLNVAL + return + } + + if ch == nil { + defer file.DecRef() + } else { + state.file = file + state.waiter, _ = waiter.NewChannelEntry(ch) + file.EventRegister(&state.waiter, waiter.EventMaskFromLinux(uint32(pfd.Events))) + } + + r := file.Readiness(waiter.EventMaskFromLinux(uint32(pfd.Events))) + pfd.REvents = int16(r.ToLinux()) & pfd.Events +} + +// releaseState releases all the pollState in "state". +func releaseState(state []pollState) { + for i := range state { + if state[i].file != nil { + state[i].file.EventUnregister(&state[i].waiter) + state[i].file.DecRef() + } + } +} + +// pollBlock polls the PollFDs in "pfd" with a bounded time specified in "timeout" +// when "timeout" is greater than zero. +// +// pollBlock returns the remaining timeout, which is always 0 on a timeout; and 0 or +// positive if interrupted by a signal. +func pollBlock(t *kernel.Task, pfd []linux.PollFD, timeout time.Duration) (time.Duration, uintptr, error) { + var ch chan struct{} + if timeout != 0 { + ch = make(chan struct{}, 1) + } + + // Register for event notification in the files involved if we may + // block (timeout not zero). Once we find a file that has a non-zero + // result, we stop registering for events but still go through all files + // to get their ready masks. + state := make([]pollState, len(pfd)) + defer releaseState(state) + n := uintptr(0) + for i := range pfd { + initReadiness(t, &pfd[i], &state[i], ch) + if pfd[i].REvents != 0 { + n++ + ch = nil + } + } + + if timeout == 0 { + return timeout, n, nil + } + + haveTimeout := timeout >= 0 + + for n == 0 { + var err error + // Wait for a notification. + timeout, err = t.BlockWithTimeout(ch, haveTimeout, timeout) + if err != nil { + if err == syserror.ETIMEDOUT { + err = nil + } + return timeout, 0, err + } + + // We got notified, count how many files are ready. If none, + // then this was a spurious notification, and we just go back + // to sleep with the remaining timeout. + for i := range state { + if state[i].file == nil { + continue + } + + r := state[i].file.Readiness(waiter.EventMaskFromLinux(uint32(pfd[i].Events))) + rl := int16(r.ToLinux()) & pfd[i].Events + if rl != 0 { + pfd[i].REvents = rl + n++ + } + } + } + + return timeout, n, nil +} + +// copyInPollFDs copies an array of struct pollfd unless nfds exceeds the max. +func copyInPollFDs(t *kernel.Task, addr usermem.Addr, nfds uint) ([]linux.PollFD, error) { + if uint64(nfds) > t.ThreadGroup().Limits().GetCapped(limits.NumberOfFiles, fileCap) { + return nil, syserror.EINVAL + } + + pfd := make([]linux.PollFD, nfds) + if nfds > 0 { + if _, err := t.CopyIn(addr, &pfd); err != nil { + return nil, err + } + } + + return pfd, nil +} + +func doPoll(t *kernel.Task, addr usermem.Addr, nfds uint, timeout time.Duration) (time.Duration, uintptr, error) { + pfd, err := copyInPollFDs(t, addr, nfds) + if err != nil { + return timeout, 0, err + } + + // Compatibility warning: Linux adds POLLHUP and POLLERR just before + // polling, in fs/select.c:do_pollfd(). Since pfd is copied out after + // polling, changing event masks here is an application-visible difference. + // (Linux also doesn't copy out event masks at all, only revents.) + for i := range pfd { + pfd[i].Events |= linux.POLLHUP | linux.POLLERR + } + remainingTimeout, n, err := pollBlock(t, pfd, timeout) + err = syserror.ConvertIntr(err, syserror.EINTR) + + // The poll entries are copied out regardless of whether + // any are set or not. This aligns with the Linux behavior. + if nfds > 0 && err == nil { + if _, err := t.CopyOut(addr, pfd); err != nil { + return remainingTimeout, 0, err + } + } + + return remainingTimeout, n, err +} + +// CopyInFDSet copies an fd set from select(2)/pselect(2). +func CopyInFDSet(t *kernel.Task, addr usermem.Addr, nBytes, nBitsInLastPartialByte int) ([]byte, error) { + set := make([]byte, nBytes) + + if addr != 0 { + if _, err := t.CopyIn(addr, &set); err != nil { + return nil, err + } + // If we only use part of the last byte, mask out the extraneous bits. + // + // N.B. This only works on little-endian architectures. + if nBitsInLastPartialByte != 0 { + set[nBytes-1] &^= byte(0xff) << nBitsInLastPartialByte + } + } + return set, nil +} + +func doSelect(t *kernel.Task, nfds int, readFDs, writeFDs, exceptFDs usermem.Addr, timeout time.Duration) (uintptr, error) { + if nfds < 0 || nfds > fileCap { + return 0, syserror.EINVAL + } + + // Calculate the size of the fd sets (one bit per fd). + nBytes := (nfds + 7) / 8 + nBitsInLastPartialByte := nfds % 8 + + // Capture all the provided input vectors. + r, err := CopyInFDSet(t, readFDs, nBytes, nBitsInLastPartialByte) + if err != nil { + return 0, err + } + w, err := CopyInFDSet(t, writeFDs, nBytes, nBitsInLastPartialByte) + if err != nil { + return 0, err + } + e, err := CopyInFDSet(t, exceptFDs, nBytes, nBitsInLastPartialByte) + if err != nil { + return 0, err + } + + // Count how many FDs are actually being requested so that we can build + // a PollFD array. + fdCount := 0 + for i := 0; i < nBytes; i++ { + v := r[i] | w[i] | e[i] + for v != 0 { + v &= (v - 1) + fdCount++ + } + } + + // Build the PollFD array. + pfd := make([]linux.PollFD, 0, fdCount) + var fd int32 + for i := 0; i < nBytes; i++ { + rV, wV, eV := r[i], w[i], e[i] + v := rV | wV | eV + m := byte(1) + for j := 0; j < 8; j++ { + if (v & m) != 0 { + // Make sure the fd is valid and decrement the reference + // immediately to ensure we don't leak. Note, another thread + // might be about to close fd. This is racy, but that's + // OK. Linux is racy in the same way. + file := t.GetFileVFS2(fd) + if file == nil { + return 0, syserror.EBADF + } + file.DecRef() + + var mask int16 + if (rV & m) != 0 { + mask |= selectReadEvents + } + + if (wV & m) != 0 { + mask |= selectWriteEvents + } + + if (eV & m) != 0 { + mask |= selectExceptEvents + } + + pfd = append(pfd, linux.PollFD{ + FD: fd, + Events: mask, + }) + } + + fd++ + m <<= 1 + } + } + + // Do the syscall, then count the number of bits set. + if _, _, err = pollBlock(t, pfd, timeout); err != nil { + return 0, syserror.ConvertIntr(err, syserror.EINTR) + } + + // r, w, and e are currently event mask bitsets; unset bits corresponding + // to events that *didn't* occur. + bitSetCount := uintptr(0) + for idx := range pfd { + events := pfd[idx].REvents + i, j := pfd[idx].FD/8, uint(pfd[idx].FD%8) + m := byte(1) << j + if r[i]&m != 0 { + if (events & selectReadEvents) != 0 { + bitSetCount++ + } else { + r[i] &^= m + } + } + if w[i]&m != 0 { + if (events & selectWriteEvents) != 0 { + bitSetCount++ + } else { + w[i] &^= m + } + } + if e[i]&m != 0 { + if (events & selectExceptEvents) != 0 { + bitSetCount++ + } else { + e[i] &^= m + } + } + } + + // Copy updated vectors back. + if readFDs != 0 { + if _, err := t.CopyOut(readFDs, r); err != nil { + return 0, err + } + } + + if writeFDs != 0 { + if _, err := t.CopyOut(writeFDs, w); err != nil { + return 0, err + } + } + + if exceptFDs != 0 { + if _, err := t.CopyOut(exceptFDs, e); err != nil { + return 0, err + } + } + + return bitSetCount, nil +} + +// timeoutRemaining returns the amount of time remaining for the specified +// timeout or 0 if it has elapsed. +// +// startNs must be from CLOCK_MONOTONIC. +func timeoutRemaining(t *kernel.Task, startNs ktime.Time, timeout time.Duration) time.Duration { + now := t.Kernel().MonotonicClock().Now() + remaining := timeout - now.Sub(startNs) + if remaining < 0 { + remaining = 0 + } + return remaining +} + +// copyOutTimespecRemaining copies the time remaining in timeout to timespecAddr. +// +// startNs must be from CLOCK_MONOTONIC. +func copyOutTimespecRemaining(t *kernel.Task, startNs ktime.Time, timeout time.Duration, timespecAddr usermem.Addr) error { + if timeout <= 0 { + return nil + } + remaining := timeoutRemaining(t, startNs, timeout) + tsRemaining := linux.NsecToTimespec(remaining.Nanoseconds()) + _, err := tsRemaining.CopyOut(t, timespecAddr) + return err +} + +// copyOutTimevalRemaining copies the time remaining in timeout to timevalAddr. +// +// startNs must be from CLOCK_MONOTONIC. +func copyOutTimevalRemaining(t *kernel.Task, startNs ktime.Time, timeout time.Duration, timevalAddr usermem.Addr) error { + if timeout <= 0 { + return nil + } + remaining := timeoutRemaining(t, startNs, timeout) + tvRemaining := linux.NsecToTimeval(remaining.Nanoseconds()) + _, err := tvRemaining.CopyOut(t, timevalAddr) + return err +} + +// pollRestartBlock encapsulates the state required to restart poll(2) via +// restart_syscall(2). +// +// +stateify savable +type pollRestartBlock struct { + pfdAddr usermem.Addr + nfds uint + timeout time.Duration +} + +// Restart implements kernel.SyscallRestartBlock.Restart. +func (p *pollRestartBlock) Restart(t *kernel.Task) (uintptr, error) { + return poll(t, p.pfdAddr, p.nfds, p.timeout) +} + +func poll(t *kernel.Task, pfdAddr usermem.Addr, nfds uint, timeout time.Duration) (uintptr, error) { + remainingTimeout, n, err := doPoll(t, pfdAddr, nfds, timeout) + // On an interrupt poll(2) is restarted with the remaining timeout. + if err == syserror.EINTR { + t.SetSyscallRestartBlock(&pollRestartBlock{ + pfdAddr: pfdAddr, + nfds: nfds, + timeout: remainingTimeout, + }) + return 0, kernel.ERESTART_RESTARTBLOCK + } + return n, err +} + +// Poll implements linux syscall poll(2). +func Poll(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + pfdAddr := args[0].Pointer() + nfds := uint(args[1].Uint()) // poll(2) uses unsigned long. + timeout := time.Duration(args[2].Int()) * time.Millisecond + n, err := poll(t, pfdAddr, nfds, timeout) + return n, nil, err +} + +// Ppoll implements linux syscall ppoll(2). +func Ppoll(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + pfdAddr := args[0].Pointer() + nfds := uint(args[1].Uint()) // poll(2) uses unsigned long. + timespecAddr := args[2].Pointer() + maskAddr := args[3].Pointer() + maskSize := uint(args[4].Uint()) + + timeout, err := copyTimespecInToDuration(t, timespecAddr) + if err != nil { + return 0, nil, err + } + + var startNs ktime.Time + if timeout > 0 { + startNs = t.Kernel().MonotonicClock().Now() + } + + if err := setTempSignalSet(t, maskAddr, maskSize); err != nil { + return 0, nil, err + } + + _, n, err := doPoll(t, pfdAddr, nfds, timeout) + copyErr := copyOutTimespecRemaining(t, startNs, timeout, timespecAddr) + // doPoll returns EINTR if interrupted, but ppoll is normally restartable + // if interrupted by something other than a signal handled by the + // application (i.e. returns ERESTARTNOHAND). However, if + // copyOutTimespecRemaining failed, then the restarted ppoll would use the + // wrong timeout, so the error should be left as EINTR. + // + // Note that this means that if err is nil but copyErr is not, copyErr is + // ignored. This is consistent with Linux. + if err == syserror.EINTR && copyErr == nil { + err = kernel.ERESTARTNOHAND + } + return n, nil, err +} + +// Select implements linux syscall select(2). +func Select(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + nfds := int(args[0].Int()) // select(2) uses an int. + readFDs := args[1].Pointer() + writeFDs := args[2].Pointer() + exceptFDs := args[3].Pointer() + timevalAddr := args[4].Pointer() + + // Use a negative Duration to indicate "no timeout". + timeout := time.Duration(-1) + if timevalAddr != 0 { + var timeval linux.Timeval + if _, err := timeval.CopyIn(t, timevalAddr); err != nil { + return 0, nil, err + } + if timeval.Sec < 0 || timeval.Usec < 0 { + return 0, nil, syserror.EINVAL + } + timeout = time.Duration(timeval.ToNsecCapped()) + } + startNs := t.Kernel().MonotonicClock().Now() + n, err := doSelect(t, nfds, readFDs, writeFDs, exceptFDs, timeout) + copyErr := copyOutTimevalRemaining(t, startNs, timeout, timevalAddr) + // See comment in Ppoll. + if err == syserror.EINTR && copyErr == nil { + err = kernel.ERESTARTNOHAND + } + return n, nil, err +} + +// Pselect implements linux syscall pselect(2). +func Pselect(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + nfds := int(args[0].Int()) // select(2) uses an int. + readFDs := args[1].Pointer() + writeFDs := args[2].Pointer() + exceptFDs := args[3].Pointer() + timespecAddr := args[4].Pointer() + maskWithSizeAddr := args[5].Pointer() + + timeout, err := copyTimespecInToDuration(t, timespecAddr) + if err != nil { + return 0, nil, err + } + + var startNs ktime.Time + if timeout > 0 { + startNs = t.Kernel().MonotonicClock().Now() + } + + if maskWithSizeAddr != 0 { + if t.Arch().Width() != 8 { + panic(fmt.Sprintf("unsupported sizeof(void*): %d", t.Arch().Width())) + } + var maskStruct sigSetWithSize + if _, err := maskStruct.CopyIn(t, maskWithSizeAddr); err != nil { + return 0, nil, err + } + if err := setTempSignalSet(t, usermem.Addr(maskStruct.sigsetAddr), uint(maskStruct.sizeofSigset)); err != nil { + return 0, nil, err + } + } + + n, err := doSelect(t, nfds, readFDs, writeFDs, exceptFDs, timeout) + copyErr := copyOutTimespecRemaining(t, startNs, timeout, timespecAddr) + // See comment in Ppoll. + if err == syserror.EINTR && copyErr == nil { + err = kernel.ERESTARTNOHAND + } + return n, nil, err +} + +// +marshal +type sigSetWithSize struct { + sigsetAddr uint64 + sizeofSigset uint64 +} + +// copyTimespecInToDuration copies a Timespec from the untrusted app range, +// validates it and converts it to a Duration. +// +// If the Timespec is larger than what can be represented in a Duration, the +// returned value is the maximum that Duration will allow. +// +// If timespecAddr is NULL, the returned value is negative. +func copyTimespecInToDuration(t *kernel.Task, timespecAddr usermem.Addr) (time.Duration, error) { + // Use a negative Duration to indicate "no timeout". + timeout := time.Duration(-1) + if timespecAddr != 0 { + var timespec linux.Timespec + if _, err := timespec.CopyIn(t, timespecAddr); err != nil { + return 0, err + } + if !timespec.Valid() { + return 0, syserror.EINVAL + } + timeout = time.Duration(timespec.ToNsecCapped()) + } + return timeout, nil +} + +func setTempSignalSet(t *kernel.Task, maskAddr usermem.Addr, maskSize uint) error { + if maskAddr == 0 { + return nil + } + if maskSize != linux.SignalSetSize { + return syserror.EINVAL + } + var mask linux.SignalSet + if _, err := mask.CopyIn(t, maskAddr); err != nil { + return err + } + mask &^= kernel.UnblockableSignals + oldmask := t.SignalMask() + t.SetSignalMask(mask) + t.SetSavedSignalMask(oldmask) + return nil +} diff --git a/pkg/sentry/syscalls/linux/vfs2/read_write.go b/pkg/sentry/syscalls/linux/vfs2/read_write.go new file mode 100644 index 000000000..7f9debd4a --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/read_write.go @@ -0,0 +1,608 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "time" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" + "gvisor.dev/gvisor/pkg/sentry/socket" + slinux "gvisor.dev/gvisor/pkg/sentry/syscalls/linux" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" + "gvisor.dev/gvisor/pkg/waiter" +) + +const ( + eventMaskRead = waiter.EventIn | waiter.EventHUp | waiter.EventErr + eventMaskWrite = waiter.EventOut | waiter.EventHUp | waiter.EventErr +) + +// Read implements Linux syscall read(2). +func Read(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + addr := args[1].Pointer() + size := args[2].SizeT() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Check that the size is legitimate. + si := int(size) + if si < 0 { + return 0, nil, syserror.EINVAL + } + + // Get the destination of the read. + dst, err := t.SingleIOSequence(addr, si, usermem.IOOpts{ + AddressSpaceActive: true, + }) + if err != nil { + return 0, nil, err + } + + n, err := read(t, file, dst, vfs.ReadOptions{}) + t.IOUsage().AccountReadSyscall(n) + return uintptr(n), nil, slinux.HandleIOErrorVFS2(t, n != 0, err, kernel.ERESTARTSYS, "read", file) +} + +// Readv implements Linux syscall readv(2). +func Readv(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + addr := args[1].Pointer() + iovcnt := int(args[2].Int()) + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Get the destination of the read. + dst, err := t.IovecsIOSequence(addr, iovcnt, usermem.IOOpts{ + AddressSpaceActive: true, + }) + if err != nil { + return 0, nil, err + } + + n, err := read(t, file, dst, vfs.ReadOptions{}) + t.IOUsage().AccountReadSyscall(n) + return uintptr(n), nil, slinux.HandleIOErrorVFS2(t, n != 0, err, kernel.ERESTARTSYS, "readv", file) +} + +func read(t *kernel.Task, file *vfs.FileDescription, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { + n, err := file.Read(t, dst, opts) + if err != syserror.ErrWouldBlock { + if n > 0 { + file.Dentry().InotifyWithParent(linux.IN_ACCESS, 0, vfs.PathEvent) + } + return n, err + } + + allowBlock, deadline, hasDeadline := blockPolicy(t, file) + if !allowBlock { + if n > 0 { + file.Dentry().InotifyWithParent(linux.IN_ACCESS, 0, vfs.PathEvent) + } + return n, err + } + + // Register for notifications. + w, ch := waiter.NewChannelEntry(nil) + file.EventRegister(&w, eventMaskRead) + + total := n + for { + // Shorten dst to reflect bytes previously read. + dst = dst.DropFirst(int(n)) + + // Issue the request and break out if it completes with anything other than + // "would block". + n, err = file.Read(t, dst, opts) + total += n + if err != syserror.ErrWouldBlock { + break + } + + // Wait for a notification that we should retry. + if err = t.BlockWithDeadline(ch, hasDeadline, deadline); err != nil { + if err == syserror.ETIMEDOUT { + err = syserror.ErrWouldBlock + } + break + } + } + file.EventUnregister(&w) + + if total > 0 { + file.Dentry().InotifyWithParent(linux.IN_ACCESS, 0, vfs.PathEvent) + } + return total, err +} + +// Pread64 implements Linux syscall pread64(2). +func Pread64(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + addr := args[1].Pointer() + size := args[2].SizeT() + offset := args[3].Int64() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Check that the offset is legitimate and does not overflow. + if offset < 0 || offset+int64(size) < 0 { + return 0, nil, syserror.EINVAL + } + + // Check that the size is legitimate. + si := int(size) + if si < 0 { + return 0, nil, syserror.EINVAL + } + + // Get the destination of the read. + dst, err := t.SingleIOSequence(addr, si, usermem.IOOpts{ + AddressSpaceActive: true, + }) + if err != nil { + return 0, nil, err + } + + n, err := pread(t, file, dst, offset, vfs.ReadOptions{}) + t.IOUsage().AccountReadSyscall(n) + return uintptr(n), nil, slinux.HandleIOErrorVFS2(t, n != 0, err, kernel.ERESTARTSYS, "pread64", file) +} + +// Preadv implements Linux syscall preadv(2). +func Preadv(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + addr := args[1].Pointer() + iovcnt := int(args[2].Int()) + offset := args[3].Int64() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Check that the offset is legitimate. + if offset < 0 { + return 0, nil, syserror.EINVAL + } + + // Get the destination of the read. + dst, err := t.IovecsIOSequence(addr, iovcnt, usermem.IOOpts{ + AddressSpaceActive: true, + }) + if err != nil { + return 0, nil, err + } + + n, err := pread(t, file, dst, offset, vfs.ReadOptions{}) + t.IOUsage().AccountReadSyscall(n) + return uintptr(n), nil, slinux.HandleIOErrorVFS2(t, n != 0, err, kernel.ERESTARTSYS, "preadv", file) +} + +// Preadv2 implements Linux syscall preadv2(2). +func Preadv2(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + // While the glibc signature is + // preadv2(int fd, struct iovec* iov, int iov_cnt, off_t offset, int flags) + // the actual syscall + // (https://elixir.bootlin.com/linux/v5.5/source/fs/read_write.c#L1142) + // splits the offset argument into a high/low value for compatibility with + // 32-bit architectures. The flags argument is the 6th argument (index 5). + fd := args[0].Int() + addr := args[1].Pointer() + iovcnt := int(args[2].Int()) + offset := args[3].Int64() + flags := args[5].Int() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Check that the offset is legitimate. + if offset < -1 { + return 0, nil, syserror.EINVAL + } + + // Get the destination of the read. + dst, err := t.IovecsIOSequence(addr, iovcnt, usermem.IOOpts{ + AddressSpaceActive: true, + }) + if err != nil { + return 0, nil, err + } + + opts := vfs.ReadOptions{ + Flags: uint32(flags), + } + var n int64 + if offset == -1 { + n, err = read(t, file, dst, opts) + } else { + n, err = pread(t, file, dst, offset, opts) + } + t.IOUsage().AccountReadSyscall(n) + return uintptr(n), nil, slinux.HandleIOErrorVFS2(t, n != 0, err, kernel.ERESTARTSYS, "preadv2", file) +} + +func pread(t *kernel.Task, file *vfs.FileDescription, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { + n, err := file.PRead(t, dst, offset, opts) + if err != syserror.ErrWouldBlock { + if n > 0 { + file.Dentry().InotifyWithParent(linux.IN_ACCESS, 0, vfs.PathEvent) + } + return n, err + } + + allowBlock, deadline, hasDeadline := blockPolicy(t, file) + if !allowBlock { + if n > 0 { + file.Dentry().InotifyWithParent(linux.IN_ACCESS, 0, vfs.PathEvent) + } + return n, err + } + + // Register for notifications. + w, ch := waiter.NewChannelEntry(nil) + file.EventRegister(&w, eventMaskRead) + + total := n + for { + // Shorten dst to reflect bytes previously read. + dst = dst.DropFirst(int(n)) + + // Issue the request and break out if it completes with anything other than + // "would block". + n, err = file.PRead(t, dst, offset+total, opts) + total += n + if err != syserror.ErrWouldBlock { + break + } + + // Wait for a notification that we should retry. + if err = t.BlockWithDeadline(ch, hasDeadline, deadline); err != nil { + if err == syserror.ETIMEDOUT { + err = syserror.ErrWouldBlock + } + break + } + } + file.EventUnregister(&w) + + if total > 0 { + file.Dentry().InotifyWithParent(linux.IN_ACCESS, 0, vfs.PathEvent) + } + return total, err +} + +// Write implements Linux syscall write(2). +func Write(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + addr := args[1].Pointer() + size := args[2].SizeT() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Check that the size is legitimate. + si := int(size) + if si < 0 { + return 0, nil, syserror.EINVAL + } + + // Get the source of the write. + src, err := t.SingleIOSequence(addr, si, usermem.IOOpts{ + AddressSpaceActive: true, + }) + if err != nil { + return 0, nil, err + } + + n, err := write(t, file, src, vfs.WriteOptions{}) + t.IOUsage().AccountWriteSyscall(n) + return uintptr(n), nil, slinux.HandleIOErrorVFS2(t, n != 0, err, kernel.ERESTARTSYS, "write", file) +} + +// Writev implements Linux syscall writev(2). +func Writev(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + addr := args[1].Pointer() + iovcnt := int(args[2].Int()) + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Get the source of the write. + src, err := t.IovecsIOSequence(addr, iovcnt, usermem.IOOpts{ + AddressSpaceActive: true, + }) + if err != nil { + return 0, nil, err + } + + n, err := write(t, file, src, vfs.WriteOptions{}) + t.IOUsage().AccountWriteSyscall(n) + return uintptr(n), nil, slinux.HandleIOErrorVFS2(t, n != 0, err, kernel.ERESTARTSYS, "writev", file) +} + +func write(t *kernel.Task, file *vfs.FileDescription, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) { + n, err := file.Write(t, src, opts) + if err != syserror.ErrWouldBlock { + if n > 0 { + file.Dentry().InotifyWithParent(linux.IN_MODIFY, 0, vfs.PathEvent) + } + return n, err + } + + allowBlock, deadline, hasDeadline := blockPolicy(t, file) + if !allowBlock { + if n > 0 { + file.Dentry().InotifyWithParent(linux.IN_MODIFY, 0, vfs.PathEvent) + } + return n, err + } + + // Register for notifications. + w, ch := waiter.NewChannelEntry(nil) + file.EventRegister(&w, eventMaskWrite) + + total := n + for { + // Shorten src to reflect bytes previously written. + src = src.DropFirst(int(n)) + + // Issue the request and break out if it completes with anything other than + // "would block". + n, err = file.Write(t, src, opts) + total += n + if err != syserror.ErrWouldBlock { + break + } + + // Wait for a notification that we should retry. + if err = t.BlockWithDeadline(ch, hasDeadline, deadline); err != nil { + if err == syserror.ETIMEDOUT { + err = syserror.ErrWouldBlock + } + break + } + } + file.EventUnregister(&w) + + if total > 0 { + file.Dentry().InotifyWithParent(linux.IN_MODIFY, 0, vfs.PathEvent) + } + return total, err +} + +// Pwrite64 implements Linux syscall pwrite64(2). +func Pwrite64(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + addr := args[1].Pointer() + size := args[2].SizeT() + offset := args[3].Int64() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Check that the offset is legitimate and does not overflow. + if offset < 0 || offset+int64(size) < 0 { + return 0, nil, syserror.EINVAL + } + + // Check that the size is legitimate. + si := int(size) + if si < 0 { + return 0, nil, syserror.EINVAL + } + + // Get the source of the write. + src, err := t.SingleIOSequence(addr, si, usermem.IOOpts{ + AddressSpaceActive: true, + }) + if err != nil { + return 0, nil, err + } + + n, err := pwrite(t, file, src, offset, vfs.WriteOptions{}) + t.IOUsage().AccountWriteSyscall(n) + return uintptr(n), nil, slinux.HandleIOErrorVFS2(t, n != 0, err, kernel.ERESTARTSYS, "pwrite64", file) +} + +// Pwritev implements Linux syscall pwritev(2). +func Pwritev(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + addr := args[1].Pointer() + iovcnt := int(args[2].Int()) + offset := args[3].Int64() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Check that the offset is legitimate. + if offset < 0 { + return 0, nil, syserror.EINVAL + } + + // Get the source of the write. + src, err := t.IovecsIOSequence(addr, iovcnt, usermem.IOOpts{ + AddressSpaceActive: true, + }) + if err != nil { + return 0, nil, err + } + + n, err := pwrite(t, file, src, offset, vfs.WriteOptions{}) + t.IOUsage().AccountReadSyscall(n) + return uintptr(n), nil, slinux.HandleIOErrorVFS2(t, n != 0, err, kernel.ERESTARTSYS, "pwritev", file) +} + +// Pwritev2 implements Linux syscall pwritev2(2). +func Pwritev2(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + // While the glibc signature is + // pwritev2(int fd, struct iovec* iov, int iov_cnt, off_t offset, int flags) + // the actual syscall + // (https://elixir.bootlin.com/linux/v5.5/source/fs/read_write.c#L1162) + // splits the offset argument into a high/low value for compatibility with + // 32-bit architectures. The flags argument is the 6th argument (index 5). + fd := args[0].Int() + addr := args[1].Pointer() + iovcnt := int(args[2].Int()) + offset := args[3].Int64() + flags := args[5].Int() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Check that the offset is legitimate. + if offset < -1 { + return 0, nil, syserror.EINVAL + } + + // Get the source of the write. + src, err := t.IovecsIOSequence(addr, iovcnt, usermem.IOOpts{ + AddressSpaceActive: true, + }) + if err != nil { + return 0, nil, err + } + + opts := vfs.WriteOptions{ + Flags: uint32(flags), + } + var n int64 + if offset == -1 { + n, err = write(t, file, src, opts) + } else { + n, err = pwrite(t, file, src, offset, opts) + } + t.IOUsage().AccountWriteSyscall(n) + return uintptr(n), nil, slinux.HandleIOErrorVFS2(t, n != 0, err, kernel.ERESTARTSYS, "pwritev2", file) +} + +func pwrite(t *kernel.Task, file *vfs.FileDescription, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) { + n, err := file.PWrite(t, src, offset, opts) + if err != syserror.ErrWouldBlock { + if n > 0 { + file.Dentry().InotifyWithParent(linux.IN_MODIFY, 0, vfs.PathEvent) + } + return n, err + } + + allowBlock, deadline, hasDeadline := blockPolicy(t, file) + if !allowBlock { + if n > 0 { + file.Dentry().InotifyWithParent(linux.IN_ACCESS, 0, vfs.PathEvent) + } + return n, err + } + + // Register for notifications. + w, ch := waiter.NewChannelEntry(nil) + file.EventRegister(&w, eventMaskWrite) + + total := n + for { + // Shorten src to reflect bytes previously written. + src = src.DropFirst(int(n)) + + // Issue the request and break out if it completes with anything other than + // "would block". + n, err = file.PWrite(t, src, offset+total, opts) + total += n + if err != syserror.ErrWouldBlock { + break + } + + // Wait for a notification that we should retry. + if err = t.BlockWithDeadline(ch, hasDeadline, deadline); err != nil { + if err == syserror.ETIMEDOUT { + err = syserror.ErrWouldBlock + } + break + } + } + file.EventUnregister(&w) + + if total > 0 { + file.Dentry().InotifyWithParent(linux.IN_ACCESS, 0, vfs.PathEvent) + } + return total, err +} + +func blockPolicy(t *kernel.Task, file *vfs.FileDescription) (allowBlock bool, deadline ktime.Time, hasDeadline bool) { + if file.StatusFlags()&linux.O_NONBLOCK != 0 { + return false, ktime.Time{}, false + } + // Sockets support read/write timeouts. + if s, ok := file.Impl().(socket.SocketVFS2); ok { + dl := s.RecvTimeout() + if dl < 0 { + return false, ktime.Time{}, false + } + if dl > 0 { + return true, t.Kernel().MonotonicClock().Now().Add(time.Duration(dl) * time.Nanosecond), true + } + } + return true, ktime.Time{}, false +} + +// Lseek implements Linux syscall lseek(2). +func Lseek(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + offset := args[1].Int64() + whence := args[2].Int() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + newoff, err := file.Seek(t, offset, whence) + return uintptr(newoff), nil, err +} diff --git a/pkg/sentry/syscalls/linux/vfs2/setstat.go b/pkg/sentry/syscalls/linux/vfs2/setstat.go new file mode 100644 index 000000000..09ecfed26 --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/setstat.go @@ -0,0 +1,428 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +const chmodMask = 0777 | linux.S_ISUID | linux.S_ISGID | linux.S_ISVTX + +// Chmod implements Linux syscall chmod(2). +func Chmod(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + pathAddr := args[0].Pointer() + mode := args[1].ModeT() + return 0, nil, fchmodat(t, linux.AT_FDCWD, pathAddr, mode) +} + +// Fchmodat implements Linux syscall fchmodat(2). +func Fchmodat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + dirfd := args[0].Int() + pathAddr := args[1].Pointer() + mode := args[2].ModeT() + return 0, nil, fchmodat(t, dirfd, pathAddr, mode) +} + +func fchmodat(t *kernel.Task, dirfd int32, pathAddr usermem.Addr, mode uint) error { + path, err := copyInPath(t, pathAddr) + if err != nil { + return err + } + + return setstatat(t, dirfd, path, disallowEmptyPath, followFinalSymlink, &vfs.SetStatOptions{ + Stat: linux.Statx{ + Mask: linux.STATX_MODE, + Mode: uint16(mode & chmodMask), + }, + }) +} + +// Fchmod implements Linux syscall fchmod(2). +func Fchmod(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + mode := args[1].ModeT() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + return 0, nil, file.SetStat(t, vfs.SetStatOptions{ + Stat: linux.Statx{ + Mask: linux.STATX_MODE, + Mode: uint16(mode & chmodMask), + }, + }) +} + +// Chown implements Linux syscall chown(2). +func Chown(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + pathAddr := args[0].Pointer() + owner := args[1].Int() + group := args[2].Int() + return 0, nil, fchownat(t, linux.AT_FDCWD, pathAddr, owner, group, 0 /* flags */) +} + +// Lchown implements Linux syscall lchown(2). +func Lchown(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + pathAddr := args[0].Pointer() + owner := args[1].Int() + group := args[2].Int() + return 0, nil, fchownat(t, linux.AT_FDCWD, pathAddr, owner, group, linux.AT_SYMLINK_NOFOLLOW) +} + +// Fchownat implements Linux syscall fchownat(2). +func Fchownat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + dirfd := args[0].Int() + pathAddr := args[1].Pointer() + owner := args[2].Int() + group := args[3].Int() + flags := args[4].Int() + return 0, nil, fchownat(t, dirfd, pathAddr, owner, group, flags) +} + +func fchownat(t *kernel.Task, dirfd int32, pathAddr usermem.Addr, owner, group, flags int32) error { + if flags&^(linux.AT_EMPTY_PATH|linux.AT_SYMLINK_NOFOLLOW) != 0 { + return syserror.EINVAL + } + + path, err := copyInPath(t, pathAddr) + if err != nil { + return err + } + + var opts vfs.SetStatOptions + if err := populateSetStatOptionsForChown(t, owner, group, &opts); err != nil { + return err + } + + return setstatat(t, dirfd, path, shouldAllowEmptyPath(flags&linux.AT_EMPTY_PATH != 0), shouldFollowFinalSymlink(flags&linux.AT_SYMLINK_NOFOLLOW == 0), &opts) +} + +func populateSetStatOptionsForChown(t *kernel.Task, owner, group int32, opts *vfs.SetStatOptions) error { + userns := t.UserNamespace() + if owner != -1 { + kuid := userns.MapToKUID(auth.UID(owner)) + if !kuid.Ok() { + return syserror.EINVAL + } + opts.Stat.Mask |= linux.STATX_UID + opts.Stat.UID = uint32(kuid) + } + if group != -1 { + kgid := userns.MapToKGID(auth.GID(group)) + if !kgid.Ok() { + return syserror.EINVAL + } + opts.Stat.Mask |= linux.STATX_GID + opts.Stat.GID = uint32(kgid) + } + return nil +} + +// Fchown implements Linux syscall fchown(2). +func Fchown(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + owner := args[1].Int() + group := args[2].Int() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + var opts vfs.SetStatOptions + if err := populateSetStatOptionsForChown(t, owner, group, &opts); err != nil { + return 0, nil, err + } + return 0, nil, file.SetStat(t, opts) +} + +// Truncate implements Linux syscall truncate(2). +func Truncate(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + addr := args[0].Pointer() + length := args[1].Int64() + + if length < 0 { + return 0, nil, syserror.EINVAL + } + + path, err := copyInPath(t, addr) + if err != nil { + return 0, nil, err + } + + err = setstatat(t, linux.AT_FDCWD, path, disallowEmptyPath, followFinalSymlink, &vfs.SetStatOptions{ + Stat: linux.Statx{ + Mask: linux.STATX_SIZE, + Size: uint64(length), + }, + }) + return 0, nil, handleSetSizeError(t, err) +} + +// Ftruncate implements Linux syscall ftruncate(2). +func Ftruncate(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + length := args[1].Int64() + + if length < 0 { + return 0, nil, syserror.EINVAL + } + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + err := file.SetStat(t, vfs.SetStatOptions{ + Stat: linux.Statx{ + Mask: linux.STATX_SIZE, + Size: uint64(length), + }, + }) + return 0, nil, handleSetSizeError(t, err) +} + +// Utime implements Linux syscall utime(2). +func Utime(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + pathAddr := args[0].Pointer() + timesAddr := args[1].Pointer() + + path, err := copyInPath(t, pathAddr) + if err != nil { + return 0, nil, err + } + + opts := vfs.SetStatOptions{ + Stat: linux.Statx{ + Mask: linux.STATX_ATIME | linux.STATX_MTIME, + }, + } + if timesAddr == 0 { + opts.Stat.Atime.Nsec = linux.UTIME_NOW + opts.Stat.Mtime.Nsec = linux.UTIME_NOW + } else { + var times linux.Utime + if _, err := times.CopyIn(t, timesAddr); err != nil { + return 0, nil, err + } + opts.Stat.Atime.Sec = times.Actime + opts.Stat.Mtime.Sec = times.Modtime + } + + return 0, nil, setstatat(t, linux.AT_FDCWD, path, disallowEmptyPath, followFinalSymlink, &opts) +} + +// Utimes implements Linux syscall utimes(2). +func Utimes(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + pathAddr := args[0].Pointer() + timesAddr := args[1].Pointer() + + path, err := copyInPath(t, pathAddr) + if err != nil { + return 0, nil, err + } + + var opts vfs.SetStatOptions + if err := populateSetStatOptionsForUtimes(t, timesAddr, &opts); err != nil { + return 0, nil, err + } + + return 0, nil, setstatat(t, linux.AT_FDCWD, path, disallowEmptyPath, followFinalSymlink, &opts) +} + +// Futimesat implements Linux syscall futimesat(2). +func Futimesat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + dirfd := args[0].Int() + pathAddr := args[1].Pointer() + timesAddr := args[2].Pointer() + + // "If filename is NULL and dfd refers to an open file, then operate on the + // file. Otherwise look up filename, possibly using dfd as a starting + // point." - fs/utimes.c + var path fspath.Path + shouldAllowEmptyPath := allowEmptyPath + if dirfd == linux.AT_FDCWD || pathAddr != 0 { + var err error + path, err = copyInPath(t, pathAddr) + if err != nil { + return 0, nil, err + } + shouldAllowEmptyPath = disallowEmptyPath + } + + var opts vfs.SetStatOptions + if err := populateSetStatOptionsForUtimes(t, timesAddr, &opts); err != nil { + return 0, nil, err + } + + return 0, nil, setstatat(t, dirfd, path, shouldAllowEmptyPath, followFinalSymlink, &opts) +} + +func populateSetStatOptionsForUtimes(t *kernel.Task, timesAddr usermem.Addr, opts *vfs.SetStatOptions) error { + if timesAddr == 0 { + opts.Stat.Mask = linux.STATX_ATIME | linux.STATX_MTIME + opts.Stat.Atime.Nsec = linux.UTIME_NOW + opts.Stat.Mtime.Nsec = linux.UTIME_NOW + return nil + } + var times [2]linux.Timeval + if _, err := t.CopyIn(timesAddr, ×); err != nil { + return err + } + if times[0].Usec < 0 || times[0].Usec > 999999 || times[1].Usec < 0 || times[1].Usec > 999999 { + return syserror.EINVAL + } + opts.Stat.Mask = linux.STATX_ATIME | linux.STATX_MTIME + opts.Stat.Atime = linux.StatxTimestamp{ + Sec: times[0].Sec, + Nsec: uint32(times[0].Usec * 1000), + } + opts.Stat.Mtime = linux.StatxTimestamp{ + Sec: times[1].Sec, + Nsec: uint32(times[1].Usec * 1000), + } + return nil +} + +// Utimensat implements Linux syscall utimensat(2). +func Utimensat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + dirfd := args[0].Int() + pathAddr := args[1].Pointer() + timesAddr := args[2].Pointer() + flags := args[3].Int() + + // Linux requires that the UTIME_OMIT check occur before checking path or + // flags. + var opts vfs.SetStatOptions + if err := populateSetStatOptionsForUtimens(t, timesAddr, &opts); err != nil { + return 0, nil, err + } + if opts.Stat.Mask == 0 { + return 0, nil, nil + } + + if flags&^linux.AT_SYMLINK_NOFOLLOW != 0 { + return 0, nil, syserror.EINVAL + } + + // "If filename is NULL and dfd refers to an open file, then operate on the + // file. Otherwise look up filename, possibly using dfd as a starting + // point." - fs/utimes.c + var path fspath.Path + shouldAllowEmptyPath := allowEmptyPath + if dirfd == linux.AT_FDCWD || pathAddr != 0 { + var err error + path, err = copyInPath(t, pathAddr) + if err != nil { + return 0, nil, err + } + shouldAllowEmptyPath = disallowEmptyPath + } + + return 0, nil, setstatat(t, dirfd, path, shouldAllowEmptyPath, shouldFollowFinalSymlink(flags&linux.AT_SYMLINK_NOFOLLOW == 0), &opts) +} + +func populateSetStatOptionsForUtimens(t *kernel.Task, timesAddr usermem.Addr, opts *vfs.SetStatOptions) error { + if timesAddr == 0 { + opts.Stat.Mask = linux.STATX_ATIME | linux.STATX_MTIME + opts.Stat.Atime.Nsec = linux.UTIME_NOW + opts.Stat.Mtime.Nsec = linux.UTIME_NOW + return nil + } + var times [2]linux.Timespec + if _, err := t.CopyIn(timesAddr, ×); err != nil { + return err + } + if times[0].Nsec != linux.UTIME_OMIT { + if times[0].Nsec != linux.UTIME_NOW && (times[0].Nsec < 0 || times[0].Nsec > 999999999) { + return syserror.EINVAL + } + opts.Stat.Mask |= linux.STATX_ATIME + opts.Stat.Atime = linux.StatxTimestamp{ + Sec: times[0].Sec, + Nsec: uint32(times[0].Nsec), + } + } + if times[1].Nsec != linux.UTIME_OMIT { + if times[1].Nsec != linux.UTIME_NOW && (times[1].Nsec < 0 || times[1].Nsec > 999999999) { + return syserror.EINVAL + } + opts.Stat.Mask |= linux.STATX_MTIME + opts.Stat.Mtime = linux.StatxTimestamp{ + Sec: times[1].Sec, + Nsec: uint32(times[1].Nsec), + } + } + return nil +} + +func setstatat(t *kernel.Task, dirfd int32, path fspath.Path, shouldAllowEmptyPath shouldAllowEmptyPath, shouldFollowFinalSymlink shouldFollowFinalSymlink, opts *vfs.SetStatOptions) error { + root := t.FSContext().RootDirectoryVFS2() + defer root.DecRef() + start := root + if !path.Absolute { + if !path.HasComponents() && !bool(shouldAllowEmptyPath) { + return syserror.ENOENT + } + if dirfd == linux.AT_FDCWD { + start = t.FSContext().WorkingDirectoryVFS2() + defer start.DecRef() + } else { + dirfile := t.GetFileVFS2(dirfd) + if dirfile == nil { + return syserror.EBADF + } + if !path.HasComponents() { + // Use FileDescription.SetStat() instead of + // VirtualFilesystem.SetStatAt(), since the former may be able + // to use opened file state to expedite the SetStat. + err := dirfile.SetStat(t, *opts) + dirfile.DecRef() + return err + } + start = dirfile.VirtualDentry() + start.IncRef() + defer start.DecRef() + dirfile.DecRef() + } + } + return t.Kernel().VFS().SetStatAt(t, t.Credentials(), &vfs.PathOperation{ + Root: root, + Start: start, + Path: path, + FollowFinalSymlink: bool(shouldFollowFinalSymlink), + }, opts) +} + +func handleSetSizeError(t *kernel.Task, err error) error { + if err == syserror.ErrExceedsFileSizeLimit { + // Convert error to EFBIG and send a SIGXFSZ per setrlimit(2). + t.SendSignal(kernel.SignalInfoNoInfo(linux.SIGXFSZ, t, t)) + return syserror.EFBIG + } + return err +} diff --git a/pkg/sentry/syscalls/linux/vfs2/signal.go b/pkg/sentry/syscalls/linux/vfs2/signal.go new file mode 100644 index 000000000..623992f6f --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/signal.go @@ -0,0 +1,100 @@ +// Copyright 2018 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/signalfd" + "gvisor.dev/gvisor/pkg/sentry/kernel" + slinux "gvisor.dev/gvisor/pkg/sentry/syscalls/linux" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +// sharedSignalfd is shared between the two calls. +func sharedSignalfd(t *kernel.Task, fd int32, sigset usermem.Addr, sigsetsize uint, flags int32) (uintptr, *kernel.SyscallControl, error) { + // Copy in the signal mask. + mask, err := slinux.CopyInSigSet(t, sigset, sigsetsize) + if err != nil { + return 0, nil, err + } + + // Always check for valid flags, even if not creating. + if flags&^(linux.SFD_NONBLOCK|linux.SFD_CLOEXEC) != 0 { + return 0, nil, syserror.EINVAL + } + + // Is this a change to an existing signalfd? + // + // The spec indicates that this should adjust the mask. + if fd != -1 { + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Is this a signalfd? + if sfd, ok := file.Impl().(*signalfd.SignalFileDescription); ok { + sfd.SetMask(mask) + return 0, nil, nil + } + + // Not a signalfd. + return 0, nil, syserror.EINVAL + } + + fileFlags := uint32(linux.O_RDWR) + if flags&linux.SFD_NONBLOCK != 0 { + fileFlags |= linux.O_NONBLOCK + } + + // Create a new file. + vfsObj := t.Kernel().VFS() + file, err := signalfd.New(vfsObj, t, mask, fileFlags) + if err != nil { + return 0, nil, err + } + defer file.DecRef() + + // Create a new descriptor. + fd, err = t.NewFDFromVFS2(0, file, kernel.FDFlags{ + CloseOnExec: flags&linux.SFD_CLOEXEC != 0, + }) + if err != nil { + return 0, nil, err + } + + // Done. + return uintptr(fd), nil, nil +} + +// Signalfd implements the linux syscall signalfd(2). +func Signalfd(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + sigset := args[1].Pointer() + sigsetsize := args[2].SizeT() + return sharedSignalfd(t, fd, sigset, sigsetsize, 0) +} + +// Signalfd4 implements the linux syscall signalfd4(2). +func Signalfd4(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + sigset := args[1].Pointer() + sigsetsize := args[2].SizeT() + flags := args[3].Int() + return sharedSignalfd(t, fd, sigset, sigsetsize, flags) +} diff --git a/pkg/sentry/syscalls/linux/vfs2/socket.go b/pkg/sentry/syscalls/linux/vfs2/socket.go new file mode 100644 index 000000000..10b668477 --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/socket.go @@ -0,0 +1,1139 @@ +// Copyright 2018 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "time" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/binary" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" + "gvisor.dev/gvisor/pkg/sentry/socket" + "gvisor.dev/gvisor/pkg/sentry/socket/control" + "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" + slinux "gvisor.dev/gvisor/pkg/sentry/syscalls/linux" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserr" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +// minListenBacklog is the minimum reasonable backlog for listening sockets. +const minListenBacklog = 8 + +// maxListenBacklog is the maximum allowed backlog for listening sockets. +const maxListenBacklog = 1024 + +// maxAddrLen is the maximum socket address length we're willing to accept. +const maxAddrLen = 200 + +// maxOptLen is the maximum sockopt parameter length we're willing to accept. +const maxOptLen = 1024 * 8 + +// maxControlLen is the maximum length of the msghdr.msg_control buffer we're +// willing to accept. Note that this limit is smaller than Linux, which allows +// buffers upto INT_MAX. +const maxControlLen = 10 * 1024 * 1024 + +// nameLenOffset is the offset from the start of the MessageHeader64 struct to +// the NameLen field. +const nameLenOffset = 8 + +// controlLenOffset is the offset form the start of the MessageHeader64 struct +// to the ControlLen field. +const controlLenOffset = 40 + +// flagsOffset is the offset form the start of the MessageHeader64 struct +// to the Flags field. +const flagsOffset = 48 + +const sizeOfInt32 = 4 + +// messageHeader64Len is the length of a MessageHeader64 struct. +var messageHeader64Len = uint64(binary.Size(MessageHeader64{})) + +// multipleMessageHeader64Len is the length of a multipeMessageHeader64 struct. +var multipleMessageHeader64Len = uint64(binary.Size(multipleMessageHeader64{})) + +// baseRecvFlags are the flags that are accepted across recvmsg(2), +// recvmmsg(2), and recvfrom(2). +const baseRecvFlags = linux.MSG_OOB | linux.MSG_DONTROUTE | linux.MSG_DONTWAIT | linux.MSG_NOSIGNAL | linux.MSG_WAITALL | linux.MSG_TRUNC | linux.MSG_CTRUNC + +// MessageHeader64 is the 64-bit representation of the msghdr struct used in +// the recvmsg and sendmsg syscalls. +type MessageHeader64 struct { + // Name is the optional pointer to a network address buffer. + Name uint64 + + // NameLen is the length of the buffer pointed to by Name. + NameLen uint32 + _ uint32 + + // Iov is a pointer to an array of io vectors that describe the memory + // locations involved in the io operation. + Iov uint64 + + // IovLen is the length of the array pointed to by Iov. + IovLen uint64 + + // Control is the optional pointer to ancillary control data. + Control uint64 + + // ControlLen is the length of the data pointed to by Control. + ControlLen uint64 + + // Flags on the sent/received message. + Flags int32 + _ int32 +} + +// multipleMessageHeader64 is the 64-bit representation of the mmsghdr struct used in +// the recvmmsg and sendmmsg syscalls. +type multipleMessageHeader64 struct { + msgHdr MessageHeader64 + msgLen uint32 + _ int32 +} + +// CopyInMessageHeader64 copies a message header from user to kernel memory. +func CopyInMessageHeader64(t *kernel.Task, addr usermem.Addr, msg *MessageHeader64) error { + b := t.CopyScratchBuffer(52) + if _, err := t.CopyInBytes(addr, b); err != nil { + return err + } + + msg.Name = usermem.ByteOrder.Uint64(b[0:]) + msg.NameLen = usermem.ByteOrder.Uint32(b[8:]) + msg.Iov = usermem.ByteOrder.Uint64(b[16:]) + msg.IovLen = usermem.ByteOrder.Uint64(b[24:]) + msg.Control = usermem.ByteOrder.Uint64(b[32:]) + msg.ControlLen = usermem.ByteOrder.Uint64(b[40:]) + msg.Flags = int32(usermem.ByteOrder.Uint32(b[48:])) + + return nil +} + +// CaptureAddress allocates memory for and copies a socket address structure +// from the untrusted address space range. +func CaptureAddress(t *kernel.Task, addr usermem.Addr, addrlen uint32) ([]byte, error) { + if addrlen > maxAddrLen { + return nil, syserror.EINVAL + } + + addrBuf := make([]byte, addrlen) + if _, err := t.CopyInBytes(addr, addrBuf); err != nil { + return nil, err + } + + return addrBuf, nil +} + +// writeAddress writes a sockaddr structure and its length to an output buffer +// in the unstrusted address space range. If the address is bigger than the +// buffer, it is truncated. +func writeAddress(t *kernel.Task, addr interface{}, addrLen uint32, addrPtr usermem.Addr, addrLenPtr usermem.Addr) error { + // Get the buffer length. + var bufLen uint32 + if _, err := t.CopyIn(addrLenPtr, &bufLen); err != nil { + return err + } + + if int32(bufLen) < 0 { + return syserror.EINVAL + } + + // Write the length unconditionally. + if _, err := t.CopyOut(addrLenPtr, addrLen); err != nil { + return err + } + + if addr == nil { + return nil + } + + if bufLen > addrLen { + bufLen = addrLen + } + + // Copy as much of the address as will fit in the buffer. + encodedAddr := binary.Marshal(nil, usermem.ByteOrder, addr) + if bufLen > uint32(len(encodedAddr)) { + bufLen = uint32(len(encodedAddr)) + } + _, err := t.CopyOutBytes(addrPtr, encodedAddr[:int(bufLen)]) + return err +} + +// Socket implements the linux syscall socket(2). +func Socket(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + domain := int(args[0].Int()) + stype := args[1].Int() + protocol := int(args[2].Int()) + + // Check and initialize the flags. + if stype & ^(0xf|linux.SOCK_NONBLOCK|linux.SOCK_CLOEXEC) != 0 { + return 0, nil, syserror.EINVAL + } + + // Create the new socket. + s, e := socket.NewVFS2(t, domain, linux.SockType(stype&0xf), protocol) + if e != nil { + return 0, nil, e.ToError() + } + defer s.DecRef() + + if err := s.SetStatusFlags(t, t.Credentials(), uint32(stype&linux.SOCK_NONBLOCK)); err != nil { + return 0, nil, err + } + + fd, err := t.NewFDFromVFS2(0, s, kernel.FDFlags{ + CloseOnExec: stype&linux.SOCK_CLOEXEC != 0, + }) + if err != nil { + return 0, nil, err + } + + return uintptr(fd), nil, nil +} + +// SocketPair implements the linux syscall socketpair(2). +func SocketPair(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + domain := int(args[0].Int()) + stype := args[1].Int() + protocol := int(args[2].Int()) + addr := args[3].Pointer() + + // Check and initialize the flags. + if stype & ^(0xf|linux.SOCK_NONBLOCK|linux.SOCK_CLOEXEC) != 0 { + return 0, nil, syserror.EINVAL + } + + // Create the socket pair. + s1, s2, e := socket.PairVFS2(t, domain, linux.SockType(stype&0xf), protocol) + if e != nil { + return 0, nil, e.ToError() + } + // Adding to the FD table will cause an extra reference to be acquired. + defer s1.DecRef() + defer s2.DecRef() + + nonblocking := uint32(stype & linux.SOCK_NONBLOCK) + if err := s1.SetStatusFlags(t, t.Credentials(), nonblocking); err != nil { + return 0, nil, err + } + if err := s2.SetStatusFlags(t, t.Credentials(), nonblocking); err != nil { + return 0, nil, err + } + + // Create the FDs for the sockets. + flags := kernel.FDFlags{ + CloseOnExec: stype&linux.SOCK_CLOEXEC != 0, + } + fds, err := t.NewFDsVFS2(0, []*vfs.FileDescription{s1, s2}, flags) + if err != nil { + return 0, nil, err + } + + if _, err := t.CopyOut(addr, fds); err != nil { + for _, fd := range fds { + if _, file := t.FDTable().Remove(fd); file != nil { + file.DecRef() + } + } + return 0, nil, err + } + + return 0, nil, nil +} + +// Connect implements the linux syscall connect(2). +func Connect(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + addr := args[1].Pointer() + addrlen := args[2].Uint() + + // Get socket from the file descriptor. + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Extract the socket. + s, ok := file.Impl().(socket.SocketVFS2) + if !ok { + return 0, nil, syserror.ENOTSOCK + } + + // Capture address and call syscall implementation. + a, err := CaptureAddress(t, addr, addrlen) + if err != nil { + return 0, nil, err + } + + blocking := (file.StatusFlags() & linux.SOCK_NONBLOCK) == 0 + return 0, nil, syserror.ConvertIntr(s.Connect(t, a, blocking).ToError(), kernel.ERESTARTSYS) +} + +// accept is the implementation of the accept syscall. It is called by accept +// and accept4 syscall handlers. +func accept(t *kernel.Task, fd int32, addr usermem.Addr, addrLen usermem.Addr, flags int) (uintptr, error) { + // Check that no unsupported flags are passed in. + if flags & ^(linux.SOCK_NONBLOCK|linux.SOCK_CLOEXEC) != 0 { + return 0, syserror.EINVAL + } + + // Get socket from the file descriptor. + file := t.GetFileVFS2(fd) + if file == nil { + return 0, syserror.EBADF + } + defer file.DecRef() + + // Extract the socket. + s, ok := file.Impl().(socket.SocketVFS2) + if !ok { + return 0, syserror.ENOTSOCK + } + + // Call the syscall implementation for this socket, then copy the + // output address if one is specified. + blocking := (file.StatusFlags() & linux.SOCK_NONBLOCK) == 0 + + peerRequested := addrLen != 0 + nfd, peer, peerLen, e := s.Accept(t, peerRequested, flags, blocking) + if e != nil { + return 0, syserror.ConvertIntr(e.ToError(), kernel.ERESTARTSYS) + } + if peerRequested { + // NOTE(magi): Linux does not give you an error if it can't + // write the data back out so neither do we. + if err := writeAddress(t, peer, peerLen, addr, addrLen); err == syserror.EINVAL { + return 0, err + } + } + return uintptr(nfd), nil +} + +// Accept4 implements the linux syscall accept4(2). +func Accept4(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + addr := args[1].Pointer() + addrlen := args[2].Pointer() + flags := int(args[3].Int()) + + n, err := accept(t, fd, addr, addrlen, flags) + return n, nil, err +} + +// Accept implements the linux syscall accept(2). +func Accept(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + addr := args[1].Pointer() + addrlen := args[2].Pointer() + + n, err := accept(t, fd, addr, addrlen, 0) + return n, nil, err +} + +// Bind implements the linux syscall bind(2). +func Bind(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + addr := args[1].Pointer() + addrlen := args[2].Uint() + + // Get socket from the file descriptor. + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Extract the socket. + s, ok := file.Impl().(socket.SocketVFS2) + if !ok { + return 0, nil, syserror.ENOTSOCK + } + + // Capture address and call syscall implementation. + a, err := CaptureAddress(t, addr, addrlen) + if err != nil { + return 0, nil, err + } + + return 0, nil, s.Bind(t, a).ToError() +} + +// Listen implements the linux syscall listen(2). +func Listen(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + backlog := args[1].Int() + + // Get socket from the file descriptor. + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Extract the socket. + s, ok := file.Impl().(socket.SocketVFS2) + if !ok { + return 0, nil, syserror.ENOTSOCK + } + + // Per Linux, the backlog is silently capped to reasonable values. + if backlog <= 0 { + backlog = minListenBacklog + } + if backlog > maxListenBacklog { + backlog = maxListenBacklog + } + + return 0, nil, s.Listen(t, int(backlog)).ToError() +} + +// Shutdown implements the linux syscall shutdown(2). +func Shutdown(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + how := args[1].Int() + + // Get socket from the file descriptor. + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Extract the socket. + s, ok := file.Impl().(socket.SocketVFS2) + if !ok { + return 0, nil, syserror.ENOTSOCK + } + + // Validate how, then call syscall implementation. + switch how { + case linux.SHUT_RD, linux.SHUT_WR, linux.SHUT_RDWR: + default: + return 0, nil, syserror.EINVAL + } + + return 0, nil, s.Shutdown(t, int(how)).ToError() +} + +// GetSockOpt implements the linux syscall getsockopt(2). +func GetSockOpt(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + level := args[1].Int() + name := args[2].Int() + optValAddr := args[3].Pointer() + optLenAddr := args[4].Pointer() + + // Get socket from the file descriptor. + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Extract the socket. + s, ok := file.Impl().(socket.SocketVFS2) + if !ok { + return 0, nil, syserror.ENOTSOCK + } + + // Read the length. Reject negative values. + optLen := int32(0) + if _, err := t.CopyIn(optLenAddr, &optLen); err != nil { + return 0, nil, err + } + if optLen < 0 { + return 0, nil, syserror.EINVAL + } + + // Call syscall implementation then copy both value and value len out. + v, e := getSockOpt(t, s, int(level), int(name), optValAddr, int(optLen)) + if e != nil { + return 0, nil, e.ToError() + } + + vLen := int32(binary.Size(v)) + if _, err := t.CopyOut(optLenAddr, vLen); err != nil { + return 0, nil, err + } + + if v != nil { + if _, err := t.CopyOut(optValAddr, v); err != nil { + return 0, nil, err + } + } + + return 0, nil, nil +} + +// getSockOpt tries to handle common socket options, or dispatches to a specific +// socket implementation. +func getSockOpt(t *kernel.Task, s socket.SocketVFS2, level, name int, optValAddr usermem.Addr, len int) (interface{}, *syserr.Error) { + if level == linux.SOL_SOCKET { + switch name { + case linux.SO_TYPE, linux.SO_DOMAIN, linux.SO_PROTOCOL: + if len < sizeOfInt32 { + return nil, syserr.ErrInvalidArgument + } + } + + switch name { + case linux.SO_TYPE: + _, skType, _ := s.Type() + return int32(skType), nil + case linux.SO_DOMAIN: + family, _, _ := s.Type() + return int32(family), nil + case linux.SO_PROTOCOL: + _, _, protocol := s.Type() + return int32(protocol), nil + } + } + + return s.GetSockOpt(t, level, name, optValAddr, len) +} + +// SetSockOpt implements the linux syscall setsockopt(2). +// +// Note that unlike Linux, enabling SO_PASSCRED does not autobind the socket. +func SetSockOpt(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + level := args[1].Int() + name := args[2].Int() + optValAddr := args[3].Pointer() + optLen := args[4].Int() + + // Get socket from the file descriptor. + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Extract the socket. + s, ok := file.Impl().(socket.SocketVFS2) + if !ok { + return 0, nil, syserror.ENOTSOCK + } + + if optLen < 0 { + return 0, nil, syserror.EINVAL + } + if optLen > maxOptLen { + return 0, nil, syserror.EINVAL + } + buf := t.CopyScratchBuffer(int(optLen)) + if _, err := t.CopyIn(optValAddr, &buf); err != nil { + return 0, nil, err + } + + // Call syscall implementation. + if err := s.SetSockOpt(t, int(level), int(name), buf); err != nil { + return 0, nil, err.ToError() + } + + return 0, nil, nil +} + +// GetSockName implements the linux syscall getsockname(2). +func GetSockName(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + addr := args[1].Pointer() + addrlen := args[2].Pointer() + + // Get socket from the file descriptor. + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Extract the socket. + s, ok := file.Impl().(socket.SocketVFS2) + if !ok { + return 0, nil, syserror.ENOTSOCK + } + + // Get the socket name and copy it to the caller. + v, vl, err := s.GetSockName(t) + if err != nil { + return 0, nil, err.ToError() + } + + return 0, nil, writeAddress(t, v, vl, addr, addrlen) +} + +// GetPeerName implements the linux syscall getpeername(2). +func GetPeerName(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + addr := args[1].Pointer() + addrlen := args[2].Pointer() + + // Get socket from the file descriptor. + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Extract the socket. + s, ok := file.Impl().(socket.SocketVFS2) + if !ok { + return 0, nil, syserror.ENOTSOCK + } + + // Get the socket peer name and copy it to the caller. + v, vl, err := s.GetPeerName(t) + if err != nil { + return 0, nil, err.ToError() + } + + return 0, nil, writeAddress(t, v, vl, addr, addrlen) +} + +// RecvMsg implements the linux syscall recvmsg(2). +func RecvMsg(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + msgPtr := args[1].Pointer() + flags := args[2].Int() + + if t.Arch().Width() != 8 { + // We only handle 64-bit for now. + return 0, nil, syserror.EINVAL + } + + // Get socket from the file descriptor. + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Extract the socket. + s, ok := file.Impl().(socket.SocketVFS2) + if !ok { + return 0, nil, syserror.ENOTSOCK + } + + // Reject flags that we don't handle yet. + if flags & ^(baseRecvFlags|linux.MSG_PEEK|linux.MSG_CMSG_CLOEXEC|linux.MSG_ERRQUEUE) != 0 { + return 0, nil, syserror.EINVAL + } + + if (file.StatusFlags() & linux.SOCK_NONBLOCK) != 0 { + flags |= linux.MSG_DONTWAIT + } + + var haveDeadline bool + var deadline ktime.Time + if dl := s.RecvTimeout(); dl > 0 { + deadline = t.Kernel().MonotonicClock().Now().Add(time.Duration(dl) * time.Nanosecond) + haveDeadline = true + } else if dl < 0 { + flags |= linux.MSG_DONTWAIT + } + + n, err := recvSingleMsg(t, s, msgPtr, flags, haveDeadline, deadline) + return n, nil, err +} + +// RecvMMsg implements the linux syscall recvmmsg(2). +func RecvMMsg(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + msgPtr := args[1].Pointer() + vlen := args[2].Uint() + flags := args[3].Int() + toPtr := args[4].Pointer() + + if t.Arch().Width() != 8 { + // We only handle 64-bit for now. + return 0, nil, syserror.EINVAL + } + + // Reject flags that we don't handle yet. + if flags & ^(baseRecvFlags|linux.MSG_CMSG_CLOEXEC|linux.MSG_ERRQUEUE) != 0 { + return 0, nil, syserror.EINVAL + } + + // Get socket from the file descriptor. + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Extract the socket. + s, ok := file.Impl().(socket.SocketVFS2) + if !ok { + return 0, nil, syserror.ENOTSOCK + } + + if (file.StatusFlags() & linux.SOCK_NONBLOCK) != 0 { + flags |= linux.MSG_DONTWAIT + } + + var haveDeadline bool + var deadline ktime.Time + if toPtr != 0 { + var ts linux.Timespec + if _, err := ts.CopyIn(t, toPtr); err != nil { + return 0, nil, err + } + if !ts.Valid() { + return 0, nil, syserror.EINVAL + } + deadline = t.Kernel().MonotonicClock().Now().Add(ts.ToDuration()) + haveDeadline = true + } + + if !haveDeadline { + if dl := s.RecvTimeout(); dl > 0 { + deadline = t.Kernel().MonotonicClock().Now().Add(time.Duration(dl) * time.Nanosecond) + haveDeadline = true + } else if dl < 0 { + flags |= linux.MSG_DONTWAIT + } + } + + var count uint32 + var err error + for i := uint64(0); i < uint64(vlen); i++ { + mp, ok := msgPtr.AddLength(i * multipleMessageHeader64Len) + if !ok { + return 0, nil, syserror.EFAULT + } + var n uintptr + if n, err = recvSingleMsg(t, s, mp, flags, haveDeadline, deadline); err != nil { + break + } + + // Copy the received length to the caller. + lp, ok := mp.AddLength(messageHeader64Len) + if !ok { + return 0, nil, syserror.EFAULT + } + if _, err = t.CopyOut(lp, uint32(n)); err != nil { + break + } + count++ + } + + if count == 0 { + return 0, nil, err + } + return uintptr(count), nil, nil +} + +func recvSingleMsg(t *kernel.Task, s socket.SocketVFS2, msgPtr usermem.Addr, flags int32, haveDeadline bool, deadline ktime.Time) (uintptr, error) { + // Capture the message header and io vectors. + var msg MessageHeader64 + if err := CopyInMessageHeader64(t, msgPtr, &msg); err != nil { + return 0, err + } + + if msg.IovLen > linux.UIO_MAXIOV { + return 0, syserror.EMSGSIZE + } + dst, err := t.IovecsIOSequence(usermem.Addr(msg.Iov), int(msg.IovLen), usermem.IOOpts{ + AddressSpaceActive: true, + }) + if err != nil { + return 0, err + } + + // FIXME(b/63594852): Pretend we have an empty error queue. + if flags&linux.MSG_ERRQUEUE != 0 { + return 0, syserror.EAGAIN + } + + // Fast path when no control message nor name buffers are provided. + if msg.ControlLen == 0 && msg.NameLen == 0 { + n, mflags, _, _, cms, err := s.RecvMsg(t, dst, int(flags), haveDeadline, deadline, false, 0) + if err != nil { + return 0, syserror.ConvertIntr(err.ToError(), kernel.ERESTARTSYS) + } + if !cms.Unix.Empty() { + mflags |= linux.MSG_CTRUNC + cms.Release() + } + + if int(msg.Flags) != mflags { + // Copy out the flags to the caller. + if _, err := t.CopyOut(msgPtr+flagsOffset, int32(mflags)); err != nil { + return 0, err + } + } + + return uintptr(n), nil + } + + if msg.ControlLen > maxControlLen { + return 0, syserror.ENOBUFS + } + n, mflags, sender, senderLen, cms, e := s.RecvMsg(t, dst, int(flags), haveDeadline, deadline, msg.NameLen != 0, msg.ControlLen) + if e != nil { + return 0, syserror.ConvertIntr(e.ToError(), kernel.ERESTARTSYS) + } + defer cms.Release() + + controlData := make([]byte, 0, msg.ControlLen) + controlData = control.PackControlMessages(t, cms, controlData) + + if cr, ok := s.(transport.Credentialer); ok && cr.Passcred() { + creds, _ := cms.Unix.Credentials.(control.SCMCredentials) + controlData, mflags = control.PackCredentials(t, creds, controlData, mflags) + } + + if cms.Unix.Rights != nil { + controlData, mflags = control.PackRightsVFS2(t, cms.Unix.Rights.(control.SCMRightsVFS2), flags&linux.MSG_CMSG_CLOEXEC != 0, controlData, mflags) + } + + // Copy the address to the caller. + if msg.NameLen != 0 { + if err := writeAddress(t, sender, senderLen, usermem.Addr(msg.Name), usermem.Addr(msgPtr+nameLenOffset)); err != nil { + return 0, err + } + } + + // Copy the control data to the caller. + if _, err := t.CopyOut(msgPtr+controlLenOffset, uint64(len(controlData))); err != nil { + return 0, err + } + if len(controlData) > 0 { + if _, err := t.CopyOut(usermem.Addr(msg.Control), controlData); err != nil { + return 0, err + } + } + + // Copy out the flags to the caller. + if _, err := t.CopyOut(msgPtr+flagsOffset, int32(mflags)); err != nil { + return 0, err + } + + return uintptr(n), nil +} + +// recvFrom is the implementation of the recvfrom syscall. It is called by +// recvfrom and recv syscall handlers. +func recvFrom(t *kernel.Task, fd int32, bufPtr usermem.Addr, bufLen uint64, flags int32, namePtr usermem.Addr, nameLenPtr usermem.Addr) (uintptr, error) { + if int(bufLen) < 0 { + return 0, syserror.EINVAL + } + + // Reject flags that we don't handle yet. + if flags & ^(baseRecvFlags|linux.MSG_PEEK|linux.MSG_CONFIRM) != 0 { + return 0, syserror.EINVAL + } + + // Get socket from the file descriptor. + file := t.GetFileVFS2(fd) + if file == nil { + return 0, syserror.EBADF + } + defer file.DecRef() + + // Extract the socket. + s, ok := file.Impl().(socket.SocketVFS2) + if !ok { + return 0, syserror.ENOTSOCK + } + + if (file.StatusFlags() & linux.SOCK_NONBLOCK) != 0 { + flags |= linux.MSG_DONTWAIT + } + + dst, err := t.SingleIOSequence(bufPtr, int(bufLen), usermem.IOOpts{ + AddressSpaceActive: true, + }) + if err != nil { + return 0, err + } + + var haveDeadline bool + var deadline ktime.Time + if dl := s.RecvTimeout(); dl > 0 { + deadline = t.Kernel().MonotonicClock().Now().Add(time.Duration(dl) * time.Nanosecond) + haveDeadline = true + } else if dl < 0 { + flags |= linux.MSG_DONTWAIT + } + + n, _, sender, senderLen, cm, e := s.RecvMsg(t, dst, int(flags), haveDeadline, deadline, nameLenPtr != 0, 0) + cm.Release() + if e != nil { + return 0, syserror.ConvertIntr(e.ToError(), kernel.ERESTARTSYS) + } + + // Copy the address to the caller. + if nameLenPtr != 0 { + if err := writeAddress(t, sender, senderLen, namePtr, nameLenPtr); err != nil { + return 0, err + } + } + + return uintptr(n), nil +} + +// RecvFrom implements the linux syscall recvfrom(2). +func RecvFrom(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + bufPtr := args[1].Pointer() + bufLen := args[2].Uint64() + flags := args[3].Int() + namePtr := args[4].Pointer() + nameLenPtr := args[5].Pointer() + + n, err := recvFrom(t, fd, bufPtr, bufLen, flags, namePtr, nameLenPtr) + return n, nil, err +} + +// SendMsg implements the linux syscall sendmsg(2). +func SendMsg(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + msgPtr := args[1].Pointer() + flags := args[2].Int() + + if t.Arch().Width() != 8 { + // We only handle 64-bit for now. + return 0, nil, syserror.EINVAL + } + + // Get socket from the file descriptor. + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Extract the socket. + s, ok := file.Impl().(socket.SocketVFS2) + if !ok { + return 0, nil, syserror.ENOTSOCK + } + + // Reject flags that we don't handle yet. + if flags & ^(linux.MSG_DONTWAIT|linux.MSG_EOR|linux.MSG_MORE|linux.MSG_NOSIGNAL) != 0 { + return 0, nil, syserror.EINVAL + } + + if (file.StatusFlags() & linux.SOCK_NONBLOCK) != 0 { + flags |= linux.MSG_DONTWAIT + } + + n, err := sendSingleMsg(t, s, file, msgPtr, flags) + return n, nil, err +} + +// SendMMsg implements the linux syscall sendmmsg(2). +func SendMMsg(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + msgPtr := args[1].Pointer() + vlen := args[2].Uint() + flags := args[3].Int() + + if t.Arch().Width() != 8 { + // We only handle 64-bit for now. + return 0, nil, syserror.EINVAL + } + + // Get socket from the file descriptor. + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Extract the socket. + s, ok := file.Impl().(socket.SocketVFS2) + if !ok { + return 0, nil, syserror.ENOTSOCK + } + + // Reject flags that we don't handle yet. + if flags & ^(linux.MSG_DONTWAIT|linux.MSG_EOR|linux.MSG_MORE|linux.MSG_NOSIGNAL) != 0 { + return 0, nil, syserror.EINVAL + } + + if (file.StatusFlags() & linux.SOCK_NONBLOCK) != 0 { + flags |= linux.MSG_DONTWAIT + } + + var count uint32 + var err error + for i := uint64(0); i < uint64(vlen); i++ { + mp, ok := msgPtr.AddLength(i * multipleMessageHeader64Len) + if !ok { + return 0, nil, syserror.EFAULT + } + var n uintptr + if n, err = sendSingleMsg(t, s, file, mp, flags); err != nil { + break + } + + // Copy the received length to the caller. + lp, ok := mp.AddLength(messageHeader64Len) + if !ok { + return 0, nil, syserror.EFAULT + } + if _, err = t.CopyOut(lp, uint32(n)); err != nil { + break + } + count++ + } + + if count == 0 { + return 0, nil, err + } + return uintptr(count), nil, nil +} + +func sendSingleMsg(t *kernel.Task, s socket.SocketVFS2, file *vfs.FileDescription, msgPtr usermem.Addr, flags int32) (uintptr, error) { + // Capture the message header. + var msg MessageHeader64 + if err := CopyInMessageHeader64(t, msgPtr, &msg); err != nil { + return 0, err + } + + var controlData []byte + if msg.ControlLen > 0 { + // Put an upper bound to prevent large allocations. + if msg.ControlLen > maxControlLen { + return 0, syserror.ENOBUFS + } + controlData = make([]byte, msg.ControlLen) + if _, err := t.CopyIn(usermem.Addr(msg.Control), &controlData); err != nil { + return 0, err + } + } + + // Read the destination address if one is specified. + var to []byte + if msg.NameLen != 0 { + var err error + to, err = CaptureAddress(t, usermem.Addr(msg.Name), msg.NameLen) + if err != nil { + return 0, err + } + } + + // Read data then call the sendmsg implementation. + if msg.IovLen > linux.UIO_MAXIOV { + return 0, syserror.EMSGSIZE + } + src, err := t.IovecsIOSequence(usermem.Addr(msg.Iov), int(msg.IovLen), usermem.IOOpts{ + AddressSpaceActive: true, + }) + if err != nil { + return 0, err + } + + controlMessages, err := control.Parse(t, s, controlData) + if err != nil { + return 0, err + } + + var haveDeadline bool + var deadline ktime.Time + if dl := s.SendTimeout(); dl > 0 { + deadline = t.Kernel().MonotonicClock().Now().Add(time.Duration(dl) * time.Nanosecond) + haveDeadline = true + } else if dl < 0 { + flags |= linux.MSG_DONTWAIT + } + + // Call the syscall implementation. + n, e := s.SendMsg(t, src, to, int(flags), haveDeadline, deadline, controlMessages) + err = slinux.HandleIOErrorVFS2(t, n != 0, e.ToError(), kernel.ERESTARTSYS, "sendmsg", file) + if err != nil { + controlMessages.Release() + } + return uintptr(n), err +} + +// sendTo is the implementation of the sendto syscall. It is called by sendto +// and send syscall handlers. +func sendTo(t *kernel.Task, fd int32, bufPtr usermem.Addr, bufLen uint64, flags int32, namePtr usermem.Addr, nameLen uint32) (uintptr, error) { + bl := int(bufLen) + if bl < 0 { + return 0, syserror.EINVAL + } + + // Get socket from the file descriptor. + file := t.GetFileVFS2(fd) + if file == nil { + return 0, syserror.EBADF + } + defer file.DecRef() + + // Extract the socket. + s, ok := file.Impl().(socket.SocketVFS2) + if !ok { + return 0, syserror.ENOTSOCK + } + + if (file.StatusFlags() & linux.SOCK_NONBLOCK) != 0 { + flags |= linux.MSG_DONTWAIT + } + + // Read the destination address if one is specified. + var to []byte + var err error + if namePtr != 0 { + to, err = CaptureAddress(t, namePtr, nameLen) + if err != nil { + return 0, err + } + } + + src, err := t.SingleIOSequence(bufPtr, bl, usermem.IOOpts{ + AddressSpaceActive: true, + }) + if err != nil { + return 0, err + } + + var haveDeadline bool + var deadline ktime.Time + if dl := s.SendTimeout(); dl > 0 { + deadline = t.Kernel().MonotonicClock().Now().Add(time.Duration(dl) * time.Nanosecond) + haveDeadline = true + } else if dl < 0 { + flags |= linux.MSG_DONTWAIT + } + + // Call the syscall implementation. + n, e := s.SendMsg(t, src, to, int(flags), haveDeadline, deadline, socket.ControlMessages{Unix: control.New(t, s, nil)}) + return uintptr(n), slinux.HandleIOErrorVFS2(t, n != 0, e.ToError(), kernel.ERESTARTSYS, "sendto", file) +} + +// SendTo implements the linux syscall sendto(2). +func SendTo(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + bufPtr := args[1].Pointer() + bufLen := args[2].Uint64() + flags := args[3].Int() + namePtr := args[4].Pointer() + nameLen := args[5].Uint() + + n, err := sendTo(t, fd, bufPtr, bufLen, flags, namePtr, nameLen) + return n, nil, err +} diff --git a/pkg/sentry/syscalls/linux/vfs2/splice.go b/pkg/sentry/syscalls/linux/vfs2/splice.go new file mode 100644 index 000000000..945a364a7 --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/splice.go @@ -0,0 +1,291 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/kernel/pipe" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/waiter" +) + +// Splice implements Linux syscall splice(2). +func Splice(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + inFD := args[0].Int() + inOffsetPtr := args[1].Pointer() + outFD := args[2].Int() + outOffsetPtr := args[3].Pointer() + count := int64(args[4].SizeT()) + flags := args[5].Int() + + if count == 0 { + return 0, nil, nil + } + if count > int64(kernel.MAX_RW_COUNT) { + count = int64(kernel.MAX_RW_COUNT) + } + + // Check for invalid flags. + if flags&^(linux.SPLICE_F_MOVE|linux.SPLICE_F_NONBLOCK|linux.SPLICE_F_MORE|linux.SPLICE_F_GIFT) != 0 { + return 0, nil, syserror.EINVAL + } + + // Get file descriptions. + inFile := t.GetFileVFS2(inFD) + if inFile == nil { + return 0, nil, syserror.EBADF + } + defer inFile.DecRef() + outFile := t.GetFileVFS2(outFD) + if outFile == nil { + return 0, nil, syserror.EBADF + } + defer outFile.DecRef() + + // Check that both files support the required directionality. + if !inFile.IsReadable() || !outFile.IsWritable() { + return 0, nil, syserror.EBADF + } + + // The operation is non-blocking if anything is non-blocking. + // + // N.B. This is a rather simplistic heuristic that avoids some + // poor edge case behavior since the exact semantics here are + // underspecified and vary between versions of Linux itself. + nonBlock := ((inFile.StatusFlags()|outFile.StatusFlags())&linux.O_NONBLOCK != 0) || (flags&linux.SPLICE_F_NONBLOCK != 0) + + // At least one file description must represent a pipe. + inPipeFD, inIsPipe := inFile.Impl().(*pipe.VFSPipeFD) + outPipeFD, outIsPipe := outFile.Impl().(*pipe.VFSPipeFD) + if !inIsPipe && !outIsPipe { + return 0, nil, syserror.EINVAL + } + + // Copy in offsets. + inOffset := int64(-1) + if inOffsetPtr != 0 { + if inIsPipe { + return 0, nil, syserror.ESPIPE + } + if inFile.Options().DenyPRead { + return 0, nil, syserror.EINVAL + } + if _, err := t.CopyIn(inOffsetPtr, &inOffset); err != nil { + return 0, nil, err + } + if inOffset < 0 { + return 0, nil, syserror.EINVAL + } + } + outOffset := int64(-1) + if outOffsetPtr != 0 { + if outIsPipe { + return 0, nil, syserror.ESPIPE + } + if outFile.Options().DenyPWrite { + return 0, nil, syserror.EINVAL + } + if _, err := t.CopyIn(outOffsetPtr, &outOffset); err != nil { + return 0, nil, err + } + if outOffset < 0 { + return 0, nil, syserror.EINVAL + } + } + + // Move data. + var ( + n int64 + err error + inCh chan struct{} + outCh chan struct{} + ) + for { + // If both input and output are pipes, delegate to the pipe + // implementation. Otherwise, exactly one end is a pipe, which we + // ensure is consistently ordered after the non-pipe FD's locks by + // passing the pipe FD as usermem.IO to the non-pipe end. + switch { + case inIsPipe && outIsPipe: + n, err = pipe.Splice(t, outPipeFD, inPipeFD, count) + case inIsPipe: + if outOffset != -1 { + n, err = outFile.PWrite(t, inPipeFD.IOSequence(count), outOffset, vfs.WriteOptions{}) + outOffset += n + } else { + n, err = outFile.Write(t, inPipeFD.IOSequence(count), vfs.WriteOptions{}) + } + case outIsPipe: + if inOffset != -1 { + n, err = inFile.PRead(t, outPipeFD.IOSequence(count), inOffset, vfs.ReadOptions{}) + inOffset += n + } else { + n, err = inFile.Read(t, outPipeFD.IOSequence(count), vfs.ReadOptions{}) + } + } + if n != 0 || err != syserror.ErrWouldBlock || nonBlock { + break + } + + // Note that the blocking behavior here is a bit different than the + // normal pattern. Because we need to have both data to read and data + // to write simultaneously, we actually explicitly block on both of + // these cases in turn before returning to the splice operation. + if inFile.Readiness(eventMaskRead)&eventMaskRead == 0 { + if inCh == nil { + inCh = make(chan struct{}, 1) + inW, _ := waiter.NewChannelEntry(inCh) + inFile.EventRegister(&inW, eventMaskRead) + defer inFile.EventUnregister(&inW) + continue // Need to refresh readiness. + } + if err = t.Block(inCh); err != nil { + break + } + } + if outFile.Readiness(eventMaskWrite)&eventMaskWrite == 0 { + if outCh == nil { + outCh = make(chan struct{}, 1) + outW, _ := waiter.NewChannelEntry(outCh) + outFile.EventRegister(&outW, eventMaskWrite) + defer outFile.EventUnregister(&outW) + continue // Need to refresh readiness. + } + if err = t.Block(outCh); err != nil { + break + } + } + } + + // Copy updated offsets out. + if inOffsetPtr != 0 { + if _, err := t.CopyOut(inOffsetPtr, &inOffset); err != nil { + return 0, nil, err + } + } + if outOffsetPtr != 0 { + if _, err := t.CopyOut(outOffsetPtr, &outOffset); err != nil { + return 0, nil, err + } + } + + if n == 0 { + return 0, nil, err + } + + // 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(linux.IN_MODIFY, 0, vfs.PathEvent) + return uintptr(n), nil, nil +} + +// Tee implements Linux syscall tee(2). +func Tee(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + inFD := args[0].Int() + outFD := args[1].Int() + count := int64(args[2].SizeT()) + flags := args[3].Int() + + if count == 0 { + return 0, nil, nil + } + if count > int64(kernel.MAX_RW_COUNT) { + count = int64(kernel.MAX_RW_COUNT) + } + + // Check for invalid flags. + if flags&^(linux.SPLICE_F_MOVE|linux.SPLICE_F_NONBLOCK|linux.SPLICE_F_MORE|linux.SPLICE_F_GIFT) != 0 { + return 0, nil, syserror.EINVAL + } + + // Get file descriptions. + inFile := t.GetFileVFS2(inFD) + if inFile == nil { + return 0, nil, syserror.EBADF + } + defer inFile.DecRef() + outFile := t.GetFileVFS2(outFD) + if outFile == nil { + return 0, nil, syserror.EBADF + } + defer outFile.DecRef() + + // Check that both files support the required directionality. + if !inFile.IsReadable() || !outFile.IsWritable() { + return 0, nil, syserror.EBADF + } + + // The operation is non-blocking if anything is non-blocking. + // + // N.B. This is a rather simplistic heuristic that avoids some + // poor edge case behavior since the exact semantics here are + // underspecified and vary between versions of Linux itself. + nonBlock := ((inFile.StatusFlags()|outFile.StatusFlags())&linux.O_NONBLOCK != 0) || (flags&linux.SPLICE_F_NONBLOCK != 0) + + // Both file descriptions must represent pipes. + inPipeFD, inIsPipe := inFile.Impl().(*pipe.VFSPipeFD) + outPipeFD, outIsPipe := outFile.Impl().(*pipe.VFSPipeFD) + if !inIsPipe || !outIsPipe { + return 0, nil, syserror.EINVAL + } + + // Copy data. + var ( + inCh chan struct{} + outCh chan struct{} + ) + for { + n, err := pipe.Tee(t, outPipeFD, inPipeFD, count) + if n != 0 { + return uintptr(n), nil, nil + } + if err != syserror.ErrWouldBlock || nonBlock { + return 0, nil, err + } + + // Note that the blocking behavior here is a bit different than the + // normal pattern. Because we need to have both data to read and data + // to write simultaneously, we actually explicitly block on both of + // these cases in turn before returning to the tee operation. + if inFile.Readiness(eventMaskRead)&eventMaskRead == 0 { + if inCh == nil { + inCh = make(chan struct{}, 1) + inW, _ := waiter.NewChannelEntry(inCh) + inFile.EventRegister(&inW, eventMaskRead) + defer inFile.EventUnregister(&inW) + continue // Need to refresh readiness. + } + if err := t.Block(inCh); err != nil { + return 0, nil, err + } + } + if outFile.Readiness(eventMaskWrite)&eventMaskWrite == 0 { + if outCh == nil { + outCh = make(chan struct{}, 1) + outW, _ := waiter.NewChannelEntry(outCh) + outFile.EventRegister(&outW, eventMaskWrite) + defer outFile.EventUnregister(&outW) + continue // Need to refresh readiness. + } + if err := t.Block(outCh); err != nil { + return 0, nil, err + } + } + } +} diff --git a/pkg/sentry/syscalls/linux/vfs2/stat.go b/pkg/sentry/syscalls/linux/vfs2/stat.go new file mode 100644 index 000000000..bb1d5cac4 --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/stat.go @@ -0,0 +1,388 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/bits" + "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/gohacks" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +// Stat implements Linux syscall stat(2). +func Stat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + pathAddr := args[0].Pointer() + statAddr := args[1].Pointer() + return 0, nil, fstatat(t, linux.AT_FDCWD, pathAddr, statAddr, 0 /* flags */) +} + +// Lstat implements Linux syscall lstat(2). +func Lstat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + pathAddr := args[0].Pointer() + statAddr := args[1].Pointer() + return 0, nil, fstatat(t, linux.AT_FDCWD, pathAddr, statAddr, linux.AT_SYMLINK_NOFOLLOW) +} + +// Newfstatat implements Linux syscall newfstatat, which backs fstatat(2). +func Newfstatat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + dirfd := args[0].Int() + pathAddr := args[1].Pointer() + statAddr := args[2].Pointer() + flags := args[3].Int() + return 0, nil, fstatat(t, dirfd, pathAddr, statAddr, flags) +} + +func fstatat(t *kernel.Task, dirfd int32, pathAddr, statAddr usermem.Addr, flags int32) error { + if flags&^(linux.AT_EMPTY_PATH|linux.AT_SYMLINK_NOFOLLOW) != 0 { + return syserror.EINVAL + } + + opts := vfs.StatOptions{ + Mask: linux.STATX_BASIC_STATS, + } + + path, err := copyInPath(t, pathAddr) + if err != nil { + return err + } + + root := t.FSContext().RootDirectoryVFS2() + defer root.DecRef() + start := root + if !path.Absolute { + if !path.HasComponents() && flags&linux.AT_EMPTY_PATH == 0 { + return syserror.ENOENT + } + if dirfd == linux.AT_FDCWD { + start = t.FSContext().WorkingDirectoryVFS2() + defer start.DecRef() + } else { + dirfile := t.GetFileVFS2(dirfd) + if dirfile == nil { + return syserror.EBADF + } + if !path.HasComponents() { + // Use FileDescription.Stat() instead of + // VirtualFilesystem.StatAt() for fstatat(fd, ""), since the + // former may be able to use opened file state to expedite the + // Stat. + statx, err := dirfile.Stat(t, opts) + dirfile.DecRef() + if err != nil { + return err + } + var stat linux.Stat + convertStatxToUserStat(t, &statx, &stat) + _, err = stat.CopyOut(t, statAddr) + return err + } + start = dirfile.VirtualDentry() + start.IncRef() + defer start.DecRef() + dirfile.DecRef() + } + } + + statx, err := t.Kernel().VFS().StatAt(t, t.Credentials(), &vfs.PathOperation{ + Root: root, + Start: start, + Path: path, + FollowFinalSymlink: flags&linux.AT_SYMLINK_NOFOLLOW == 0, + }, &opts) + if err != nil { + return err + } + var stat linux.Stat + convertStatxToUserStat(t, &statx, &stat) + _, err = stat.CopyOut(t, statAddr) + return err +} + +func timespecFromStatxTimestamp(sxts linux.StatxTimestamp) linux.Timespec { + return linux.Timespec{ + Sec: sxts.Sec, + Nsec: int64(sxts.Nsec), + } +} + +// Fstat implements Linux syscall fstat(2). +func Fstat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + statAddr := args[1].Pointer() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + statx, err := file.Stat(t, vfs.StatOptions{ + Mask: linux.STATX_BASIC_STATS, + }) + if err != nil { + return 0, nil, err + } + var stat linux.Stat + convertStatxToUserStat(t, &statx, &stat) + _, err = stat.CopyOut(t, statAddr) + return 0, nil, err +} + +// Statx implements Linux syscall statx(2). +func Statx(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + dirfd := args[0].Int() + pathAddr := args[1].Pointer() + flags := args[2].Int() + mask := args[3].Uint() + statxAddr := args[4].Pointer() + + if flags&^(linux.AT_EMPTY_PATH|linux.AT_SYMLINK_NOFOLLOW|linux.AT_STATX_SYNC_TYPE) != 0 { + return 0, nil, syserror.EINVAL + } + // Make sure that only one sync type option is set. + syncType := uint32(flags & linux.AT_STATX_SYNC_TYPE) + if syncType != 0 && !bits.IsPowerOfTwo32(syncType) { + return 0, nil, syserror.EINVAL + } + if mask&linux.STATX__RESERVED != 0 { + return 0, nil, syserror.EINVAL + } + + opts := vfs.StatOptions{ + Mask: mask, + Sync: uint32(flags & linux.AT_STATX_SYNC_TYPE), + } + + path, err := copyInPath(t, pathAddr) + if err != nil { + return 0, nil, err + } + + root := t.FSContext().RootDirectoryVFS2() + defer root.DecRef() + start := root + if !path.Absolute { + if !path.HasComponents() && flags&linux.AT_EMPTY_PATH == 0 { + return 0, nil, syserror.ENOENT + } + if dirfd == linux.AT_FDCWD { + start = t.FSContext().WorkingDirectoryVFS2() + defer start.DecRef() + } else { + dirfile := t.GetFileVFS2(dirfd) + if dirfile == nil { + return 0, nil, syserror.EBADF + } + if !path.HasComponents() { + // Use FileDescription.Stat() instead of + // VirtualFilesystem.StatAt() for statx(fd, ""), since the + // former may be able to use opened file state to expedite the + // Stat. + statx, err := dirfile.Stat(t, opts) + dirfile.DecRef() + if err != nil { + return 0, nil, err + } + userifyStatx(t, &statx) + _, err = statx.CopyOut(t, statxAddr) + return 0, nil, err + } + start = dirfile.VirtualDentry() + start.IncRef() + defer start.DecRef() + dirfile.DecRef() + } + } + + statx, err := t.Kernel().VFS().StatAt(t, t.Credentials(), &vfs.PathOperation{ + Root: root, + Start: start, + Path: path, + FollowFinalSymlink: flags&linux.AT_SYMLINK_NOFOLLOW == 0, + }, &opts) + if err != nil { + return 0, nil, err + } + userifyStatx(t, &statx) + _, err = statx.CopyOut(t, statxAddr) + return 0, nil, err +} + +func userifyStatx(t *kernel.Task, statx *linux.Statx) { + userns := t.UserNamespace() + statx.UID = uint32(auth.KUID(statx.UID).In(userns).OrOverflow()) + statx.GID = uint32(auth.KGID(statx.GID).In(userns).OrOverflow()) +} + +// Readlink implements Linux syscall readlink(2). +func Readlink(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + pathAddr := args[0].Pointer() + bufAddr := args[1].Pointer() + size := args[2].SizeT() + return readlinkat(t, linux.AT_FDCWD, pathAddr, bufAddr, size) +} + +// Access implements Linux syscall access(2). +func Access(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + addr := args[0].Pointer() + mode := args[1].ModeT() + + return 0, nil, accessAt(t, linux.AT_FDCWD, addr, mode) +} + +// Faccessat implements Linux syscall faccessat(2). +// +// Note that the faccessat() system call does not take a flags argument: +// "The raw faccessat() system call takes only the first three arguments. The +// AT_EACCESS and AT_SYMLINK_NOFOLLOW flags are actually implemented within +// the glibc wrapper function for faccessat(). If either of these flags is +// specified, then the wrapper function employs fstatat(2) to determine access +// permissions." - faccessat(2) +func Faccessat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + dirfd := args[0].Int() + addr := args[1].Pointer() + mode := args[2].ModeT() + + return 0, nil, accessAt(t, dirfd, addr, mode) +} + +func accessAt(t *kernel.Task, dirfd int32, pathAddr usermem.Addr, mode uint) error { + const rOK = 4 + const wOK = 2 + const xOK = 1 + + // Sanity check the mode. + if mode&^(rOK|wOK|xOK) != 0 { + return syserror.EINVAL + } + + path, err := copyInPath(t, pathAddr) + if err != nil { + return err + } + tpop, err := getTaskPathOperation(t, dirfd, path, disallowEmptyPath, followFinalSymlink) + if err != nil { + return err + } + defer tpop.Release() + + // access(2) and faccessat(2) check permissions using real + // UID/GID, not effective UID/GID. + // + // "access() needs to use the real uid/gid, not the effective + // uid/gid. We do this by temporarily clearing all FS-related + // capabilities and switching the fsuid/fsgid around to the + // real ones." -fs/open.c:faccessat + creds := t.Credentials().Fork() + creds.EffectiveKUID = creds.RealKUID + creds.EffectiveKGID = creds.RealKGID + if creds.EffectiveKUID.In(creds.UserNamespace) == auth.RootUID { + creds.EffectiveCaps = creds.PermittedCaps + } else { + creds.EffectiveCaps = 0 + } + + return t.Kernel().VFS().AccessAt(t, creds, vfs.AccessTypes(mode), &tpop.pop) +} + +// Readlinkat implements Linux syscall mknodat(2). +func Readlinkat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + dirfd := args[0].Int() + pathAddr := args[1].Pointer() + bufAddr := args[2].Pointer() + size := args[3].SizeT() + return readlinkat(t, dirfd, pathAddr, bufAddr, size) +} + +func readlinkat(t *kernel.Task, dirfd int32, pathAddr, bufAddr usermem.Addr, size uint) (uintptr, *kernel.SyscallControl, error) { + if int(size) <= 0 { + return 0, nil, syserror.EINVAL + } + + path, err := copyInPath(t, pathAddr) + if err != nil { + return 0, nil, err + } + // "Since Linux 2.6.39, pathname can be an empty string, in which case the + // call operates on the symbolic link referred to by dirfd ..." - + // readlinkat(2) + tpop, err := getTaskPathOperation(t, dirfd, path, allowEmptyPath, nofollowFinalSymlink) + if err != nil { + return 0, nil, err + } + defer tpop.Release() + + target, err := t.Kernel().VFS().ReadlinkAt(t, t.Credentials(), &tpop.pop) + if err != nil { + return 0, nil, err + } + + if len(target) > int(size) { + target = target[:size] + } + n, err := t.CopyOutBytes(bufAddr, gohacks.ImmutableBytesFromString(target)) + if n == 0 { + return 0, nil, err + } + return uintptr(n), nil, nil +} + +// Statfs implements Linux syscall statfs(2). +func Statfs(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + pathAddr := args[0].Pointer() + bufAddr := args[1].Pointer() + + path, err := copyInPath(t, pathAddr) + if err != nil { + return 0, nil, err + } + tpop, err := getTaskPathOperation(t, linux.AT_FDCWD, path, disallowEmptyPath, followFinalSymlink) + if err != nil { + return 0, nil, err + } + defer tpop.Release() + + statfs, err := t.Kernel().VFS().StatFSAt(t, t.Credentials(), &tpop.pop) + if err != nil { + return 0, nil, err + } + _, err = statfs.CopyOut(t, bufAddr) + return 0, nil, err +} + +// Fstatfs implements Linux syscall fstatfs(2). +func Fstatfs(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + bufAddr := args[1].Pointer() + + tpop, err := getTaskPathOperation(t, fd, fspath.Path{}, allowEmptyPath, nofollowFinalSymlink) + if err != nil { + return 0, nil, err + } + defer tpop.Release() + + statfs, err := t.Kernel().VFS().StatFSAt(t, t.Credentials(), &tpop.pop) + if err != nil { + return 0, nil, err + } + _, err = statfs.CopyOut(t, bufAddr) + return 0, nil, err +} diff --git a/pkg/sentry/syscalls/linux/vfs2/stat_amd64.go b/pkg/sentry/syscalls/linux/vfs2/stat_amd64.go new file mode 100644 index 000000000..2da538fc6 --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/stat_amd64.go @@ -0,0 +1,46 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// +build amd64 + +package vfs2 + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" +) + +// This takes both input and output as pointer arguments to avoid copying large +// structs. +func convertStatxToUserStat(t *kernel.Task, statx *linux.Statx, stat *linux.Stat) { + // Linux just copies fields from struct kstat without regard to struct + // kstat::result_mask (fs/stat.c:cp_new_stat()), so we do too. + userns := t.UserNamespace() + *stat = linux.Stat{ + Dev: uint64(linux.MakeDeviceID(uint16(statx.DevMajor), statx.DevMinor)), + Ino: statx.Ino, + Nlink: uint64(statx.Nlink), + Mode: uint32(statx.Mode), + UID: uint32(auth.KUID(statx.UID).In(userns).OrOverflow()), + GID: uint32(auth.KGID(statx.GID).In(userns).OrOverflow()), + Rdev: uint64(linux.MakeDeviceID(uint16(statx.RdevMajor), statx.RdevMinor)), + Size: int64(statx.Size), + Blksize: int64(statx.Blksize), + Blocks: int64(statx.Blocks), + ATime: timespecFromStatxTimestamp(statx.Atime), + MTime: timespecFromStatxTimestamp(statx.Mtime), + CTime: timespecFromStatxTimestamp(statx.Ctime), + } +} diff --git a/pkg/sentry/syscalls/linux/vfs2/stat_arm64.go b/pkg/sentry/syscalls/linux/vfs2/stat_arm64.go new file mode 100644 index 000000000..88b9c7627 --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/stat_arm64.go @@ -0,0 +1,46 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// +build arm64 + +package vfs2 + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" +) + +// This takes both input and output as pointer arguments to avoid copying large +// structs. +func convertStatxToUserStat(t *kernel.Task, statx *linux.Statx, stat *linux.Stat) { + // Linux just copies fields from struct kstat without regard to struct + // kstat::result_mask (fs/stat.c:cp_new_stat()), so we do too. + userns := t.UserNamespace() + *stat = linux.Stat{ + Dev: uint64(linux.MakeDeviceID(uint16(statx.DevMajor), statx.DevMinor)), + Ino: statx.Ino, + Nlink: uint32(statx.Nlink), + Mode: uint32(statx.Mode), + UID: uint32(auth.KUID(statx.UID).In(userns).OrOverflow()), + GID: uint32(auth.KGID(statx.GID).In(userns).OrOverflow()), + Rdev: uint64(linux.MakeDeviceID(uint16(statx.RdevMajor), statx.RdevMinor)), + Size: int64(statx.Size), + Blksize: int32(statx.Blksize), + Blocks: int64(statx.Blocks), + ATime: timespecFromStatxTimestamp(statx.Atime), + MTime: timespecFromStatxTimestamp(statx.Mtime), + CTime: timespecFromStatxTimestamp(statx.Ctime), + } +} diff --git a/pkg/sentry/syscalls/linux/vfs2/sync.go b/pkg/sentry/syscalls/linux/vfs2/sync.go new file mode 100644 index 000000000..365250b0b --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/sync.go @@ -0,0 +1,87 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/syserror" +) + +// Sync implements Linux syscall sync(2). +func Sync(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + return 0, nil, t.Kernel().VFS().SyncAllFilesystems(t) +} + +// Syncfs implements Linux syscall syncfs(2). +func Syncfs(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + return 0, nil, file.SyncFS(t) +} + +// Fsync implements Linux syscall fsync(2). +func Fsync(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + return 0, nil, file.Sync(t) +} + +// Fdatasync implements Linux syscall fdatasync(2). +func Fdatasync(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + // TODO(gvisor.dev/issue/1897): Avoid writeback of unnecessary metadata. + return Fsync(t, args) +} + +// SyncFileRange implements Linux syscall sync_file_range(2). +func SyncFileRange(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + offset := args[1].Int64() + nbytes := args[2].Int64() + flags := args[3].Uint() + + if offset < 0 { + return 0, nil, syserror.EINVAL + } + if nbytes < 0 { + return 0, nil, syserror.EINVAL + } + if flags&^(linux.SYNC_FILE_RANGE_WAIT_BEFORE|linux.SYNC_FILE_RANGE_WRITE|linux.SYNC_FILE_RANGE_WAIT_AFTER) != 0 { + return 0, nil, syserror.EINVAL + } + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // TODO(gvisor.dev/issue/1897): Avoid writeback of data ranges outside of + // [offset, offset+nbytes). + return 0, nil, file.Sync(t) +} diff --git a/pkg/sentry/syscalls/linux/vfs2/sys_read.go b/pkg/sentry/syscalls/linux/vfs2/sys_read.go deleted file mode 100644 index 7667524c7..000000000 --- a/pkg/sentry/syscalls/linux/vfs2/sys_read.go +++ /dev/null @@ -1,95 +0,0 @@ -// Copyright 2020 The gVisor Authors. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -package vfs2 - -import ( - "gvisor.dev/gvisor/pkg/sentry/arch" - "gvisor.dev/gvisor/pkg/sentry/kernel" - "gvisor.dev/gvisor/pkg/sentry/syscalls/linux" - "gvisor.dev/gvisor/pkg/sentry/vfs" - "gvisor.dev/gvisor/pkg/syserror" - "gvisor.dev/gvisor/pkg/usermem" - "gvisor.dev/gvisor/pkg/waiter" -) - -const ( - // EventMaskRead contains events that can be triggered on reads. - EventMaskRead = waiter.EventIn | waiter.EventHUp | waiter.EventErr -) - -// Read implements linux syscall read(2). Note that we try to get a buffer that -// is exactly the size requested because some applications like qemu expect -// they can do large reads all at once. Bug for bug. Same for other read -// calls below. -func Read(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { - fd := args[0].Int() - addr := args[1].Pointer() - size := args[2].SizeT() - - file := t.GetFileVFS2(fd) - if file == nil { - return 0, nil, syserror.EBADF - } - defer file.DecRef() - - // Check that the size is legitimate. - si := int(size) - if si < 0 { - return 0, nil, syserror.EINVAL - } - - // Get the destination of the read. - dst, err := t.SingleIOSequence(addr, si, usermem.IOOpts{ - AddressSpaceActive: true, - }) - if err != nil { - return 0, nil, err - } - - n, err := read(t, file, dst, vfs.ReadOptions{}) - t.IOUsage().AccountReadSyscall(n) - return uintptr(n), nil, linux.HandleIOErrorVFS2(t, n != 0, err, kernel.ERESTARTSYS, "read", file) -} - -func read(t *kernel.Task, file *vfs.FileDescription, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { - n, err := file.Read(t, dst, opts) - if err != syserror.ErrWouldBlock { - return n, err - } - - // Register for notifications. - w, ch := waiter.NewChannelEntry(nil) - file.EventRegister(&w, EventMaskRead) - - total := n - for { - // Shorten dst to reflect bytes previously read. - dst = dst.DropFirst(int(n)) - - // Issue the request and break out if it completes with anything other than - // "would block". - n, err := file.Read(t, dst, opts) - total += n - if err != syserror.ErrWouldBlock { - break - } - if err := t.Block(ch); err != nil { - break - } - } - file.EventUnregister(&w) - - return total, err -} diff --git a/pkg/sentry/syscalls/linux/vfs2/timerfd.go b/pkg/sentry/syscalls/linux/vfs2/timerfd.go new file mode 100644 index 000000000..5ac79bc09 --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/timerfd.go @@ -0,0 +1,127 @@ +// Copyright 2018 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/timerfd" + "gvisor.dev/gvisor/pkg/sentry/kernel" + ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" + "gvisor.dev/gvisor/pkg/syserror" +) + +// TimerfdCreate implements Linux syscall timerfd_create(2). +func TimerfdCreate(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + clockID := args[0].Int() + flags := args[1].Int() + + if flags&^(linux.TFD_CLOEXEC|linux.TFD_NONBLOCK) != 0 { + return 0, nil, syserror.EINVAL + } + + // Timerfds aren't writable per se (their implementation of Write just + // returns EINVAL), but they are "opened for writing", which is necessary + // to actually reach said implementation of Write. + fileFlags := uint32(linux.O_RDWR) + if flags&linux.TFD_NONBLOCK != 0 { + fileFlags |= linux.O_NONBLOCK + } + + var clock ktime.Clock + switch clockID { + case linux.CLOCK_REALTIME: + clock = t.Kernel().RealtimeClock() + case linux.CLOCK_MONOTONIC, linux.CLOCK_BOOTTIME: + clock = t.Kernel().MonotonicClock() + default: + return 0, nil, syserror.EINVAL + } + vfsObj := t.Kernel().VFS() + file, err := timerfd.New(vfsObj, clock, fileFlags) + if err != nil { + return 0, nil, err + } + defer file.DecRef() + fd, err := t.NewFDFromVFS2(0, file, kernel.FDFlags{ + CloseOnExec: flags&linux.TFD_CLOEXEC != 0, + }) + if err != nil { + return 0, nil, err + } + return uintptr(fd), nil, nil +} + +// TimerfdSettime implements Linux syscall timerfd_settime(2). +func TimerfdSettime(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + flags := args[1].Int() + newValAddr := args[2].Pointer() + oldValAddr := args[3].Pointer() + + if flags&^(linux.TFD_TIMER_ABSTIME) != 0 { + return 0, nil, syserror.EINVAL + } + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + tfd, ok := file.Impl().(*timerfd.TimerFileDescription) + if !ok { + return 0, nil, syserror.EINVAL + } + + var newVal linux.Itimerspec + if _, err := t.CopyIn(newValAddr, &newVal); err != nil { + return 0, nil, err + } + newS, err := ktime.SettingFromItimerspec(newVal, flags&linux.TFD_TIMER_ABSTIME != 0, tfd.Clock()) + if err != nil { + return 0, nil, err + } + tm, oldS := tfd.SetTime(newS) + if oldValAddr != 0 { + oldVal := ktime.ItimerspecFromSetting(tm, oldS) + if _, err := t.CopyOut(oldValAddr, &oldVal); err != nil { + return 0, nil, err + } + } + return 0, nil, nil +} + +// TimerfdGettime implements Linux syscall timerfd_gettime(2). +func TimerfdGettime(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + curValAddr := args[1].Pointer() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + tfd, ok := file.Impl().(*timerfd.TimerFileDescription) + if !ok { + return 0, nil, syserror.EINVAL + } + + tm, s := tfd.GetTime() + curVal := ktime.ItimerspecFromSetting(tm, s) + _, err := t.CopyOut(curValAddr, &curVal) + return 0, nil, err +} diff --git a/pkg/sentry/syscalls/linux/vfs2/vfs2.go b/pkg/sentry/syscalls/linux/vfs2/vfs2.go new file mode 100644 index 000000000..7b6e7571a --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/vfs2.go @@ -0,0 +1,172 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// Package vfs2 provides syscall implementations that use VFS2. +package vfs2 + +import ( + "gvisor.dev/gvisor/pkg/sentry/syscalls" + "gvisor.dev/gvisor/pkg/sentry/syscalls/linux" +) + +// Override syscall table to add syscalls implementations from this package. +func Override() { + // Override AMD64. + s := linux.AMD64 + s.Table[0] = syscalls.Supported("read", Read) + s.Table[1] = syscalls.Supported("write", Write) + s.Table[2] = syscalls.Supported("open", Open) + s.Table[3] = syscalls.Supported("close", Close) + s.Table[4] = syscalls.Supported("stat", Stat) + s.Table[5] = syscalls.Supported("fstat", Fstat) + s.Table[6] = syscalls.Supported("lstat", Lstat) + s.Table[7] = syscalls.Supported("poll", Poll) + s.Table[8] = syscalls.Supported("lseek", Lseek) + s.Table[9] = syscalls.Supported("mmap", Mmap) + s.Table[16] = syscalls.Supported("ioctl", Ioctl) + s.Table[17] = syscalls.Supported("pread64", Pread64) + s.Table[18] = syscalls.Supported("pwrite64", Pwrite64) + s.Table[19] = syscalls.Supported("readv", Readv) + s.Table[20] = syscalls.Supported("writev", Writev) + s.Table[21] = syscalls.Supported("access", Access) + s.Table[22] = syscalls.Supported("pipe", Pipe) + s.Table[23] = syscalls.Supported("select", Select) + s.Table[32] = syscalls.Supported("dup", Dup) + s.Table[33] = syscalls.Supported("dup2", Dup2) + delete(s.Table, 40) // sendfile + s.Table[41] = syscalls.Supported("socket", Socket) + s.Table[42] = syscalls.Supported("connect", Connect) + s.Table[43] = syscalls.Supported("accept", Accept) + s.Table[44] = syscalls.Supported("sendto", SendTo) + s.Table[45] = syscalls.Supported("recvfrom", RecvFrom) + s.Table[46] = syscalls.Supported("sendmsg", SendMsg) + s.Table[47] = syscalls.Supported("recvmsg", RecvMsg) + s.Table[48] = syscalls.Supported("shutdown", Shutdown) + s.Table[49] = syscalls.Supported("bind", Bind) + s.Table[50] = syscalls.Supported("listen", Listen) + s.Table[51] = syscalls.Supported("getsockname", GetSockName) + s.Table[52] = syscalls.Supported("getpeername", GetPeerName) + s.Table[53] = syscalls.Supported("socketpair", SocketPair) + s.Table[54] = syscalls.Supported("setsockopt", SetSockOpt) + s.Table[55] = syscalls.Supported("getsockopt", GetSockOpt) + s.Table[59] = syscalls.Supported("execve", Execve) + s.Table[72] = syscalls.Supported("fcntl", Fcntl) + delete(s.Table, 73) // flock + s.Table[74] = syscalls.Supported("fsync", Fsync) + s.Table[75] = syscalls.Supported("fdatasync", Fdatasync) + s.Table[76] = syscalls.Supported("truncate", Truncate) + s.Table[77] = syscalls.Supported("ftruncate", Ftruncate) + s.Table[78] = syscalls.Supported("getdents", Getdents) + s.Table[79] = syscalls.Supported("getcwd", Getcwd) + s.Table[80] = syscalls.Supported("chdir", Chdir) + s.Table[81] = syscalls.Supported("fchdir", Fchdir) + s.Table[82] = syscalls.Supported("rename", Rename) + s.Table[83] = syscalls.Supported("mkdir", Mkdir) + s.Table[84] = syscalls.Supported("rmdir", Rmdir) + s.Table[85] = syscalls.Supported("creat", Creat) + s.Table[86] = syscalls.Supported("link", Link) + s.Table[87] = syscalls.Supported("unlink", Unlink) + s.Table[88] = syscalls.Supported("symlink", Symlink) + s.Table[89] = syscalls.Supported("readlink", Readlink) + s.Table[90] = syscalls.Supported("chmod", Chmod) + s.Table[91] = syscalls.Supported("fchmod", Fchmod) + s.Table[92] = syscalls.Supported("chown", Chown) + s.Table[93] = syscalls.Supported("fchown", Fchown) + s.Table[94] = syscalls.Supported("lchown", Lchown) + s.Table[132] = syscalls.Supported("utime", Utime) + s.Table[133] = syscalls.Supported("mknod", Mknod) + s.Table[137] = syscalls.Supported("statfs", Statfs) + s.Table[138] = syscalls.Supported("fstatfs", Fstatfs) + s.Table[161] = syscalls.Supported("chroot", Chroot) + s.Table[162] = syscalls.Supported("sync", Sync) + s.Table[165] = syscalls.Supported("mount", Mount) + s.Table[166] = syscalls.Supported("umount2", Umount2) + delete(s.Table, 187) // readahead + s.Table[188] = syscalls.Supported("setxattr", Setxattr) + s.Table[189] = syscalls.Supported("lsetxattr", Lsetxattr) + s.Table[190] = syscalls.Supported("fsetxattr", Fsetxattr) + s.Table[191] = syscalls.Supported("getxattr", Getxattr) + s.Table[192] = syscalls.Supported("lgetxattr", Lgetxattr) + s.Table[193] = syscalls.Supported("fgetxattr", Fgetxattr) + s.Table[194] = syscalls.Supported("listxattr", Listxattr) + s.Table[195] = syscalls.Supported("llistxattr", Llistxattr) + s.Table[196] = syscalls.Supported("flistxattr", Flistxattr) + s.Table[197] = syscalls.Supported("removexattr", Removexattr) + s.Table[198] = syscalls.Supported("lremovexattr", Lremovexattr) + s.Table[199] = syscalls.Supported("fremovexattr", Fremovexattr) + delete(s.Table, 206) // io_setup + delete(s.Table, 207) // io_destroy + delete(s.Table, 208) // io_getevents + delete(s.Table, 209) // io_submit + delete(s.Table, 210) // io_cancel + s.Table[213] = syscalls.Supported("epoll_create", EpollCreate) + s.Table[217] = syscalls.Supported("getdents64", Getdents64) + delete(s.Table, 221) // fdavise64 + s.Table[232] = syscalls.Supported("epoll_wait", EpollWait) + s.Table[233] = syscalls.Supported("epoll_ctl", EpollCtl) + s.Table[235] = syscalls.Supported("utimes", Utimes) + s.Table[253] = syscalls.PartiallySupported("inotify_init", InotifyInit, "inotify events are only available inside the sandbox.", nil) + s.Table[254] = syscalls.PartiallySupported("inotify_add_watch", InotifyAddWatch, "inotify events are only available inside the sandbox.", nil) + s.Table[255] = syscalls.PartiallySupported("inotify_rm_watch", InotifyRmWatch, "inotify events are only available inside the sandbox.", nil) + s.Table[257] = syscalls.Supported("openat", Openat) + s.Table[258] = syscalls.Supported("mkdirat", Mkdirat) + s.Table[259] = syscalls.Supported("mknodat", Mknodat) + s.Table[260] = syscalls.Supported("fchownat", Fchownat) + s.Table[261] = syscalls.Supported("futimesat", Futimesat) + s.Table[262] = syscalls.Supported("newfstatat", Newfstatat) + s.Table[263] = syscalls.Supported("unlinkat", Unlinkat) + s.Table[264] = syscalls.Supported("renameat", Renameat) + s.Table[265] = syscalls.Supported("linkat", Linkat) + s.Table[266] = syscalls.Supported("symlinkat", Symlinkat) + s.Table[267] = syscalls.Supported("readlinkat", Readlinkat) + s.Table[268] = syscalls.Supported("fchmodat", Fchmodat) + s.Table[269] = syscalls.Supported("faccessat", Faccessat) + s.Table[270] = syscalls.Supported("pselect", Pselect) + s.Table[271] = syscalls.Supported("ppoll", Ppoll) + s.Table[275] = syscalls.Supported("splice", Splice) + s.Table[276] = syscalls.Supported("tee", Tee) + s.Table[277] = syscalls.Supported("sync_file_range", SyncFileRange) + s.Table[280] = syscalls.Supported("utimensat", Utimensat) + s.Table[281] = syscalls.Supported("epoll_pwait", EpollPwait) + s.Table[282] = syscalls.Supported("signalfd", Signalfd) + s.Table[283] = syscalls.Supported("timerfd_create", TimerfdCreate) + s.Table[284] = syscalls.Supported("eventfd", Eventfd) + delete(s.Table, 285) // fallocate + s.Table[286] = syscalls.Supported("timerfd_settime", TimerfdSettime) + s.Table[287] = syscalls.Supported("timerfd_gettime", TimerfdGettime) + s.Table[288] = syscalls.Supported("accept4", Accept4) + s.Table[289] = syscalls.Supported("signalfd4", Signalfd4) + s.Table[290] = syscalls.Supported("eventfd2", Eventfd2) + s.Table[291] = syscalls.Supported("epoll_create1", EpollCreate1) + s.Table[292] = syscalls.Supported("dup3", Dup3) + s.Table[293] = syscalls.Supported("pipe2", Pipe2) + s.Table[294] = syscalls.PartiallySupported("inotify_init1", InotifyInit1, "inotify events are only available inside the sandbox.", nil) + s.Table[295] = syscalls.Supported("preadv", Preadv) + s.Table[296] = syscalls.Supported("pwritev", Pwritev) + s.Table[299] = syscalls.Supported("recvmmsg", RecvMMsg) + s.Table[306] = syscalls.Supported("syncfs", Syncfs) + s.Table[307] = syscalls.Supported("sendmmsg", SendMMsg) + s.Table[316] = syscalls.Supported("renameat2", Renameat2) + s.Table[319] = syscalls.Supported("memfd_create", MemfdCreate) + s.Table[322] = syscalls.Supported("execveat", Execveat) + s.Table[327] = syscalls.Supported("preadv2", Preadv2) + s.Table[328] = syscalls.Supported("pwritev2", Pwritev2) + s.Table[332] = syscalls.Supported("statx", Statx) + s.Init() + + // Override ARM64. + s = linux.ARM64 + s.Table[63] = syscalls.Supported("read", Read) + s.Init() +} diff --git a/pkg/sentry/syscalls/linux/vfs2/xattr.go b/pkg/sentry/syscalls/linux/vfs2/xattr.go new file mode 100644 index 000000000..af455d5c1 --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/xattr.go @@ -0,0 +1,356 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "bytes" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/gohacks" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +// Listxattr implements Linux syscall listxattr(2). +func Listxattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + return listxattr(t, args, followFinalSymlink) +} + +// Llistxattr implements Linux syscall llistxattr(2). +func Llistxattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + return listxattr(t, args, nofollowFinalSymlink) +} + +func listxattr(t *kernel.Task, args arch.SyscallArguments, shouldFollowFinalSymlink shouldFollowFinalSymlink) (uintptr, *kernel.SyscallControl, error) { + pathAddr := args[0].Pointer() + listAddr := args[1].Pointer() + size := args[2].SizeT() + + path, err := copyInPath(t, pathAddr) + if err != nil { + return 0, nil, err + } + tpop, err := getTaskPathOperation(t, linux.AT_FDCWD, path, disallowEmptyPath, shouldFollowFinalSymlink) + if err != nil { + return 0, nil, err + } + defer tpop.Release() + + names, err := t.Kernel().VFS().ListxattrAt(t, t.Credentials(), &tpop.pop, uint64(size)) + if err != nil { + return 0, nil, err + } + n, err := copyOutXattrNameList(t, listAddr, size, names) + if err != nil { + return 0, nil, err + } + return uintptr(n), nil, nil +} + +// Flistxattr implements Linux syscall flistxattr(2). +func Flistxattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + listAddr := args[1].Pointer() + size := args[2].SizeT() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + names, err := file.Listxattr(t, uint64(size)) + if err != nil { + return 0, nil, err + } + n, err := copyOutXattrNameList(t, listAddr, size, names) + if err != nil { + return 0, nil, err + } + return uintptr(n), nil, nil +} + +// Getxattr implements Linux syscall getxattr(2). +func Getxattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + return getxattr(t, args, followFinalSymlink) +} + +// Lgetxattr implements Linux syscall lgetxattr(2). +func Lgetxattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + return getxattr(t, args, nofollowFinalSymlink) +} + +func getxattr(t *kernel.Task, args arch.SyscallArguments, shouldFollowFinalSymlink shouldFollowFinalSymlink) (uintptr, *kernel.SyscallControl, error) { + pathAddr := args[0].Pointer() + nameAddr := args[1].Pointer() + valueAddr := args[2].Pointer() + size := args[3].SizeT() + + path, err := copyInPath(t, pathAddr) + if err != nil { + return 0, nil, err + } + tpop, err := getTaskPathOperation(t, linux.AT_FDCWD, path, disallowEmptyPath, shouldFollowFinalSymlink) + if err != nil { + return 0, nil, err + } + defer tpop.Release() + + name, err := copyInXattrName(t, nameAddr) + if err != nil { + return 0, nil, err + } + + value, err := t.Kernel().VFS().GetxattrAt(t, t.Credentials(), &tpop.pop, &vfs.GetxattrOptions{ + Name: name, + Size: uint64(size), + }) + if err != nil { + return 0, nil, err + } + n, err := copyOutXattrValue(t, valueAddr, size, value) + if err != nil { + return 0, nil, err + } + return uintptr(n), nil, nil +} + +// Fgetxattr implements Linux syscall fgetxattr(2). +func Fgetxattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + nameAddr := args[1].Pointer() + valueAddr := args[2].Pointer() + size := args[3].SizeT() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + name, err := copyInXattrName(t, nameAddr) + if err != nil { + return 0, nil, err + } + + value, err := file.Getxattr(t, &vfs.GetxattrOptions{Name: name, Size: uint64(size)}) + if err != nil { + return 0, nil, err + } + n, err := copyOutXattrValue(t, valueAddr, size, value) + if err != nil { + return 0, nil, err + } + return uintptr(n), nil, nil +} + +// Setxattr implements Linux syscall setxattr(2). +func Setxattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + return 0, nil, setxattr(t, args, followFinalSymlink) +} + +// Lsetxattr implements Linux syscall lsetxattr(2). +func Lsetxattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + return 0, nil, setxattr(t, args, nofollowFinalSymlink) +} + +func setxattr(t *kernel.Task, args arch.SyscallArguments, shouldFollowFinalSymlink shouldFollowFinalSymlink) error { + pathAddr := args[0].Pointer() + nameAddr := args[1].Pointer() + valueAddr := args[2].Pointer() + size := args[3].SizeT() + flags := args[4].Int() + + if flags&^(linux.XATTR_CREATE|linux.XATTR_REPLACE) != 0 { + return syserror.EINVAL + } + + path, err := copyInPath(t, pathAddr) + if err != nil { + return err + } + tpop, err := getTaskPathOperation(t, linux.AT_FDCWD, path, disallowEmptyPath, shouldFollowFinalSymlink) + if err != nil { + return err + } + defer tpop.Release() + + name, err := copyInXattrName(t, nameAddr) + if err != nil { + return err + } + value, err := copyInXattrValue(t, valueAddr, size) + if err != nil { + return err + } + + return t.Kernel().VFS().SetxattrAt(t, t.Credentials(), &tpop.pop, &vfs.SetxattrOptions{ + Name: name, + Value: value, + Flags: uint32(flags), + }) +} + +// Fsetxattr implements Linux syscall fsetxattr(2). +func Fsetxattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + nameAddr := args[1].Pointer() + valueAddr := args[2].Pointer() + size := args[3].SizeT() + flags := args[4].Int() + + if flags&^(linux.XATTR_CREATE|linux.XATTR_REPLACE) != 0 { + return 0, nil, syserror.EINVAL + } + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + name, err := copyInXattrName(t, nameAddr) + if err != nil { + return 0, nil, err + } + value, err := copyInXattrValue(t, valueAddr, size) + if err != nil { + return 0, nil, err + } + + return 0, nil, file.Setxattr(t, &vfs.SetxattrOptions{ + Name: name, + Value: value, + Flags: uint32(flags), + }) +} + +// Removexattr implements Linux syscall removexattr(2). +func Removexattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + return 0, nil, removexattr(t, args, followFinalSymlink) +} + +// Lremovexattr implements Linux syscall lremovexattr(2). +func Lremovexattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + return 0, nil, removexattr(t, args, nofollowFinalSymlink) +} + +func removexattr(t *kernel.Task, args arch.SyscallArguments, shouldFollowFinalSymlink shouldFollowFinalSymlink) error { + pathAddr := args[0].Pointer() + nameAddr := args[1].Pointer() + + path, err := copyInPath(t, pathAddr) + if err != nil { + return err + } + tpop, err := getTaskPathOperation(t, linux.AT_FDCWD, path, disallowEmptyPath, shouldFollowFinalSymlink) + if err != nil { + return err + } + defer tpop.Release() + + name, err := copyInXattrName(t, nameAddr) + if err != nil { + return err + } + + return t.Kernel().VFS().RemovexattrAt(t, t.Credentials(), &tpop.pop, name) +} + +// Fremovexattr implements Linux syscall fremovexattr(2). +func Fremovexattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + nameAddr := args[1].Pointer() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + name, err := copyInXattrName(t, nameAddr) + if err != nil { + return 0, nil, err + } + + return 0, nil, file.Removexattr(t, name) +} + +func copyInXattrName(t *kernel.Task, nameAddr usermem.Addr) (string, error) { + name, err := t.CopyInString(nameAddr, linux.XATTR_NAME_MAX+1) + if err != nil { + if err == syserror.ENAMETOOLONG { + return "", syserror.ERANGE + } + return "", err + } + if len(name) == 0 { + return "", syserror.ERANGE + } + return name, nil +} + +func copyOutXattrNameList(t *kernel.Task, listAddr usermem.Addr, size uint, names []string) (int, error) { + if size > linux.XATTR_LIST_MAX { + size = linux.XATTR_LIST_MAX + } + var buf bytes.Buffer + for _, name := range names { + buf.WriteString(name) + buf.WriteByte(0) + } + if size == 0 { + // Return the size that would be required to accomodate the list. + return buf.Len(), nil + } + if buf.Len() > int(size) { + if size >= linux.XATTR_LIST_MAX { + return 0, syserror.E2BIG + } + return 0, syserror.ERANGE + } + return t.CopyOutBytes(listAddr, buf.Bytes()) +} + +func copyInXattrValue(t *kernel.Task, valueAddr usermem.Addr, size uint) (string, error) { + if size > linux.XATTR_SIZE_MAX { + return "", syserror.E2BIG + } + buf := make([]byte, size) + if _, err := t.CopyInBytes(valueAddr, buf); err != nil { + return "", err + } + return gohacks.StringFromImmutableBytes(buf), nil +} + +func copyOutXattrValue(t *kernel.Task, valueAddr usermem.Addr, size uint, value string) (int, error) { + if size > linux.XATTR_SIZE_MAX { + size = linux.XATTR_SIZE_MAX + } + if size == 0 { + // Return the size that would be required to accomodate the value. + return len(value), nil + } + if len(value) > int(size) { + if size >= linux.XATTR_SIZE_MAX { + return 0, syserror.E2BIG + } + return 0, syserror.ERANGE + } + return t.CopyOutBytes(valueAddr, gohacks.ImmutableBytesFromString(value)) +} diff --git a/pkg/sentry/usage/memory.go b/pkg/sentry/usage/memory.go index 4320ad17f..ab1d140d2 100644 --- a/pkg/sentry/usage/memory.go +++ b/pkg/sentry/usage/memory.go @@ -252,18 +252,23 @@ func (m *MemoryLocked) Copy() (MemoryStats, uint64) { return ms, m.totalLocked() } -// MinimumTotalMemoryBytes is the minimum reported total system memory. -// -// This can be configured through options provided to the Sentry at start. -// This number is purely synthetic. This is only set before the application -// starts executing, and must not be modified. -var MinimumTotalMemoryBytes uint64 = 2 << 30 // 2 GB +// These options control how much total memory the is reported to the application. +// They may only be set before the application starts executing, and must not +// be modified. +var ( + // MinimumTotalMemoryBytes is the minimum reported total system memory. + MinimumTotalMemoryBytes uint64 = 2 << 30 // 2 GB + + // MaximumTotalMemoryBytes is the maximum reported total system memory. + // The 0 value indicates no maximum. + MaximumTotalMemoryBytes uint64 +) // TotalMemory returns the "total usable memory" available. // // This number doesn't really have a true value so it's based on the following -// inputs and further bounded to be above some minimum guaranteed value (2GB), -// additionally ensuring that total memory reported is always less than used. +// inputs and further bounded to be above the MinumumTotalMemoryBytes and below +// MaximumTotalMemoryBytes. // // memSize should be the platform.Memory size reported by platform.Memory.TotalSize() // used is the total memory reported by MemoryLocked.Total() @@ -279,5 +284,8 @@ func TotalMemory(memSize, used uint64) uint64 { memSize = uint64(1) << (uint(msb) + 1) } } + if MaximumTotalMemoryBytes > 0 && memSize > MaximumTotalMemoryBytes { + memSize = MaximumTotalMemoryBytes + } return memSize } diff --git a/pkg/sentry/vfs/BUILD b/pkg/sentry/vfs/BUILD index 14b39eb9d..774cc66cc 100644 --- a/pkg/sentry/vfs/BUILD +++ b/pkg/sentry/vfs/BUILD @@ -15,6 +15,18 @@ go_template_instance( }, ) +go_template_instance( + name = "event_list", + out = "event_list.go", + package = "vfs", + prefix = "event", + template = "//pkg/ilist:generic_list", + types = { + "Element": "*Event", + "Linker": "*Event", + }, +) + go_library( name = "vfs", srcs = [ @@ -25,11 +37,13 @@ go_library( "device.go", "epoll.go", "epoll_interest_list.go", + "event_list.go", "file_description.go", "file_description_impl_util.go", "filesystem.go", "filesystem_impl_util.go", "filesystem_type.go", + "inotify.go", "mount.go", "mount_unsafe.go", "options.go", @@ -41,16 +55,28 @@ go_library( visibility = ["//pkg/sentry:internal"], deps = [ "//pkg/abi/linux", + "//pkg/binary", "//pkg/context", + "//pkg/fd", + "//pkg/fdnotifier", "//pkg/fspath", + "//pkg/gohacks", + "//pkg/log", + "//pkg/safemem", "//pkg/sentry/arch", + "//pkg/sentry/fs", "//pkg/sentry/fs/lock", "//pkg/sentry/kernel/auth", + "//pkg/sentry/kernel/time", + "//pkg/sentry/limits", "//pkg/sentry/memmap", + "//pkg/sentry/socket/unix/transport", + "//pkg/sentry/uniqueid", "//pkg/sync", "//pkg/syserror", "//pkg/usermem", "//pkg/waiter", + "@org_golang_x_sys//unix:go_default_library", ], ) diff --git a/pkg/sentry/vfs/README.md b/pkg/sentry/vfs/README.md index 9aa133bcb..66f3105bd 100644 --- a/pkg/sentry/vfs/README.md +++ b/pkg/sentry/vfs/README.md @@ -39,8 +39,8 @@ Mount references are held by: - Mount: Each referenced Mount holds a reference on its parent, which is the mount containing its mount point. -- VirtualFilesystem: A reference is held on each Mount that has not been - umounted. +- VirtualFilesystem: A reference is held on each Mount that has been connected + to a mount point, but not yet umounted. MountNamespace and FileDescription references are held by users of VFS. The expectation is that each `kernel.Task` holds a reference on its corresponding diff --git a/pkg/sentry/vfs/anonfs.go b/pkg/sentry/vfs/anonfs.go index 2db25be49..b7c6b60b8 100644 --- a/pkg/sentry/vfs/anonfs.go +++ b/pkg/sentry/vfs/anonfs.go @@ -21,6 +21,7 @@ import ( "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/fspath" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" ) @@ -41,7 +42,27 @@ func (vfs *VirtualFilesystem) NewAnonVirtualDentry(name string) VirtualDentry { } } -const anonfsBlockSize = usermem.PageSize // via fs/libfs.c:pseudo_fs_fill_super() +const ( + anonfsBlockSize = usermem.PageSize // via fs/libfs.c:pseudo_fs_fill_super() + + // Mode, UID, and GID for a generic anonfs file. + anonFileMode = 0600 // no type is correct + anonFileUID = auth.RootKUID + anonFileGID = auth.RootKGID +) + +// anonFilesystemType implements FilesystemType. +type anonFilesystemType struct{} + +// GetFilesystem implements FilesystemType.GetFilesystem. +func (anonFilesystemType) GetFilesystem(context.Context, *VirtualFilesystem, *auth.Credentials, string, GetFilesystemOptions) (*Filesystem, *Dentry, error) { + panic("cannot instaniate an anon filesystem") +} + +// Name implemenents FilesystemType.Name. +func (anonFilesystemType) Name() string { + return "none" +} // anonFilesystem is the implementation of FilesystemImpl that backs // VirtualDentries returned by VirtualFilesystem.NewAnonVirtualDentry(). @@ -69,6 +90,14 @@ func (fs *anonFilesystem) Sync(ctx context.Context) error { return nil } +// AccessAt implements vfs.Filesystem.Impl.AccessAt. +func (fs *anonFilesystem) AccessAt(ctx context.Context, rp *ResolvingPath, creds *auth.Credentials, ats AccessTypes) error { + if !rp.Done() { + return syserror.ENOTDIR + } + return GenericCheckPermissions(creds, ats, anonFileMode, anonFileUID, anonFileGID) +} + // GetDentryAt implements FilesystemImpl.GetDentryAt. func (fs *anonFilesystem) GetDentryAt(ctx context.Context, rp *ResolvingPath, opts GetDentryOptions) (*Dentry, error) { if !rp.Done() { @@ -167,13 +196,13 @@ func (fs *anonFilesystem) StatAt(ctx context.Context, rp *ResolvingPath, opts St Mask: linux.STATX_TYPE | linux.STATX_MODE | linux.STATX_NLINK | linux.STATX_UID | linux.STATX_GID | linux.STATX_INO | linux.STATX_SIZE | linux.STATX_BLOCKS, Blksize: anonfsBlockSize, Nlink: 1, - UID: uint32(auth.RootKUID), - GID: uint32(auth.RootKGID), - Mode: 0600, // no type is correct + UID: uint32(anonFileUID), + GID: uint32(anonFileGID), + Mode: anonFileMode, Ino: 1, Size: 0, Blocks: 0, - DevMajor: 0, + DevMajor: linux.UNNAMED_MAJOR, DevMinor: fs.devMinor, }, nil } @@ -205,8 +234,19 @@ func (fs *anonFilesystem) UnlinkAt(ctx context.Context, rp *ResolvingPath) error return syserror.EPERM } +// BoundEndpointAt implements FilesystemImpl.BoundEndpointAt. +func (fs *anonFilesystem) BoundEndpointAt(ctx context.Context, rp *ResolvingPath, opts BoundEndpointOptions) (transport.BoundEndpoint, error) { + if !rp.Final() { + return nil, syserror.ENOTDIR + } + if err := GenericCheckPermissions(rp.Credentials(), MayWrite, anonFileMode, anonFileUID, anonFileGID); err != nil { + return nil, err + } + return nil, syserror.ECONNREFUSED +} + // ListxattrAt implements FilesystemImpl.ListxattrAt. -func (fs *anonFilesystem) ListxattrAt(ctx context.Context, rp *ResolvingPath) ([]string, error) { +func (fs *anonFilesystem) ListxattrAt(ctx context.Context, rp *ResolvingPath, size uint64) ([]string, error) { if !rp.Done() { return nil, syserror.ENOTDIR } @@ -214,7 +254,7 @@ func (fs *anonFilesystem) ListxattrAt(ctx context.Context, rp *ResolvingPath) ([ } // GetxattrAt implements FilesystemImpl.GetxattrAt. -func (fs *anonFilesystem) GetxattrAt(ctx context.Context, rp *ResolvingPath, name string) (string, error) { +func (fs *anonFilesystem) GetxattrAt(ctx context.Context, rp *ResolvingPath, opts GetxattrOptions) (string, error) { if !rp.Done() { return "", syserror.ENOTDIR } @@ -257,3 +297,15 @@ func (d *anonDentry) TryIncRef() bool { func (d *anonDentry) DecRef() { // no-op } + +// InotifyWithParent implements DentryImpl.InotifyWithParent. +// +// TODO(gvisor.dev/issue/1479): Implement inotify. +func (d *anonDentry) InotifyWithParent(events uint32, cookie uint32, et EventType) {} + +// Watches implements DentryImpl.Watches. +// +// TODO(gvisor.dev/issue/1479): Implement inotify. +func (d *anonDentry) Watches() *Watches { + return nil +} diff --git a/pkg/sentry/vfs/context.go b/pkg/sentry/vfs/context.go index d97362b9a..c9e724fef 100644 --- a/pkg/sentry/vfs/context.go +++ b/pkg/sentry/vfs/context.go @@ -29,9 +29,10 @@ const ( CtxRoot ) -// MountNamespaceFromContext returns the MountNamespace used by ctx. It does -// not take a reference on the returned MountNamespace. If ctx is not -// associated with a MountNamespace, MountNamespaceFromContext returns nil. +// MountNamespaceFromContext returns the MountNamespace used by ctx. If ctx is +// not associated with a MountNamespace, MountNamespaceFromContext returns nil. +// +// A reference is taken on the returned MountNamespace. func MountNamespaceFromContext(ctx context.Context) *MountNamespace { if v := ctx.Value(CtxMountNamespace); v != nil { return v.(*MountNamespace) @@ -48,3 +49,27 @@ func RootFromContext(ctx context.Context) VirtualDentry { } return VirtualDentry{} } + +type rootContext struct { + context.Context + root VirtualDentry +} + +// WithRoot returns a copy of ctx with the given root. +func WithRoot(ctx context.Context, root VirtualDentry) context.Context { + return &rootContext{ + Context: ctx, + root: root, + } +} + +// Value implements Context.Value. +func (rc rootContext) Value(key interface{}) interface{} { + switch key { + case CtxRoot: + rc.root.IncRef() + return rc.root + default: + return rc.Context.Value(key) + } +} diff --git a/pkg/sentry/vfs/dentry.go b/pkg/sentry/vfs/dentry.go index 486a76475..24af13eb1 100644 --- a/pkg/sentry/vfs/dentry.go +++ b/pkg/sentry/vfs/dentry.go @@ -15,34 +15,17 @@ package vfs import ( - "fmt" "sync/atomic" "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/syserror" ) -// Dentry represents a node in a Filesystem tree which may represent a file. +// Dentry represents a node in a Filesystem tree at which a file exists. // // Dentries are reference-counted. Unless otherwise specified, all Dentry // methods require that a reference is held. // -// A Dentry transitions through up to 3 different states through its lifetime: -// -// - Dentries are initially "independent". Independent Dentries have no parent, -// and consequently no name. -// -// - Dentry.InsertChild() causes an independent Dentry to become a "child" of -// another Dentry. A child node has a parent node, and a name in that parent, -// both of which are mutable by DentryMoveChild(). Each child Dentry's name is -// unique within its parent. -// -// - Dentry.RemoveChild() causes a child Dentry to become "disowned". A -// disowned Dentry can still refer to its former parent and its former name in -// said parent, but the disowned Dentry is no longer reachable from its parent, -// and a new Dentry with the same name may become a child of the parent. (This -// is analogous to a struct dentry being "unhashed" in Linux.) -// // Dentry is loosely analogous to Linux's struct dentry, but: // // - VFS does not associate Dentries with inodes. gVisor interacts primarily @@ -57,9 +40,6 @@ import ( // and/or FileDescriptionImpl methods in gVisor's VFS. Filesystems that do // support inodes may store appropriate state in implementations of DentryImpl. // -// - VFS does not provide synchronization for mutable Dentry fields, other than -// mount-related ones. -// // - VFS does not require that Dentries are instantiated for all paths accessed // through VFS, only those that are tracked beyond the scope of a single // Filesystem operation. This includes file descriptions, mount points, mount @@ -67,41 +47,34 @@ import ( // of Dentries for operations on mutable remote filesystems that can't actually // cache any state in the Dentry. // +// - VFS does not track filesystem structure (i.e. relationships between +// Dentries), since both the relevant state and synchronization are +// filesystem-specific. +// // - For the reasons above, VFS is not directly responsible for managing Dentry // lifetime. Dentry reference counts only indicate the extent to which VFS // requires Dentries to exist; Filesystems may elect to cache or discard // Dentries with zero references. +// +// +stateify savable type Dentry struct { - // parent is this Dentry's parent in this Filesystem. If this Dentry is - // independent, parent is nil. - parent *Dentry + // mu synchronizes deletion/invalidation and mounting over this Dentry. + mu sync.Mutex `state:"nosave"` - // name is this Dentry's name in parent. - name string - - flags uint32 + // dead is true if the file represented by this Dentry has been deleted (by + // CommitDeleteDentry or CommitRenameReplaceDentry) or invalidated (by + // InvalidateDentry). dead is protected by mu. + dead bool // mounts is the number of Mounts for which this Dentry is Mount.point. // mounts is accessed using atomic memory operations. mounts uint32 - // children are child Dentries. - children map[string]*Dentry - - // mu synchronizes disowning and mounting over this Dentry. - mu sync.Mutex - // impl is the DentryImpl associated with this Dentry. impl is immutable. // This should be the last field in Dentry. impl DentryImpl } -const ( - // dflagsDisownedMask is set in Dentry.flags if the Dentry has been - // disowned. - dflagsDisownedMask = 1 << iota -) - // Init must be called before first use of d. func (d *Dentry) Init(impl DentryImpl) { d.impl = impl @@ -130,20 +103,22 @@ type DentryImpl interface { // DecRef decrements the Dentry's reference count. DecRef() -} - -// IsDisowned returns true if d is disowned. -func (d *Dentry) IsDisowned() bool { - return atomic.LoadUint32(&d.flags)&dflagsDisownedMask != 0 -} - -// Preconditions: !d.IsDisowned(). -func (d *Dentry) setDisowned() { - atomic.AddUint32(&d.flags, dflagsDisownedMask) -} -func (d *Dentry) isMounted() bool { - return atomic.LoadUint32(&d.mounts) != 0 + // InotifyWithParent notifies all watches on the targets represented by this + // dentry and its parent. The parent's watches are notified first, followed + // by this dentry's. + // + // InotifyWithParent automatically adds the IN_ISDIR flag for dentries + // representing directories. + // + // Note that the events may not actually propagate up to the user, depending + // on the event masks. + InotifyWithParent(events uint32, cookie uint32, et EventType) + + // Watches returns the set of inotify watches for the file corresponding to + // the Dentry. Dentries that are hard links to the same underlying file + // share the same watches. + Watches() *Watches } // IncRef increments d's reference count. @@ -162,104 +137,37 @@ func (d *Dentry) DecRef() { d.impl.DecRef() } -// These functions are exported so that filesystem implementations can use -// them. The vfs package, and users of VFS, should not call these functions. -// Unless otherwise specified, these methods require that there are no -// concurrent mutators of d. - -// Name returns d's name in its parent in its owning Filesystem. If d is -// independent, Name returns an empty string. -func (d *Dentry) Name() string { - return d.name -} - -// Parent returns d's parent in its owning Filesystem. It does not take a -// reference on the returned Dentry. If d is independent, Parent returns nil. -func (d *Dentry) Parent() *Dentry { - return d.parent +// IsDead returns true if d has been deleted or invalidated by its owning +// filesystem. +func (d *Dentry) IsDead() bool { + d.mu.Lock() + defer d.mu.Unlock() + return d.dead } -// ParentOrSelf is equivalent to Parent, but returns d if d is independent. -func (d *Dentry) ParentOrSelf() *Dentry { - if d.parent == nil { - return d - } - return d.parent +func (d *Dentry) isMounted() bool { + return atomic.LoadUint32(&d.mounts) != 0 } -// Child returns d's child with the given name in its owning Filesystem. It -// does not take a reference on the returned Dentry. If no such child exists, -// Child returns nil. -func (d *Dentry) Child(name string) *Dentry { - return d.children[name] +// InotifyWithParent notifies all watches on the inodes for this dentry and +// its parent of events. +func (d *Dentry) InotifyWithParent(events uint32, cookie uint32, et EventType) { + d.impl.InotifyWithParent(events, cookie, et) } -// HasChildren returns true if d has any children. -func (d *Dentry) HasChildren() bool { - return len(d.children) != 0 +// Watches returns the set of inotify watches associated with d. +func (d *Dentry) Watches() *Watches { + return d.impl.Watches() } -// Children returns a map containing all of d's children. -func (d *Dentry) Children() map[string]*Dentry { - if !d.HasChildren() { - return nil - } - m := make(map[string]*Dentry) - for name, child := range d.children { - m[name] = child - } - return m -} - -// InsertChild makes child a child of d with the given name. -// -// InsertChild is a mutator of d and child. -// -// Preconditions: child must be an independent Dentry. d and child must be from -// the same Filesystem. d must not already have a child with the given name. -func (d *Dentry) InsertChild(child *Dentry, name string) { - if checkInvariants { - if _, ok := d.children[name]; ok { - panic(fmt.Sprintf("parent already contains a child named %q", name)) - } - if child.parent != nil || child.name != "" { - panic(fmt.Sprintf("child is not independent: parent = %v, name = %q", child.parent, child.name)) - } - } - if d.children == nil { - d.children = make(map[string]*Dentry) - } - d.children[name] = child - child.parent = d - child.name = name -} - -// IsAncestorOf returns true if d is an ancestor of d2; that is, d is either -// d2's parent or an ancestor of d2's parent. -func (d *Dentry) IsAncestorOf(d2 *Dentry) bool { - for d2.parent != nil { - if d2.parent == d { - return true - } - d2 = d2.parent - } - return false -} +// The following functions are exported so that filesystem implementations can +// use them. The vfs package, and users of VFS, should not call these +// functions. // PrepareDeleteDentry must be called before attempting to delete the file // represented by d. If PrepareDeleteDentry succeeds, the caller must call // AbortDeleteDentry or CommitDeleteDentry depending on the deletion's outcome. -// -// Preconditions: d is a child Dentry. func (vfs *VirtualFilesystem) PrepareDeleteDentry(mntns *MountNamespace, d *Dentry) error { - if checkInvariants { - if d.parent == nil { - panic("d is independent") - } - if d.IsDisowned() { - panic("d is already disowned") - } - } vfs.mountMu.Lock() if mntns.mountpoints[d] != 0 { vfs.mountMu.Unlock() @@ -278,42 +186,27 @@ func (vfs *VirtualFilesystem) AbortDeleteDentry(d *Dentry) { d.mu.Unlock() } -// CommitDeleteDentry must be called after the file represented by d is -// deleted, and causes d to become disowned. -// -// CommitDeleteDentry is a mutator of d and d.Parent(). -// -// Preconditions: PrepareDeleteDentry was previously called on d. +// CommitDeleteDentry must be called after PrepareDeleteDentry if the deletion +// succeeds. func (vfs *VirtualFilesystem) CommitDeleteDentry(d *Dentry) { - if d.parent != nil { - delete(d.parent.children, d.name) - } - d.setDisowned() + d.dead = true d.mu.Unlock() if d.isMounted() { - vfs.forgetDisownedMountpoint(d) + vfs.forgetDeadMountpoint(d) } } -// ForceDeleteDentry causes d to become disowned. It should only be used in -// cases where VFS has no ability to stop the deletion (e.g. d represents the -// local state of a file on a remote filesystem on which the file has already -// been deleted). -// -// ForceDeleteDentry is a mutator of d and d.Parent(). -// -// Preconditions: d is a child Dentry. -func (vfs *VirtualFilesystem) ForceDeleteDentry(d *Dentry) { - if checkInvariants { - if d.parent == nil { - panic("d is independent") - } - if d.IsDisowned() { - panic("d is already disowned") - } - } +// InvalidateDentry is called when d ceases to represent the file it formerly +// did for reasons outside of VFS' control (e.g. d represents the local state +// of a file on a remote filesystem on which the file has already been +// deleted). +func (vfs *VirtualFilesystem) InvalidateDentry(d *Dentry) { d.mu.Lock() - vfs.CommitDeleteDentry(d) + d.dead = true + d.mu.Unlock() + if d.isMounted() { + vfs.forgetDeadMountpoint(d) + } } // PrepareRenameDentry must be called before attempting to rename the file @@ -322,25 +215,9 @@ func (vfs *VirtualFilesystem) ForceDeleteDentry(d *Dentry) { // caller must call AbortRenameDentry, CommitRenameReplaceDentry, or // CommitRenameExchangeDentry depending on the rename's outcome. // -// Preconditions: from is a child Dentry. If to is not nil, it must be a child -// Dentry from the same Filesystem. from != to. +// Preconditions: If to is not nil, it must be a child Dentry from the same +// Filesystem. from != to. func (vfs *VirtualFilesystem) PrepareRenameDentry(mntns *MountNamespace, from, to *Dentry) error { - if checkInvariants { - if from.parent == nil { - panic("from is independent") - } - if from.IsDisowned() { - panic("from is already disowned") - } - if to != nil { - if to.parent == nil { - panic("to is independent") - } - if to.IsDisowned() { - panic("to is already disowned") - } - } - } vfs.mountMu.Lock() if mntns.mountpoints[from] != 0 { vfs.mountMu.Unlock() @@ -374,24 +251,14 @@ func (vfs *VirtualFilesystem) AbortRenameDentry(from, to *Dentry) { // is renamed without RENAME_EXCHANGE. If to is not nil, it represents the file // that was replaced by from. // -// CommitRenameReplaceDentry is a mutator of from, to, from.Parent(), and -// to.Parent(). -// // Preconditions: PrepareRenameDentry was previously called on from and to. -// newParent.Child(newName) == to. -func (vfs *VirtualFilesystem) CommitRenameReplaceDentry(from, newParent *Dentry, newName string, to *Dentry) { - if newParent.children == nil { - newParent.children = make(map[string]*Dentry) - } - newParent.children[newName] = from - from.parent = newParent - from.name = newName +func (vfs *VirtualFilesystem) CommitRenameReplaceDentry(from, to *Dentry) { from.mu.Unlock() if to != nil { - to.setDisowned() + to.dead = true to.mu.Unlock() if to.isMounted() { - vfs.forgetDisownedMountpoint(to) + vfs.forgetDeadMountpoint(to) } } } @@ -399,25 +266,18 @@ func (vfs *VirtualFilesystem) CommitRenameReplaceDentry(from, newParent *Dentry, // CommitRenameExchangeDentry must be called after the files represented by // from and to are exchanged by rename(RENAME_EXCHANGE). // -// CommitRenameExchangeDentry is a mutator of from, to, from.Parent(), and -// to.Parent(). -// // Preconditions: PrepareRenameDentry was previously called on from and to. func (vfs *VirtualFilesystem) CommitRenameExchangeDentry(from, to *Dentry) { - from.parent, to.parent = to.parent, from.parent - from.name, to.name = to.name, from.name - from.parent.children[from.name] = from - to.parent.children[to.name] = to from.mu.Unlock() to.mu.Unlock() } -// forgetDisownedMountpoint is called when a mount point is deleted to umount -// all mounts using it in all other mount namespaces. +// forgetDeadMountpoint is called when a mount point is deleted or invalidated +// to umount all mounts using it in all other mount namespaces. // -// forgetDisownedMountpoint is analogous to Linux's +// forgetDeadMountpoint is analogous to Linux's // fs/namespace.c:__detach_mounts(). -func (vfs *VirtualFilesystem) forgetDisownedMountpoint(d *Dentry) { +func (vfs *VirtualFilesystem) forgetDeadMountpoint(d *Dentry) { var ( vdsToDecRef []VirtualDentry mountsToDecRef []*Mount diff --git a/pkg/sentry/vfs/device.go b/pkg/sentry/vfs/device.go index 3af2aa58d..1e9dffc8f 100644 --- a/pkg/sentry/vfs/device.go +++ b/pkg/sentry/vfs/device.go @@ -56,6 +56,7 @@ type Device interface { Open(ctx context.Context, mnt *Mount, d *Dentry, opts OpenOptions) (*FileDescription, error) } +// +stateify savable type registeredDevice struct { dev Device opts RegisterDeviceOptions @@ -63,6 +64,8 @@ type registeredDevice struct { // RegisterDeviceOptions contains options to // VirtualFilesystem.RegisterDevice(). +// +// +stateify savable type RegisterDeviceOptions struct { // GroupName is the name shown for this device registration in // /proc/devices. If GroupName is empty, this registration will not be @@ -100,7 +103,7 @@ func (vfs *VirtualFilesystem) OpenDeviceSpecialFile(ctx context.Context, mnt *Mo } // GetAnonBlockDevMinor allocates and returns an unused minor device number for -// an "anonymous" block device with major number 0. +// an "anonymous" block device with major number UNNAMED_MAJOR. func (vfs *VirtualFilesystem) GetAnonBlockDevMinor() (uint32, error) { vfs.anonBlockDevMinorMu.Lock() defer vfs.anonBlockDevMinorMu.Unlock() diff --git a/pkg/sentry/vfs/epoll.go b/pkg/sentry/vfs/epoll.go index 7c83f9a5a..8297f964b 100644 --- a/pkg/sentry/vfs/epoll.go +++ b/pkg/sentry/vfs/epoll.go @@ -85,8 +85,8 @@ type epollInterest struct { ready bool epollInterestEntry - // userData is the epoll_data_t associated with this epollInterest. - // userData is protected by epoll.mu. + // userData is the struct epoll_event::data associated with this + // epollInterest. userData is protected by epoll.mu. userData [2]int32 } @@ -99,6 +99,8 @@ func (vfs *VirtualFilesystem) NewEpollInstanceFD() (*FileDescription, error) { interest: make(map[epollInterestKey]*epollInterest), } if err := ep.vfsfd.Init(ep, linux.O_RDWR, vd.Mount(), vd.Dentry(), &FileDescriptionOptions{ + DenyPRead: true, + DenyPWrite: true, UseDentryMetadata: true, }); err != nil { return nil, err @@ -157,7 +159,7 @@ func (ep *EpollInstance) Seek(ctx context.Context, offset int64, whence int32) ( // AddInterest implements the semantics of EPOLL_CTL_ADD. // // Preconditions: A reference must be held on file. -func (ep *EpollInstance) AddInterest(file *FileDescription, num int32, mask uint32, userData [2]int32) error { +func (ep *EpollInstance) AddInterest(file *FileDescription, num int32, event linux.EpollEvent) error { // Check for cyclic polling if necessary. subep, _ := file.impl.(*EpollInstance) if subep != nil { @@ -183,13 +185,14 @@ func (ep *EpollInstance) AddInterest(file *FileDescription, num int32, mask uint } // Register interest in file. - mask |= linux.EPOLLERR | linux.EPOLLRDHUP + mask := event.Events | linux.EPOLLERR | linux.EPOLLRDHUP epi := &epollInterest{ epoll: ep, key: key, mask: mask, - userData: userData, + userData: event.Data, } + epi.waiter.Callback = epi ep.interest[key] = epi wmask := waiter.EventMaskFromLinux(mask) file.EventRegister(&epi.waiter, wmask) @@ -202,6 +205,9 @@ func (ep *EpollInstance) AddInterest(file *FileDescription, num int32, mask uint // Add epi to file.epolls so that it is removed when the last // FileDescription reference is dropped. file.epollMu.Lock() + if file.epolls == nil { + file.epolls = make(map[*epollInterest]struct{}) + } file.epolls[epi] = struct{}{} file.epollMu.Unlock() @@ -236,7 +242,7 @@ func (ep *EpollInstance) mightPollRecursive(ep2 *EpollInstance, remainingRecursi // ModifyInterest implements the semantics of EPOLL_CTL_MOD. // // Preconditions: A reference must be held on file. -func (ep *EpollInstance) ModifyInterest(file *FileDescription, num int32, mask uint32, userData [2]int32) error { +func (ep *EpollInstance) ModifyInterest(file *FileDescription, num int32, event linux.EpollEvent) error { ep.interestMu.Lock() defer ep.interestMu.Unlock() @@ -250,13 +256,13 @@ func (ep *EpollInstance) ModifyInterest(file *FileDescription, num int32, mask u } // Update epi for the next call to ep.ReadEvents(). + mask := event.Events | linux.EPOLLERR | linux.EPOLLRDHUP ep.mu.Lock() epi.mask = mask - epi.userData = userData + epi.userData = event.Data ep.mu.Unlock() // Re-register with the new mask. - mask |= linux.EPOLLERR | linux.EPOLLRDHUP file.EventUnregister(&epi.waiter) wmask := waiter.EventMaskFromLinux(mask) file.EventRegister(&epi.waiter, wmask) @@ -363,8 +369,7 @@ func (ep *EpollInstance) ReadEvents(events []linux.EpollEvent) int { // Report ievents. events[i] = linux.EpollEvent{ Events: ievents.ToLinux(), - Fd: epi.userData[0], - Data: epi.userData[1], + Data: epi.userData, } i++ if i == len(events) { diff --git a/pkg/sentry/vfs/file_description.go b/pkg/sentry/vfs/file_description.go index 5bac660c7..bb294563d 100644 --- a/pkg/sentry/vfs/file_description.go +++ b/pkg/sentry/vfs/file_description.go @@ -84,6 +84,13 @@ type FileDescriptionOptions struct { // usually only the case if O_DIRECT would actually have an effect. AllowDirectIO bool + // If DenyPRead is true, calls to FileDescription.PRead() return ESPIPE. + DenyPRead bool + + // If DenyPWrite is true, calls to FileDescription.PWrite() return + // ESPIPE. + DenyPWrite bool + // If UseDentryMetadata is true, calls to FileDescription methods that // interact with file and filesystem metadata (Stat, SetStat, StatFS, // Listxattr, Getxattr, Setxattr, Removexattr) are implemented by calling @@ -100,10 +107,10 @@ type FileDescriptionOptions struct { } // Init must be called before first use of fd. If it succeeds, it takes -// references on mnt and d. statusFlags is the initial file description status -// flags, which is usually the full set of flags passed to open(2). -func (fd *FileDescription) Init(impl FileDescriptionImpl, statusFlags uint32, mnt *Mount, d *Dentry, opts *FileDescriptionOptions) error { - writable := MayWriteFileWithOpenFlags(statusFlags) +// references on mnt and d. flags is the initial file description flags, which +// is usually the full set of flags passed to open(2). +func (fd *FileDescription) Init(impl FileDescriptionImpl, flags uint32, mnt *Mount, d *Dentry, opts *FileDescriptionOptions) error { + writable := MayWriteFileWithOpenFlags(flags) if writable { if err := mnt.CheckBeginWrite(); err != nil { return err @@ -111,14 +118,18 @@ func (fd *FileDescription) Init(impl FileDescriptionImpl, statusFlags uint32, mn } fd.refs = 1 - fd.statusFlags = statusFlags | linux.O_LARGEFILE + + // Remove "file creation flags" to mirror the behavior from file.f_flags in + // fs/open.c:do_dentry_open + fd.statusFlags = flags &^ (linux.O_CREAT | linux.O_EXCL | linux.O_NOCTTY | linux.O_TRUNC) fd.vd = VirtualDentry{ mount: mnt, dentry: d, } - fd.vd.IncRef() + mnt.IncRef() + d.IncRef() fd.opts = *opts - fd.readable = MayReadFileWithOpenFlags(statusFlags) + fd.readable = MayReadFileWithOpenFlags(flags) fd.writable = writable fd.impl = impl return nil @@ -175,6 +186,12 @@ func (fd *FileDescription) DecRef() { } } +// Refs returns the current number of references. The returned count +// is inherently racy and is unsafe to use without external synchronization. +func (fd *FileDescription) Refs() int64 { + return atomic.LoadInt64(&fd.refs) +} + // Mount returns the mount on which fd was opened. It does not take a reference // on the returned Mount. func (fd *FileDescription) Mount() *Mount { @@ -193,6 +210,11 @@ func (fd *FileDescription) VirtualDentry() VirtualDentry { return fd.vd } +// Options returns the options passed to fd.Init(). +func (fd *FileDescription) Options() FileDescriptionOptions { + return fd.opts +} + // StatusFlags returns file description status flags, as for fcntl(F_GETFL). func (fd *FileDescription) StatusFlags() uint32 { return atomic.LoadUint32(&fd.statusFlags) @@ -286,7 +308,8 @@ type FileDescriptionImpl interface { Stat(ctx context.Context, opts StatOptions) (linux.Statx, error) // SetStat updates metadata for the file represented by the - // FileDescription. + // FileDescription. Implementations are responsible for checking if the + // operation can be performed (see vfs.CheckSetStat() for common checks). SetStat(ctx context.Context, opts SetStatOptions) error // StatFS returns metadata for the filesystem containing the file @@ -305,6 +328,7 @@ type FileDescriptionImpl interface { // - If opts.Flags specifies unsupported options, PRead returns EOPNOTSUPP. // // Preconditions: The FileDescription was opened for reading. + // FileDescriptionOptions.DenyPRead == false. PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts ReadOptions) (int64, error) // Read is similar to PRead, but does not specify an offset. @@ -336,6 +360,7 @@ type FileDescriptionImpl interface { // EOPNOTSUPP. // // Preconditions: The FileDescription was opened for writing. + // FileDescriptionOptions.DenyPWrite == false. PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts WriteOptions) (int64, error) // Write is similar to PWrite, but does not specify an offset, which is @@ -381,11 +406,11 @@ type FileDescriptionImpl interface { Ioctl(ctx context.Context, uio usermem.IO, args arch.SyscallArguments) (uintptr, error) // Listxattr returns all extended attribute names for the file. - Listxattr(ctx context.Context) ([]string, error) + Listxattr(ctx context.Context, size uint64) ([]string, error) // Getxattr returns the value associated with the given extended attribute // for the file. - Getxattr(ctx context.Context, name string) (string, error) + Getxattr(ctx context.Context, opts GetxattrOptions) (string, error) // Setxattr changes the value associated with the given extended attribute // for the file. @@ -435,11 +460,11 @@ type Dirent struct { // IterDirentsCallback receives Dirents from FileDescriptionImpl.IterDirents. type IterDirentsCallback interface { - // Handle handles the given iterated Dirent. It returns true if iteration - // should continue, and false if FileDescriptionImpl.IterDirents should - // terminate now and restart with the same Dirent the next time it is - // called. - Handle(dirent Dirent) bool + // Handle handles the given iterated Dirent. If Handle returns a non-nil + // error, FileDescriptionImpl.IterDirents must stop iteration and return + // the error; the next call to FileDescriptionImpl.IterDirents should + // restart with the same Dirent. + Handle(dirent Dirent) error } // OnClose is called when a file descriptor representing the FileDescription is @@ -514,6 +539,9 @@ func (fd *FileDescription) EventUnregister(e *waiter.Entry) { // offset, and returns the number of bytes read. PRead is permitted to return // partial reads with a nil error. func (fd *FileDescription) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts ReadOptions) (int64, error) { + if fd.opts.DenyPRead { + return 0, syserror.ESPIPE + } if !fd.readable { return 0, syserror.EBADF } @@ -532,6 +560,9 @@ func (fd *FileDescription) Read(ctx context.Context, dst usermem.IOSequence, opt // offset, and returns the number of bytes written. PWrite is permitted to // return partial writes with a nil error. func (fd *FileDescription) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts WriteOptions) (int64, error) { + if fd.opts.DenyPWrite { + return 0, syserror.ESPIPE + } if !fd.writable { return 0, syserror.EBADF } @@ -576,18 +607,23 @@ func (fd *FileDescription) Ioctl(ctx context.Context, uio usermem.IO, args arch. // Listxattr returns all extended attribute names for the file represented by // fd. -func (fd *FileDescription) Listxattr(ctx context.Context) ([]string, error) { +// +// If the size of the list (including a NUL terminating byte after every entry) +// would exceed size, ERANGE may be returned. Note that implementations +// are free to ignore size entirely and return without error). In all cases, +// if size is 0, the list should be returned without error, regardless of size. +func (fd *FileDescription) Listxattr(ctx context.Context, size uint64) ([]string, error) { if fd.opts.UseDentryMetadata { vfsObj := fd.vd.mount.vfs rp := vfsObj.getResolvingPath(auth.CredentialsFromContext(ctx), &PathOperation{ Root: fd.vd, Start: fd.vd, }) - names, err := fd.vd.mount.fs.impl.ListxattrAt(ctx, rp) + names, err := fd.vd.mount.fs.impl.ListxattrAt(ctx, rp, size) vfsObj.putResolvingPath(rp) return names, err } - names, err := fd.impl.Listxattr(ctx) + names, err := fd.impl.Listxattr(ctx, size) if err == syserror.ENOTSUP { // Linux doesn't actually return ENOTSUP in this case; instead, // fs/xattr.c:vfs_listxattr() falls back to allowing the security @@ -600,34 +636,39 @@ func (fd *FileDescription) Listxattr(ctx context.Context) ([]string, error) { // Getxattr returns the value associated with the given extended attribute for // the file represented by fd. -func (fd *FileDescription) Getxattr(ctx context.Context, name string) (string, error) { +// +// If the size of the return value exceeds opts.Size, ERANGE may be returned +// (note that implementations are free to ignore opts.Size entirely and return +// without error). In all cases, if opts.Size is 0, the value should be +// returned without error, regardless of size. +func (fd *FileDescription) Getxattr(ctx context.Context, opts *GetxattrOptions) (string, error) { if fd.opts.UseDentryMetadata { vfsObj := fd.vd.mount.vfs rp := vfsObj.getResolvingPath(auth.CredentialsFromContext(ctx), &PathOperation{ Root: fd.vd, Start: fd.vd, }) - val, err := fd.vd.mount.fs.impl.GetxattrAt(ctx, rp, name) + val, err := fd.vd.mount.fs.impl.GetxattrAt(ctx, rp, *opts) vfsObj.putResolvingPath(rp) return val, err } - return fd.impl.Getxattr(ctx, name) + return fd.impl.Getxattr(ctx, *opts) } // Setxattr changes the value associated with the given extended attribute for // the file represented by fd. -func (fd *FileDescription) Setxattr(ctx context.Context, opts SetxattrOptions) error { +func (fd *FileDescription) Setxattr(ctx context.Context, opts *SetxattrOptions) error { if fd.opts.UseDentryMetadata { vfsObj := fd.vd.mount.vfs rp := vfsObj.getResolvingPath(auth.CredentialsFromContext(ctx), &PathOperation{ Root: fd.vd, Start: fd.vd, }) - err := fd.vd.mount.fs.impl.SetxattrAt(ctx, rp, opts) + err := fd.vd.mount.fs.impl.SetxattrAt(ctx, rp, *opts) vfsObj.putResolvingPath(rp) return err } - return fd.impl.Setxattr(ctx, opts) + return fd.impl.Setxattr(ctx, *opts) } // Removexattr removes the given extended attribute from the file represented diff --git a/pkg/sentry/vfs/file_description_impl_util.go b/pkg/sentry/vfs/file_description_impl_util.go index c2a52ec1b..f4c111926 100644 --- a/pkg/sentry/vfs/file_description_impl_util.go +++ b/pkg/sentry/vfs/file_description_impl_util.go @@ -33,8 +33,8 @@ import ( // implementations to adapt: // - Have a local fileDescription struct (containing FileDescription) which // embeds FileDescriptionDefaultImpl and overrides the default methods -// which are common to all fd implementations for that for that filesystem -// like StatusFlags, SetStatusFlags, Stat, SetStat, StatFS, etc. +// which are common to all fd implementations for that filesystem like +// StatusFlags, SetStatusFlags, Stat, SetStat, StatFS, etc. // - This should be embedded in all file description implementations as the // first field by value. // - Directory FDs would also embed DirectoryFileDescriptionDefaultImpl. @@ -130,14 +130,14 @@ func (FileDescriptionDefaultImpl) Ioctl(ctx context.Context, uio usermem.IO, arg // Listxattr implements FileDescriptionImpl.Listxattr analogously to // inode_operations::listxattr == NULL in Linux. -func (FileDescriptionDefaultImpl) Listxattr(ctx context.Context) ([]string, error) { +func (FileDescriptionDefaultImpl) Listxattr(ctx context.Context, size uint64) ([]string, error) { // This isn't exactly accurate; see FileDescription.Listxattr. return nil, syserror.ENOTSUP } // Getxattr implements FileDescriptionImpl.Getxattr analogously to // inode::i_opflags & IOP_XATTR == 0 in Linux. -func (FileDescriptionDefaultImpl) Getxattr(ctx context.Context, name string) (string, error) { +func (FileDescriptionDefaultImpl) Getxattr(ctx context.Context, opts GetxattrOptions) (string, error) { return "", syserror.ENOTSUP } @@ -339,6 +339,11 @@ func (fd *DynamicBytesFileDescriptionImpl) pwriteLocked(ctx context.Context, src if opts.Flags&^(linux.RWF_HIPRI|linux.RWF_DSYNC|linux.RWF_SYNC) != 0 { return 0, syserror.EOPNOTSUPP } + limit, err := CheckLimit(ctx, offset, src.NumBytes()) + if err != nil { + return 0, err + } + src = src.TakeFirst64(limit) writable, ok := fd.data.(WritableDynamicBytesSource) if !ok { diff --git a/pkg/sentry/vfs/file_description_impl_util_test.go b/pkg/sentry/vfs/file_description_impl_util_test.go index 8fa26418e..3a75d4d62 100644 --- a/pkg/sentry/vfs/file_description_impl_util_test.go +++ b/pkg/sentry/vfs/file_description_impl_util_test.go @@ -107,7 +107,10 @@ func (fd *testFD) SetStat(ctx context.Context, opts SetStatOptions) error { func TestGenCountFD(t *testing.T) { ctx := contexttest.Context(t) - vfsObj := New() // vfs.New() + vfsObj := &VirtualFilesystem{} + if err := vfsObj.Init(); err != nil { + t.Fatalf("VFS init: %v", err) + } fd := newTestFD(vfsObj, linux.O_RDWR, &genCount{}) defer fd.DecRef() @@ -162,7 +165,10 @@ func TestGenCountFD(t *testing.T) { func TestWritable(t *testing.T) { ctx := contexttest.Context(t) - vfsObj := New() // vfs.New() + vfsObj := &VirtualFilesystem{} + if err := vfsObj.Init(); err != nil { + t.Fatalf("VFS init: %v", err) + } fd := newTestFD(vfsObj, linux.O_RDWR, &storeData{data: "init"}) defer fd.DecRef() diff --git a/pkg/sentry/vfs/filesystem.go b/pkg/sentry/vfs/filesystem.go index a06a6caf3..1edd584c9 100644 --- a/pkg/sentry/vfs/filesystem.go +++ b/pkg/sentry/vfs/filesystem.go @@ -20,6 +20,8 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" ) // A Filesystem is a tree of nodes represented by Dentries, which forms part of @@ -29,6 +31,8 @@ import ( // Filesystem methods require that a reference is held. // // Filesystem is analogous to Linux's struct super_block. +// +// +stateify savable type Filesystem struct { // refs is the reference count. refs is accessed using atomic memory // operations. @@ -38,21 +42,30 @@ type Filesystem struct { // immutable. vfs *VirtualFilesystem + // fsType is the FilesystemType of this Filesystem. + fsType FilesystemType + // impl is the FilesystemImpl associated with this Filesystem. impl is // immutable. This should be the last field in Dentry. impl FilesystemImpl } // Init must be called before first use of fs. -func (fs *Filesystem) Init(vfsObj *VirtualFilesystem, impl FilesystemImpl) { +func (fs *Filesystem) Init(vfsObj *VirtualFilesystem, fsType FilesystemType, impl FilesystemImpl) { fs.refs = 1 fs.vfs = vfsObj + fs.fsType = fsType fs.impl = impl vfsObj.filesystemsMu.Lock() vfsObj.filesystems[fs] = struct{}{} vfsObj.filesystemsMu.Unlock() } +// FilesystemType returns the FilesystemType for this Filesystem. +func (fs *Filesystem) FilesystemType() FilesystemType { + return fs.fsType +} + // VirtualFilesystem returns the containing VirtualFilesystem. func (fs *Filesystem) VirtualFilesystem() *VirtualFilesystem { return fs.vfs @@ -142,6 +155,9 @@ type FilesystemImpl interface { // file data to be written to the underlying [filesystem]", as by syncfs(2). Sync(ctx context.Context) error + // AccessAt checks whether a user with creds can access the file at rp. + AccessAt(ctx context.Context, rp *ResolvingPath, creds *auth.Credentials, ats AccessTypes) error + // GetDentryAt returns a Dentry representing the file at rp. A reference is // taken on the returned Dentry. // @@ -330,7 +346,10 @@ type FilesystemImpl interface { // ENOTEMPTY. // // Preconditions: !rp.Done(). For the final path component in rp, - // !rp.ShouldFollowSymlink(). oldName is not "." or "..". + // !rp.ShouldFollowSymlink(). oldParentVD.Dentry() was obtained from a + // previous call to + // oldParentVD.Mount().Filesystem().Impl().GetParentDentryAt(). oldName is + // not "." or "..". // // Postconditions: If RenameAt returns an error returned by // ResolvingPath.Resolve*(), then !rp.Done(). @@ -360,7 +379,9 @@ type FilesystemImpl interface { // ResolvingPath.Resolve*(), then !rp.Done(). RmdirAt(ctx context.Context, rp *ResolvingPath) error - // SetStatAt updates metadata for the file at the given path. + // SetStatAt updates metadata for the file at the given path. Implementations + // are responsible for checking if the operation can be performed + // (see vfs.CheckSetStat() for common checks). // // Errors: // @@ -422,9 +443,14 @@ type FilesystemImpl interface { // Errors: // // - If extended attributes are not supported by the filesystem, - // ListxattrAt returns nil. (See FileDescription.Listxattr for an - // explanation.) - ListxattrAt(ctx context.Context, rp *ResolvingPath) ([]string, error) + // ListxattrAt returns ENOTSUP. + // + // - If the size of the list (including a NUL terminating byte after every + // entry) would exceed size, ERANGE may be returned. Note that + // implementations are free to ignore size entirely and return without + // error). In all cases, if size is 0, the list should be returned without + // error, regardless of size. + ListxattrAt(ctx context.Context, rp *ResolvingPath, size uint64) ([]string, error) // GetxattrAt returns the value associated with the given extended // attribute for the file at rp. @@ -433,7 +459,15 @@ type FilesystemImpl interface { // // - If extended attributes are not supported by the filesystem, GetxattrAt // returns ENOTSUP. - GetxattrAt(ctx context.Context, rp *ResolvingPath, name string) (string, error) + // + // - If an extended attribute named opts.Name does not exist, ENODATA is + // returned. + // + // - If the size of the return value exceeds opts.Size, ERANGE may be + // returned (note that implementations are free to ignore opts.Size entirely + // and return without error). In all cases, if opts.Size is 0, the value + // should be returned without error, regardless of size. + GetxattrAt(ctx context.Context, rp *ResolvingPath, opts GetxattrOptions) (string, error) // SetxattrAt changes the value associated with the given extended // attribute for the file at rp. @@ -442,6 +476,10 @@ type FilesystemImpl interface { // // - If extended attributes are not supported by the filesystem, SetxattrAt // returns ENOTSUP. + // + // - If XATTR_CREATE is set in opts.Flag and opts.Name already exists, + // EEXIST is returned. If XATTR_REPLACE is set and opts.Name does not exist, + // ENODATA is returned. SetxattrAt(ctx context.Context, rp *ResolvingPath, opts SetxattrOptions) error // RemovexattrAt removes the given extended attribute from the file at rp. @@ -450,8 +488,21 @@ type FilesystemImpl interface { // // - If extended attributes are not supported by the filesystem, // RemovexattrAt returns ENOTSUP. + // + // - If name does not exist, ENODATA is returned. RemovexattrAt(ctx context.Context, rp *ResolvingPath, name string) error + // BoundEndpointAt returns the Unix socket endpoint bound at the path rp. + // + // Errors: + // + // - If the file does not have write permissions, then BoundEndpointAt + // returns EACCES. + // + // - If a non-socket file exists at rp, then BoundEndpointAt returns + // ECONNREFUSED. + BoundEndpointAt(ctx context.Context, rp *ResolvingPath, opts BoundEndpointOptions) (transport.BoundEndpoint, error) + // PrependPath prepends a path from vd to vd.Mount().Root() to b. // // If vfsroot.Ok(), it is the contextual VFS root; if it is encountered @@ -474,7 +525,7 @@ type FilesystemImpl interface { // Preconditions: vd.Mount().Filesystem().Impl() == this FilesystemImpl. PrependPath(ctx context.Context, vfsroot, vd VirtualDentry, b *fspath.Builder) error - // TODO: inotify_add_watch(); bind() + // TODO(gvisor.dev/issue/1479): inotify_add_watch() } // PrependPathAtVFSRootError is returned by implementations of diff --git a/pkg/sentry/vfs/filesystem_impl_util.go b/pkg/sentry/vfs/filesystem_impl_util.go index 7315a588e..465e610e0 100644 --- a/pkg/sentry/vfs/filesystem_impl_util.go +++ b/pkg/sentry/vfs/filesystem_impl_util.go @@ -16,8 +16,6 @@ package vfs import ( "strings" - - "gvisor.dev/gvisor/pkg/fspath" ) // GenericParseMountOptions parses a comma-separated list of options of the @@ -43,27 +41,3 @@ func GenericParseMountOptions(str string) map[string]string { } return m } - -// GenericPrependPath may be used by implementations of -// FilesystemImpl.PrependPath() for which a single statically-determined lock -// or set of locks is sufficient to ensure its preconditions (as opposed to -// e.g. per-Dentry locks). -// -// Preconditions: Dentry.Name() and Dentry.Parent() must be held constant for -// vd.Dentry() and all of its ancestors. -func GenericPrependPath(vfsroot, vd VirtualDentry, b *fspath.Builder) error { - mnt, d := vd.mount, vd.dentry - for { - if mnt == vfsroot.mount && d == vfsroot.dentry { - return PrependPathAtVFSRootError{} - } - if d == mnt.root { - return nil - } - if d.parent == nil { - return PrependPathAtNonMountRootError{} - } - b.PrependComponent(d.name) - d = d.parent - } -} diff --git a/pkg/sentry/vfs/filesystem_type.go b/pkg/sentry/vfs/filesystem_type.go index c58b70728..f2298f7f6 100644 --- a/pkg/sentry/vfs/filesystem_type.go +++ b/pkg/sentry/vfs/filesystem_type.go @@ -30,6 +30,9 @@ type FilesystemType interface { // along with its mount root. A reference is taken on the returned // Filesystem and Dentry. GetFilesystem(ctx context.Context, vfsObj *VirtualFilesystem, creds *auth.Credentials, source string, opts GetFilesystemOptions) (*Filesystem, *Dentry, error) + + // Name returns the name of this FilesystemType. + Name() string } // GetFilesystemOptions contains options to FilesystemType.GetFilesystem. @@ -44,6 +47,7 @@ type GetFilesystemOptions struct { InternalData interface{} } +// +stateify savable type registeredFilesystemType struct { fsType FilesystemType opts RegisterFilesystemTypeOptions diff --git a/pkg/sentry/vfs/genericfstree/BUILD b/pkg/sentry/vfs/genericfstree/BUILD new file mode 100644 index 000000000..d8fd92677 --- /dev/null +++ b/pkg/sentry/vfs/genericfstree/BUILD @@ -0,0 +1,16 @@ +load("//tools/go_generics:defs.bzl", "go_template") + +package( + default_visibility = ["//:sandbox"], + licenses = ["notice"], +) + +go_template( + name = "generic_fstree", + srcs = [ + "genericfstree.go", + ], + types = [ + "Dentry", + ], +) diff --git a/pkg/sentry/vfs/genericfstree/genericfstree.go b/pkg/sentry/vfs/genericfstree/genericfstree.go new file mode 100644 index 000000000..8882fa84a --- /dev/null +++ b/pkg/sentry/vfs/genericfstree/genericfstree.go @@ -0,0 +1,81 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// Package genericfstree provides tools for implementing vfs.FilesystemImpls +// where a single statically-determined lock or set of locks is sufficient to +// ensure that a Dentry's name and parent are contextually immutable. +// +// Clients using this package must use the go_template_instance rule in +// tools/go_generics/defs.bzl to create an instantiation of this template +// package, providing types to use in place of Dentry. +package genericfstree + +import ( + "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/sentry/vfs" +) + +// Dentry is a required type parameter that is a struct with the given fields. +type Dentry struct { + // vfsd is the embedded vfs.Dentry corresponding to this vfs.DentryImpl. + vfsd vfs.Dentry + + // parent is the parent of this Dentry in the filesystem's tree. If this + // Dentry is a filesystem root, parent is nil. + parent *Dentry + + // name is the name of this Dentry in its parent. If this Dentry is a + // filesystem root, name is unspecified. + name string +} + +// IsAncestorDentry returns true if d is an ancestor of d2; that is, d is +// either d2's parent or an ancestor of d2's parent. +func IsAncestorDentry(d, d2 *Dentry) bool { + for d2 != nil { // Stop at root, where d2.parent == nil. + if d2.parent == d { + return true + } + if d2.parent == d2 { + return false + } + d2 = d2.parent + } + return false +} + +// ParentOrSelf returns d.parent. If d.parent is nil, ParentOrSelf returns d. +func ParentOrSelf(d *Dentry) *Dentry { + if d.parent != nil { + return d.parent + } + return d +} + +// PrependPath is a generic implementation of FilesystemImpl.PrependPath(). +func PrependPath(vfsroot vfs.VirtualDentry, mnt *vfs.Mount, d *Dentry, b *fspath.Builder) error { + for { + if mnt == vfsroot.Mount() && &d.vfsd == vfsroot.Dentry() { + return vfs.PrependPathAtVFSRootError{} + } + if &d.vfsd == mnt.Root() { + return nil + } + if d.parent == nil { + return vfs.PrependPathAtNonMountRootError{} + } + b.PrependComponent(d.name) + d = d.parent + } +} diff --git a/pkg/sentry/vfs/inotify.go b/pkg/sentry/vfs/inotify.go new file mode 100644 index 000000000..05a3051a4 --- /dev/null +++ b/pkg/sentry/vfs/inotify.go @@ -0,0 +1,697 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs + +import ( + "bytes" + "fmt" + "sync/atomic" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/uniqueid" + "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" + "gvisor.dev/gvisor/pkg/waiter" +) + +// inotifyEventBaseSize is the base size of linux's struct inotify_event. This +// must be a power 2 for rounding below. +const inotifyEventBaseSize = 16 + +// EventType defines different kinds of inotfiy events. +// +// The way events are labelled appears somewhat arbitrary, but they must match +// Linux so that IN_EXCL_UNLINK behaves as it does in Linux. +type EventType uint8 + +// PathEvent and InodeEvent correspond to FSNOTIFY_EVENT_PATH and +// FSNOTIFY_EVENT_INODE in Linux. +const ( + PathEvent EventType = iota + InodeEvent EventType = iota +) + +// Inotify represents an inotify instance created by inotify_init(2) or +// inotify_init1(2). Inotify implements FileDescriptionImpl. +// +// Lock ordering: +// Inotify.mu -> Watches.mu -> Inotify.evMu +// +// +stateify savable +type Inotify struct { + vfsfd FileDescription + FileDescriptionDefaultImpl + DentryMetadataFileDescriptionImpl + + // Unique identifier for this inotify instance. We don't just reuse the + // inotify fd because fds can be duped. These should not be exposed to the + // user, since we may aggressively reuse an id on S/R. + id uint64 + + // queue is used to notify interested parties when the inotify instance + // becomes readable or writable. + queue waiter.Queue `state:"nosave"` + + // evMu *only* protects the events list. We need a separate lock while + // queuing events: using mu may violate lock ordering, since at that point + // the calling goroutine may already hold Watches.mu. + evMu sync.Mutex `state:"nosave"` + + // A list of pending events for this inotify instance. Protected by evMu. + events eventList + + // A scratch buffer, used to serialize inotify events. Allocate this + // ahead of time for the sake of performance. Protected by evMu. + scratch []byte + + // mu protects the fields below. + mu sync.Mutex `state:"nosave"` + + // nextWatchMinusOne is used to allocate watch descriptors on this Inotify + // instance. Note that Linux starts numbering watch descriptors from 1. + nextWatchMinusOne int32 + + // Map from watch descriptors to watch objects. + watches map[int32]*Watch +} + +var _ FileDescriptionImpl = (*Inotify)(nil) + +// NewInotifyFD constructs a new Inotify instance. +func NewInotifyFD(ctx context.Context, vfsObj *VirtualFilesystem, flags uint32) (*FileDescription, error) { + // O_CLOEXEC affects file descriptors, so it must be handled outside of vfs. + flags &^= linux.O_CLOEXEC + if flags&^linux.O_NONBLOCK != 0 { + return nil, syserror.EINVAL + } + + id := uniqueid.GlobalFromContext(ctx) + vd := vfsObj.NewAnonVirtualDentry(fmt.Sprintf("[inotifyfd:%d]", id)) + defer vd.DecRef() + fd := &Inotify{ + id: id, + scratch: make([]byte, inotifyEventBaseSize), + watches: make(map[int32]*Watch), + } + if err := fd.vfsfd.Init(fd, flags, vd.Mount(), vd.Dentry(), &FileDescriptionOptions{ + UseDentryMetadata: true, + DenyPRead: true, + DenyPWrite: true, + }); err != nil { + return nil, err + } + return &fd.vfsfd, nil +} + +// Release implements FileDescriptionImpl.Release. Release removes all +// watches and frees all resources for an inotify instance. +func (i *Inotify) Release() { + // We need to hold i.mu to avoid a race with concurrent calls to + // Inotify.handleDeletion from Watches. There's no risk of Watches + // accessing this Inotify after the destructor ends, because we remove all + // references to it below. + i.mu.Lock() + defer i.mu.Unlock() + for _, w := range i.watches { + // Remove references to the watch from the watches set on the target. We + // don't need to worry about the references from i.watches, since this + // file description is about to be destroyed. + w.set.Remove(i.id) + } +} + +// EventRegister implements waiter.Waitable. +func (i *Inotify) EventRegister(e *waiter.Entry, mask waiter.EventMask) { + i.queue.EventRegister(e, mask) +} + +// EventUnregister implements waiter.Waitable. +func (i *Inotify) EventUnregister(e *waiter.Entry) { + i.queue.EventUnregister(e) +} + +// Readiness implements waiter.Waitable.Readiness. +// +// Readiness indicates whether there are pending events for an inotify instance. +func (i *Inotify) Readiness(mask waiter.EventMask) waiter.EventMask { + ready := waiter.EventMask(0) + + i.evMu.Lock() + defer i.evMu.Unlock() + + if !i.events.Empty() { + ready |= waiter.EventIn + } + + return mask & ready +} + +// PRead implements FileDescriptionImpl. +func (*Inotify) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts ReadOptions) (int64, error) { + return 0, syserror.ESPIPE +} + +// PWrite implements FileDescriptionImpl. +func (*Inotify) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts WriteOptions) (int64, error) { + return 0, syserror.ESPIPE +} + +// Write implements FileDescriptionImpl.Write. +func (*Inotify) Write(ctx context.Context, src usermem.IOSequence, opts WriteOptions) (int64, error) { + return 0, syserror.EBADF +} + +// Read implements FileDescriptionImpl.Read. +func (i *Inotify) Read(ctx context.Context, dst usermem.IOSequence, opts ReadOptions) (int64, error) { + if dst.NumBytes() < inotifyEventBaseSize { + return 0, syserror.EINVAL + } + + i.evMu.Lock() + defer i.evMu.Unlock() + + if i.events.Empty() { + // Nothing to read yet, tell caller to block. + return 0, syserror.ErrWouldBlock + } + + var writeLen int64 + for it := i.events.Front(); it != nil; { + // Advance `it` before the element is removed from the list, or else + // it.Next() will always be nil. + event := it + it = it.Next() + + // Does the buffer have enough remaining space to hold the event we're + // about to write out? + if dst.NumBytes() < int64(event.sizeOf()) { + if writeLen > 0 { + // Buffer wasn't big enough for all pending events, but we did + // write some events out. + return writeLen, nil + } + return 0, syserror.EINVAL + } + + // Linux always dequeues an available event as long as there's enough + // buffer space to copy it out, even if the copy below fails. Emulate + // this behaviour. + i.events.Remove(event) + + // Buffer has enough space, copy event to the read buffer. + n, err := event.CopyTo(ctx, i.scratch, dst) + if err != nil { + return 0, err + } + + writeLen += n + dst = dst.DropFirst64(n) + } + return writeLen, nil +} + +// Ioctl implements fs.FileOperations.Ioctl. +func (i *Inotify) Ioctl(ctx context.Context, uio usermem.IO, args arch.SyscallArguments) (uintptr, error) { + switch args[1].Int() { + case linux.FIONREAD: + i.evMu.Lock() + defer i.evMu.Unlock() + var n uint32 + for e := i.events.Front(); e != nil; e = e.Next() { + n += uint32(e.sizeOf()) + } + var buf [4]byte + usermem.ByteOrder.PutUint32(buf[:], n) + _, err := uio.CopyOut(ctx, args[2].Pointer(), buf[:], usermem.IOOpts{}) + return 0, err + + default: + return 0, syserror.ENOTTY + } +} + +func (i *Inotify) queueEvent(ev *Event) { + i.evMu.Lock() + + // Check if we should coalesce the event we're about to queue with the last + // one currently in the queue. Events are coalesced if they are identical. + if last := i.events.Back(); last != nil { + if ev.equals(last) { + // "Coalesce" the two events by simply not queuing the new one. We + // don't need to raise a waiter.EventIn notification because no new + // data is available for reading. + i.evMu.Unlock() + return + } + } + + i.events.PushBack(ev) + + // Release mutex before notifying waiters because we don't control what they + // can do. + i.evMu.Unlock() + + i.queue.Notify(waiter.EventIn) +} + +// newWatchLocked creates and adds a new watch to target. +// +// Precondition: i.mu must be locked. +func (i *Inotify) newWatchLocked(target *Dentry, mask uint32) *Watch { + targetWatches := target.Watches() + w := &Watch{ + owner: i, + wd: i.nextWatchIDLocked(), + set: targetWatches, + mask: mask, + } + + // Hold the watch in this inotify instance as well as the watch set on the + // target. + i.watches[w.wd] = w + targetWatches.Add(w) + return w +} + +// newWatchIDLocked allocates and returns a new watch descriptor. +// +// Precondition: i.mu must be locked. +func (i *Inotify) nextWatchIDLocked() int32 { + i.nextWatchMinusOne++ + return i.nextWatchMinusOne +} + +// handleDeletion handles the deletion of the target of watch w. It removes w +// from i.watches and a watch removal event is generated. +func (i *Inotify) handleDeletion(w *Watch) { + i.mu.Lock() + _, found := i.watches[w.wd] + delete(i.watches, w.wd) + i.mu.Unlock() + + if found { + i.queueEvent(newEvent(w.wd, "", linux.IN_IGNORED, 0)) + } +} + +// AddWatch constructs a new inotify watch and adds it to the target. It +// returns the watch descriptor returned by inotify_add_watch(2). +func (i *Inotify) AddWatch(target *Dentry, mask uint32) int32 { + // Note: Locking this inotify instance protects the result returned by + // Lookup() below. With the lock held, we know for sure the lookup result + // won't become stale because it's impossible for *this* instance to + // add/remove watches on target. + i.mu.Lock() + defer i.mu.Unlock() + + // Does the target already have a watch from this inotify instance? + if existing := target.Watches().Lookup(i.id); existing != nil { + newmask := mask + if mask&linux.IN_MASK_ADD != 0 { + // "Add (OR) events to watch mask for this pathname if it already + // exists (instead of replacing mask)." -- inotify(7) + newmask |= atomic.LoadUint32(&existing.mask) + } + atomic.StoreUint32(&existing.mask, newmask) + return existing.wd + } + + // No existing watch, create a new watch. + w := i.newWatchLocked(target, mask) + return w.wd +} + +// RmWatch looks up an inotify watch for the given 'wd' and configures the +// target to stop sending events to this inotify instance. +func (i *Inotify) RmWatch(wd int32) error { + i.mu.Lock() + + // Find the watch we were asked to removed. + w, ok := i.watches[wd] + if !ok { + i.mu.Unlock() + return syserror.EINVAL + } + + // Remove the watch from this instance. + delete(i.watches, wd) + + // Remove the watch from the watch target. + w.set.Remove(w.OwnerID()) + i.mu.Unlock() + + // Generate the event for the removal. + i.queueEvent(newEvent(wd, "", linux.IN_IGNORED, 0)) + + return nil +} + +// Watches is the collection of all inotify watches on a single file. +// +// +stateify savable +type Watches struct { + // mu protects the fields below. + mu sync.RWMutex `state:"nosave"` + + // ws is the map of active watches in this collection, keyed by the inotify + // instance id of the owner. + ws map[uint64]*Watch +} + +// Lookup returns the watch owned by an inotify instance with the given id. +// Returns nil if no such watch exists. +// +// Precondition: the inotify instance with the given id must be locked to +// prevent the returned watch from being concurrently modified or replaced in +// Inotify.watches. +func (w *Watches) Lookup(id uint64) *Watch { + w.mu.Lock() + defer w.mu.Unlock() + return w.ws[id] +} + +// Add adds watch into this set of watches. +// +// Precondition: the inotify instance with the given id must be locked. +func (w *Watches) Add(watch *Watch) { + w.mu.Lock() + defer w.mu.Unlock() + + owner := watch.OwnerID() + // Sanity check, we should never have two watches for one owner on the + // same target. + if _, exists := w.ws[owner]; exists { + panic(fmt.Sprintf("Watch collision with ID %+v", owner)) + } + if w.ws == nil { + w.ws = make(map[uint64]*Watch) + } + w.ws[owner] = watch +} + +// Remove removes a watch with the given id from this set of watches and +// releases it. The caller is responsible for generating any watch removal +// event, as appropriate. The provided id must match an existing watch in this +// collection. +// +// Precondition: the inotify instance with the given id must be locked. +func (w *Watches) Remove(id uint64) { + w.mu.Lock() + defer w.mu.Unlock() + + if w.ws == nil { + // This watch set is being destroyed. The thread executing the + // destructor is already in the process of deleting all our watches. We + // got here with no references on the target because we raced with the + // destructor notifying all the watch owners of destruction. See the + // comment in Watches.HandleDeletion for why this race exists. + return + } + + if _, ok := w.ws[id]; !ok { + // While there's technically no problem with silently ignoring a missing + // watch, this is almost certainly a bug. + panic(fmt.Sprintf("Attempt to remove a watch, but no watch found with provided id %+v.", id)) + } + delete(w.ws, id) +} + +// Notify queues a new event with all watches in this set. +func (w *Watches) Notify(name string, events, cookie uint32, et EventType) { + w.NotifyWithExclusions(name, events, cookie, et, false) +} + +// NotifyWithExclusions queues a new event with watches in this set. Watches +// with IN_EXCL_UNLINK are skipped if the event is coming from a child that +// has been unlinked. +func (w *Watches) NotifyWithExclusions(name string, events, cookie uint32, et EventType, unlinked bool) { + // N.B. We don't defer the unlocks because Notify is in the hot path of + // all IO operations, and the defer costs too much for small IO + // operations. + w.mu.RLock() + for _, watch := range w.ws { + if unlinked && watch.ExcludeUnlinkedChildren() && et == PathEvent { + continue + } + watch.Notify(name, events, cookie) + } + w.mu.RUnlock() +} + +// HandleDeletion is called when the watch target is destroyed to emit +// the appropriate events. +func (w *Watches) HandleDeletion() { + w.Notify("", linux.IN_DELETE_SELF, 0, InodeEvent) + + // TODO(gvisor.dev/issue/1479): This doesn't work because maps are not copied + // by value. Ideally, we wouldn't have this circular locking so we can just + // notify of IN_DELETE_SELF in the same loop below. + // + // We can't hold w.mu while calling watch.handleDeletion to preserve lock + // ordering w.r.t to the owner inotify instances. Instead, atomically move + // the watches map into a local variable so we can iterate over it safely. + // + // Because of this however, it is possible for the watches' owners to reach + // this inode while the inode has no refs. This is still safe because the + // owners can only reach the inode until this function finishes calling + // watch.handleDeletion below and the inode is guaranteed to exist in the + // meantime. But we still have to be very careful not to rely on inode state + // that may have been already destroyed. + var ws map[uint64]*Watch + w.mu.Lock() + ws = w.ws + w.ws = nil + w.mu.Unlock() + + for _, watch := range ws { + // TODO(gvisor.dev/issue/1479): consider refactoring this. + watch.handleDeletion() + } +} + +// Watch represent a particular inotify watch created by inotify_add_watch. +// +// +stateify savable +type Watch struct { + // Inotify instance which owns this watch. + owner *Inotify + + // Descriptor for this watch. This is unique across an inotify instance. + wd int32 + + // set is the watch set containing this watch. It belongs to the target file + // of this watch. + set *Watches + + // Events being monitored via this watch. Must be accessed with atomic + // memory operations. + mask uint32 +} + +// OwnerID returns the id of the inotify instance that owns this watch. +func (w *Watch) OwnerID() uint64 { + return w.owner.id +} + +// ExcludeUnlinkedChildren indicates whether the watched object should continue +// to be notified of events of its children after they have been unlinked, e.g. +// for an open file descriptor. +// +// TODO(gvisor.dev/issue/1479): Implement IN_EXCL_UNLINK. +// We can do this by keeping track of the set of unlinked children in Watches +// to skip notification. +func (w *Watch) ExcludeUnlinkedChildren() bool { + return atomic.LoadUint32(&w.mask)&linux.IN_EXCL_UNLINK != 0 +} + +// Notify queues a new event on this watch. +func (w *Watch) Notify(name string, events uint32, cookie uint32) { + mask := atomic.LoadUint32(&w.mask) + if mask&events == 0 { + // We weren't watching for this event. + return + } + + // Event mask should include bits matched from the watch plus all control + // event bits. + unmaskableBits := ^uint32(0) &^ linux.IN_ALL_EVENTS + effectiveMask := unmaskableBits | mask + matchedEvents := effectiveMask & events + w.owner.queueEvent(newEvent(w.wd, name, matchedEvents, cookie)) +} + +// handleDeletion handles the deletion of w's target. +func (w *Watch) handleDeletion() { + w.owner.handleDeletion(w) +} + +// Event represents a struct inotify_event from linux. +// +// +stateify savable +type Event struct { + eventEntry + + wd int32 + mask uint32 + cookie uint32 + + // len is computed based on the name field is set automatically by + // Event.setName. It should be 0 when no name is set; otherwise it is the + // length of the name slice. + len uint32 + + // The name field has special padding requirements and should only be set by + // calling Event.setName. + name []byte +} + +func newEvent(wd int32, name string, events, cookie uint32) *Event { + e := &Event{ + wd: wd, + mask: events, + cookie: cookie, + } + if name != "" { + e.setName(name) + } + return e +} + +// paddedBytes converts a go string to a null-terminated c-string, padded with +// null bytes to a total size of 'l'. 'l' must be large enough for all the bytes +// in the 's' plus at least one null byte. +func paddedBytes(s string, l uint32) []byte { + if l < uint32(len(s)+1) { + panic("Converting string to byte array results in truncation, this can lead to buffer-overflow due to the missing null-byte!") + } + b := make([]byte, l) + copy(b, s) + + // b was zero-value initialized during make(), so the rest of the slice is + // already filled with null bytes. + + return b +} + +// setName sets the optional name for this event. +func (e *Event) setName(name string) { + // We need to pad the name such that the entire event length ends up a + // multiple of inotifyEventBaseSize. + unpaddedLen := len(name) + 1 + // Round up to nearest multiple of inotifyEventBaseSize. + e.len = uint32((unpaddedLen + inotifyEventBaseSize - 1) & ^(inotifyEventBaseSize - 1)) + // Make sure we haven't overflowed and wrapped around when rounding. + if unpaddedLen > int(e.len) { + panic("Overflow when rounding inotify event size, the 'name' field was too big.") + } + e.name = paddedBytes(name, e.len) +} + +func (e *Event) sizeOf() int { + s := inotifyEventBaseSize + int(e.len) + if s < inotifyEventBaseSize { + panic("overflow") + } + return s +} + +// CopyTo serializes this event to dst. buf is used as a scratch buffer to +// construct the output. We use a buffer allocated ahead of time for +// performance. buf must be at least inotifyEventBaseSize bytes. +func (e *Event) CopyTo(ctx context.Context, buf []byte, dst usermem.IOSequence) (int64, error) { + usermem.ByteOrder.PutUint32(buf[0:], uint32(e.wd)) + usermem.ByteOrder.PutUint32(buf[4:], e.mask) + usermem.ByteOrder.PutUint32(buf[8:], e.cookie) + usermem.ByteOrder.PutUint32(buf[12:], e.len) + + writeLen := 0 + + n, err := dst.CopyOut(ctx, buf) + if err != nil { + return 0, err + } + writeLen += n + dst = dst.DropFirst(n) + + if e.len > 0 { + n, err = dst.CopyOut(ctx, e.name) + if err != nil { + return 0, err + } + writeLen += n + } + + // Santiy check. + if writeLen != e.sizeOf() { + panic(fmt.Sprintf("Serialized unexpected amount of data for an event, expected %d, wrote %d.", e.sizeOf(), writeLen)) + } + + return int64(writeLen), nil +} + +func (e *Event) equals(other *Event) bool { + return e.wd == other.wd && + e.mask == other.mask && + e.cookie == other.cookie && + e.len == other.len && + bytes.Equal(e.name, other.name) +} + +// InotifyEventFromStatMask generates the appropriate events for an operation +// that set the stats specified in mask. +func InotifyEventFromStatMask(mask uint32) uint32 { + var ev uint32 + if mask&(linux.STATX_UID|linux.STATX_GID|linux.STATX_MODE) != 0 { + ev |= linux.IN_ATTRIB + } + if mask&linux.STATX_SIZE != 0 { + ev |= linux.IN_MODIFY + } + + if (mask & (linux.STATX_ATIME | linux.STATX_MTIME)) == (linux.STATX_ATIME | linux.STATX_MTIME) { + // Both times indicates a utime(s) call. + ev |= linux.IN_ATTRIB + } else if mask&linux.STATX_ATIME != 0 { + ev |= linux.IN_ACCESS + } else if mask&linux.STATX_MTIME != 0 { + mask |= linux.IN_MODIFY + } + return ev +} + +// InotifyRemoveChild sends the appriopriate notifications to the watch sets of +// the child being removed and its parent. +func InotifyRemoveChild(self, parent *Watches, name string) { + self.Notify("", linux.IN_ATTRIB, 0, InodeEvent) + parent.Notify(name, linux.IN_DELETE, 0, InodeEvent) + // TODO(gvisor.dev/issue/1479): implement IN_EXCL_UNLINK. +} + +// InotifyRename sends the appriopriate notifications to the watch sets of the +// file being renamed and its old/new parents. +func InotifyRename(ctx context.Context, renamed, oldParent, newParent *Watches, oldName, newName string, isDir bool) { + var dirEv uint32 + if isDir { + dirEv = linux.IN_ISDIR + } + cookie := uniqueid.InotifyCookie(ctx) + oldParent.Notify(oldName, dirEv|linux.IN_MOVED_FROM, cookie, InodeEvent) + newParent.Notify(newName, dirEv|linux.IN_MOVED_TO, cookie, InodeEvent) + // Somewhat surprisingly, self move events do not have a cookie. + renamed.Notify("", linux.IN_MOVE_SELF, 0, InodeEvent) +} diff --git a/pkg/sentry/vfs/memxattr/BUILD b/pkg/sentry/vfs/memxattr/BUILD new file mode 100644 index 000000000..d8c4d27b9 --- /dev/null +++ b/pkg/sentry/vfs/memxattr/BUILD @@ -0,0 +1,15 @@ +load("//tools:defs.bzl", "go_library") + +package(licenses = ["notice"]) + +go_library( + name = "memxattr", + srcs = ["xattr.go"], + visibility = ["//pkg/sentry:internal"], + deps = [ + "//pkg/abi/linux", + "//pkg/sentry/vfs", + "//pkg/sync", + "//pkg/syserror", + ], +) diff --git a/pkg/sentry/vfs/memxattr/xattr.go b/pkg/sentry/vfs/memxattr/xattr.go new file mode 100644 index 000000000..cc1e7d764 --- /dev/null +++ b/pkg/sentry/vfs/memxattr/xattr.go @@ -0,0 +1,102 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// Package memxattr provides a default, in-memory extended attribute +// implementation. +package memxattr + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/pkg/syserror" +) + +// SimpleExtendedAttributes implements extended attributes using a map of +// names to values. +// +// +stateify savable +type SimpleExtendedAttributes struct { + // mu protects the below fields. + mu sync.RWMutex `state:"nosave"` + xattrs map[string]string +} + +// Getxattr returns the value at 'name'. +func (x *SimpleExtendedAttributes) Getxattr(opts *vfs.GetxattrOptions) (string, error) { + x.mu.RLock() + value, ok := x.xattrs[opts.Name] + x.mu.RUnlock() + if !ok { + return "", syserror.ENODATA + } + // Check that the size of the buffer provided in getxattr(2) is large enough + // to contain the value. + if opts.Size != 0 && uint64(len(value)) > opts.Size { + return "", syserror.ERANGE + } + return value, nil +} + +// Setxattr sets 'value' at 'name'. +func (x *SimpleExtendedAttributes) Setxattr(opts *vfs.SetxattrOptions) error { + x.mu.Lock() + defer x.mu.Unlock() + if x.xattrs == nil { + if opts.Flags&linux.XATTR_REPLACE != 0 { + return syserror.ENODATA + } + x.xattrs = make(map[string]string) + } + + _, ok := x.xattrs[opts.Name] + if ok && opts.Flags&linux.XATTR_CREATE != 0 { + return syserror.EEXIST + } + if !ok && opts.Flags&linux.XATTR_REPLACE != 0 { + return syserror.ENODATA + } + + x.xattrs[opts.Name] = opts.Value + return nil +} + +// Listxattr returns all names in xattrs. +func (x *SimpleExtendedAttributes) Listxattr(size uint64) ([]string, error) { + // Keep track of the size of the buffer needed in listxattr(2) for the list. + listSize := 0 + x.mu.RLock() + names := make([]string, 0, len(x.xattrs)) + for n := range x.xattrs { + names = append(names, n) + // Add one byte per null terminator. + listSize += len(n) + 1 + } + x.mu.RUnlock() + if size != 0 && uint64(listSize) > size { + return nil, syserror.ERANGE + } + return names, nil +} + +// Removexattr removes the xattr at 'name'. +func (x *SimpleExtendedAttributes) Removexattr(name string) error { + x.mu.Lock() + defer x.mu.Unlock() + if _, ok := x.xattrs[name]; !ok { + return syserror.ENODATA + } + delete(x.xattrs, name) + return nil +} diff --git a/pkg/sentry/vfs/mount.go b/pkg/sentry/vfs/mount.go index 1fbb420f9..32f901bd8 100644 --- a/pkg/sentry/vfs/mount.go +++ b/pkg/sentry/vfs/mount.go @@ -15,7 +15,11 @@ package vfs import ( + "bytes" + "fmt" "math" + "sort" + "strings" "sync/atomic" "gvisor.dev/gvisor/pkg/abi/linux" @@ -38,14 +42,23 @@ import ( // // Mount is analogous to Linux's struct mount. (gVisor does not distinguish // between struct mount and struct vfsmount.) +// +// +stateify savable type Mount struct { - // vfs, fs, and root are immutable. References are held on fs and root. + // vfs, fs, root are immutable. References are held on fs and root. // // Invariant: root belongs to fs. vfs *VirtualFilesystem fs *Filesystem root *Dentry + // ID is the immutable mount ID. + ID uint64 + + // Flags contains settings as specified for mount(2), e.g. MS_NOEXEC, except + // for MS_RDONLY which is tracked in "writers". Immutable. + Flags MountFlags + // key is protected by VirtualFilesystem.mountMu and // VirtualFilesystem.mounts.seq, and may be nil. References are held on // key.parent and key.point if they are not nil. @@ -79,13 +92,43 @@ type Mount struct { writers int64 } -// A MountNamespace is a collection of Mounts. -// +func newMount(vfs *VirtualFilesystem, fs *Filesystem, root *Dentry, mntns *MountNamespace, opts *MountOptions) *Mount { + mnt := &Mount{ + ID: atomic.AddUint64(&vfs.lastMountID, 1), + Flags: opts.Flags, + vfs: vfs, + fs: fs, + root: root, + ns: mntns, + refs: 1, + } + if opts.ReadOnly { + mnt.setReadOnlyLocked(true) + } + return mnt +} + +// Options returns a copy of the MountOptions currently applicable to mnt. +func (mnt *Mount) Options() MountOptions { + mnt.vfs.mountMu.Lock() + defer mnt.vfs.mountMu.Unlock() + return MountOptions{ + Flags: mnt.Flags, + ReadOnly: mnt.readOnly(), + } +} + +// A MountNamespace is a collection of Mounts.// // MountNamespaces are reference-counted. Unless otherwise specified, all // MountNamespace methods require that a reference is held. // // MountNamespace is analogous to Linux's struct mnt_namespace. +// +// +stateify savable type MountNamespace struct { + // Owner is the usernamespace that owns this mount namespace. + Owner *auth.UserNamespace + // root is the MountNamespace's root mount. root is immutable. root *Mount @@ -114,6 +157,7 @@ type MountNamespace struct { func (vfs *VirtualFilesystem) NewMountNamespace(ctx context.Context, creds *auth.Credentials, source, fsTypeName string, opts *GetFilesystemOptions) (*MountNamespace, error) { rft := vfs.getFilesystemType(fsTypeName) if rft == nil { + ctx.Warningf("Unknown filesystem type: %s", fsTypeName) return nil, syserror.ENODEV } fs, root, err := rft.fsType.GetFilesystem(ctx, vfs, creds, source, *opts) @@ -121,49 +165,63 @@ func (vfs *VirtualFilesystem) NewMountNamespace(ctx context.Context, creds *auth return nil, err } mntns := &MountNamespace{ + Owner: creds.UserNamespace, refs: 1, mountpoints: make(map[*Dentry]uint32), } - mntns.root = &Mount{ - vfs: vfs, - fs: fs, - root: root, - ns: mntns, - refs: 1, - } + mntns.root = newMount(vfs, fs, root, mntns, &MountOptions{}) return mntns, nil } -// MountAt creates and mounts a Filesystem configured by the given arguments. -func (vfs *VirtualFilesystem) MountAt(ctx context.Context, creds *auth.Credentials, source string, target *PathOperation, fsTypeName string, opts *MountOptions) error { +// NewDisconnectedMount returns a Mount representing fs with the given root +// (which may be nil). The new Mount is not associated with any MountNamespace +// and is not connected to any other Mounts. References are taken on fs and +// root. +func (vfs *VirtualFilesystem) NewDisconnectedMount(fs *Filesystem, root *Dentry, opts *MountOptions) (*Mount, error) { + fs.IncRef() + if root != nil { + root.IncRef() + } + return newMount(vfs, fs, root, nil /* mntns */, opts), nil +} + +// MountDisconnected creates a Filesystem configured by the given arguments, +// then returns a Mount representing it. The new Mount is not associated with +// any MountNamespace and is not connected to any other Mounts. +func (vfs *VirtualFilesystem) MountDisconnected(ctx context.Context, creds *auth.Credentials, source string, fsTypeName string, opts *MountOptions) (*Mount, error) { rft := vfs.getFilesystemType(fsTypeName) if rft == nil { - return syserror.ENODEV + return nil, syserror.ENODEV } if !opts.InternalMount && !rft.opts.AllowUserMount { - return syserror.ENODEV + return nil, syserror.ENODEV } fs, root, err := rft.fsType.GetFilesystem(ctx, vfs, creds, source, opts.GetFilesystemOptions) if err != nil { - return err + return nil, err } + defer root.DecRef() + defer fs.DecRef() + return vfs.NewDisconnectedMount(fs, root, opts) +} + +// ConnectMountAt connects mnt at the path represented by target. +// +// Preconditions: mnt must be disconnected. +func (vfs *VirtualFilesystem) ConnectMountAt(ctx context.Context, creds *auth.Credentials, mnt *Mount, target *PathOperation) error { // We can't hold vfs.mountMu while calling FilesystemImpl methods due to // lock ordering. vd, err := vfs.GetDentryAt(ctx, creds, target, &GetDentryOptions{}) if err != nil { - root.DecRef() - fs.DecRef() return err } vfs.mountMu.Lock() vd.dentry.mu.Lock() for { - if vd.dentry.IsDisowned() { + if vd.dentry.dead { vd.dentry.mu.Unlock() vfs.mountMu.Unlock() vd.DecRef() - root.DecRef() - fs.DecRef() return syserror.ENOENT } // vd might have been mounted over between vfs.GetDentryAt() and @@ -192,17 +250,10 @@ func (vfs *VirtualFilesystem) MountAt(ctx context.Context, creds *auth.Credentia } vd.dentry.mu.Lock() } - // TODO: Linux requires that either both the mount point and the mount root - // are directories, or neither are, and returns ENOTDIR if this is not the - // case. + // TODO(gvisor.dev/issue/1035): Linux requires that either both the mount + // point and the mount root are directories, or neither are, and returns + // ENOTDIR if this is not the case. mntns := vd.mount.ns - mnt := &Mount{ - vfs: vfs, - fs: fs, - root: root, - ns: mntns, - refs: 1, - } vfs.mounts.seq.BeginWrite() vfs.connectLocked(mnt, vd, mntns) vfs.mounts.seq.EndWrite() @@ -211,6 +262,19 @@ func (vfs *VirtualFilesystem) MountAt(ctx context.Context, creds *auth.Credentia return nil } +// MountAt creates and mounts a Filesystem configured by the given arguments. +func (vfs *VirtualFilesystem) MountAt(ctx context.Context, creds *auth.Credentials, source string, target *PathOperation, fsTypeName string, opts *MountOptions) error { + mnt, err := vfs.MountDisconnected(ctx, creds, source, fsTypeName, opts) + if err != nil { + return err + } + defer mnt.DecRef() + if err := vfs.ConnectMountAt(ctx, creds, mnt, target); err != nil { + return err + } + return nil +} + // UmountAt removes the Mount at the given path. func (vfs *VirtualFilesystem) UmountAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, opts *UmountOptions) error { if opts.Flags&^(linux.MNT_FORCE|linux.MNT_DETACH) != 0 { @@ -218,6 +282,9 @@ func (vfs *VirtualFilesystem) UmountAt(ctx context.Context, creds *auth.Credenti } // MNT_FORCE is currently unimplemented except for the permission check. + // Force unmounting specifically requires CAP_SYS_ADMIN in the root user + // namespace, and not in the owner user namespace for the target mount. See + // fs/namespace.c:SYSCALL_DEFINE2(umount, ...) if opts.Flags&linux.MNT_FORCE != 0 && creds.HasCapabilityIn(linux.CAP_SYS_ADMIN, creds.UserNamespace.Root()) { return syserror.EPERM } @@ -231,14 +298,17 @@ func (vfs *VirtualFilesystem) UmountAt(ctx context.Context, creds *auth.Credenti return syserror.EINVAL } vfs.mountMu.Lock() - if mntns := MountNamespaceFromContext(ctx); mntns != nil && mntns != vd.mount.ns { - vfs.mountMu.Unlock() - return syserror.EINVAL + if mntns := MountNamespaceFromContext(ctx); mntns != nil { + defer mntns.DecRef() + if mntns != vd.mount.ns { + vfs.mountMu.Unlock() + return syserror.EINVAL + } } - // TODO(jamieliu): Linux special-cases umount of the caller's root, which - // we don't implement yet (we'll just fail it since the caller holds a - // reference on it). + // TODO(gvisor.dev/issue/1035): Linux special-cases umount of the caller's + // root, which we don't implement yet (we'll just fail it since the caller + // holds a reference on it). vfs.mounts.seq.BeginWrite() if opts.Flags&linux.MNT_DETACH == 0 { @@ -330,14 +400,22 @@ func (vfs *VirtualFilesystem) umountRecursiveLocked(mnt *Mount, opts *umountRecu // references held by vd. // // Preconditions: vfs.mountMu must be locked. vfs.mounts.seq must be in a -// writer critical section. d.mu must be locked. mnt.parent() == nil. +// writer critical section. d.mu must be locked. mnt.parent() == nil, i.e. mnt +// must not already be connected. func (vfs *VirtualFilesystem) connectLocked(mnt *Mount, vd VirtualDentry, mntns *MountNamespace) { + if checkInvariants { + if mnt.parent() != nil { + panic("VFS.connectLocked called on connected mount") + } + } + mnt.IncRef() // dropped by callers of umountRecursiveLocked mnt.storeKey(vd) if vd.mount.children == nil { vd.mount.children = make(map[*Mount]struct{}) } vd.mount.children[mnt] = struct{}{} atomic.AddUint32(&vd.dentry.mounts, 1) + mnt.ns = mntns mntns.mountpoints[vd.dentry]++ vfs.mounts.insertSeqed(mnt) vfsmpmounts, ok := vfs.mountpoints[vd.dentry] @@ -355,6 +433,11 @@ func (vfs *VirtualFilesystem) connectLocked(mnt *Mount, vd VirtualDentry, mntns // writer critical section. mnt.parent() != nil. func (vfs *VirtualFilesystem) disconnectLocked(mnt *Mount) VirtualDentry { vd := mnt.loadKey() + if checkInvariants { + if vd.mount != nil { + panic("VFS.disconnectLocked called on disconnected mount") + } + } mnt.storeKey(VirtualDentry{}) delete(vd.mount.children, mnt) atomic.AddUint32(&vd.dentry.mounts, math.MaxUint32) // -1 @@ -605,12 +688,34 @@ func (mnt *Mount) setReadOnlyLocked(ro bool) error { return nil } +func (mnt *Mount) readOnly() bool { + return atomic.LoadInt64(&mnt.writers) < 0 +} + // Filesystem returns the mounted Filesystem. It does not take a reference on // the returned Filesystem. func (mnt *Mount) Filesystem() *Filesystem { return mnt.fs } +// submountsLocked returns this Mount and all Mounts that are descendents of +// it. +// +// Precondition: mnt.vfs.mountMu must be held. +func (mnt *Mount) submountsLocked() []*Mount { + mounts := []*Mount{mnt} + for m := range mnt.children { + mounts = append(mounts, m.submountsLocked()...) + } + return mounts +} + +// Root returns the mount's root. It does not take a reference on the returned +// Dentry. +func (mnt *Mount) Root() *Dentry { + return mnt.root +} + // Root returns mntns' root. A reference is taken on the returned // VirtualDentry. func (mntns *MountNamespace) Root() VirtualDentry { @@ -621,3 +726,178 @@ func (mntns *MountNamespace) Root() VirtualDentry { vd.IncRef() return vd } + +// GenerateProcMounts emits the contents of /proc/[pid]/mounts for vfs to buf. +// +// Preconditions: taskRootDir.Ok(). +func (vfs *VirtualFilesystem) GenerateProcMounts(ctx context.Context, taskRootDir VirtualDentry, buf *bytes.Buffer) { + vfs.mountMu.Lock() + defer vfs.mountMu.Unlock() + rootMnt := taskRootDir.mount + mounts := rootMnt.submountsLocked() + sort.Slice(mounts, func(i, j int) bool { return mounts[i].ID < mounts[j].ID }) + for _, mnt := range mounts { + // Get the path to this mount relative to task root. + mntRootVD := VirtualDentry{ + mount: mnt, + dentry: mnt.root, + } + path, err := vfs.PathnameReachable(ctx, taskRootDir, mntRootVD) + if err != nil { + // For some reason we didn't get a path. Log a warning + // and run with empty path. + ctx.Warningf("Error getting pathname for mount root %+v: %v", mnt.root, err) + path = "" + } + if path == "" { + // Either an error occurred, or path is not reachable + // from root. + break + } + + opts := "rw" + if mnt.readOnly() { + opts = "ro" + } + if mnt.Flags.NoATime { + opts = ",noatime" + } + if mnt.Flags.NoExec { + opts += ",noexec" + } + + // Format: + // <special device or remote filesystem> <mount point> <filesystem type> <mount options> <needs dump> <fsck order> + // + // The "needs dump" and "fsck order" flags are always 0, which + // is allowed. + fmt.Fprintf(buf, "%s %s %s %s %d %d\n", "none", path, mnt.fs.FilesystemType().Name(), opts, 0, 0) + } +} + +// GenerateProcMountInfo emits the contents of /proc/[pid]/mountinfo for vfs to +// buf. +// +// Preconditions: taskRootDir.Ok(). +func (vfs *VirtualFilesystem) GenerateProcMountInfo(ctx context.Context, taskRootDir VirtualDentry, buf *bytes.Buffer) { + vfs.mountMu.Lock() + defer vfs.mountMu.Unlock() + rootMnt := taskRootDir.mount + mounts := rootMnt.submountsLocked() + sort.Slice(mounts, func(i, j int) bool { return mounts[i].ID < mounts[j].ID }) + for _, mnt := range mounts { + // Get the path to this mount relative to task root. + mntRootVD := VirtualDentry{ + mount: mnt, + dentry: mnt.root, + } + path, err := vfs.PathnameReachable(ctx, taskRootDir, mntRootVD) + if err != nil { + // For some reason we didn't get a path. Log a warning + // and run with empty path. + ctx.Warningf("Error getting pathname for mount root %+v: %v", mnt.root, err) + path = "" + } + if path == "" { + // Either an error occurred, or path is not reachable + // from root. + break + } + // Stat the mount root to get the major/minor device numbers. + pop := &PathOperation{ + Root: mntRootVD, + Start: mntRootVD, + } + statx, err := vfs.StatAt(ctx, auth.NewAnonymousCredentials(), pop, &StatOptions{}) + if err != nil { + // Well that's not good. Ignore this mount. + break + } + + // Format: + // 36 35 98:0 /mnt1 /mnt2 rw,noatime master:1 - ext3 /dev/root rw,errors=continue + // (1)(2)(3) (4) (5) (6) (7) (8) (9) (10) (11) + + // (1) Mount ID. + fmt.Fprintf(buf, "%d ", mnt.ID) + + // (2) Parent ID (or this ID if there is no parent). + pID := mnt.ID + if p := mnt.parent(); p != nil { + pID = p.ID + } + fmt.Fprintf(buf, "%d ", pID) + + // (3) Major:Minor device ID. We don't have a superblock, so we + // just use the root inode device number. + fmt.Fprintf(buf, "%d:%d ", statx.DevMajor, statx.DevMinor) + + // (4) Root: the pathname of the directory in the filesystem + // which forms the root of this mount. + // + // NOTE(b/78135857): This will always be "/" until we implement + // bind mounts. + fmt.Fprintf(buf, "/ ") + + // (5) Mount point (relative to process root). + fmt.Fprintf(buf, "%s ", manglePath(path)) + + // (6) Mount options. + opts := "rw" + if mnt.readOnly() { + opts = "ro" + } + if mnt.Flags.NoATime { + opts = ",noatime" + } + if mnt.Flags.NoExec { + opts += ",noexec" + } + fmt.Fprintf(buf, "%s ", opts) + + // (7) Optional fields: zero or more fields of the form "tag[:value]". + // (8) Separator: the end of the optional fields is marked by a single hyphen. + fmt.Fprintf(buf, "- ") + + // (9) Filesystem type. + fmt.Fprintf(buf, "%s ", mnt.fs.FilesystemType().Name()) + + // (10) Mount source: filesystem-specific information or "none". + fmt.Fprintf(buf, "none ") + + // (11) Superblock options, and final newline. + fmt.Fprintf(buf, "%s\n", superBlockOpts(path, mnt)) + } +} + +// manglePath replaces ' ', '\t', '\n', and '\\' with their octal equivalents. +// See Linux fs/seq_file.c:mangle_path. +func manglePath(p string) string { + r := strings.NewReplacer(" ", "\\040", "\t", "\\011", "\n", "\\012", "\\", "\\134") + return r.Replace(p) +} + +// superBlockOpts returns the super block options string for the the mount at +// the given path. +func superBlockOpts(mountPath string, mnt *Mount) string { + // gVisor doesn't (yet) have a concept of super block options, so we + // use the ro/rw bit from the mount flag. + opts := "rw" + if mnt.readOnly() { + opts = "ro" + } + + // NOTE(b/147673608): If the mount is a cgroup, we also need to include + // the cgroup name in the options. For now we just read that from the + // path. + // + // TODO(gvisor.dev/issue/190): Once gVisor has full cgroup support, we + // should get this value from the cgroup itself, and not rely on the + // path. + if mnt.fs.FilesystemType().Name() == "cgroup" { + splitPath := strings.Split(mountPath, "/") + cgroupType := splitPath[len(splitPath)-1] + opts += "," + cgroupType + } + return opts +} diff --git a/pkg/sentry/vfs/mount_test.go b/pkg/sentry/vfs/mount_test.go index 3b933468d..3335e4057 100644 --- a/pkg/sentry/vfs/mount_test.go +++ b/pkg/sentry/vfs/mount_test.go @@ -55,7 +55,7 @@ func TestMountTableInsertLookup(t *testing.T) { } } -// TODO: concurrent lookup/insertion/removal +// TODO(gvisor.dev/issue/1035): concurrent lookup/insertion/removal. // must be powers of 2 var benchNumMounts = []int{1 << 2, 1 << 5, 1 << 8} diff --git a/pkg/sentry/vfs/mount_unsafe.go b/pkg/sentry/vfs/mount_unsafe.go index bd90d36c4..70f850ca4 100644 --- a/pkg/sentry/vfs/mount_unsafe.go +++ b/pkg/sentry/vfs/mount_unsafe.go @@ -13,7 +13,7 @@ // limitations under the License. // +build go1.12 -// +build !go1.15 +// +build !go1.16 // Check go:linkname function signatures when updating Go version. @@ -26,6 +26,7 @@ import ( "sync/atomic" "unsafe" + "gvisor.dev/gvisor/pkg/gohacks" "gvisor.dev/gvisor/pkg/sync" ) @@ -64,6 +65,8 @@ func (mnt *Mount) storeKey(vd VirtualDentry) { // (provided mutation is sufficiently uncommon). // // mountTable.Init() must be called on new mountTables before use. +// +// +stateify savable type mountTable struct { // mountTable is implemented as a seqcount-protected hash table that // resolves collisions with linear probing, featuring Robin Hood insertion @@ -75,8 +78,8 @@ type mountTable struct { // intrinsics and inline assembly, limiting the performance of this // approach.) - seq sync.SeqCount - seed uint32 // for hashing keys + seq sync.SeqCount `state:"nosave"` + seed uint32 // for hashing keys // size holds both length (number of elements) and capacity (number of // slots): capacity is stored as its base-2 log (referred to as order) in @@ -89,7 +92,7 @@ type mountTable struct { // length and cap in separate uint32s) for ~free. size uint64 - slots unsafe.Pointer // []mountSlot; never nil after Init + slots unsafe.Pointer `state:"nosave"` // []mountSlot; never nil after Init } type mountSlot struct { @@ -158,7 +161,7 @@ func newMountTableSlots(cap uintptr) unsafe.Pointer { // Lookup may be called even if there are concurrent mutators of mt. func (mt *mountTable) Lookup(parent *Mount, point *Dentry) *Mount { key := mountKey{parent: unsafe.Pointer(parent), point: unsafe.Pointer(point)} - hash := memhash(noescape(unsafe.Pointer(&key)), uintptr(mt.seed), mountKeyBytes) + hash := memhash(gohacks.Noescape(unsafe.Pointer(&key)), uintptr(mt.seed), mountKeyBytes) loop: for { @@ -359,12 +362,3 @@ func memhash(p unsafe.Pointer, seed, s uintptr) uintptr //go:linkname rand32 runtime.fastrand func rand32() uint32 - -// This is copy/pasted from runtime.noescape(), and is needed because arguments -// apparently escape from all functions defined by linkname. -// -//go:nosplit -func noescape(p unsafe.Pointer) unsafe.Pointer { - x := uintptr(p) - return unsafe.Pointer(x ^ 0) -} diff --git a/pkg/sentry/vfs/options.go b/pkg/sentry/vfs/options.go index fdf8be157..f223aeda8 100644 --- a/pkg/sentry/vfs/options.go +++ b/pkg/sentry/vfs/options.go @@ -16,6 +16,7 @@ package vfs import ( "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" ) // GetDentryOptions contains options to VirtualFilesystem.GetDentryAt() and @@ -32,6 +33,25 @@ type GetDentryOptions struct { type MkdirOptions struct { // Mode is the file mode bits for the created directory. Mode linux.FileMode + + // If ForSyntheticMountpoint is true, FilesystemImpl.MkdirAt() may create + // the given directory in memory only (as opposed to persistent storage). + // The created directory should be able to support the creation of + // subdirectories with ForSyntheticMountpoint == true. It does not need to + // support the creation of subdirectories with ForSyntheticMountpoint == + // false, or files of other types. + // + // FilesystemImpls are permitted to ignore the ForSyntheticMountpoint + // option. + // + // The ForSyntheticMountpoint option exists because, unlike mount(2), the + // OCI Runtime Specification permits the specification of mount points that + // do not exist, under the expectation that container runtimes will create + // them. (More accurately, the OCI Runtime Specification completely fails + // to document this feature, but it's implemented by runc.) + // ForSyntheticMountpoint allows such mount points to be created even when + // the underlying persistent filesystem is immutable. + ForSyntheticMountpoint bool } // MknodOptions contains options to VirtualFilesystem.MknodAt() and @@ -44,10 +64,31 @@ type MknodOptions struct { // DevMinor are the major and minor device numbers for the created device. DevMajor uint32 DevMinor uint32 + + // Endpoint is the endpoint to bind to the created file, if a socket file is + // being created for bind(2) on a Unix domain socket. + Endpoint transport.BoundEndpoint +} + +// MountFlags contains flags as specified for mount(2), e.g. MS_NOEXEC. +// MS_RDONLY is not part of MountFlags because it's tracked in Mount.writers. +type MountFlags struct { + // NoExec is equivalent to MS_NOEXEC. + NoExec bool + + // NoATime is equivalent to MS_NOATIME and indicates that the + // filesystem should not update access time in-place. + NoATime bool } // MountOptions contains options to VirtualFilesystem.MountAt(). type MountOptions struct { + // Flags contains flags as specified for mount(2), e.g. MS_NOEXEC. + Flags MountFlags + + // ReadOnly is equivalent to MS_RDONLY. + ReadOnly bool + // GetFilesystemOptions contains options to FilesystemType.GetFilesystem(). GetFilesystemOptions GetFilesystemOptions @@ -61,7 +102,7 @@ type MountOptions struct { type OpenOptions struct { // Flags contains access mode and flags as specified for open(2). // - // FilesystemImpls is reponsible for implementing the following flags: + // FilesystemImpls are responsible for implementing the following flags: // O_RDONLY, O_WRONLY, O_RDWR, O_APPEND, O_CREAT, O_DIRECT, O_DSYNC, // O_EXCL, O_NOATIME, O_NOCTTY, O_NONBLOCK, O_PATH, O_SYNC, O_TMPFILE, and // O_TRUNC. VFS is responsible for handling O_DIRECTORY, O_LARGEFILE, and @@ -75,7 +116,8 @@ type OpenOptions struct { // FileExec is set when the file is being opened to be executed. // VirtualFilesystem.OpenAt() checks that the caller has execute permissions - // on the file, and that the file is a regular file. + // on the file, that the file is a regular file, and that the mount doesn't + // have MS_NOEXEC set. FileExec bool } @@ -113,6 +155,38 @@ type SetStatOptions struct { Stat linux.Statx } +// BoundEndpointOptions contains options to VirtualFilesystem.BoundEndpointAt() +// and FilesystemImpl.BoundEndpointAt(). +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 + // was created through a bind syscall, so the path can be stored on creation. + // However, if the endpoint was created in FilesystemImpl.BoundEndpointAt(), + // then we may not know what the original bind address was. + // + // For example, if connect(2) is called with address "foo" which corresponds + // a remote named socket in goferfs, we need to generate an endpoint wrapping + // that file. In this case, we can use Addr to set the endpoint address to + // "foo". Note that Addr is only a best-effort attempt--we still do not know + // the exact address that was used on the remote fs to bind the socket (it + // may have been "foo", "./foo", etc.). + Addr string +} + +// GetxattrOptions contains options to VirtualFilesystem.GetxattrAt(), +// FilesystemImpl.GetxattrAt(), FileDescription.Getxattr(), and +// FileDescriptionImpl.Getxattr(). +type GetxattrOptions struct { + // Name is the name of the extended attribute to retrieve. + Name string + + // Size is the maximum value size that the caller will tolerate. If the value + // is larger than size, getxattr methods may return ERANGE, but they are also + // free to ignore the hint entirely (i.e. the value returned may be larger + // than size). All size checking is done independently at the syscall layer. + Size uint64 +} + // SetxattrOptions contains options to VirtualFilesystem.SetxattrAt(), // FilesystemImpl.SetxattrAt(), FileDescription.Setxattr(), and // FileDescriptionImpl.Setxattr(). diff --git a/pkg/sentry/vfs/pathname.go b/pkg/sentry/vfs/pathname.go index b318c681a..cd78d66bc 100644 --- a/pkg/sentry/vfs/pathname.go +++ b/pkg/sentry/vfs/pathname.go @@ -58,7 +58,7 @@ loop: switch err.(type) { case nil: if vd.mount == vfsroot.mount && vd.mount.root == vfsroot.dentry { - // GenericPrependPath() will have returned + // genericfstree.PrependPath() will have returned // PrependPathAtVFSRootError in this case since it checks // against vfsroot before mnt.root, but other implementations // of FilesystemImpl.PrependPath() may return nil instead. @@ -84,16 +84,59 @@ loop: } } b.PrependByte('/') - if origD.IsDisowned() { + if origD.IsDead() { b.AppendString(" (deleted)") } return b.String(), nil } +// PathnameReachable returns an absolute pathname to vd, consistent with +// Linux's __d_path() (as used by seq_path_root()). If vfsroot.Ok() and vd is +// not reachable from vfsroot, such that seq_path_root() would return SEQ_SKIP +// (causing the entire containing entry to be skipped), PathnameReachable +// returns ("", nil). +func (vfs *VirtualFilesystem) PathnameReachable(ctx context.Context, vfsroot, vd VirtualDentry) (string, error) { + b := getFSPathBuilder() + defer putFSPathBuilder(b) + haveRef := false + defer func() { + if haveRef { + vd.DecRef() + } + }() +loop: + for { + err := vd.mount.fs.impl.PrependPath(ctx, vfsroot, vd, b) + switch err.(type) { + case nil: + if vd.mount == vfsroot.mount && vd.mount.root == vfsroot.dentry { + break loop + } + nextVD := vfs.getMountpointAt(vd.mount, vfsroot) + if !nextVD.Ok() { + return "", nil + } + if haveRef { + vd.DecRef() + } + vd = nextVD + haveRef = true + case PrependPathAtVFSRootError: + break loop + case PrependPathAtNonMountRootError, PrependPathSyntheticError: + return "", nil + default: + return "", err + } + } + b.PrependByte('/') + return b.String(), nil +} + // PathnameForGetcwd returns an absolute pathname to vd, consistent with // Linux's sys_getcwd(). func (vfs *VirtualFilesystem) PathnameForGetcwd(ctx context.Context, vfsroot, vd VirtualDentry) (string, error) { - if vd.dentry.IsDisowned() { + if vd.dentry.IsDead() { return "", syserror.ENOENT } diff --git a/pkg/sentry/vfs/permissions.go b/pkg/sentry/vfs/permissions.go index 8e250998a..f9647f90e 100644 --- a/pkg/sentry/vfs/permissions.go +++ b/pkg/sentry/vfs/permissions.go @@ -15,8 +15,12 @@ package vfs import ( + "math" + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/limits" "gvisor.dev/gvisor/pkg/syserror" ) @@ -25,9 +29,9 @@ type AccessTypes uint16 // Bits in AccessTypes. const ( + MayExec AccessTypes = 1 + MayWrite AccessTypes = 2 MayRead AccessTypes = 4 - MayWrite = 2 - MayExec = 1 ) // OnlyRead returns true if access _only_ allows read. @@ -52,16 +56,17 @@ func (a AccessTypes) MayExec() bool { // GenericCheckPermissions checks that creds has the given access rights on a // file with the given permissions, UID, and GID, subject to the rules of -// fs/namei.c:generic_permission(). isDir is true if the file is a directory. -func GenericCheckPermissions(creds *auth.Credentials, ats AccessTypes, isDir bool, mode uint16, kuid auth.KUID, kgid auth.KGID) error { +// fs/namei.c:generic_permission(). +func GenericCheckPermissions(creds *auth.Credentials, ats AccessTypes, mode linux.FileMode, kuid auth.KUID, kgid auth.KGID) error { // Check permission bits. - perms := mode + perms := uint16(mode.Permissions()) if creds.EffectiveKUID == kuid { perms >>= 6 } else if creds.InGroup(kgid) { perms >>= 3 } if uint16(ats)&perms == uint16(ats) { + // All permission bits match, access granted. return nil } @@ -73,7 +78,7 @@ func GenericCheckPermissions(creds *auth.Credentials, ats AccessTypes, isDir boo } // CAP_DAC_READ_SEARCH allows the caller to read and search arbitrary // directories, and read arbitrary non-directory files. - if (isDir && !ats.MayWrite()) || ats.OnlyRead() { + if (mode.IsDir() && !ats.MayWrite()) || ats.OnlyRead() { if creds.HasCapability(linux.CAP_DAC_READ_SEARCH) { return nil } @@ -81,7 +86,7 @@ func GenericCheckPermissions(creds *auth.Credentials, ats AccessTypes, isDir boo // CAP_DAC_OVERRIDE allows arbitrary access to directories, read/write // access to non-directory files, and execute access to non-directory files // for which at least one execute bit is set. - if isDir || !ats.MayExec() || (mode&0111 != 0) { + if mode.IsDir() || !ats.MayExec() || (mode.Permissions()&0111 != 0) { if creds.HasCapability(linux.CAP_DAC_OVERRIDE) { return nil } @@ -147,7 +152,16 @@ func MayWriteFileWithOpenFlags(flags uint32) bool { // CheckSetStat checks that creds has permission to change the metadata of a // file with the given permissions, UID, and GID as specified by stat, subject // to the rules of Linux's fs/attr.c:setattr_prepare(). -func CheckSetStat(creds *auth.Credentials, stat *linux.Statx, mode uint16, kuid auth.KUID, kgid auth.KGID) error { +func CheckSetStat(ctx context.Context, creds *auth.Credentials, stat *linux.Statx, mode linux.FileMode, kuid auth.KUID, kgid auth.KGID) error { + if stat.Mask&linux.STATX_SIZE != 0 { + limit, err := CheckLimit(ctx, 0, int64(stat.Size)) + if err != nil { + return err + } + if limit < int64(stat.Size) { + return syserror.ErrExceedsFileSizeLimit + } + } if stat.Mask&linux.STATX_MODE != 0 { if !CanActAsOwner(creds, kuid) { return syserror.EPERM @@ -177,11 +191,7 @@ func CheckSetStat(creds *auth.Credentials, stat *linux.Statx, mode uint16, kuid (stat.Mask&linux.STATX_CTIME != 0 && stat.Ctime.Nsec != linux.UTIME_NOW) { return syserror.EPERM } - // isDir is irrelevant in the following call to - // GenericCheckPermissions since ats == MayWrite means that - // CAP_DAC_READ_SEARCH does not apply, and CAP_DAC_OVERRIDE - // applies, regardless of isDir. - if err := GenericCheckPermissions(creds, MayWrite, false /* isDir */, mode, kuid, kgid); err != nil { + if err := GenericCheckPermissions(creds, MayWrite, mode, kuid, kgid); err != nil { return err } } @@ -205,3 +215,21 @@ func CanActAsOwner(creds *auth.Credentials, kuid auth.KUID) bool { func HasCapabilityOnFile(creds *auth.Credentials, cp linux.Capability, kuid auth.KUID, kgid auth.KGID) bool { return creds.HasCapability(cp) && creds.UserNamespace.MapFromKUID(kuid).Ok() && creds.UserNamespace.MapFromKGID(kgid).Ok() } + +// CheckLimit enforces file size rlimits. It returns error if the write +// operation must not proceed. Otherwise it returns the max length allowed to +// without violating the limit. +func CheckLimit(ctx context.Context, offset, size int64) (int64, error) { + fileSizeLimit := limits.FromContext(ctx).Get(limits.FileSize).Cur + if fileSizeLimit > math.MaxInt64 { + return size, nil + } + if offset >= int64(fileSizeLimit) { + return 0, syserror.ErrExceedsFileSizeLimit + } + remaining := int64(fileSizeLimit) - offset + if remaining < size { + return remaining, nil + } + return size, nil +} diff --git a/pkg/sentry/vfs/resolving_path.go b/pkg/sentry/vfs/resolving_path.go index 8a0b382f6..9d047ff88 100644 --- a/pkg/sentry/vfs/resolving_path.go +++ b/pkg/sentry/vfs/resolving_path.go @@ -29,7 +29,9 @@ import ( // // From the perspective of FilesystemImpl methods, a ResolvingPath represents a // starting Dentry on the associated Filesystem (on which a reference is -// already held) and a stream of path components relative to that Dentry. +// already held), a stream of path components relative to that Dentry, and +// elements of the invoking Context that are commonly required by +// FilesystemImpl methods. // // ResolvingPath is loosely analogous to Linux's struct nameidata. type ResolvingPath struct { @@ -228,7 +230,7 @@ func (rp *ResolvingPath) Advance() { rp.pit = next } else { // at end of path segment, continue with next one rp.curPart-- - rp.pit = rp.parts[rp.curPart-1] + rp.pit = rp.parts[rp.curPart] } } @@ -251,18 +253,17 @@ func (rp *ResolvingPath) relpathCommit() { rp.origParts[rp.curPart] = rp.pit } -// ResolveParent returns the VFS parent of d. It does not take a reference on -// the returned Dentry. -// -// Preconditions: There are no concurrent mutators of d. -// -// Postconditions: If the returned error is nil, then the returned Dentry is -// not nil. -func (rp *ResolvingPath) ResolveParent(d *Dentry) (*Dentry, error) { - var parent *Dentry +// CheckRoot is called before resolving the parent of the Dentry d. If the +// Dentry is contextually a VFS root, such that path resolution should treat +// d's parent as itself, CheckRoot returns (true, nil). If the Dentry is the +// root of a non-root mount, such that path resolution should switch to another +// Mount, CheckRoot returns (unspecified, non-nil error). Otherwise, path +// resolution should resolve d's parent normally, and CheckRoot returns (false, +// nil). +func (rp *ResolvingPath) CheckRoot(d *Dentry) (bool, error) { if d == rp.root.dentry && rp.mount == rp.root.mount { - // At contextual VFS root. - parent = d + // At contextual VFS root (due to e.g. chroot(2)). + return true, nil } else if d == rp.mount.root { // At mount root ... vd := rp.vfs.getMountpointAt(rp.mount, rp.root) @@ -270,69 +271,49 @@ func (rp *ResolvingPath) ResolveParent(d *Dentry) (*Dentry, error) { // ... of non-root mount. rp.nextMount = vd.mount rp.nextStart = vd.dentry - return nil, resolveMountRootOrJumpError{} + return false, resolveMountRootOrJumpError{} } // ... of root mount. - parent = d - } else if d.parent == nil { - // At filesystem root. - parent = d - } else { - parent = d.parent - } - if parent.isMounted() { - if mnt := rp.vfs.getMountAt(rp.mount, parent); mnt != nil { - rp.nextMount = mnt - return nil, resolveMountPointError{} - } + return true, nil } - return parent, nil + return false, nil } -// ResolveChild returns the VFS child of d with the given name. It does not -// take a reference on the returned Dentry. If no such child exists, -// ResolveChild returns (nil, nil). -// -// Preconditions: There are no concurrent mutators of d. -func (rp *ResolvingPath) ResolveChild(d *Dentry, name string) (*Dentry, error) { - child := d.children[name] - if child == nil { - return nil, nil +// CheckMount is called after resolving the parent or child of another Dentry +// to d. If d is a mount point, such that path resolution should switch to +// another Mount, CheckMount returns a non-nil error. Otherwise, CheckMount +// returns nil. +func (rp *ResolvingPath) CheckMount(d *Dentry) error { + if !d.isMounted() { + return nil } - if child.isMounted() { - if mnt := rp.vfs.getMountAt(rp.mount, child); mnt != nil { - rp.nextMount = mnt - return nil, resolveMountPointError{} - } - } - return child, nil -} - -// ResolveComponent returns the Dentry reached by starting at d and resolving -// the current path component in the stream represented by rp. It does not -// advance the stream. It does not take a reference on the returned Dentry. If -// no such Dentry exists, ResolveComponent returns (nil, nil). -// -// Preconditions: !rp.Done(). There are no concurrent mutators of d. -func (rp *ResolvingPath) ResolveComponent(d *Dentry) (*Dentry, error) { - switch pc := rp.Component(); pc { - case ".": - return d, nil - case "..": - return rp.ResolveParent(d) - default: - return rp.ResolveChild(d, pc) + if mnt := rp.vfs.getMountAt(rp.mount, d); mnt != nil { + rp.nextMount = mnt + return resolveMountPointError{} } + return nil } // ShouldFollowSymlink returns true if, supposing that the current path // component in pcs represents a symbolic link, the symbolic link should be // followed. // +// If path is terminated with '/', the '/' is considered the last element and +// any symlink before that is followed: +// - For most non-creating walks, the last path component is handled by +// fs/namei.c:lookup_last(), which sets LOOKUP_FOLLOW if the first byte +// after the path component is non-NULL (which is only possible if it's '/') +// and the path component is of type LAST_NORM. +// +// - For open/openat/openat2 without O_CREAT, the last path component is +// handled by fs/namei.c:do_last(), which does the same, though without the +// LAST_NORM check. +// // Preconditions: !rp.Done(). func (rp *ResolvingPath) ShouldFollowSymlink() bool { - // Non-final symlinks are always followed. - return rp.flags&rpflagsFollowFinalSymlink != 0 || !rp.Final() + // Non-final symlinks are always followed. Paths terminated with '/' are also + // always followed. + return rp.flags&rpflagsFollowFinalSymlink != 0 || !rp.Final() || rp.MustBeDir() } // HandleSymlink is called when the current path component is a symbolic link diff --git a/pkg/sentry/vfs/vfs.go b/pkg/sentry/vfs/vfs.go index 9629afee9..9acca8bc7 100644 --- a/pkg/sentry/vfs/vfs.go +++ b/pkg/sentry/vfs/vfs.go @@ -38,6 +38,7 @@ import ( "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/fspath" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/syserror" ) @@ -46,11 +47,13 @@ import ( // // There is no analogue to the VirtualFilesystem type in Linux, as the // equivalent state in Linux is global. +// +// +stateify savable type VirtualFilesystem struct { // mountMu serializes mount mutations. // // mountMu is analogous to Linux's namespace_sem. - mountMu sync.Mutex + mountMu sync.Mutex `state:"nosave"` // mounts maps (mount parent, mount point) pairs to mounts. (Since mounts // are uniquely namespaced, including mount parent in the key correctly @@ -79,6 +82,10 @@ type VirtualFilesystem struct { // mountpoints is analogous to Linux's mountpoint_hashtable. mountpoints map[*Dentry]map[*Mount]struct{} + // lastMountID is the last allocated mount ID. lastMountID is accessed + // using atomic memory operations. + lastMountID uint64 + // anonMount is a Mount, not included in mounts or mountpoints, // representing an anonFilesystem. anonMount is used to back // VirtualDentries returned by VirtualFilesystem.NewAnonVirtualDentry(). @@ -89,56 +96,59 @@ type VirtualFilesystem struct { // devices contains all registered Devices. devices is protected by // devicesMu. - devicesMu sync.RWMutex + devicesMu sync.RWMutex `state:"nosave"` devices map[devTuple]*registeredDevice // anonBlockDevMinor contains all allocated anonymous block device minor // numbers. anonBlockDevMinorNext is a lower bound for the smallest // unallocated anonymous block device number. anonBlockDevMinorNext and // anonBlockDevMinor are protected by anonBlockDevMinorMu. - anonBlockDevMinorMu sync.Mutex + anonBlockDevMinorMu sync.Mutex `state:"nosave"` anonBlockDevMinorNext uint32 anonBlockDevMinor map[uint32]struct{} // fsTypes contains all registered FilesystemTypes. fsTypes is protected by // fsTypesMu. - fsTypesMu sync.RWMutex + fsTypesMu sync.RWMutex `state:"nosave"` fsTypes map[string]*registeredFilesystemType // filesystems contains all Filesystems. filesystems is protected by // filesystemsMu. - filesystemsMu sync.Mutex + filesystemsMu sync.Mutex `state:"nosave"` filesystems map[*Filesystem]struct{} } -// New returns a new VirtualFilesystem with no mounts or FilesystemTypes. -func New() *VirtualFilesystem { - vfs := &VirtualFilesystem{ - mountpoints: make(map[*Dentry]map[*Mount]struct{}), - devices: make(map[devTuple]*registeredDevice), - anonBlockDevMinorNext: 1, - anonBlockDevMinor: make(map[uint32]struct{}), - fsTypes: make(map[string]*registeredFilesystemType), - filesystems: make(map[*Filesystem]struct{}), - } +// Init initializes a new VirtualFilesystem with no mounts or FilesystemTypes. +func (vfs *VirtualFilesystem) Init() error { + vfs.mountpoints = make(map[*Dentry]map[*Mount]struct{}) + vfs.devices = make(map[devTuple]*registeredDevice) + vfs.anonBlockDevMinorNext = 1 + vfs.anonBlockDevMinor = make(map[uint32]struct{}) + vfs.fsTypes = make(map[string]*registeredFilesystemType) + vfs.filesystems = make(map[*Filesystem]struct{}) vfs.mounts.Init() // Construct vfs.anonMount. anonfsDevMinor, err := vfs.GetAnonBlockDevMinor() if err != nil { - panic(fmt.Sprintf("VirtualFilesystem.GetAnonBlockDevMinor() failed during VirtualFilesystem construction: %v", err)) + // This shouldn't be possible since anonBlockDevMinorNext was + // initialized to 1 above (no device numbers have been allocated yet). + panic(fmt.Sprintf("VirtualFilesystem.Init: device number allocation for anonfs failed: %v", err)) } anonfs := anonFilesystem{ devMinor: anonfsDevMinor, } - anonfs.vfsfs.Init(vfs, &anonfs) - vfs.anonMount = &Mount{ - vfs: vfs, - fs: &anonfs.vfsfs, - refs: 1, + anonfs.vfsfs.Init(vfs, &anonFilesystemType{}, &anonfs) + defer anonfs.vfsfs.DecRef() + anonMount, err := vfs.NewDisconnectedMount(&anonfs.vfsfs, nil, &MountOptions{}) + if err != nil { + // We should not be passing any MountOptions that would cause + // construction of this mount to fail. + panic(fmt.Sprintf("VirtualFilesystem.Init: anonfs mount failed: %v", err)) } + vfs.anonMount = anonMount - return vfs + return nil } // PathOperation specifies the path operated on by a VFS method. @@ -169,6 +179,23 @@ type PathOperation struct { FollowFinalSymlink bool } +// AccessAt checks whether a user with creds has access to the file at +// the given path. +func (vfs *VirtualFilesystem) AccessAt(ctx context.Context, creds *auth.Credentials, ats AccessTypes, pop *PathOperation) error { + rp := vfs.getResolvingPath(creds, pop) + for { + err := rp.mount.fs.impl.AccessAt(ctx, rp, creds, ats) + if err == nil { + vfs.putResolvingPath(rp) + return nil + } + if !rp.handleError(err) { + vfs.putResolvingPath(rp) + return err + } + } +} + // GetDentryAt returns a VirtualDentry representing the given path, at which a // file must exist. A reference is taken on the returned VirtualDentry. func (vfs *VirtualFilesystem) GetDentryAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, opts *GetDentryOptions) (VirtualDentry, error) { @@ -208,7 +235,7 @@ func (vfs *VirtualFilesystem) getParentDirAndName(ctx context.Context, creds *au } if checkInvariants { if rp.canHandleError(err) && rp.Done() { - panic(fmt.Sprintf("%T.GetParentDentryAt() consumed all path components and returned %T", rp.mount.fs.impl, err)) + panic(fmt.Sprintf("%T.GetParentDentryAt() consumed all path components and returned %v", rp.mount.fs.impl, err)) } } if !rp.handleError(err) { @@ -249,7 +276,7 @@ func (vfs *VirtualFilesystem) LinkAt(ctx context.Context, creds *auth.Credential } if checkInvariants { if rp.canHandleError(err) && rp.Done() { - panic(fmt.Sprintf("%T.LinkAt() consumed all path components and returned %T", rp.mount.fs.impl, err)) + panic(fmt.Sprintf("%T.LinkAt() consumed all path components and returned %v", rp.mount.fs.impl, err)) } } if !rp.handleError(err) { @@ -285,7 +312,7 @@ func (vfs *VirtualFilesystem) MkdirAt(ctx context.Context, creds *auth.Credentia } if checkInvariants { if rp.canHandleError(err) && rp.Done() { - panic(fmt.Sprintf("%T.MkdirAt() consumed all path components and returned %T", rp.mount.fs.impl, err)) + panic(fmt.Sprintf("%T.MkdirAt() consumed all path components and returned %v", rp.mount.fs.impl, err)) } } if !rp.handleError(err) { @@ -312,13 +339,13 @@ func (vfs *VirtualFilesystem) MknodAt(ctx context.Context, creds *auth.Credentia rp := vfs.getResolvingPath(creds, pop) for { err := rp.mount.fs.impl.MknodAt(ctx, rp, *opts) - if err != nil { + if err == nil { vfs.putResolvingPath(rp) return nil } if checkInvariants { if rp.canHandleError(err) && rp.Done() { - panic(fmt.Sprintf("%T.MknodAt() consumed all path components and returned %T", rp.mount.fs.impl, err)) + panic(fmt.Sprintf("%T.MknodAt() consumed all path components and returned %v", rp.mount.fs.impl, err)) } } if !rp.handleError(err) { @@ -333,14 +360,11 @@ func (vfs *VirtualFilesystem) MknodAt(ctx context.Context, creds *auth.Credentia func (vfs *VirtualFilesystem) OpenAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, opts *OpenOptions) (*FileDescription, error) { // Remove: // - // - O_LARGEFILE, which we always report in FileDescription status flags - // since only 64-bit architectures are supported at this time. - // // - O_CLOEXEC, which affects file descriptors and therefore must be // handled outside of VFS. // // - Unknown flags. - opts.Flags &= linux.O_ACCMODE | linux.O_CREAT | linux.O_EXCL | linux.O_NOCTTY | linux.O_TRUNC | linux.O_APPEND | linux.O_NONBLOCK | linux.O_DSYNC | linux.O_ASYNC | linux.O_DIRECT | linux.O_DIRECTORY | linux.O_NOFOLLOW | linux.O_NOATIME | linux.O_SYNC | linux.O_PATH | linux.O_TMPFILE + opts.Flags &= linux.O_ACCMODE | linux.O_CREAT | linux.O_EXCL | linux.O_NOCTTY | linux.O_TRUNC | linux.O_APPEND | linux.O_NONBLOCK | linux.O_DSYNC | linux.O_ASYNC | linux.O_DIRECT | linux.O_LARGEFILE | linux.O_DIRECTORY | linux.O_NOFOLLOW | linux.O_NOATIME | linux.O_SYNC | linux.O_PATH | linux.O_TMPFILE // Linux's __O_SYNC (which we call linux.O_SYNC) implies O_DSYNC. if opts.Flags&linux.O_SYNC != 0 { opts.Flags |= linux.O_DSYNC @@ -380,24 +404,25 @@ func (vfs *VirtualFilesystem) OpenAt(ctx context.Context, creds *auth.Credential if err == nil { vfs.putResolvingPath(rp) - // TODO(gvisor.dev/issue/1193): Move inside fsimpl to avoid another call - // to FileDescription.Stat(). if opts.FileExec { + if fd.Mount().Flags.NoExec { + fd.DecRef() + return nil, syserror.EACCES + } + // Only a regular file can be executed. stat, err := fd.Stat(ctx, StatOptions{Mask: linux.STATX_TYPE}) if err != nil { + fd.DecRef() return nil, err } - if stat.Mask&linux.STATX_TYPE != 0 { - // This shouldn't happen, but if type can't be retrieved, file can't - // be executed. - return nil, syserror.EACCES - } - if linux.FileMode(stat.Mode).FileType() != linux.ModeRegular { + if stat.Mask&linux.STATX_TYPE == 0 || stat.Mode&linux.S_IFMT != linux.S_IFREG { + fd.DecRef() return nil, syserror.EACCES } } + fd.Dentry().InotifyWithParent(linux.IN_OPEN, 0, PathEvent) return fd, nil } if !rp.handleError(err) { @@ -472,7 +497,7 @@ func (vfs *VirtualFilesystem) RenameAt(ctx context.Context, creds *auth.Credenti } if checkInvariants { if rp.canHandleError(err) && rp.Done() { - panic(fmt.Sprintf("%T.RenameAt() consumed all path components and returned %T", rp.mount.fs.impl, err)) + panic(fmt.Sprintf("%T.RenameAt() consumed all path components and returned %v", rp.mount.fs.impl, err)) } } if !rp.handleError(err) { @@ -505,7 +530,7 @@ func (vfs *VirtualFilesystem) RmdirAt(ctx context.Context, creds *auth.Credentia } if checkInvariants { if rp.canHandleError(err) && rp.Done() { - panic(fmt.Sprintf("%T.RmdirAt() consumed all path components and returned %T", rp.mount.fs.impl, err)) + panic(fmt.Sprintf("%T.RmdirAt() consumed all path components and returned %v", rp.mount.fs.impl, err)) } } if !rp.handleError(err) { @@ -586,7 +611,7 @@ func (vfs *VirtualFilesystem) SymlinkAt(ctx context.Context, creds *auth.Credent } if checkInvariants { if rp.canHandleError(err) && rp.Done() { - panic(fmt.Sprintf("%T.SymlinkAt() consumed all path components and returned %T", rp.mount.fs.impl, err)) + panic(fmt.Sprintf("%T.SymlinkAt() consumed all path components and returned %v", rp.mount.fs.impl, err)) } } if !rp.handleError(err) { @@ -618,7 +643,7 @@ func (vfs *VirtualFilesystem) UnlinkAt(ctx context.Context, creds *auth.Credenti } if checkInvariants { if rp.canHandleError(err) && rp.Done() { - panic(fmt.Sprintf("%T.UnlinkAt() consumed all path components and returned %T", rp.mount.fs.impl, err)) + panic(fmt.Sprintf("%T.UnlinkAt() consumed all path components and returned %v", rp.mount.fs.impl, err)) } } if !rp.handleError(err) { @@ -628,12 +653,39 @@ func (vfs *VirtualFilesystem) UnlinkAt(ctx context.Context, creds *auth.Credenti } } +// BoundEndpointAt gets the bound endpoint at the given path, if one exists. +func (vfs *VirtualFilesystem) BoundEndpointAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, opts *BoundEndpointOptions) (transport.BoundEndpoint, error) { + if !pop.Path.Begin.Ok() { + if pop.Path.Absolute { + return nil, syserror.ECONNREFUSED + } + return nil, syserror.ENOENT + } + rp := vfs.getResolvingPath(creds, pop) + for { + bep, err := rp.mount.fs.impl.BoundEndpointAt(ctx, rp, *opts) + if err == nil { + vfs.putResolvingPath(rp) + return bep, nil + } + if checkInvariants { + if rp.canHandleError(err) && rp.Done() { + panic(fmt.Sprintf("%T.BoundEndpointAt() consumed all path components and returned %v", rp.mount.fs.impl, err)) + } + } + if !rp.handleError(err) { + vfs.putResolvingPath(rp) + return nil, err + } + } +} + // ListxattrAt returns all extended attribute names for the file at the given // path. -func (vfs *VirtualFilesystem) ListxattrAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation) ([]string, error) { +func (vfs *VirtualFilesystem) ListxattrAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, size uint64) ([]string, error) { rp := vfs.getResolvingPath(creds, pop) for { - names, err := rp.mount.fs.impl.ListxattrAt(ctx, rp) + names, err := rp.mount.fs.impl.ListxattrAt(ctx, rp, size) if err == nil { vfs.putResolvingPath(rp) return names, nil @@ -655,10 +707,10 @@ func (vfs *VirtualFilesystem) ListxattrAt(ctx context.Context, creds *auth.Crede // GetxattrAt returns the value associated with the given extended attribute // for the file at the given path. -func (vfs *VirtualFilesystem) GetxattrAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, name string) (string, error) { +func (vfs *VirtualFilesystem) GetxattrAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, opts *GetxattrOptions) (string, error) { rp := vfs.getResolvingPath(creds, pop) for { - val, err := rp.mount.fs.impl.GetxattrAt(ctx, rp, name) + val, err := rp.mount.fs.impl.GetxattrAt(ctx, rp, *opts) if err == nil { vfs.putResolvingPath(rp) return val, nil @@ -743,11 +795,21 @@ func (vfs *VirtualFilesystem) SyncAllFilesystems(ctx context.Context) error { // VirtualDentry methods require that a reference is held on the VirtualDentry. // // VirtualDentry is analogous to Linux's struct path. +// +// +stateify savable type VirtualDentry struct { mount *Mount dentry *Dentry } +// MakeVirtualDentry creates a VirtualDentry. +func MakeVirtualDentry(mount *Mount, dentry *Dentry) VirtualDentry { + return VirtualDentry{ + mount: mount, + dentry: dentry, + } +} + // Ok returns true if vd is not empty. It does not require that a reference is // held. func (vd VirtualDentry) Ok() bool { diff --git a/pkg/sentry/watchdog/watchdog.go b/pkg/sentry/watchdog/watchdog.go index bfb2fac26..e2894f9f5 100644 --- a/pkg/sentry/watchdog/watchdog.go +++ b/pkg/sentry/watchdog/watchdog.go @@ -77,7 +77,10 @@ var DefaultOpts = Opts{ // trigger it. const descheduleThreshold = 1 * time.Second -var stuckTasks = metric.MustCreateNewUint64Metric("/watchdog/stuck_tasks_detected", true /* sync */, "Cumulative count of stuck tasks detected") +var ( + stuckStartup = metric.MustCreateNewUint64Metric("/watchdog/stuck_startup_detected", true /* sync */, "Incremented once on startup watchdog timeout") + stuckTasks = metric.MustCreateNewUint64Metric("/watchdog/stuck_tasks_detected", true /* sync */, "Cumulative count of stuck tasks detected") +) // Amount of time to wait before dumping the stack to the log again when the same task(s) remains stuck. var stackDumpSameTaskPeriod = time.Minute @@ -220,8 +223,11 @@ func (w *Watchdog) waitForStart() { // We are fine. return } + + stuckStartup.Increment() + var buf bytes.Buffer - buf.WriteString("Watchdog.Start() not called within %s:\n") + buf.WriteString(fmt.Sprintf("Watchdog.Start() not called within %s", w.StartupTimeout)) w.doAction(w.StartupTimeoutAction, false, &buf) } @@ -255,7 +261,7 @@ func (w *Watchdog) runTurn() { case <-done: case <-time.After(w.TaskTimeout): // Report if the watchdog is not making progress. - // No one is wathching the watchdog watcher though. + // No one is watching the watchdog watcher though. w.reportStuckWatchdog() <-done } @@ -317,28 +323,28 @@ func (w *Watchdog) report(offenders map[*kernel.Task]*offender, newTaskFound boo buf.WriteString("Search for '(*Task).run(0x..., 0x<tid>)' in the stack dump to find the offending goroutine") - // Dump stack only if a new task is detected or if it sometime has - // passed since the last time a stack dump was generated. - skipStack := newTaskFound || time.Since(w.lastStackDump) >= stackDumpSameTaskPeriod - w.doAction(w.TaskTimeoutAction, skipStack, &buf) + // Force stack dump only if a new task is detected. + w.doAction(w.TaskTimeoutAction, newTaskFound, &buf) } func (w *Watchdog) reportStuckWatchdog() { var buf bytes.Buffer - buf.WriteString("Watchdog goroutine is stuck:\n") + buf.WriteString("Watchdog goroutine is stuck:") w.doAction(w.TaskTimeoutAction, false, &buf) } -// doAction will take the given action. If the action is LogWarnind and -// skipStack is true, then the stack printing will be skipped. -func (w *Watchdog) doAction(action Action, skipStack bool, msg *bytes.Buffer) { +// doAction will take the given action. If the action is LogWarning, the stack +// is not always dumped to the log to prevent log flooding. "forceStack" +// guarantees that the stack will be dumped regardless. +func (w *Watchdog) doAction(action Action, forceStack bool, msg *bytes.Buffer) { switch action { case LogWarning: - if skipStack { + // Dump stack only if forced or sometime has passed since the last time a + // stack dump was generated. + if !forceStack && time.Since(w.lastStackDump) < stackDumpSameTaskPeriod { msg.WriteString("\n...[stack dump skipped]...") log.Warningf(msg.String()) return - } log.TracebackAll(msg.String()) w.lastStackDump = time.Now() @@ -359,7 +365,8 @@ func (w *Watchdog) doAction(action Action, skipStack bool, msg *bytes.Buffer) { case <-metricsEmitted: case <-time.After(1 * time.Second): } - panic(fmt.Sprintf("Stack for running G's are skipped while panicking.\n%s", msg.String())) + panic(fmt.Sprintf("%s\nStack for running G's are skipped while panicking.", msg.String())) + default: panic(fmt.Sprintf("Unknown watchdog action %v", action)) |