diff options
Diffstat (limited to 'pkg/sentry')
141 files changed, 7009 insertions, 1081 deletions
diff --git a/pkg/sentry/arch/arch_aarch64.go b/pkg/sentry/arch/arch_aarch64.go index 3b6987665..b998f84fc 100644 --- a/pkg/sentry/arch/arch_aarch64.go +++ b/pkg/sentry/arch/arch_aarch64.go @@ -32,29 +32,38 @@ import ( const ( // SyscallWidth is the width of insturctions. SyscallWidth = 4 + + // fpsimdMagic is the magic number which is used in fpsimd_context. + fpsimdMagic = 0x46508001 + + // fpsimdContextSize is the size of fpsimd_context. + fpsimdContextSize = 0x210 ) +// ARMTrapFlag is the mask for the trap flag. +const ARMTrapFlag = uint64(1) << 21 + // aarch64FPState is aarch64 floating point state. type aarch64FPState []byte -// initAarch64FPState (defined in asm files) sets up initial state. -func initAarch64FPState(data *FloatingPointData) { - // TODO(gvisor.dev/issue/1238): floating-point is not supported. +// initAarch64FPState sets up initial state. +func initAarch64FPState(data aarch64FPState) { + binary.LittleEndian.PutUint32(data, fpsimdMagic) + binary.LittleEndian.PutUint32(data[4:], fpsimdContextSize) } func newAarch64FPStateSlice() []byte { - return alignedBytes(4096, 32)[:4096] + return alignedBytes(4096, 16)[:fpsimdContextSize] } // newAarch64FPState returns an initialized floating point state. // // The returned state is large enough to store all floating point state // supported by host, even if the app won't use much of it due to a restricted -// FeatureSet. Since they may still be able to see state not advertised by -// CPUID we must ensure it does not contain any sentry state. +// FeatureSet. func newAarch64FPState() aarch64FPState { f := aarch64FPState(newAarch64FPStateSlice()) - initAarch64FPState(f.FloatingPointData()) + initAarch64FPState(f) return f } @@ -86,6 +95,9 @@ type State struct { // Our floating point state. aarch64FPState `state:"wait"` + // TLS pointer + TPValue uint64 + // FeatureSet is a pointer to the currently active feature set. FeatureSet *cpuid.FeatureSet } @@ -133,10 +145,11 @@ 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, } } @@ -246,6 +259,7 @@ func (s *State) PtraceSetFPRegs(src io.Reader) (int, error) { const ( _NT_PRSTATUS = 1 _NT_PRFPREG = 2 + _NT_ARM_TLS = 0x401 ) // PtraceGetRegSet implements Context.PtraceGetRegSet. @@ -285,8 +299,10 @@ func New(arch Arch, fs *cpuid.FeatureSet) Context { case ARM64: return &context64{ State{ - FeatureSet: fs, + aarch64FPState: newAarch64FPState(), + FeatureSet: fs, }, + []aarch64FPState(nil), } } panic(fmt.Sprintf("unknown architecture %v", arch)) diff --git a/pkg/sentry/arch/arch_arm64.go b/pkg/sentry/arch/arch_arm64.go index ac98897b5..db99c5acb 100644 --- a/pkg/sentry/arch/arch_arm64.go +++ b/pkg/sentry/arch/arch_arm64.go @@ -53,6 +53,11 @@ const ( preferredPIELoadAddr usermem.Addr = maxAddr64 / 6 * 5 ) +var ( + // CPUIDInstruction doesn't exist on ARM64. + CPUIDInstruction = []byte{} +) + // These constants are selected as heuristics to help make the Platform's // potentially limited address space conform as closely to Linux as possible. const ( @@ -68,6 +73,7 @@ const ( // context64 represents an ARM64 context. type context64 struct { State + sigFPState []aarch64FPState // fpstate to be restored on sigreturn. } // Arch implements Context.Arch. @@ -75,10 +81,19 @@ func (c *context64) Arch() Arch { return ARM64 } +func (c *context64) copySigFPState() []aarch64FPState { + var sigfps []aarch64FPState + for _, s := range c.sigFPState { + sigfps = append(sigfps, s.fork()) + } + return sigfps +} + // Fork returns an exact copy of this context. func (c *context64) Fork() Context { return &context64{ - State: c.State.Fork(), + State: c.State.Fork(), + sigFPState: c.copySigFPState(), } } @@ -125,16 +140,17 @@ func (c *context64) SetStack(value uintptr) { // TLS returns the current TLS pointer. func (c *context64) TLS() uintptr { - // TODO(gvisor.dev/issue/1238): TLS is not supported. - // MRS_TPIDR_EL0 - return 0 + return uintptr(c.TPValue) } // SetTLS sets the current TLS pointer. Returns false if value is invalid. func (c *context64) SetTLS(value uintptr) bool { - // TODO(gvisor.dev/issue/1238): TLS is not supported. - // MSR_TPIDR_EL0 - return false + if value >= uintptr(maxAddr64) { + return false + } + + c.TPValue = uint64(value) + return true } // SetOldRSeqInterruptedIP implements Context.SetOldRSeqInterruptedIP. diff --git a/pkg/sentry/arch/signal_arm64.go b/pkg/sentry/arch/signal_arm64.go index 4f4cc46a8..0c1db4b13 100644 --- a/pkg/sentry/arch/signal_arm64.go +++ b/pkg/sentry/arch/signal_arm64.go @@ -30,14 +30,29 @@ type SignalContext64 struct { Sp uint64 Pc uint64 Pstate uint64 - _pad [8]byte // __attribute__((__aligned__(16))) - Reserved [4096]uint8 + _pad [8]byte // __attribute__((__aligned__(16))) + Fpsimd64 FpsimdContext // size = 528 + Reserved [3568]uint8 +} + +type aarch64Ctx struct { + Magic uint32 + Size uint32 +} + +// FpsimdContext is equivalent to struct fpsimd_context on arm64 +// (arch/arm64/include/uapi/asm/sigcontext.h). +type FpsimdContext struct { + Head aarch64Ctx + Fpsr uint32 + Fpcr uint32 + Vregs [64]uint64 // actually [32]uint128 } // UContext64 is equivalent to ucontext on arm64(arch/arm64/include/uapi/asm/ucontext.h). type UContext64 struct { Flags uint64 - Link *UContext64 + Link uint64 Stack SignalStack Sigset linux.SignalSet // glibc uses a 1024-bit sigset_t diff --git a/pkg/sentry/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/fs/dev/BUILD b/pkg/sentry/fs/dev/BUILD index 9b6bb26d0..9379a4d7b 100644 --- a/pkg/sentry/fs/dev/BUILD +++ b/pkg/sentry/fs/dev/BUILD @@ -26,6 +26,7 @@ go_library( "//pkg/sentry/fs/fsutil", "//pkg/sentry/fs/ramfs", "//pkg/sentry/fs/tmpfs", + "//pkg/sentry/inet", "//pkg/sentry/kernel", "//pkg/sentry/memmap", "//pkg/sentry/mm", diff --git a/pkg/sentry/fs/dev/dev.go b/pkg/sentry/fs/dev/dev.go index 7e66c29b0..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" ) @@ -124,10 +125,12 @@ func New(ctx context.Context, msrc *fs.MountSource) *fs.Inode { "ptmx": newSymlink(ctx, "pts/ptmx", msrc), "tty": newCharacterDevice(ctx, newTTYDevice(ctx, fs.RootOwner, 0666), msrc, ttyDevMajor, ttyDevMinor), + } - "net": newDirectory(ctx, map[string]*fs.Inode{ + 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), + }, msrc) } iops := ramfs.NewDir(ctx, contents, fs.RootOwner, fs.FilePermsFromMode(0555)) diff --git a/pkg/sentry/fs/dev/net_tun.go b/pkg/sentry/fs/dev/net_tun.go index 755644488..dc7ad075a 100644 --- a/pkg/sentry/fs/dev/net_tun.go +++ b/pkg/sentry/fs/dev/net_tun.go @@ -20,6 +20,7 @@ import ( "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" @@ -168,3 +169,9 @@ func (fops *netTunFileOperations) EventRegister(e *waiter.Entry, mask waiter.Eve 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..e0b32e1c1 100644 --- a/pkg/sentry/fs/dirent.go +++ b/pkg/sentry/fs/dirent.go @@ -1438,8 +1438,8 @@ 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 +1465,33 @@ 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 } @@ -1542,7 +1545,7 @@ 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 +1583,7 @@ 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/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/host/control.go b/pkg/sentry/fs/host/control.go index 1658979fc..cd84e1337 100644 --- a/pkg/sentry/fs/host/control.go +++ b/pkg/sentry/fs/host/control.go @@ -32,6 +32,8 @@ func newSCMRights(fds []int) control.SCMRights { } // Files implements control.SCMRights.Files. +// +// TODO(gvisor.dev/issue/2017): Port to VFS2. func (c *scmRights) Files(ctx context.Context, max int) (control.RightsFiles, bool) { n := max var trunc bool diff --git a/pkg/sentry/fs/host/util.go b/pkg/sentry/fs/host/util.go index e37e687c6..7c60dc1db 100644 --- a/pkg/sentry/fs/host/util.go +++ b/pkg/sentry/fs/host/util.go @@ -24,7 +24,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/device" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" - ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" + "gvisor.dev/gvisor/pkg/sentry/kernel/time" "gvisor.dev/gvisor/pkg/syserror" ) @@ -152,9 +152,9 @@ func unstableAttr(mo *superOperations, s *syscall.Stat_t) fs.UnstableAttr { Usage: s.Blocks * 512, Perms: fs.FilePermsFromMode(linux.FileMode(s.Mode)), Owner: owner(mo, 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), + AccessTime: time.FromUnix(s.Atim.Sec, s.Atim.Nsec), + ModificationTime: time.FromUnix(s.Mtim.Sec, s.Mtim.Nsec), + StatusChangeTime: time.FromUnix(s.Ctim.Sec, s.Ctim.Nsec), Links: uint64(s.Nlink), } } @@ -165,6 +165,8 @@ 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 +179,8 @@ 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/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..c7981f66e 100644 --- a/pkg/sentry/fs/mounts.go +++ b/pkg/sentry/fs/mounts.go @@ -100,10 +100,14 @@ func newUndoMount(d *Dirent) *Mount { } } -// Root returns the root dirent of this mount. Callers must call DecRef on the -// returned dirent. +// Root returns the root dirent of this mount. +// +// This may return nil if the mount has already been free. Callers must handle this +// case appropriately. If non-nil, callers must call DecRef on the returned *Dirent. func (m *Mount) Root() *Dirent { - m.root.IncRef() + if !m.root.TryIncRef() { + return nil + } return m.root } diff --git a/pkg/sentry/fs/proc/mounts.go b/pkg/sentry/fs/proc/mounts.go index c10888100..94deb553b 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) @@ -207,6 +212,9 @@ 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 95d5817ff..bd18177d4 100644 --- a/pkg/sentry/fs/proc/net.go +++ b/pkg/sentry/fs/proc/net.go @@ -40,47 +40,48 @@ import ( // LINT.IfChange -// newNet creates a new proc net entry. -func (p *proc) newNetDir(ctx context.Context, k *kernel.Kernel, msrc *fs.MountSource) *fs.Inode { +// newNetDir creates a new proc net entry. +func newNetDir(t *kernel.Task, msrc *fs.MountSource) *fs.Inode { + k := t.Kernel() + var contents map[string]*fs.Inode - // TODO(gvisor.dev/issue/1833): Support for using the network stack in the - // network namespace of the calling process. We should make this per-process, - // a.k.a. /proc/PID/net, and make /proc/net a symlink to /proc/self/net. - if s := p.k.RootNetworkNamespace().Stack(); s != nil { + if s := t.NetworkNamespace().Stack(); s != nil { + // TODO(gvisor.dev/issue/1833): Make sure file contents reflect the task + // network namespace. contents = map[string]*fs.Inode{ - "dev": seqfile.NewSeqFileInode(ctx, &netDev{s: s}, msrc), - "snmp": seqfile.NewSeqFileInode(ctx, &netSnmp{s: s}, msrc), + "dev": seqfile.NewSeqFileInode(t, &netDev{s: s}, msrc), + "snmp": seqfile.NewSeqFileInode(t, &netSnmp{s: s}, msrc), // The following files are simple stubs until they are // implemented in netstack, if the file contains a // header the stub is just the header otherwise it is // an empty file. - "arp": newStaticProcInode(ctx, msrc, []byte("IP address HW type Flags HW address Mask Device\n")), + "arp": newStaticProcInode(t, msrc, []byte("IP address HW type Flags HW address Mask Device\n")), - "netlink": newStaticProcInode(ctx, msrc, []byte("sk Eth Pid Groups Rmem Wmem Dump Locks Drops Inode\n")), - "netstat": newStaticProcInode(ctx, msrc, []byte("TcpExt: SyncookiesSent SyncookiesRecv SyncookiesFailed EmbryonicRsts PruneCalled RcvPruned OfoPruned OutOfWindowIcmps LockDroppedIcmps ArpFilter TW TWRecycled TWKilled PAWSPassive PAWSActive PAWSEstab DelayedACKs DelayedACKLocked DelayedACKLost ListenOverflows ListenDrops TCPPrequeued TCPDirectCopyFromBacklog TCPDirectCopyFromPrequeue TCPPrequeueDropped TCPHPHits TCPHPHitsToUser TCPPureAcks TCPHPAcks TCPRenoRecovery TCPSackRecovery TCPSACKReneging TCPFACKReorder TCPSACKReorder TCPRenoReorder TCPTSReorder TCPFullUndo TCPPartialUndo TCPDSACKUndo TCPLossUndo TCPLostRetransmit TCPRenoFailures TCPSackFailures TCPLossFailures TCPFastRetrans TCPForwardRetrans TCPSlowStartRetrans TCPTimeouts TCPLossProbes TCPLossProbeRecovery TCPRenoRecoveryFail TCPSackRecoveryFail TCPSchedulerFailed TCPRcvCollapsed TCPDSACKOldSent TCPDSACKOfoSent TCPDSACKRecv TCPDSACKOfoRecv TCPAbortOnData TCPAbortOnClose TCPAbortOnMemory TCPAbortOnTimeout TCPAbortOnLinger TCPAbortFailed TCPMemoryPressures TCPSACKDiscard TCPDSACKIgnoredOld TCPDSACKIgnoredNoUndo TCPSpuriousRTOs TCPMD5NotFound TCPMD5Unexpected TCPMD5Failure TCPSackShifted TCPSackMerged TCPSackShiftFallback TCPBacklogDrop TCPMinTTLDrop TCPDeferAcceptDrop IPReversePathFilter TCPTimeWaitOverflow TCPReqQFullDoCookies TCPReqQFullDrop TCPRetransFail TCPRcvCoalesce TCPOFOQueue TCPOFODrop TCPOFOMerge TCPChallengeACK TCPSYNChallenge TCPFastOpenActive TCPFastOpenActiveFail TCPFastOpenPassive TCPFastOpenPassiveFail TCPFastOpenListenOverflow TCPFastOpenCookieReqd TCPSpuriousRtxHostQueues BusyPollRxPackets TCPAutoCorking TCPFromZeroWindowAdv TCPToZeroWindowAdv TCPWantZeroWindowAdv TCPSynRetrans TCPOrigDataSent TCPHystartTrainDetect TCPHystartTrainCwnd TCPHystartDelayDetect TCPHystartDelayCwnd TCPACKSkippedSynRecv TCPACKSkippedPAWS TCPACKSkippedSeq TCPACKSkippedFinWait2 TCPACKSkippedTimeWait TCPACKSkippedChallenge TCPWinProbe TCPKeepAlive TCPMTUPFail TCPMTUPSuccess\n")), - "packet": newStaticProcInode(ctx, msrc, []byte("sk RefCnt Type Proto Iface R Rmem User Inode\n")), - "protocols": newStaticProcInode(ctx, msrc, []byte("protocol size sockets memory press maxhdr slab module cl co di ac io in de sh ss gs se re sp bi br ha uh gp em\n")), + "netlink": newStaticProcInode(t, msrc, []byte("sk Eth Pid Groups Rmem Wmem Dump Locks Drops Inode\n")), + "netstat": newStaticProcInode(t, msrc, []byte("TcpExt: SyncookiesSent SyncookiesRecv SyncookiesFailed EmbryonicRsts PruneCalled RcvPruned OfoPruned OutOfWindowIcmps LockDroppedIcmps ArpFilter TW TWRecycled TWKilled PAWSPassive PAWSActive PAWSEstab DelayedACKs DelayedACKLocked DelayedACKLost ListenOverflows ListenDrops TCPPrequeued TCPDirectCopyFromBacklog TCPDirectCopyFromPrequeue TCPPrequeueDropped TCPHPHits TCPHPHitsToUser TCPPureAcks TCPHPAcks TCPRenoRecovery TCPSackRecovery TCPSACKReneging TCPFACKReorder TCPSACKReorder TCPRenoReorder TCPTSReorder TCPFullUndo TCPPartialUndo TCPDSACKUndo TCPLossUndo TCPLostRetransmit TCPRenoFailures TCPSackFailures TCPLossFailures TCPFastRetrans TCPForwardRetrans TCPSlowStartRetrans TCPTimeouts TCPLossProbes TCPLossProbeRecovery TCPRenoRecoveryFail TCPSackRecoveryFail TCPSchedulerFailed TCPRcvCollapsed TCPDSACKOldSent TCPDSACKOfoSent TCPDSACKRecv TCPDSACKOfoRecv TCPAbortOnData TCPAbortOnClose TCPAbortOnMemory TCPAbortOnTimeout TCPAbortOnLinger TCPAbortFailed TCPMemoryPressures TCPSACKDiscard TCPDSACKIgnoredOld TCPDSACKIgnoredNoUndo TCPSpuriousRTOs TCPMD5NotFound TCPMD5Unexpected TCPMD5Failure TCPSackShifted TCPSackMerged TCPSackShiftFallback TCPBacklogDrop TCPMinTTLDrop TCPDeferAcceptDrop IPReversePathFilter TCPTimeWaitOverflow TCPReqQFullDoCookies TCPReqQFullDrop TCPRetransFail TCPRcvCoalesce TCPOFOQueue TCPOFODrop TCPOFOMerge TCPChallengeACK TCPSYNChallenge TCPFastOpenActive TCPFastOpenActiveFail TCPFastOpenPassive TCPFastOpenPassiveFail TCPFastOpenListenOverflow TCPFastOpenCookieReqd TCPSpuriousRtxHostQueues BusyPollRxPackets TCPAutoCorking TCPFromZeroWindowAdv TCPToZeroWindowAdv TCPWantZeroWindowAdv TCPSynRetrans TCPOrigDataSent TCPHystartTrainDetect TCPHystartTrainCwnd TCPHystartDelayDetect TCPHystartDelayCwnd TCPACKSkippedSynRecv TCPACKSkippedPAWS TCPACKSkippedSeq TCPACKSkippedFinWait2 TCPACKSkippedTimeWait TCPACKSkippedChallenge TCPWinProbe TCPKeepAlive TCPMTUPFail TCPMTUPSuccess\n")), + "packet": newStaticProcInode(t, msrc, []byte("sk RefCnt Type Proto Iface R Rmem User Inode\n")), + "protocols": newStaticProcInode(t, msrc, []byte("protocol size sockets memory press maxhdr slab module cl co di ac io in de sh ss gs se re sp bi br ha uh gp em\n")), // Linux sets psched values to: nsec per usec, psched // tick in ns, 1000000, high res timer ticks per sec // (ClockGetres returns 1ns resolution). - "psched": newStaticProcInode(ctx, msrc, []byte(fmt.Sprintf("%08x %08x %08x %08x\n", uint64(time.Microsecond/time.Nanosecond), 64, 1000000, uint64(time.Second/time.Nanosecond)))), - "ptype": newStaticProcInode(ctx, msrc, []byte("Type Device Function\n")), - "route": seqfile.NewSeqFileInode(ctx, &netRoute{s: s}, msrc), - "tcp": seqfile.NewSeqFileInode(ctx, &netTCP{k: k}, msrc), - "udp": seqfile.NewSeqFileInode(ctx, &netUDP{k: k}, msrc), - "unix": seqfile.NewSeqFileInode(ctx, &netUnix{k: k}, msrc), + "psched": newStaticProcInode(t, msrc, []byte(fmt.Sprintf("%08x %08x %08x %08x\n", uint64(time.Microsecond/time.Nanosecond), 64, 1000000, uint64(time.Second/time.Nanosecond)))), + "ptype": newStaticProcInode(t, msrc, []byte("Type Device Function\n")), + "route": seqfile.NewSeqFileInode(t, &netRoute{s: s}, msrc), + "tcp": seqfile.NewSeqFileInode(t, &netTCP{k: k}, msrc), + "udp": seqfile.NewSeqFileInode(t, &netUDP{k: k}, msrc), + "unix": seqfile.NewSeqFileInode(t, &netUnix{k: k}, msrc), } if s.SupportsIPv6() { - contents["if_inet6"] = seqfile.NewSeqFileInode(ctx, &ifinet6{s: s}, msrc) - contents["ipv6_route"] = newStaticProcInode(ctx, msrc, []byte("")) - contents["tcp6"] = seqfile.NewSeqFileInode(ctx, &netTCP6{k: k}, msrc) - contents["udp6"] = newStaticProcInode(ctx, msrc, []byte(" sl local_address remote_address st tx_queue rx_queue tr tm->when retrnsmt uid timeout inode\n")) + contents["if_inet6"] = seqfile.NewSeqFileInode(t, &ifinet6{s: s}, msrc) + contents["ipv6_route"] = newStaticProcInode(t, msrc, []byte("")) + contents["tcp6"] = seqfile.NewSeqFileInode(t, &netTCP6{k: k}, msrc) + contents["udp6"] = newStaticProcInode(t, msrc, []byte(" sl local_address remote_address st tx_queue rx_queue tr tm->when retrnsmt uid timeout inode\n")) } } - d := ramfs.NewDir(ctx, contents, fs.RootOwner, fs.FilePermsFromMode(0555)) - return newProcInode(ctx, d, msrc, fs.SpecialDirectory, nil) + d := ramfs.NewDir(t, contents, fs.RootOwner, fs.FilePermsFromMode(0555)) + return newProcInode(t, d, msrc, fs.SpecialDirectory, t) } // ifinet6 implements seqfile.SeqSource for /proc/net/if_inet6. @@ -837,4 +838,4 @@ func (n *netUDP) ReadSeqFileData(ctx context.Context, h seqfile.SeqHandle) ([]se return data, 0 } -// LINT.ThenChange(../../fsimpl/proc/tasks_net.go) +// LINT.ThenChange(../../fsimpl/proc/task_net.go) diff --git a/pkg/sentry/fs/proc/proc.go b/pkg/sentry/fs/proc/proc.go index c8abb5052..c659224a7 100644 --- a/pkg/sentry/fs/proc/proc.go +++ b/pkg/sentry/fs/proc/proc.go @@ -70,6 +70,7 @@ func New(ctx context.Context, msrc *fs.MountSource, cgroupControllers map[string "loadavg": seqfile.NewSeqFileInode(ctx, &loadavgData{}, msrc), "meminfo": seqfile.NewSeqFileInode(ctx, &meminfoData{k}, msrc), "mounts": newProcInode(ctx, ramfs.NewSymlink(ctx, fs.RootOwner, "self/mounts"), msrc, fs.Symlink, nil), + "net": newProcInode(ctx, ramfs.NewSymlink(ctx, fs.RootOwner, "self/net"), msrc, fs.Symlink, nil), "self": newSelf(ctx, pidns, msrc), "stat": seqfile.NewSeqFileInode(ctx, &statData{k}, msrc), "thread-self": newThreadSelf(ctx, pidns, msrc), @@ -86,7 +87,6 @@ func New(ctx context.Context, msrc *fs.MountSource, cgroupControllers map[string } // Add more contents that need proc to be initialized. - p.AddChild(ctx, "net", p.newNetDir(ctx, k, msrc)) p.AddChild(ctx, "sys", p.newSysDir(ctx, msrc)) return newProcInode(ctx, p, msrc, fs.SpecialDirectory, nil), nil diff --git a/pkg/sentry/fs/proc/task.go b/pkg/sentry/fs/proc/task.go index 8ab8d8a02..d6c5dd2c1 100644 --- a/pkg/sentry/fs/proc/task.go +++ b/pkg/sentry/fs/proc/task.go @@ -72,24 +72,27 @@ 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) @@ -796,4 +799,95 @@ 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/fsbridge/vfs.go b/pkg/sentry/fsbridge/vfs.go index e657c39bc..79b808359 100644 --- a/pkg/sentry/fsbridge/vfs.go +++ b/pkg/sentry/fsbridge/vfs.go @@ -115,17 +115,19 @@ func NewVFSLookup(mntns *vfs.MountNamespace, root, workingDir vfs.VirtualDentry) // // remainingTraversals is not configurable in VFS2, all callers are using the // default anyways. -// -// TODO(gvisor.dev/issue/1623): Check mount has read and exec permission. -func (l *vfsLookup) OpenPath(ctx context.Context, path string, opts vfs.OpenOptions, _ *uint, resolveFinal bool) (File, error) { +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.root, - Path: fspath.Parse(path), + 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 diff --git a/pkg/sentry/fsimpl/ext/filesystem.go b/pkg/sentry/fsimpl/ext/filesystem.go index e05429d41..8497be615 100644 --- a/pkg/sentry/fsimpl/ext/filesystem.go +++ b/pkg/sentry/fsimpl/ext/filesystem.go @@ -22,6 +22,7 @@ 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/vfs" "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/syserror" @@ -255,6 +256,15 @@ 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) diff --git a/pkg/sentry/fsimpl/gofer/filesystem.go b/pkg/sentry/fsimpl/gofer/filesystem.go index 5cfb0dc4c..38e4cdbc5 100644 --- a/pkg/sentry/fsimpl/gofer/filesystem.go +++ b/pkg/sentry/fsimpl/gofer/filesystem.go @@ -21,6 +21,7 @@ import ( "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/fspath" "gvisor.dev/gvisor/pkg/p9" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserror" ) @@ -499,6 +500,18 @@ 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, d.isDir()) +} + // GetDentryAt implements vfs.FilesystemImpl.GetDentryAt. func (fs *filesystem) GetDentryAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.GetDentryOptions) (*vfs.Dentry, error) { var ds *[]*dentry diff --git a/pkg/sentry/fsimpl/gofer/gofer.go b/pkg/sentry/fsimpl/gofer/gofer.go index d00850e25..c4a8f0b38 100644 --- a/pkg/sentry/fsimpl/gofer/gofer.go +++ b/pkg/sentry/fsimpl/gofer/gofer.go @@ -1045,13 +1045,13 @@ 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), 0); 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) diff --git a/pkg/sentry/fsimpl/gofer/regular_file.go b/pkg/sentry/fsimpl/gofer/regular_file.go index 54c1031a7..e95209661 100644 --- a/pkg/sentry/fsimpl/gofer/regular_file.go +++ b/pkg/sentry/fsimpl/gofer/regular_file.go @@ -361,8 +361,15 @@ 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 } diff --git a/pkg/sentry/fsimpl/host/BUILD b/pkg/sentry/fsimpl/host/BUILD new file mode 100644 index 000000000..5d67f88e3 --- /dev/null +++ b/pkg/sentry/fsimpl/host/BUILD @@ -0,0 +1,29 @@ +load("//tools:defs.bzl", "go_library") + +licenses(["notice"]) + +go_library( + name = "host", + srcs = [ + "default_file.go", + "host.go", + "util.go", + ], + visibility = ["//pkg/sentry:internal"], + deps = [ + "//pkg/abi/linux", + "//pkg/context", + "//pkg/fd", + "//pkg/log", + "//pkg/refs", + "//pkg/safemem", + "//pkg/sentry/fsimpl/kernfs", + "//pkg/sentry/kernel/auth", + "//pkg/sentry/memmap", + "//pkg/sentry/vfs", + "//pkg/sync", + "//pkg/syserror", + "//pkg/usermem", + "@org_golang_x_sys//unix:go_default_library", + ], +) diff --git a/pkg/sentry/fsimpl/host/default_file.go b/pkg/sentry/fsimpl/host/default_file.go new file mode 100644 index 000000000..98682ba5e --- /dev/null +++ b/pkg/sentry/fsimpl/host/default_file.go @@ -0,0 +1,242 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package host + +import ( + "math" + "syscall" + + "golang.org/x/sys/unix" + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/fd" + "gvisor.dev/gvisor/pkg/safemem" + "gvisor.dev/gvisor/pkg/sentry/memmap" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +// defaultFileFD implements FileDescriptionImpl for non-socket, non-TTY files. +type defaultFileFD struct { + fileDescription + + // canMap specifies whether we allow the file to be memory mapped. + canMap bool + + // mu protects the fields below. + mu sync.Mutex + + // offset specifies the current file offset. + offset int64 +} + +// TODO(gvisor.dev/issue/1672): Implement Waitable interface. + +// PRead implements FileDescriptionImpl. +func (f *defaultFileFD) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { + // TODO(b/34716638): Some char devices do support offsets, e.g. /dev/null. + if f.inode.isStream { + return 0, syserror.ESPIPE + } + + return readFromHostFD(ctx, f.inode.hostFD, dst, offset, int(opts.Flags)) +} + +// Read implements FileDescriptionImpl. +func (f *defaultFileFD) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { + // TODO(b/34716638): Some char devices do support offsets, e.g. /dev/null. + if f.inode.isStream { + // These files can't be memory mapped, assert this. + if f.canMap { + panic("files that can return EWOULDBLOCK (sockets, pipes, etc.) cannot be memory mapped") + } + + n, err := readFromHostFD(ctx, f.inode.hostFD, dst, -1, int(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.mu.Lock() + n, err := readFromHostFD(ctx, f.inode.hostFD, dst, f.offset, int(opts.Flags)) + f.offset += n + f.mu.Unlock() + return n, err +} + +func readFromHostFD(ctx context.Context, hostFD int, dst usermem.IOSequence, offset int64, flags int) (int64, error) { + // TODO(gvisor.dev/issue/1672): Support select preadv2 flags. + if flags != 0 { + return 0, syserror.EOPNOTSUPP + } + + var reader safemem.Reader + if offset == -1 { + reader = safemem.FromIOReader{fd.NewReadWriter(hostFD)} + } else { + reader = safemem.FromVecReaderFunc{ + func(srcs [][]byte) (int64, error) { + n, err := unix.Preadv(hostFD, srcs, offset) + return int64(n), err + }, + } + } + n, err := dst.CopyOutFrom(ctx, reader) + return int64(n), err +} + +// PWrite implements FileDescriptionImpl. +func (f *defaultFileFD) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) { + // TODO(b/34716638): Some char devices do support offsets, e.g. /dev/null. + if f.inode.isStream { + return 0, syserror.ESPIPE + } + + return writeToHostFD(ctx, f.inode.hostFD, src, offset, int(opts.Flags)) +} + +// Write implements FileDescriptionImpl. +func (f *defaultFileFD) Write(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) { + // TODO(b/34716638): Some char devices do support offsets, e.g. /dev/null. + if f.inode.isStream { + // These files can't be memory mapped, assert this. + if f.canMap { + panic("files that can return EWOULDBLOCK (sockets, pipes, etc.) cannot be memory mapped") + } + + n, err := writeToHostFD(ctx, f.inode.hostFD, src, -1, int(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.mu.Lock() + n, err := writeToHostFD(ctx, f.inode.hostFD, src, f.offset, int(opts.Flags)) + f.offset += n + f.mu.Unlock() + return n, err +} + +func writeToHostFD(ctx context.Context, hostFD int, src usermem.IOSequence, offset int64, flags int) (int64, error) { + // TODO(gvisor.dev/issue/1672): Support select pwritev2 flags. + if flags != 0 { + return 0, syserror.EOPNOTSUPP + } + + var writer safemem.Writer + if offset == -1 { + writer = safemem.FromIOWriter{fd.NewReadWriter(hostFD)} + } else { + writer = safemem.FromVecWriterFunc{ + func(srcs [][]byte) (int64, error) { + n, err := unix.Pwritev(hostFD, srcs, offset) + return int64(n), err + }, + } + } + n, err := src.CopyInTo(ctx, writer) + return int64(n), err +} + +// Seek implements FileDescriptionImpl. +// +// Note that we do not support seeking on directories, since we do not even +// allow directory fds to be imported at all. +func (f *defaultFileFD) Seek(_ context.Context, offset int64, whence int32) (int64, error) { + // TODO(b/34716638): Some char devices do support seeking, e.g. /dev/null. + if f.inode.isStream { + return 0, syserror.ESPIPE + } + + f.mu.Lock() + defer f.mu.Unlock() + + switch whence { + case linux.SEEK_SET: + if offset < 0 { + return f.offset, syserror.EINVAL + } + f.offset = offset + + case linux.SEEK_CUR: + // Check for overflow. Note that underflow cannot occur, since f.offset >= 0. + if offset > math.MaxInt64-f.offset { + return f.offset, syserror.EOVERFLOW + } + if f.offset+offset < 0 { + return f.offset, syserror.EINVAL + } + f.offset += offset + + case linux.SEEK_END: + var s syscall.Stat_t + if err := syscall.Fstat(f.inode.hostFD, &s); err != nil { + return f.offset, err + } + size := s.Size + + // Check for overflow. Note that underflow cannot occur, since size >= 0. + if offset > math.MaxInt64-size { + return f.offset, syserror.EOVERFLOW + } + if size+offset < 0 { + return f.offset, syserror.EINVAL + } + f.offset = size + offset + + case linux.SEEK_DATA, linux.SEEK_HOLE: + // Modifying the offset in the host file table should not matter, since + // this is the only place where we use it. + // + // For reading and writing, we always rely on our internal offset. + n, err := unix.Seek(f.inode.hostFD, offset, int(whence)) + if err != nil { + return f.offset, err + } + f.offset = n + + default: + // Invalid whence. + return f.offset, syserror.EINVAL + } + + return f.offset, nil +} + +// Sync implements FileDescriptionImpl. +func (f *defaultFileFD) Sync(context.Context) error { + // TODO(gvisor.dev/issue/1672): Currently we do not support the SyncData optimization, so we always sync everything. + return unix.Fsync(f.inode.hostFD) +} + +// ConfigureMMap implements FileDescriptionImpl. +func (f *defaultFileFD) ConfigureMMap(_ context.Context, opts *memmap.MMapOpts) error { + if !f.canMap { + return syserror.ENODEV + } + // TODO(gvisor.dev/issue/1672): Implement ConfigureMMap and Mappable interface. + return syserror.ENODEV +} diff --git a/pkg/sentry/fsimpl/host/host.go b/pkg/sentry/fsimpl/host/host.go new file mode 100644 index 000000000..0be812d13 --- /dev/null +++ b/pkg/sentry/fsimpl/host/host.go @@ -0,0 +1,390 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// Package host provides a filesystem implementation for host files imported as +// file descriptors. +package host + +import ( + "errors" + "fmt" + "syscall" + + "golang.org/x/sys/unix" + "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/fsimpl/kernfs" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/pkg/syserror" +) + +// filesystem implements vfs.FilesystemImpl. +type filesystem struct { + kernfs.Filesystem +} + +// NewMount returns a new disconnected mount in vfsObj that may be passed to ImportFD. +func NewMount(vfsObj *vfs.VirtualFilesystem) (*vfs.Mount, error) { + fs := &filesystem{} + fs.Init(vfsObj) + vfsfs := fs.VFSFilesystem() + // NewDisconnectedMount will take an additional reference on vfsfs. + defer vfsfs.DecRef() + return vfsObj.NewDisconnectedMount(vfsfs, nil, &vfs.MountOptions{}) +} + +// ImportFD sets up and returns a vfs.FileDescription from a donated fd. +func ImportFD(mnt *vfs.Mount, hostFD int, ownerUID auth.KUID, ownerGID auth.KGID, isTTY bool) (*vfs.FileDescription, error) { + fs, ok := mnt.Filesystem().Impl().(*kernfs.Filesystem) + if !ok { + return nil, fmt.Errorf("can't import host FDs into filesystems of type %T", mnt.Filesystem().Impl()) + } + + // Retrieve metadata. + var s syscall.Stat_t + if err := syscall.Fstat(hostFD, &s); err != nil { + return nil, err + } + + fileMode := linux.FileMode(s.Mode) + fileType := fileMode.FileType() + // Pipes, character devices, and sockets. + isStream := fileType == syscall.S_IFIFO || fileType == syscall.S_IFCHR || fileType == syscall.S_IFSOCK + + i := &inode{ + hostFD: hostFD, + isStream: isStream, + isTTY: isTTY, + ino: fs.NextIno(), + mode: fileMode, + uid: ownerUID, + gid: ownerGID, + } + + d := &kernfs.Dentry{} + d.Init(i) + // i.open will take a reference on d. + defer d.DecRef() + + return i.open(d.VFSDentry(), mnt) +} + +// inode implements kernfs.Inode. +type inode struct { + kernfs.InodeNotDirectory + kernfs.InodeNotSymlink + + // When the reference count reaches zero, the host fd is closed. + refs.AtomicRefCount + + // hostFD contains the host fd that this file was originally created from, + // which must be available at time of restore. + // + // This field is initialized at creation time and is immutable. + hostFD int + + // isStream is true if the host fd points to a file representing a stream, + // e.g. a socket or a pipe. Such files are not seekable and can return + // EWOULDBLOCK for I/O operations. + // + // This field is initialized at creation time and is immutable. + isStream bool + + // isTTY is true if this file represents a TTY. + // + // This field is initialized at creation time and is immutable. + isTTY bool + + // ino is an inode number unique within this filesystem. + ino uint64 + + // mu protects the inode metadata below. + mu sync.Mutex + + // mode is the file mode of this inode. Note that this value may become out + // of date if the mode is changed on the host, e.g. with chmod. + mode linux.FileMode + + // uid and gid of the file owner. Note that these refer to the owner of the + // file created on import, not the fd on the host. + uid auth.KUID + gid auth.KGID +} + +// Note that these flags may become out of date, since they can be modified +// on the host, e.g. with fcntl. +func fileFlagsFromHostFD(fd int) (int, error) { + flags, err := unix.FcntlInt(uintptr(fd), syscall.F_GETFL, 0) + if err != nil { + log.Warningf("Failed to get file flags for donated FD %d: %v", fd, err) + return 0, err + } + // TODO(gvisor.dev/issue/1672): implement behavior corresponding to these allowed flags. + flags &= syscall.O_ACCMODE | syscall.O_DIRECT | syscall.O_NONBLOCK | syscall.O_DSYNC | syscall.O_SYNC | syscall.O_APPEND + return flags, nil +} + +// CheckPermissions implements kernfs.Inode. +func (i *inode) CheckPermissions(ctx context.Context, creds *auth.Credentials, atx vfs.AccessTypes) error { + return vfs.GenericCheckPermissions(creds, atx, false /* isDir */, uint16(i.mode), i.uid, i.gid) +} + +// Mode implements kernfs.Inode. +func (i *inode) Mode() linux.FileMode { + return i.mode +} + +// Stat implements kernfs.Inode. +func (i *inode) Stat(_ *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) { + if opts.Mask&linux.STATX__RESERVED != 0 { + return linux.Statx{}, syserror.EINVAL + } + if opts.Sync&linux.AT_STATX_SYNC_TYPE == linux.AT_STATX_SYNC_TYPE { + return linux.Statx{}, syserror.EINVAL + } + + // Limit our host call only to known flags. + mask := opts.Mask & linux.STATX_ALL + var s unix.Statx_t + err := unix.Statx(i.hostFD, "", int(unix.AT_EMPTY_PATH|opts.Sync), int(mask), &s) + // Fallback to fstat(2), if statx(2) is not supported on the host. + // + // TODO(b/151263641): Remove fallback. + if err == syserror.ENOSYS { + return i.fstat(opts) + } else if err != nil { + return linux.Statx{}, err + } + + ls := linux.Statx{Mask: mask} + // Unconditionally fill blksize, attributes, and device numbers, as indicated + // by /include/uapi/linux/stat.h. + // + // RdevMajor/RdevMinor are left as zero, so as not to expose host device + // numbers. + // + // TODO(gvisor.dev/issue/1672): Use kernfs-specific, internally defined + // device numbers. If we use the device number from the host, it may collide + // with another sentry-internal device number. We handle device/inode + // numbers without relying on the host to prevent collisions. + ls.Blksize = s.Blksize + ls.Attributes = s.Attributes + ls.AttributesMask = s.Attributes_mask + + if mask|linux.STATX_TYPE != 0 { + ls.Mode |= s.Mode & linux.S_IFMT + } + if mask|linux.STATX_MODE != 0 { + ls.Mode |= s.Mode &^ linux.S_IFMT + } + if mask|linux.STATX_NLINK != 0 { + ls.Nlink = s.Nlink + } + if mask|linux.STATX_ATIME != 0 { + ls.Atime = unixToLinuxStatxTimestamp(s.Atime) + } + if mask|linux.STATX_BTIME != 0 { + ls.Btime = unixToLinuxStatxTimestamp(s.Btime) + } + if mask|linux.STATX_CTIME != 0 { + ls.Ctime = unixToLinuxStatxTimestamp(s.Ctime) + } + if mask|linux.STATX_MTIME != 0 { + ls.Mtime = unixToLinuxStatxTimestamp(s.Mtime) + } + if mask|linux.STATX_SIZE != 0 { + ls.Size = s.Size + } + if mask|linux.STATX_BLOCKS != 0 { + ls.Blocks = s.Blocks + } + + // Use our own internal inode number and file owner. + if mask|linux.STATX_INO != 0 { + ls.Ino = i.ino + } + if mask|linux.STATX_UID != 0 { + ls.UID = uint32(i.uid) + } + if mask|linux.STATX_GID != 0 { + ls.GID = uint32(i.gid) + } + + return ls, nil +} + +// fstat is a best-effort fallback for inode.Stat() if the host does not +// support statx(2). +// +// We ignore the mask and sync flags in opts and simply supply +// STATX_BASIC_STATS, as fstat(2) itself does not allow the specification +// of a mask or sync flags. fstat(2) does not provide any metadata +// equivalent to Statx.Attributes, Statx.AttributesMask, or Statx.Btime, so +// those fields remain empty. +func (i *inode) fstat(opts vfs.StatOptions) (linux.Statx, error) { + var s unix.Stat_t + if err := unix.Fstat(i.hostFD, &s); err != nil { + return linux.Statx{}, err + } + + // Note that rdev numbers are left as 0; do not expose host device numbers. + ls := linux.Statx{ + Mask: linux.STATX_BASIC_STATS, + Blksize: uint32(s.Blksize), + Nlink: uint32(s.Nlink), + Mode: uint16(s.Mode), + Size: uint64(s.Size), + Blocks: uint64(s.Blocks), + Atime: timespecToStatxTimestamp(s.Atim), + Ctime: timespecToStatxTimestamp(s.Ctim), + Mtime: timespecToStatxTimestamp(s.Mtim), + } + + // Use our own internal inode number and file owner. + // + // TODO(gvisor.dev/issue/1672): Use a kernfs-specific device number as well. + // If we use the device number from the host, it may collide with another + // sentry-internal device number. We handle device/inode numbers without + // relying on the host to prevent collisions. + ls.Ino = i.ino + ls.UID = uint32(i.uid) + ls.GID = uint32(i.gid) + + return ls, nil +} + +// SetStat implements kernfs.Inode. +func (i *inode) SetStat(_ *vfs.Filesystem, opts vfs.SetStatOptions) error { + s := opts.Stat + + m := s.Mask + if m == 0 { + return nil + } + if m&(linux.STATX_UID|linux.STATX_GID) != 0 { + return syserror.EPERM + } + if m&linux.STATX_MODE != 0 { + if err := syscall.Fchmod(i.hostFD, uint32(s.Mode)); err != nil { + return err + } + i.mode = linux.FileMode(s.Mode) + } + if m&linux.STATX_SIZE != 0 { + if err := syscall.Ftruncate(i.hostFD, int64(s.Size)); err != nil { + return err + } + } + if m&(linux.STATX_ATIME|linux.STATX_MTIME) != 0 { + timestamps := []unix.Timespec{ + toTimespec(s.Atime, m&linux.STATX_ATIME == 0), + toTimespec(s.Mtime, m&linux.STATX_MTIME == 0), + } + if err := unix.UtimesNanoAt(i.hostFD, "", timestamps, unix.AT_EMPTY_PATH); err != nil { + return err + } + } + return nil +} + +// DecRef implements kernfs.Inode. +func (i *inode) DecRef() { + i.AtomicRefCount.DecRefWithDestructor(i.Destroy) +} + +// Destroy implements kernfs.Inode. +func (i *inode) Destroy() { + if err := unix.Close(i.hostFD); err != nil { + log.Warningf("failed to close host fd %d: %v", i.hostFD, err) + } +} + +// Open implements kernfs.Inode. +func (i *inode) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + return i.open(vfsd, rp.Mount()) +} + +func (i *inode) open(d *vfs.Dentry, mnt *vfs.Mount) (*vfs.FileDescription, error) { + fileType := i.mode.FileType() + if fileType == syscall.S_IFSOCK { + if i.isTTY { + return nil, errors.New("cannot use host socket as TTY") + } + // TODO(gvisor.dev/issue/1672): support importing sockets. + return nil, errors.New("importing host sockets not supported") + } + + // TODO(gvisor.dev/issue/1672): Whitelist specific file types here, so that + // we don't allow importing arbitrary file types without proper support. + if i.isTTY { + // TODO(gvisor.dev/issue/1672): support importing host fd as TTY. + return nil, errors.New("importing host fd as TTY not supported") + } + + // For simplicity, set offset to 0. Technically, we should + // only set to 0 on files that are not seekable (sockets, pipes, etc.), + // and use the offset from the host fd otherwise. + fd := &defaultFileFD{ + fileDescription: fileDescription{ + inode: i, + }, + canMap: canMap(uint32(fileType)), + mu: sync.Mutex{}, + offset: 0, + } + + vfsfd := &fd.vfsfd + flags, err := fileFlagsFromHostFD(i.hostFD) + if err != nil { + return nil, err + } + + if err := vfsfd.Init(fd, uint32(flags), mnt, d, &vfs.FileDescriptionOptions{}); err != nil { + 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 +} + +// SetStat implements vfs.FileDescriptionImpl. +func (f *fileDescription) SetStat(_ context.Context, opts vfs.SetStatOptions) error { + return f.inode.SetStat(nil, opts) +} + +// Stat implements vfs.FileDescriptionImpl. +func (f *fileDescription) Stat(_ context.Context, opts vfs.StatOptions) (linux.Statx, error) { + return f.inode.Stat(nil, opts) +} + +// Release implements vfs.FileDescriptionImpl. +func (f *fileDescription) Release() { + // noop +} diff --git a/pkg/sentry/fsimpl/host/util.go b/pkg/sentry/fsimpl/host/util.go new file mode 100644 index 000000000..d519feef5 --- /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) unix.Timespec { + if omit { + return unix.Timespec{ + Sec: 0, + Nsec: unix.UTIME_OMIT, + } + } + return unix.Timespec{ + Sec: int64(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/fsimpl/kernfs/dynamic_bytes_file.go b/pkg/sentry/fsimpl/kernfs/dynamic_bytes_file.go index d092ccb2a..0d27a8867 100644 --- a/pkg/sentry/fsimpl/kernfs/dynamic_bytes_file.go +++ b/pkg/sentry/fsimpl/kernfs/dynamic_bytes_file.go @@ -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(*vfs.Filesystem, 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 5650512e0..da821d524 100644 --- a/pkg/sentry/fsimpl/kernfs/fd_impl_util.go +++ b/pkg/sentry/fsimpl/kernfs/fd_impl_util.go @@ -107,9 +107,13 @@ func (fd *GenericDirectoryFD) IterDirents(ctx context.Context, cb vfs.IterDirent fs.mu.Lock() defer fs.mu.Unlock() + opts := vfs.StatOptions{Mask: linux.STATX_INO} // Handle ".". if fd.off == 0 { - stat := fd.inode().Stat(vfsFS) + stat, err := fd.inode().Stat(vfsFS, opts) + if err != nil { + return err + } dirent := vfs.Dirent{ Name: ".", Type: linux.DT_DIR, @@ -125,7 +129,10 @@ func (fd *GenericDirectoryFD) IterDirents(ctx context.Context, cb vfs.IterDirent // Handle "..". if fd.off == 1 { parentInode := vfsd.ParentOrSelf().Impl().(*Dentry).inode - stat := parentInode.Stat(vfsFS) + stat, err := parentInode.Stat(vfsFS, opts) + if err != nil { + return err + } dirent := vfs.Dirent{ Name: "..", Type: linux.FileMode(stat.Mode).DirentType(), @@ -146,7 +153,10 @@ func (fd *GenericDirectoryFD) IterDirents(ctx context.Context, cb vfs.IterDirent childIdx := fd.off - 2 for it := fd.children.nthLocked(childIdx); it != nil; it = it.Next() { inode := it.Dentry.Impl().(*Dentry).inode - stat := inode.Stat(vfsFS) + stat, err := inode.Stat(vfsFS, opts) + if err != nil { + return err + } dirent := vfs.Dirent{ Name: it.Name, Type: linux.FileMode(stat.Mode).DirentType(), @@ -190,7 +200,7 @@ 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. diff --git a/pkg/sentry/fsimpl/kernfs/filesystem.go b/pkg/sentry/fsimpl/kernfs/filesystem.go index 292f58afd..3288de290 100644 --- a/pkg/sentry/fsimpl/kernfs/filesystem.go +++ b/pkg/sentry/fsimpl/kernfs/filesystem.go @@ -12,16 +12,17 @@ // 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/vfs" "gvisor.dev/gvisor/pkg/syserror" ) @@ -229,6 +230,19 @@ func (fs *Filesystem) Sync(ctx context.Context) error { return nil } +// AccessAt implements vfs.Filesystem.Impl.AccessAt. +func (fs *Filesystem) AccessAt(ctx context.Context, rp *vfs.ResolvingPath, creds *auth.Credentials, ats vfs.AccessTypes) error { + fs.mu.RLock() + defer fs.mu.RUnlock() + defer fs.processDeferredDecRefs() + + _, inode, err := fs.walkExistingLocked(ctx, rp) + if err != nil { + return err + } + return inode.CheckPermissions(ctx, creds, ats) +} + // GetDentryAt implements vfs.FilesystemImpl.GetDentryAt. func (fs *Filesystem) GetDentryAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.GetDentryOptions) (*vfs.Dentry, error) { fs.mu.RLock() @@ -634,7 +648,7 @@ 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. diff --git a/pkg/sentry/fsimpl/kernfs/inode_impl_util.go b/pkg/sentry/fsimpl/kernfs/inode_impl_util.go index 099d70a16..4ed41326d 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 } @@ -177,7 +177,7 @@ 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 } @@ -219,7 +219,7 @@ func (a *InodeAttrs) Mode() linux.FileMode { // Stat partially implements Inode.Stat. Note that this function doesn't provide // all the stat fields, and the embedder should consider extending the result // with filesystem-specific fields. -func (a *InodeAttrs) Stat(*vfs.Filesystem) linux.Statx { +func (a *InodeAttrs) Stat(*vfs.Filesystem, vfs.StatOptions) (linux.Statx, error) { var stat linux.Statx stat.Mask = linux.STATX_TYPE | linux.STATX_MODE | linux.STATX_UID | linux.STATX_GID | linux.STATX_INO | linux.STATX_NLINK stat.Ino = atomic.LoadUint64(&a.ino) @@ -228,9 +228,9 @@ func (a *InodeAttrs) Stat(*vfs.Filesystem) linux.Statx { stat.GID = atomic.LoadUint32(&a.gid) stat.Nlink = atomic.LoadUint32(&a.nlink) - // TODO: Implement other stat fields like timestamps. + // TODO(gvisor.dev/issue/1193): Implement other stat fields like timestamps. - return stat + return stat, nil } // SetStat implements Inode.SetStat. @@ -256,7 +256,7 @@ 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 } @@ -554,3 +554,16 @@ func (s *StaticDirectory) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs fd.Init(rp.Mount(), vfsd, &s.OrderedChildren, &opts) return fd.VFSFileDescription(), nil } + +// SetStat implements Inode.SetStat not allowing inode attributes to be changed. +func (*StaticDirectory) SetStat(*vfs.Filesystem, 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..18a34a590 100644 --- a/pkg/sentry/fsimpl/kernfs/kernfs.go +++ b/pkg/sentry/fsimpl/kernfs/kernfs.go @@ -176,8 +176,6 @@ type Dentry struct { vfsd vfs.Dentry inode Inode - refs uint64 - // flags caches useful information about the dentry from the inode. See the // dflags* consts above. Must be accessed by atomic ops. flags uint32 @@ -302,7 +300,8 @@ 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. + // Precondition: rp.Done(). vfsd.Impl() must be the kernfs Dentry containing + // the inode on which Open() is being called. Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) } @@ -320,7 +319,7 @@ 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. @@ -328,7 +327,7 @@ 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. diff --git a/pkg/sentry/fsimpl/kernfs/symlink.go b/pkg/sentry/fsimpl/kernfs/symlink.go index 0ee7eb9b7..41c5a3099 100644 --- a/pkg/sentry/fsimpl/kernfs/symlink.go +++ b/pkg/sentry/fsimpl/kernfs/symlink.go @@ -18,6 +18,8 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" ) // StaticSymlink provides an Inode implementation for symlinks that point to @@ -52,3 +54,8 @@ func (s *StaticSymlink) Init(creds *auth.Credentials, ino uint64, target string) func (s *StaticSymlink) Readlink(_ context.Context) (string, error) { return s.target, nil } + +// SetStat implements Inode.SetStat not allowing inode attributes to be changed. +func (*StaticSymlink) SetStat(*vfs.Filesystem, vfs.SetStatOptions) error { + return syserror.EPERM +} diff --git a/pkg/sentry/fsimpl/proc/BUILD b/pkg/sentry/fsimpl/proc/BUILD index a83245866..8156984eb 100644 --- a/pkg/sentry/fsimpl/proc/BUILD +++ b/pkg/sentry/fsimpl/proc/BUILD @@ -8,10 +8,11 @@ go_library( "filesystem.go", "subtasks.go", "task.go", + "task_fds.go", "task_files.go", + "task_net.go", "tasks.go", "tasks_files.go", - "tasks_net.go", "tasks_sys.go", ], visibility = ["//pkg/sentry:internal"], @@ -19,8 +20,10 @@ go_library( "//pkg/abi/linux", "//pkg/context", "//pkg/log", + "//pkg/refs", "//pkg/safemem", "//pkg/sentry/fs", + "//pkg/sentry/fsbridge", "//pkg/sentry/fsimpl/kernfs", "//pkg/sentry/inet", "//pkg/sentry/kernel", @@ -53,6 +56,7 @@ 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/subtasks.go b/pkg/sentry/fsimpl/proc/subtasks.go index f3f4e49b4..ea6d60f6e 100644 --- a/pkg/sentry/fsimpl/proc/subtasks.go +++ b/pkg/sentry/fsimpl/proc/subtasks.go @@ -34,6 +34,7 @@ type subtasksInode struct { kernfs.InodeDirectoryNoNewChildren kernfs.InodeAttrs kernfs.OrderedChildren + kernfs.AlwaysValid task *kernel.Task pidns *kernel.PIDNamespace @@ -61,11 +62,6 @@ 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) @@ -121,8 +117,18 @@ func (i *subtasksInode) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.O } // 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(*vfs.Filesystem, vfs.SetStatOptions) error { + return syserror.EPERM } diff --git a/pkg/sentry/fsimpl/proc/task.go b/pkg/sentry/fsimpl/proc/task.go index 2d814668a..fae3fc5aa 100644 --- a/pkg/sentry/fsimpl/proc/task.go +++ b/pkg/sentry/fsimpl/proc/task.go @@ -45,28 +45,31 @@ var _ kernfs.Inode = (*taskInode)(nil) func newTaskInode(inoGen InoGenerator, task *kernel.Task, pidns *kernel.PIDNamespace, isThreadGroup bool, cgroupControllers map[string]string) *kernfs.Dentry { contents := map[string]*kernfs.Dentry{ - "auxv": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &auxvData{task: task}), - "cmdline": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &cmdlineData{task: task, arg: cmdlineDataArg}), - "comm": newComm(task, inoGen.NextIno(), 0444), - "environ": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &cmdlineData{task: task, arg: environDataArg}), - //"exe": newExe(t, msrc), - //"fd": newFdDir(t, msrc), - //"fdinfo": newFdInfoDir(t, msrc), - "gid_map": newTaskOwnedFile(task, inoGen.NextIno(), 0644, &idMapData{task: task, gids: true}), - "io": newTaskOwnedFile(task, inoGen.NextIno(), 0400, newIO(task, isThreadGroup)), - "maps": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &mapsData{task: task}), - //"mountinfo": seqfile.NewSeqFileInode(t, &mountInfoFile{t: t}, msrc), - //"mounts": seqfile.NewSeqFileInode(t, &mountsFile{t: t}, msrc), + "auxv": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &auxvData{task: task}), + "cmdline": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &cmdlineData{task: task, arg: cmdlineDataArg}), + "comm": newComm(task, inoGen.NextIno(), 0444), + "environ": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &cmdlineData{task: task, arg: environDataArg}), + "exe": newExeSymlink(task, inoGen.NextIno()), + "fd": newFDDirInode(task, inoGen), + "fdinfo": newFDInfoDirInode(task, inoGen), + "gid_map": newTaskOwnedFile(task, inoGen.NextIno(), 0644, &idMapData{task: task, gids: true}), + "io": newTaskOwnedFile(task, inoGen.NextIno(), 0400, newIO(task, isThreadGroup)), + "maps": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &mapsData{task: task}), + "mountinfo": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &mountInfoData{task: task}), + "mounts": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &mountsData{task: task}), + "net": newTaskNetDir(task, inoGen), "ns": newTaskOwnedDir(task, inoGen.NextIno(), 0511, map[string]*kernfs.Dentry{ "net": newNamespaceSymlink(task, inoGen.NextIno(), "net"), "pid": newNamespaceSymlink(task, inoGen.NextIno(), "pid"), "user": newNamespaceSymlink(task, inoGen.NextIno(), "user"), }), - "smaps": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &smapsData{task: task}), - "stat": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &taskStatData{task: task, pidns: pidns, tgstats: isThreadGroup}), - "statm": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &statmData{task: task}), - "status": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &statusData{task: task, pidns: pidns}), - "uid_map": newTaskOwnedFile(task, inoGen.NextIno(), 0644, &idMapData{task: task, gids: false}), + "oom_score": newTaskOwnedFile(task, inoGen.NextIno(), 0444, newStaticFile("0\n")), + "oom_score_adj": newTaskOwnedFile(task, inoGen.NextIno(), 0644, &oomScoreAdj{task: task}), + "smaps": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &smapsData{task: task}), + "stat": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &taskStatData{task: task, pidns: pidns, tgstats: isThreadGroup}), + "statm": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &statmData{task: task}), + "status": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &statusData{task: task, pidns: pidns}), + "uid_map": newTaskOwnedFile(task, inoGen.NextIno(), 0644, &idMapData{task: task, gids: false}), } if isThreadGroup { contents["task"] = newSubtasks(task, pidns, inoGen, cgroupControllers) @@ -104,13 +107,9 @@ func (i *taskInode) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenO 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(*vfs.Filesystem, vfs.SetStatOptions) error { + return syserror.EPERM } // taskOwnedInode implements kernfs.Inode and overrides inode owner with task @@ -152,12 +151,21 @@ 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. @@ -234,7 +242,7 @@ func newNamespaceSymlink(task *kernel.Task, ino uint64, ns string) *kernfs.Dentr // 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..76bfc5307 --- /dev/null +++ b/pkg/sentry/fsimpl/proc/task_fds.go @@ -0,0 +1,287 @@ +// 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" +) + +type fdDir struct { + inoGen InoGenerator + task *kernel.Task + + // When produceSymlinks is set, dirents produces for the FDs are reported + // as symlink. Otherwise, they are reported as regular files. + produceSymlink bool +} + +func (i *fdDir) lookup(name string) (*vfs.FileDescription, kernel.FDFlags, error) { + fd, err := strconv.ParseUint(name, 10, 64) + if err != nil { + return nil, kernel.FDFlags{}, syserror.ENOENT + } + + var ( + file *vfs.FileDescription + flags kernel.FDFlags + ) + i.task.WithMuLocked(func(t *kernel.Task) { + if fdTable := t.FDTable(); fdTable != nil { + file, flags = fdTable.GetVFS2(int32(fd)) + } + }) + if file == nil { + return nil, kernel.FDFlags{}, syserror.ENOENT + } + return file, flags, nil +} + +// IterDirents implements kernfs.inodeDynamicLookup. +func (i *fdDir) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback, absOffset, relOffset int64) (int64, error) { + var fds []int32 + i.task.WithMuLocked(func(t *kernel.Task) { + if fdTable := t.FDTable(); fdTable != nil { + fds = fdTable.GetFDs() + } + }) + + offset := absOffset + relOffset + typ := uint8(linux.DT_REG) + if i.produceSymlink { + typ = linux.DT_LNK + } + + // Find the appropriate starting point. + idx := sort.Search(len(fds), func(i int) bool { return fds[i] >= int32(relOffset) }) + if idx >= len(fds) { + return offset, nil + } + for _, fd := range fds[idx:] { + dirent := vfs.Dirent{ + Name: strconv.FormatUint(uint64(fd), 10), + Type: typ, + Ino: i.inoGen.NextIno(), + NextOff: offset + 1, + } + if err := cb.Handle(dirent); err != nil { + return offset, err + } + offset++ + } + return offset, nil +} + +// fdDirInode represents the inode for /proc/[pid]/fd directory. +// +// +stateify savable +type fdDirInode struct { + kernfs.InodeNotSymlink + kernfs.InodeDirectoryNoNewChildren + kernfs.InodeAttrs + kernfs.OrderedChildren + kernfs.AlwaysValid + fdDir +} + +var _ kernfs.Inode = (*fdDirInode)(nil) + +func newFDDirInode(task *kernel.Task, inoGen InoGenerator) *kernfs.Dentry { + inode := &fdDirInode{ + fdDir: fdDir{ + inoGen: inoGen, + task: task, + produceSymlink: true, + }, + } + inode.InodeAttrs.Init(task.Credentials(), inoGen.NextIno(), linux.ModeDirectory|0555) + + dentry := &kernfs.Dentry{} + dentry.Init(inode) + inode.OrderedChildren.Init(kernfs.OrderedChildrenOptions{}) + + return dentry +} + +// Lookup implements kernfs.inodeDynamicLookup. +func (i *fdDirInode) Lookup(ctx context.Context, name string) (*vfs.Dentry, error) { + file, _, err := i.lookup(name) + if err != nil { + return nil, err + } + taskDentry := newFDSymlink(i.task.Credentials(), file, i.inoGen.NextIno()) + return taskDentry.VFSDentry(), nil +} + +// Open implements kernfs.Inode. +func (i *fdDirInode) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd := &kernfs.GenericDirectoryFD{} + fd.Init(rp.Mount(), vfsd, &i.OrderedChildren, &opts) + return fd.VFSFileDescription(), nil +} + +// CheckPermissions implements kernfs.Inode. +// +// This is to match Linux, which uses a special permission handler to guarantee +// that a process can still access /proc/self/fd after it has executed +// setuid. See fs/proc/fd.c:proc_fd_permission. +func (i *fdDirInode) CheckPermissions(ctx context.Context, creds *auth.Credentials, ats vfs.AccessTypes) error { + err := i.InodeAttrs.CheckPermissions(ctx, creds, ats) + if err == nil { + // Access granted, no extra check needed. + return nil + } + if t := kernel.TaskFromContext(ctx); t != nil { + // Allow access if the task trying to access it is in the thread group + // corresponding to this directory. + if i.task.ThreadGroup() == t.ThreadGroup() { + // Access granted (overridden). + return nil + } + } + return err +} + +// fdSymlink is an symlink for the /proc/[pid]/fd/[fd] file. +// +// +stateify savable +type fdSymlink struct { + refs.AtomicRefCount + kernfs.InodeAttrs + kernfs.InodeSymlink + + file *vfs.FileDescription +} + +var _ kernfs.Inode = (*fdSymlink)(nil) + +func newFDSymlink(creds *auth.Credentials, file *vfs.FileDescription, ino uint64) *kernfs.Dentry { + file.IncRef() + inode := &fdSymlink{file: file} + inode.Init(creds, ino, linux.ModeSymlink|0777) + + d := &kernfs.Dentry{} + d.Init(inode) + return d +} + +func (s *fdSymlink) Readlink(ctx context.Context) (string, error) { + root := vfs.RootFromContext(ctx) + defer root.DecRef() + + vfsObj := s.file.VirtualDentry().Mount().Filesystem().VirtualFilesystem() + return vfsObj.PathnameWithDeleted(ctx, root, s.file.VirtualDentry()) +} + +func (s *fdSymlink) DecRef() { + s.AtomicRefCount.DecRefWithDestructor(func() { + s.Destroy() + }) +} + +func (s *fdSymlink) Destroy() { + s.file.DecRef() +} + +// fdInfoDirInode represents the inode for /proc/[pid]/fdinfo directory. +// +// +stateify savable +type fdInfoDirInode struct { + kernfs.InodeNotSymlink + kernfs.InodeDirectoryNoNewChildren + kernfs.InodeAttrs + kernfs.OrderedChildren + kernfs.AlwaysValid + fdDir +} + +var _ kernfs.Inode = (*fdInfoDirInode)(nil) + +func newFDInfoDirInode(task *kernel.Task, inoGen InoGenerator) *kernfs.Dentry { + inode := &fdInfoDirInode{ + fdDir: fdDir{ + inoGen: inoGen, + task: task, + }, + } + inode.InodeAttrs.Init(task.Credentials(), inoGen.NextIno(), linux.ModeDirectory|0555) + + dentry := &kernfs.Dentry{} + dentry.Init(inode) + inode.OrderedChildren.Init(kernfs.OrderedChildrenOptions{}) + + return dentry +} + +// Lookup implements kernfs.inodeDynamicLookup. +func (i *fdInfoDirInode) Lookup(ctx context.Context, name string) (*vfs.Dentry, error) { + file, flags, err := i.lookup(name) + if err != nil { + return nil, err + } + + data := &fdInfoData{file: file, flags: flags} + dentry := newTaskOwnedFile(i.task, i.inoGen.NextIno(), 0444, data) + return dentry.VFSDentry(), nil +} + +// Open implements kernfs.Inode. +func (i *fdInfoDirInode) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd := &kernfs.GenericDirectoryFD{} + fd.Init(rp.Mount(), vfsd, &i.OrderedChildren, &opts) + return fd.VFSFileDescription(), nil +} + +// fdInfoData implements vfs.DynamicBytesSource for /proc/[pid]/fdinfo/[fd]. +// +// +stateify savable +type fdInfoData struct { + kernfs.DynamicBytesFile + refs.AtomicRefCount + + file *vfs.FileDescription + flags kernel.FDFlags +} + +var _ dynamicInode = (*fdInfoData)(nil) + +func (d *fdInfoData) DecRef() { + d.AtomicRefCount.DecRefWithDestructor(d.destroy) +} + +func (d *fdInfoData) destroy() { + d.file.DecRef() +} + +// Generate implements vfs.DynamicBytesSource.Generate. +func (d *fdInfoData) Generate(ctx context.Context, buf *bytes.Buffer) error { + // 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(d.file.StatusFlags()) | d.flags.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..8c743df8d 100644 --- a/pkg/sentry/fsimpl/proc/task_files.go +++ b/pkg/sentry/fsimpl/proc/task_files.go @@ -18,10 +18,14 @@ import ( "bytes" "fmt" "io" + "sort" + "strings" "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/fsbridge" "gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" @@ -496,7 +500,7 @@ 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 +529,293 @@ func (i *ioData) Generate(ctx context.Context, buf *bytes.Buffer) error { fmt.Fprintf(buf, "cancelled_write_bytes: %d\n", io.BytesWriteCancelled) return nil } + +// oomScoreAdj is a stub of the /proc/<pid>/oom_score_adj file. +// +// +stateify savable +type oomScoreAdj struct { + kernfs.DynamicBytesFile + + task *kernel.Task +} + +var _ vfs.WritableDynamicBytesSource = (*oomScoreAdj)(nil) + +// Generate implements vfs.DynamicBytesSource.Generate. +func (o *oomScoreAdj) Generate(ctx context.Context, buf *bytes.Buffer) error { + if o.task.ExitState() == kernel.TaskExitDead { + return syserror.ESRCH + } + fmt.Fprintf(buf, "%d\n", o.task.OOMScoreAdj()) + return nil +} + +// Write implements vfs.WritableDynamicBytesSource.Write. +func (o *oomScoreAdj) Write(ctx context.Context, src usermem.IOSequence, offset int64) (int64, error) { + if src.NumBytes() == 0 { + return 0, nil + } + + // Limit input size so as not to impact performance if input size is large. + src = src.TakeFirst(usermem.PageSize - 1) + + var v int32 + n, err := usermem.CopyInt32StringInVec(ctx, src.IO, src.Addrs, &v, src.Opts) + if err != nil { + return 0, err + } + + if o.task.ExitState() == kernel.TaskExitDead { + return 0, syserror.ESRCH + } + if err := o.task.SetOOMScoreAdj(v); err != nil { + return 0, err + } + + return n, nil +} + +// exeSymlink is an symlink for the /proc/[pid]/exe file. +// +// +stateify savable +type exeSymlink struct { + kernfs.InodeAttrs + kernfs.InodeNoopRefCount + kernfs.InodeSymlink + + task *kernel.Task +} + +var _ kernfs.Inode = (*exeSymlink)(nil) + +func newExeSymlink(task *kernel.Task, ino uint64) *kernfs.Dentry { + inode := &exeSymlink{task: task} + inode.Init(task.Credentials(), ino, linux.ModeSymlink|0777) + + d := &kernfs.Dentry{} + d.Init(inode) + return d +} + +// Readlink implements kernfs.Inode. +func (s *exeSymlink) Readlink(ctx context.Context) (string, error) { + if !kernel.ContextCanTrace(ctx, s.task, false) { + return "", syserror.EACCES + } + + // Pull out the executable for /proc/[pid]/exe. + exec, err := s.executable() + if err != nil { + return "", err + } + defer exec.DecRef() + + return exec.PathnameWithDeleted(ctx), nil +} + +func (s *exeSymlink) executable() (file fsbridge.File, err error) { + s.task.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 + } + + // 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.ENOENT + } + }) + return +} + +// forEachMountSource runs f for the process root mount and each mount that is +// a descendant of the root. +func forEachMount(t *kernel.Task, fn func(string, *fs.Mount)) { + var fsctx *kernel.FSContext + t.WithMuLocked(func(t *kernel.Task) { + fsctx = t.FSContext() + }) + if fsctx == nil { + // The task has been destroyed. Nothing to show here. + return + } + + // All mount points must be relative to the rootDir, and mounts outside + // will be excluded. + rootDir := fsctx.RootDirectory() + if rootDir == nil { + // The task has been destroyed. Nothing to show here. + return + } + defer rootDir.DecRef() + + mnt := t.MountNamespace().FindMount(rootDir) + if mnt == nil { + // Has it just been unmounted? + return + } + ms := t.MountNamespace().AllMountsUnder(mnt) + sort.Slice(ms, func(i, j int) bool { + return ms[i].ID < ms[j].ID + }) + 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) + } +} + +// 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 { + forEachMount(i.task, 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) + + // (1) MountSource ID. + fmt.Fprintf(buf, "%d ", m.ID) + + // (2) Parent ID (or this ID if there is no parent). + pID := m.ID + if !m.IsRoot() && !m.IsUndo() { + pID = m.ParentID + } + 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. + sa := mroot.Inode.StableAttr + fmt.Fprintf(buf, "%d:%d ", sa.DeviceFileMajor, sa.DeviceFileMinor) + + // (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 ", mountPath) + + // (6) Mount options. + flags := mroot.Inode.MountSource.Flags + opts := "rw" + if flags.ReadOnly { + opts = "ro" + } + if flags.NoAtime { + opts += ",noatime" + } + if 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 ", mroot.Inode.MountSource.FilesystemType) + + // (10) Mount source: filesystem-specific information or "none". + fmt.Fprintf(buf, "none ") + + // (11) Superblock options, and final newline. + fmt.Fprintf(buf, "%s\n", superBlockOpts(mountPath, mroot.Inode.MountSource)) + }) + return nil +} + +func superBlockOpts(mountPath string, msrc *fs.MountSource) 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 msrc.Flags.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/issues/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" { + splitPath := strings.Split(mountPath, "/") + cgroupType := splitPath[len(splitPath)-1] + opts += "," + cgroupType + } + return opts +} + +// mountsData is used to implement /proc/[pid]/mounts. +// +// +stateify savable +type mountsData struct { + kernfs.DynamicBytesFile + + task *kernel.Task +} + +var _ dynamicInode = (*mountInfoData)(nil) + +// Generate implements vfs.DynamicBytesSource.Generate. +func (i *mountsData) Generate(ctx context.Context, buf *bytes.Buffer) error { + forEachMount(i.task, func(mountPath string, m *fs.Mount) { + // Format: + // <special device or remote filesystem> <mount point> <filesystem type> <mount options> <needs dump> <fsck order> + // + // We use the filesystem name as the first field, since there + // is no real block device we can point to, and we also should + // not expose anything about the remote filesystem. + // + // Only ro/rw option is supported for now. + // + // 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 + opts := "rw" + if flags.ReadOnly { + opts = "ro" + } + fmt.Fprintf(buf, "%s %s %s %s %d %d\n", "none", mountPath, root.Inode.MountSource.FilesystemType, opts, 0, 0) + }) + return nil +} diff --git a/pkg/sentry/fsimpl/proc/tasks_net.go b/pkg/sentry/fsimpl/proc/task_net.go index d4e1812d8..373a7b17d 100644 --- a/pkg/sentry/fsimpl/proc/tasks_net.go +++ b/pkg/sentry/fsimpl/proc/task_net.go @@ -37,12 +37,13 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) -func newNetDir(root *auth.Credentials, inoGen InoGenerator, k *kernel.Kernel) *kernfs.Dentry { +func newTaskNetDir(task *kernel.Task, inoGen InoGenerator) *kernfs.Dentry { + k := task.Kernel() + pidns := task.PIDNamespace() + root := auth.NewRootCredentials(pidns.UserNamespace()) + var contents map[string]*kernfs.Dentry - // TODO(gvisor.dev/issue/1833): Support for using the network stack in the - // network namespace of the calling process. We should make this per-process, - // a.k.a. /proc/PID/net, and make /proc/net a symlink to /proc/self/net. - if stack := k.RootNetworkNamespace().Stack(); stack != nil { + if stack := task.NetworkNamespace().Stack(); stack != nil { const ( arp = "IP address HW type Flags HW address Mask Device\n" netlink = "sk Eth Pid Groups Rmem Wmem Dump Locks Drops Inode\n" @@ -53,6 +54,8 @@ func newNetDir(root *auth.Credentials, inoGen InoGenerator, k *kernel.Kernel) *k ) psched := fmt.Sprintf("%08x %08x %08x %08x\n", uint64(time.Microsecond/time.Nanosecond), 64, 1000000, uint64(time.Second/time.Nanosecond)) + // TODO(gvisor.dev/issue/1833): Make sure file contents reflect the task + // network namespace. contents = map[string]*kernfs.Dentry{ "dev": newDentry(root, inoGen.NextIno(), 0444, &netDevData{stack: stack}), "snmp": newDentry(root, inoGen.NextIno(), 0444, &netSnmpData{stack: stack}), @@ -84,7 +87,7 @@ func newNetDir(root *auth.Credentials, inoGen InoGenerator, k *kernel.Kernel) *k } } - return kernfs.NewStaticDir(root, inoGen.NextIno(), 0555, contents) + return newTaskOwnedDir(task, inoGen.NextIno(), 0555, contents) } // ifinet6 implements vfs.DynamicBytesSource for /proc/net/if_inet6. diff --git a/pkg/sentry/fsimpl/proc/tasks.go b/pkg/sentry/fsimpl/proc/tasks.go index ce08a7d53..9f2ef8200 100644 --- a/pkg/sentry/fsimpl/proc/tasks.go +++ b/pkg/sentry/fsimpl/proc/tasks.go @@ -46,6 +46,7 @@ type tasksInode struct { kernfs.InodeDirectoryNoNewChildren kernfs.InodeAttrs kernfs.OrderedChildren + kernfs.AlwaysValid inoGen InoGenerator pidns *kernel.PIDNamespace @@ -66,23 +67,23 @@ var _ kernfs.Inode = (*tasksInode)(nil) func newTasksInode(inoGen InoGenerator, k *kernel.Kernel, pidns *kernel.PIDNamespace, cgroupControllers map[string]string) (*tasksInode, *kernfs.Dentry) { root := auth.NewRootCredentials(pidns.UserNamespace()) contents := map[string]*kernfs.Dentry{ - "cpuinfo": newDentry(root, inoGen.NextIno(), 0444, newStaticFile(cpuInfoData(k))), - //"filesystems": newDentry(root, inoGen.NextIno(), 0444, &filesystemsData{}), - "loadavg": newDentry(root, inoGen.NextIno(), 0444, &loadavgData{}), - "sys": newSysDir(root, inoGen, k), - "meminfo": newDentry(root, inoGen.NextIno(), 0444, &meminfoData{}), - "mounts": kernfs.NewStaticSymlink(root, inoGen.NextIno(), "self/mounts"), - "net": newNetDir(root, inoGen, k), - "stat": newDentry(root, inoGen.NextIno(), 0444, &statData{}), - "uptime": newDentry(root, inoGen.NextIno(), 0444, &uptimeData{}), - "version": newDentry(root, inoGen.NextIno(), 0444, &versionData{}), + "cpuinfo": newDentry(root, inoGen.NextIno(), 0444, newStaticFileSetStat(cpuInfoData(k))), + "filesystems": newDentry(root, inoGen.NextIno(), 0444, &filesystemsData{}), + "loadavg": newDentry(root, inoGen.NextIno(), 0444, &loadavgData{}), + "sys": newSysDir(root, inoGen, k), + "meminfo": newDentry(root, inoGen.NextIno(), 0444, &meminfoData{}), + "mounts": kernfs.NewStaticSymlink(root, inoGen.NextIno(), "self/mounts"), + "net": kernfs.NewStaticSymlink(root, inoGen.NextIno(), "self/net"), + "stat": newDentry(root, inoGen.NextIno(), 0444, &statData{}), + "uptime": newDentry(root, inoGen.NextIno(), 0444, &uptimeData{}), + "version": newDentry(root, inoGen.NextIno(), 0444, &versionData{}), } inode := &tasksInode{ pidns: pidns, inoGen: inoGen, - selfSymlink: newSelfSymlink(root, inoGen.NextIno(), 0444, pidns).VFSDentry(), - threadSelfSymlink: newThreadSelfSymlink(root, inoGen.NextIno(), 0444, pidns).VFSDentry(), + selfSymlink: newSelfSymlink(root, inoGen.NextIno(), pidns).VFSDentry(), + threadSelfSymlink: newThreadSelfSymlink(root, inoGen.NextIno(), pidns).VFSDentry(), cgroupControllers: cgroupControllers, } inode.InodeAttrs.Init(root, inoGen.NextIno(), linux.ModeDirectory|0555) @@ -121,11 +122,6 @@ func (i *tasksInode) Lookup(ctx context.Context, name string) (*vfs.Dentry, erro return taskDentry.VFSDentry(), nil } -// Valid implements kernfs.inodeDynamicLookup. -func (i *tasksInode) Valid(ctx context.Context) bool { - return true -} - // IterDirents implements kernfs.inodeDynamicLookup. func (i *tasksInode) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback, offset, _ int64) (int64, error) { // fs/proc/internal.h: #define FIRST_PROCESS_ENTRY 256 @@ -211,17 +207,36 @@ func (i *tasksInode) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.Open 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..20085bb39 100644 --- a/pkg/sentry/fsimpl/proc/tasks_files.go +++ b/pkg/sentry/fsimpl/proc/tasks_files.go @@ -26,6 +26,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/kernel/time" "gvisor.dev/gvisor/pkg/sentry/usage" + "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" ) @@ -40,9 +41,9 @@ type selfSymlink struct { var _ kernfs.Inode = (*selfSymlink)(nil) -func newSelfSymlink(creds *auth.Credentials, ino uint64, perm linux.FileMode, pidns *kernel.PIDNamespace) *kernfs.Dentry { +func newSelfSymlink(creds *auth.Credentials, ino uint64, pidns *kernel.PIDNamespace) *kernfs.Dentry { inode := &selfSymlink{pidns: pidns} - inode.Init(creds, ino, linux.ModeSymlink|perm) + inode.Init(creds, ino, linux.ModeSymlink|0777) d := &kernfs.Dentry{} d.Init(inode) @@ -62,6 +63,11 @@ func (s *selfSymlink) Readlink(ctx context.Context) (string, error) { return strconv.FormatUint(uint64(tgid), 10), nil } +// SetStat implements Inode.SetStat not allowing inode attributes to be changed. +func (*selfSymlink) SetStat(*vfs.Filesystem, vfs.SetStatOptions) error { + return syserror.EPERM +} + type threadSelfSymlink struct { kernfs.InodeAttrs kernfs.InodeNoopRefCount @@ -72,9 +78,9 @@ type threadSelfSymlink struct { var _ kernfs.Inode = (*threadSelfSymlink)(nil) -func newThreadSelfSymlink(creds *auth.Credentials, ino uint64, perm linux.FileMode, pidns *kernel.PIDNamespace) *kernfs.Dentry { +func newThreadSelfSymlink(creds *auth.Credentials, ino uint64, pidns *kernel.PIDNamespace) *kernfs.Dentry { inode := &threadSelfSymlink{pidns: pidns} - inode.Init(creds, ino, linux.ModeSymlink|perm) + inode.Init(creds, ino, linux.ModeSymlink|0777) d := &kernfs.Dentry{} d.Init(inode) @@ -95,6 +101,23 @@ func (s *threadSelfSymlink) Readlink(ctx context.Context) (string, error) { return fmt.Sprintf("%d/task/%d", tgid, tid), nil } +// SetStat implements Inode.SetStat not allowing inode attributes to be changed. +func (*threadSelfSymlink) SetStat(*vfs.Filesystem, 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(fs *vfs.Filesystem, opts vfs.SetStatOptions) error { + return d.DynamicBytesFile.InodeAttrs.SetStat(fs, 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 +160,20 @@ 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 +197,7 @@ 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 +224,13 @@ 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 +243,15 @@ func (d *loadavgData) Generate(ctx context.Context, buf *bytes.Buffer) error { // // +stateify savable type meminfoData struct { - kernfs.DynamicBytesFile - - // k is the owning Kernel. - k *kernel.Kernel + dynamicBytesFileSetAttr } var _ dynamicInode = (*meminfoData)(nil) // Generate implements vfs.DynamicBytesSource.Generate. -func (d *meminfoData) Generate(ctx context.Context, buf *bytes.Buffer) error { - mf := d.k.MemoryFile() +func (*meminfoData) Generate(ctx context.Context, buf *bytes.Buffer) error { + k := kernel.KernelFromContext(ctx) + mf := k.MemoryFile() mf.UpdateUsage() snapshot, totalUsage := usage.MemoryAccounting.Copy() totalSize := usage.TotalMemory(mf.TotalSize(), totalUsage) @@ -275,7 +294,7 @@ 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 +313,15 @@ 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 +352,19 @@ 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_test.go b/pkg/sentry/fsimpl/proc/tasks_test.go index c5d531fe0..d0f97c137 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, } ) @@ -93,17 +103,37 @@ func setup(t *testing.T) *testutil.System { k.VFS().MustRegisterFilesystemType(Name, &FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ AllowUserMount: true, }) - fsOpts := vfs.GetFilesystemOptions{ - InternalData: &InternalData{ - Cgroups: map[string]string{ - "cpuset": "/foo/cpuset", - "memory": "/foo/memory", + + mntns, err := k.VFS().NewMountNamespace(ctx, creds, "", tmpfs.Name, &vfs.GetFilesystemOptions{}) + if err != nil { + t.Fatalf("NewMountNamespace(): %v", err) + } + pop := &vfs.PathOperation{ + Root: mntns.Root(), + Start: mntns.Root(), + Path: fspath.Parse("/proc"), + } + if err := k.VFS().MkdirAt(ctx, creds, pop, &vfs.MkdirOptions{Mode: 0777}); err != nil { + t.Fatalf("MkDir(/proc): %v", err) + } + + pop = &vfs.PathOperation{ + Root: mntns.Root(), + Start: mntns.Root(), + Path: fspath.Parse("/proc"), + } + mntOpts := &vfs.MountOptions{ + GetFilesystemOptions: vfs.GetFilesystemOptions{ + InternalData: &InternalData{ + Cgroups: map[string]string{ + "cpuset": "/foo/cpuset", + "memory": "/foo/memory", + }, }, }, } - mntns, err := k.VFS().NewMountNamespace(ctx, creds, "", Name, &fsOpts) - if err != nil { - t.Fatalf("NewMountNamespace(): %v", err) + if err := k.VFS().MountAt(ctx, creds, "", pop, Name, mntOpts); err != nil { + t.Fatalf("MountAt(/proc): %v", err) } return testutil.NewSystem(ctx, t, k.VFS(), mntns) } @@ -112,7 +142,7 @@ func TestTasksEmpty(t *testing.T) { s := setup(t) defer s.Destroy() - collector := s.ListDirents(s.PathOpAtRoot("/")) + collector := s.ListDirents(s.PathOpAtRoot("/proc")) s.AssertAllDirentTypes(collector, tasksStaticFiles) s.AssertDirentOffsets(collector, tasksStaticFilesNextOffs) } @@ -138,7 +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) @@ -178,7 +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, @@ -188,6 +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 { @@ -198,10 +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) } } @@ -299,12 +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) } @@ -341,7 +373,7 @@ func TestTask(t *testing.T) { t.Fatalf("CreateTask(): %v", err) } - collector := s.ListDirents(s.PathOpAtRoot("/1")) + collector := s.ListDirents(s.PathOpAtRoot("/proc/1")) s.AssertAllDirentTypes(collector, taskStaticFiles) } @@ -359,14 +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 { @@ -409,6 +441,7 @@ func iterateDir(ctx context.Context, t *testing.T, s *testutil.System, fd *vfs.F t.Errorf("vfsfs.OpenAt(%v) failed: %v", childPath, err) continue } + defer child.DecRef() stat, err := child.Stat(ctx, vfs.StatOptions{}) if err != nil { t.Errorf("Stat(%v) failed: %v", childPath, err) @@ -429,6 +462,22 @@ func TestTree(t *testing.T) { defer s.Destroy() k := kernel.KernelFromContext(s.Ctx) + + pop := &vfs.PathOperation{ + Root: s.Root, + Start: s.Root, + Path: fspath.Parse("test-file"), + } + opts := &vfs.OpenOptions{ + Flags: linux.O_RDONLY | linux.O_CREAT, + Mode: 0777, + } + file, err := s.VFS.OpenAt(s.Ctx, s.Creds, pop, opts) + if err != nil { + t.Fatalf("failed to create test file: %v", err) + } + defer file.DecRef() + var tasks []*kernel.Task for i := 0; i < 5; i++ { tc := k.NewThreadGroup(nil, k.RootPIDNamespace(), kernel.NewSignalHandlers(), linux.SIGCHLD, k.GlobalInit().Limits()) @@ -436,6 +485,8 @@ func TestTree(t *testing.T) { if err != nil { t.Fatalf("CreateTask(): %v", err) } + // Add file to populate /proc/[pid]/fd and fdinfo directories. + task.FDTable().NewFDVFS2(task, 0, file, kernel.FDFlags{}) tasks = append(tasks, task) } @@ -443,11 +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/sys/sys.go b/pkg/sentry/fsimpl/sys/sys.go index c36c4fa11..3928ff2c8 100644 --- a/pkg/sentry/fsimpl/sys/sys.go +++ b/pkg/sentry/fsimpl/sys/sys.go @@ -94,15 +94,17 @@ 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(*vfs.Filesystem, 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) + if err := fd.Init(rp.Mount(), vfsd, &d.OrderedChildren, &opts); err != nil { + return nil, err + } return fd.VFSFileDescription(), nil } diff --git a/pkg/sentry/fsimpl/testutil/BUILD b/pkg/sentry/fsimpl/testutil/BUILD index e4f36f4ae..0e4053a46 100644 --- a/pkg/sentry/fsimpl/testutil/BUILD +++ b/pkg/sentry/fsimpl/testutil/BUILD @@ -16,12 +16,14 @@ go_library( "//pkg/cpuid", "//pkg/fspath", "//pkg/memutil", + "//pkg/sentry/fsbridge", "//pkg/sentry/fsimpl/tmpfs", "//pkg/sentry/kernel", "//pkg/sentry/kernel/auth", "//pkg/sentry/kernel/sched", "//pkg/sentry/limits", "//pkg/sentry/loader", + "//pkg/sentry/mm", "//pkg/sentry/pgalloc", "//pkg/sentry/platform", "//pkg/sentry/platform/kvm", diff --git a/pkg/sentry/fsimpl/testutil/kernel.go b/pkg/sentry/fsimpl/testutil/kernel.go index 488478e29..c16a36cdb 100644 --- a/pkg/sentry/fsimpl/testutil/kernel.go +++ b/pkg/sentry/fsimpl/testutil/kernel.go @@ -23,13 +23,16 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/cpuid" + "gvisor.dev/gvisor/pkg/fspath" "gvisor.dev/gvisor/pkg/memutil" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" "gvisor.dev/gvisor/pkg/sentry/fsimpl/tmpfs" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/kernel/sched" "gvisor.dev/gvisor/pkg/sentry/limits" "gvisor.dev/gvisor/pkg/sentry/loader" + "gvisor.dev/gvisor/pkg/sentry/mm" "gvisor.dev/gvisor/pkg/sentry/pgalloc" "gvisor.dev/gvisor/pkg/sentry/platform" "gvisor.dev/gvisor/pkg/sentry/time" @@ -123,10 +126,17 @@ func Boot() (*kernel.Kernel, error) { // CreateTask creates a new bare bones task for tests. func CreateTask(ctx context.Context, name string, tc *kernel.ThreadGroup, mntns *vfs.MountNamespace, root, cwd vfs.VirtualDentry) (*kernel.Task, error) { k := kernel.KernelFromContext(ctx) + exe, err := newFakeExecutable(ctx, k.VFS(), auth.CredentialsFromContext(ctx), root) + if err != nil { + return nil, err + } + m := mm.NewMemoryManager(k, k, k.SleepForAddressSpaceActivation) + m.SetExecutable(fsbridge.NewVFSFile(exe)) + config := &kernel.TaskConfig{ Kernel: k, ThreadGroup: tc, - TaskContext: &kernel.TaskContext{Name: name}, + TaskContext: &kernel.TaskContext{Name: name, MemoryManager: m}, Credentials: auth.CredentialsFromContext(ctx), NetworkNamespace: k.RootNetworkNamespace(), AllowedCPUMask: sched.NewFullCPUSet(k.ApplicationCores()), @@ -135,10 +145,25 @@ func CreateTask(ctx context.Context, name string, tc *kernel.ThreadGroup, mntns AbstractSocketNamespace: kernel.NewAbstractSocketNamespace(), MountNamespaceVFS2: mntns, FSContext: kernel.NewFSContextVFS2(root, cwd, 0022), + FDTable: k.NewFDTable(), } return k.TaskSet().NewTask(config) } +func newFakeExecutable(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, root vfs.VirtualDentry) (*vfs.FileDescription, error) { + const name = "executable" + pop := &vfs.PathOperation{ + Root: root, + Start: root, + Path: fspath.Parse(name), + } + opts := &vfs.OpenOptions{ + Flags: linux.O_RDONLY | linux.O_CREAT, + Mode: 0777, + } + return vfsObj.OpenAt(ctx, creds, pop, opts) +} + func createMemoryFile() (*pgalloc.MemoryFile, error) { const memfileName = "test-memory" memfd, err := memutil.CreateMemFD(memfileName, 0) diff --git a/pkg/sentry/fsimpl/tmpfs/filesystem.go b/pkg/sentry/fsimpl/tmpfs/filesystem.go index e1b551422..02637fca6 100644 --- a/pkg/sentry/fsimpl/tmpfs/filesystem.go +++ b/pkg/sentry/fsimpl/tmpfs/filesystem.go @@ -20,6 +20,7 @@ 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/vfs" "gvisor.dev/gvisor/pkg/syserror" ) @@ -154,6 +155,17 @@ func (fs *filesystem) doCreateAt(rp *vfs.ResolvingPath, dir bool, create func(pa return create(parent, name) } +// 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, d.inode.isDir()) +} + // GetDentryAt implements vfs.FilesystemImpl.GetDentryAt. func (fs *filesystem) GetDentryAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.GetDentryOptions) (*vfs.Dentry, error) { fs.mu.RLock() diff --git a/pkg/sentry/inet/namespace.go b/pkg/sentry/inet/namespace.go index c16667e7f..029af3025 100644 --- a/pkg/sentry/inet/namespace.go +++ b/pkg/sentry/inet/namespace.go @@ -23,7 +23,10 @@ type Namespace struct { // creator allows kernel to create new network stack for network namespaces. // If nil, no networking will function if network is namespaced. - creator NetworkStackCreator + // + // 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 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..7de2e509e 100644 --- a/pkg/sentry/kernel/fd_table.go +++ b/pkg/sentry/kernel/fd_table.go @@ -191,10 +191,12 @@ func (f *FDTable) Size() int { return int(size) } -// forEach iterates over all non-nil files. +// forEach iterates over all non-nil files in sorted order. // // It is the caller's responsibility to acquire an appropriate lock. func (f *FDTable) forEach(fn func(fd int32, file *fs.File, fileVFS2 *vfs.FileDescription, flags FDFlags)) { + // retries tracks the number of failed TryIncRef attempts for the same FD. + retries := 0 fd := int32(0) for { file, fileVFS2, flags, ok := f.getAll(fd) @@ -204,17 +206,26 @@ func (f *FDTable) forEach(fn func(fd int32, file *fs.File, fileVFS2 *vfs.FileDes switch { case file != nil: if !file.TryIncRef() { + retries++ + if retries > 1000 { + panic(fmt.Sprintf("File in FD table has been destroyed. FD: %d, File: %+v, FileOps: %+v", fd, file, file.FileOperations)) + } continue // Race caught. } fn(fd, file, nil, flags) file.DecRef() case fileVFS2 != nil: if !fileVFS2.TryIncRef() { + retries++ + if retries > 1000 { + panic(fmt.Sprintf("File in FD table has been destroyed. FD: %d, File: %+v, Impl: %+v", fd, fileVFS2, fileVFS2.Impl())) + } continue // Race caught. } fn(fd, nil, fileVFS2, flags) fileVFS2.DecRef() } + retries = 0 fd++ } } @@ -296,6 +307,50 @@ func (f *FDTable) NewFDs(ctx context.Context, fd int32, files []*fs.File, flags 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.get(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 +371,6 @@ 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 +379,8 @@ 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 +458,10 @@ 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) { diff --git a/pkg/sentry/kernel/fs_context.go b/pkg/sentry/kernel/fs_context.go index 7218aa24e..47f78df9a 100644 --- a/pkg/sentry/kernel/fs_context.go +++ b/pkg/sentry/kernel/fs_context.go @@ -244,6 +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 c62fd6eb1..6feda8fa1 100644 --- a/pkg/sentry/kernel/kernel.go +++ b/pkg/sentry/kernel/kernel.go @@ -247,6 +247,10 @@ type Kernel struct { // VFS keeps the filesystem state used across the kernel. vfs vfs.VirtualFilesystem + + // 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. @@ -463,6 +467,11 @@ 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 { @@ -480,7 +489,7 @@ 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 @@ -529,6 +538,11 @@ 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() for t := range ts.Root.tids { @@ -751,6 +765,8 @@ func (ctx *createProcessContext) Value(key interface{}) interface{} { 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: @@ -999,11 +1015,14 @@ 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() - } - }) + // TODO(gvisor.dev/issue/1663): Add save support for VFS2. + if !VFS2Enabled { + t.fdTable.forEach(func(_ int32, file *fs.File, _ *vfs.FileDescription, _ FDFlags) { + if tfd, ok := file.FileOperations.(*timerfd.TimerOperations); ok { + tfd.PauseTimer() + } + }) + } } } k.timekeeper.PauseUpdates() @@ -1028,12 +1047,15 @@ func (k *Kernel) resumeTimeLocked() { it.ResumeTimer() } } - 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() - } - }) + // TODO(gvisor.dev/issue/1663): Add save support for VFS2. + if !VFS2Enabled { + 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() + } + }) + } } } } @@ -1477,6 +1499,8 @@ func (ctx supervisorContext) Value(key interface{}) interface{} { 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: diff --git a/pkg/sentry/kernel/pipe/BUILD b/pkg/sentry/kernel/pipe/BUILD index 4c049d5b4..f29dc0472 100644 --- a/pkg/sentry/kernel/pipe/BUILD +++ b/pkg/sentry/kernel/pipe/BUILD @@ -1,25 +1,10 @@ load("//tools:defs.bzl", "go_library", "go_test") -load("//tools/go_generics:defs.bzl", "go_template_instance") package(licenses = ["notice"]) -go_template_instance( - name = "buffer_list", - out = "buffer_list.go", - package = "pipe", - prefix = "buffer", - template = "//pkg/ilist:generic_list", - types = { - "Element": "*buffer", - "Linker": "*buffer", - }, -) - go_library( name = "pipe", srcs = [ - "buffer.go", - "buffer_list.go", "device.go", "node.go", "pipe.go", @@ -33,8 +18,8 @@ go_library( deps = [ "//pkg/abi/linux", "//pkg/amutex", + "//pkg/buffer", "//pkg/context", - "//pkg/safemem", "//pkg/sentry/arch", "//pkg/sentry/device", "//pkg/sentry/fs", @@ -51,7 +36,6 @@ go_test( name = "pipe_test", size = "small", srcs = [ - "buffer_test.go", "node_test.go", "pipe_test.go", ], diff --git a/pkg/sentry/kernel/pipe/buffer.go b/pkg/sentry/kernel/pipe/buffer.go deleted file mode 100644 index fe3be5dbd..000000000 --- a/pkg/sentry/kernel/pipe/buffer.go +++ /dev/null @@ -1,115 +0,0 @@ -// Copyright 2018 The gVisor Authors. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -package pipe - -import ( - "io" - - "gvisor.dev/gvisor/pkg/safemem" - "gvisor.dev/gvisor/pkg/sync" -) - -// buffer encapsulates a queueable byte buffer. -// -// Note that the total size is slightly less than two pages. This -// is done intentionally to ensure that the buffer object aligns -// with runtime internals. We have no hard size or alignment -// requirements. This two page size will effectively minimize -// internal fragmentation, but still have a large enough chunk -// to limit excessive segmentation. -// -// +stateify savable -type buffer struct { - data [8144]byte - read int - write int - bufferEntry -} - -// Reset resets internal data. -// -// This must be called before use. -func (b *buffer) Reset() { - b.read = 0 - b.write = 0 -} - -// Empty indicates the buffer is empty. -// -// This indicates there is no data left to read. -func (b *buffer) Empty() bool { - return b.read == b.write -} - -// Full indicates the buffer is full. -// -// This indicates there is no capacity left to write. -func (b *buffer) Full() bool { - return b.write == len(b.data) -} - -// WriteFromBlocks implements safemem.Writer.WriteFromBlocks. -func (b *buffer) WriteFromBlocks(srcs safemem.BlockSeq) (uint64, error) { - dst := safemem.BlockSeqOf(safemem.BlockFromSafeSlice(b.data[b.write:])) - n, err := safemem.CopySeq(dst, srcs) - b.write += int(n) - return n, err -} - -// WriteFromReader writes to the buffer from an io.Reader. -func (b *buffer) WriteFromReader(r io.Reader, count int64) (int64, error) { - dst := b.data[b.write:] - if count < int64(len(dst)) { - dst = b.data[b.write:][:count] - } - n, err := r.Read(dst) - b.write += n - return int64(n), err -} - -// ReadToBlocks implements safemem.Reader.ReadToBlocks. -func (b *buffer) ReadToBlocks(dsts safemem.BlockSeq) (uint64, error) { - src := safemem.BlockSeqOf(safemem.BlockFromSafeSlice(b.data[b.read:b.write])) - n, err := safemem.CopySeq(dsts, src) - b.read += int(n) - return n, err -} - -// ReadToWriter reads from the buffer into an io.Writer. -func (b *buffer) ReadToWriter(w io.Writer, count int64, dup bool) (int64, error) { - src := b.data[b.read:b.write] - if count < int64(len(src)) { - src = b.data[b.read:][:count] - } - n, err := w.Write(src) - if !dup { - b.read += n - } - return int64(n), err -} - -// bufferPool is a pool for buffers. -var bufferPool = sync.Pool{ - New: func() interface{} { - return new(buffer) - }, -} - -// newBuffer grabs a new buffer from the pool. -func newBuffer() *buffer { - b := bufferPool.Get().(*buffer) - b.Reset() - return b -} diff --git a/pkg/sentry/kernel/pipe/pipe.go b/pkg/sentry/kernel/pipe/pipe.go index 08410283f..725e9db7d 100644 --- a/pkg/sentry/kernel/pipe/pipe.go +++ b/pkg/sentry/kernel/pipe/pipe.go @@ -20,6 +20,7 @@ import ( "sync/atomic" "syscall" + "gvisor.dev/gvisor/pkg/buffer" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sync" @@ -70,10 +71,10 @@ type Pipe struct { // mu protects all pipe internal state below. mu sync.Mutex `state:"nosave"` - // data is the buffer queue of pipe contents. + // view is the underlying set of buffers. // // This is protected by mu. - data bufferList + view buffer.View // max is the maximum size of the pipe in bytes. When this max has been // reached, writers will get EWOULDBLOCK. @@ -81,11 +82,6 @@ type Pipe struct { // This is protected by mu. max int64 - // size is the current size of the pipe in bytes. - // - // This is protected by mu. - size int64 - // hadWriter indicates if this pipe ever had a writer. Note that this // does not necessarily indicate there is *currently* a writer, just // that there has been a writer at some point since the pipe was @@ -196,7 +192,7 @@ type readOps struct { limit func(int64) // read performs the actual read operation. - read func(*buffer) (int64, error) + read func(*buffer.View) (int64, error) } // read reads data from the pipe into dst and returns the number of bytes @@ -213,7 +209,7 @@ func (p *Pipe) read(ctx context.Context, ops readOps) (int64, error) { defer p.mu.Unlock() // Is the pipe empty? - if p.size == 0 { + if p.view.Size() == 0 { if !p.HasWriters() { // There are no writers, return EOF. return 0, nil @@ -222,71 +218,13 @@ func (p *Pipe) read(ctx context.Context, ops readOps) (int64, error) { } // Limit how much we consume. - if ops.left() > p.size { - ops.limit(p.size) + if ops.left() > p.view.Size() { + ops.limit(p.view.Size()) } - done := int64(0) - for ops.left() > 0 { - // Pop the first buffer. - first := p.data.Front() - if first == nil { - break - } - - // Copy user data. - n, err := ops.read(first) - done += int64(n) - p.size -= n - - // Empty buffer? - if first.Empty() { - // Push to the free list. - p.data.Remove(first) - bufferPool.Put(first) - } - - // Handle errors. - if err != nil { - return done, err - } - } - - return done, nil -} - -// dup duplicates all data from this pipe into the given writer. -// -// There is no blocking behavior implemented here. The writer may propagate -// some blocking error. All the writes must be complete writes. -func (p *Pipe) dup(ctx context.Context, ops readOps) (int64, error) { - p.mu.Lock() - defer p.mu.Unlock() - - // Is the pipe empty? - if p.size == 0 { - if !p.HasWriters() { - // See above. - return 0, nil - } - return 0, syserror.ErrWouldBlock - } - - // Limit how much we consume. - if ops.left() > p.size { - ops.limit(p.size) - } - - done := int64(0) - for buf := p.data.Front(); buf != nil; buf = buf.Next() { - n, err := ops.read(buf) - done += n - if err != nil { - return done, err - } - } - - return done, nil + // Copy user data; the read op is responsible for trimming. + done, err := ops.read(&p.view) + return done, err } type writeOps struct { @@ -297,7 +235,7 @@ type writeOps struct { limit func(int64) // write should write to the provided buffer. - write func(*buffer) (int64, error) + write func(*buffer.View) (int64, error) } // write writes data from sv into the pipe and returns the number of bytes @@ -317,33 +255,19 @@ 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 { + if avail := p.max - p.view.Size(); 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 - - // Handle errors. - if err != nil { - return done, err - } + // Copy user data. + done, err := ops.write(&p.view) + if err != nil { + return done, err } + if wanted > done { // Partial write due to full pipe. return done, syserror.ErrWouldBlock @@ -396,7 +320,7 @@ 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 +346,7 @@ 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 +375,7 @@ 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 +398,7 @@ func (p *Pipe) SetFifoSize(size int64) (int64, error) { } p.mu.Lock() defer p.mu.Unlock() - if size < p.size { + if size < p.view.Size() { return 0, syserror.EBUSY } p.max = size diff --git a/pkg/sentry/kernel/pipe/pipe_util.go b/pkg/sentry/kernel/pipe/pipe_util.go index 80158239e..5a1d4fd57 100644 --- a/pkg/sentry/kernel/pipe/pipe_util.go +++ b/pkg/sentry/kernel/pipe/pipe_util.go @@ -21,6 +21,7 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/amutex" + "gvisor.dev/gvisor/pkg/buffer" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sync" @@ -49,9 +50,10 @@ func (p *Pipe) Read(ctx context.Context, dst usermem.IOSequence) (int64, error) limit: func(l int64) { dst = dst.TakeFirst64(l) }, - read: func(buf *buffer) (int64, error) { - n, err := dst.CopyOutFrom(ctx, buf) + read: func(view *buffer.View) (int64, error) { + n, err := dst.CopyOutFrom(ctx, view) dst = dst.DropFirst64(n) + view.TrimFront(n) return n, err }, }) @@ -70,16 +72,15 @@ func (p *Pipe) WriteTo(ctx context.Context, w io.Writer, count int64, dup bool) limit: func(l int64) { count = l }, - read: func(buf *buffer) (int64, error) { - n, err := buf.ReadToWriter(w, count, dup) + read: func(view *buffer.View) (int64, error) { + n, err := view.ReadToWriter(w, count) + if !dup { + view.TrimFront(n) + } count -= n return n, err }, } - if dup { - // There is no notification for dup operations. - return p.dup(ctx, ops) - } n, err := p.read(ctx, ops) if n > 0 { p.Notify(waiter.EventOut) @@ -96,8 +97,8 @@ func (p *Pipe) Write(ctx context.Context, src usermem.IOSequence) (int64, error) limit: func(l int64) { src = src.TakeFirst64(l) }, - write: func(buf *buffer) (int64, error) { - n, err := src.CopyInTo(ctx, buf) + write: func(view *buffer.View) (int64, error) { + n, err := src.CopyInTo(ctx, view) src = src.DropFirst64(n) return n, err }, @@ -117,8 +118,8 @@ func (p *Pipe) ReadFrom(ctx context.Context, r io.Reader, count int64) (int64, e limit: func(l int64) { count = l }, - write: func(buf *buffer) (int64, error) { - n, err := buf.WriteFromReader(r, count) + write: func(view *buffer.View) (int64, error) { + n, err := view.WriteFromReader(r, count) count -= n return n, err }, diff --git a/pkg/sentry/kernel/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/task.go b/pkg/sentry/kernel/task.go index e37e23231..8452ddf5b 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" ) @@ -789,6 +790,15 @@ 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. @@ -798,6 +808,15 @@ 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() @@ -829,3 +848,18 @@ 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 +} diff --git a/pkg/sentry/kernel/task_clone.go b/pkg/sentry/kernel/task_clone.go index 78866f280..e1ecca99e 100644 --- a/pkg/sentry/kernel/task_clone.go +++ b/pkg/sentry/kernel/task_clone.go @@ -15,6 +15,8 @@ package kernel import ( + "sync/atomic" + "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/bpf" "gvisor.dev/gvisor/pkg/sentry/inet" @@ -260,6 +262,7 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) { sh = sh.Fork() } tg = t.k.NewThreadGroup(tg.mounts, pidns, sh, opts.TerminationSignal, tg.limits.GetCopy()) + tg.oomScoreAdj = atomic.LoadInt32(&t.tg.oomScoreAdj) rseqAddr = t.rseqAddr rseqSignature = t.rseqSignature } diff --git a/pkg/sentry/kernel/task_context.go b/pkg/sentry/kernel/task_context.go index 2be982684..0158b1788 100644 --- a/pkg/sentry/kernel/task_context.go +++ b/pkg/sentry/kernel/task_context.go @@ -140,7 +140,7 @@ func (k *Kernel) LoadTaskImage(ctx context.Context, args loader.LoadArgs) (*Task } // Prepare a new user address space to load into. - m := mm.NewMemoryManager(k, k) + m := mm.NewMemoryManager(k, k, k.SleepForAddressSpaceActivation) defer m.DecUsers(ctx) args.MemoryManager = m diff --git a/pkg/sentry/kernel/task_exec.go b/pkg/sentry/kernel/task_exec.go index 8f57a34a6..00c425cca 100644 --- a/pkg/sentry/kernel/task_exec.go +++ b/pkg/sentry/kernel/task_exec.go @@ -220,7 +220,7 @@ func (r *runSyscallAfterExecStop) execute(t *Task) taskRunState { t.mu.Unlock() t.unstopVforkParent() // NOTE(b/30316266): All locks must be dropped prior to calling Activate. - t.MemoryManager().Activate() + t.MemoryManager().Activate(t) t.ptraceExec(oldTID) return (*runSyscallExit)(nil) diff --git a/pkg/sentry/kernel/task_log.go b/pkg/sentry/kernel/task_log.go index 6d737d3e5..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...) } } diff --git a/pkg/sentry/kernel/task_run.go b/pkg/sentry/kernel/task_run.go index 5568c91bc..799cbcd93 100644 --- a/pkg/sentry/kernel/task_run.go +++ b/pkg/sentry/kernel/task_run.go @@ -126,13 +126,39 @@ func (t *Task) doStop() { } } +func (*runApp) handleCPUIDInstruction(t *Task) error { + if len(arch.CPUIDInstruction) == 0 { + // CPUID emulation isn't supported, but this code can be + // executed, because the ptrace platform returns + // ErrContextSignalCPUID on page faults too. Look at + // pkg/sentry/platform/ptrace/ptrace.go:context.Switch for more + // details. + return platform.ErrContextSignal + } + // Is this a CPUID instruction? + region := trace.StartRegion(t.traceContext, cpuidRegion) + expected := arch.CPUIDInstruction[:] + found := make([]byte, len(expected)) + _, err := t.CopyIn(usermem.Addr(t.Arch().IP()), &found) + if err == nil && bytes.Equal(expected, found) { + // Skip the cpuid instruction. + t.Arch().CPUIDEmulate(t) + t.Arch().SetIP(t.Arch().IP() + uintptr(len(expected))) + region.End() + + return nil + } + region.End() // Not an actual CPUID, but required copy-in. + return platform.ErrContextSignal +} + // The runApp state checks for interrupts before executing untrusted // application code. // // +stateify savable type runApp struct{} -func (*runApp) execute(t *Task) taskRunState { +func (app *runApp) execute(t *Task) taskRunState { if t.interrupted() { // Checkpointing instructs tasks to stop by sending an interrupt, so we // must check for stops before entering runInterrupt (instead of @@ -237,21 +263,10 @@ func (*runApp) execute(t *Task) taskRunState { return (*runApp)(nil) case platform.ErrContextSignalCPUID: - // Is this a CPUID instruction? - region := trace.StartRegion(t.traceContext, cpuidRegion) - expected := arch.CPUIDInstruction[:] - found := make([]byte, len(expected)) - _, err := t.CopyIn(usermem.Addr(t.Arch().IP()), &found) - if err == nil && bytes.Equal(expected, found) { - // Skip the cpuid instruction. - t.Arch().CPUIDEmulate(t) - t.Arch().SetIP(t.Arch().IP() + uintptr(len(expected))) - region.End() - + if err := app.handleCPUIDInstruction(t); err == nil { // Resume execution. return (*runApp)(nil) } - region.End() // Not an actual CPUID, but required copy-in. // The instruction at the given RIP was not a CPUID, and we // fallthrough to the default signal deliver behavior below. 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 268f62e9d..52849f5b3 100644 --- a/pkg/sentry/kernel/thread_group.go +++ b/pkg/sentry/kernel/thread_group.go @@ -254,6 +254,13 @@ type ThreadGroup struct { // // tty is protected by the signal mutex. tty *TTY + + // oomScoreAdj is the thread group's OOM score adjustment. This is + // currently not used but is maintained for consistency. + // TODO(gvisor.dev/issue/1967) + // + // oomScoreAdj is accessed using atomic memory operations. + oomScoreAdj int32 } // NewThreadGroup returns a new, empty thread group in PID namespace pidns. The diff --git a/pkg/sentry/mm/address_space.go b/pkg/sentry/mm/address_space.go index 94d39af60..0332fc71c 100644 --- a/pkg/sentry/mm/address_space.go +++ b/pkg/sentry/mm/address_space.go @@ -18,6 +18,7 @@ import ( "fmt" "sync/atomic" + "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/platform" "gvisor.dev/gvisor/pkg/usermem" ) @@ -38,7 +39,7 @@ 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. for { @@ -91,16 +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 } diff --git a/pkg/sentry/mm/lifecycle.go b/pkg/sentry/mm/lifecycle.go index 3c263ebaa..d8a5b9d29 100644 --- a/pkg/sentry/mm/lifecycle.go +++ b/pkg/sentry/mm/lifecycle.go @@ -28,16 +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, } } @@ -79,9 +80,10 @@ 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, } // Copy vmas. diff --git a/pkg/sentry/mm/mm.go b/pkg/sentry/mm/mm.go index 637383c7a..c2195ae11 100644 --- a/pkg/sentry/mm/mm.go +++ b/pkg/sentry/mm/mm.go @@ -226,6 +226,11 @@ 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 } // vma represents a virtual memory area. diff --git a/pkg/sentry/mm/mm_test.go b/pkg/sentry/mm/mm_test.go index edacca741..fdc308542 100644 --- a/pkg/sentry/mm/mm_test.go +++ b/pkg/sentry/mm/mm_test.go @@ -31,7 +31,7 @@ import ( func testMemoryManager(ctx context.Context) *MemoryManager { p := platform.FromContext(ctx) mfp := pgalloc.MemoryFileProviderFromContext(ctx) - mm := NewMemoryManager(p, mfp) + mm := NewMemoryManager(p, mfp, false) mm.layout = arch.MmapLayout{ MinAddr: p.MinUserAddress(), MaxAddr: p.MaxUserAddress(), diff --git a/pkg/sentry/platform/kvm/bluepill.go b/pkg/sentry/platform/kvm/bluepill.go index 35cd55fef..4b23f7803 100644 --- a/pkg/sentry/platform/kvm/bluepill.go +++ b/pkg/sentry/platform/kvm/bluepill.go @@ -81,12 +81,6 @@ func (c *vCPU) die(context *arch.SignalContext64, msg string) { // Save the death message, which will be thrown. c.dieState.message = msg - // Reload all registers to have an accurate stack trace when we return - // to host mode. This means that the stack should be unwound correctly. - if errno := c.getUserRegisters(&c.dieState.guestRegs); errno != 0 { - throw(msg) - } - // Setup the trampoline. dieArchSetup(c, context, &c.dieState.guestRegs) } diff --git a/pkg/sentry/platform/kvm/bluepill_amd64_unsafe.go b/pkg/sentry/platform/kvm/bluepill_amd64_unsafe.go index a63a6a071..99cac665d 100644 --- a/pkg/sentry/platform/kvm/bluepill_amd64_unsafe.go +++ b/pkg/sentry/platform/kvm/bluepill_amd64_unsafe.go @@ -31,6 +31,12 @@ import ( // //go:nosplit func dieArchSetup(c *vCPU, context *arch.SignalContext64, guestRegs *userRegs) { + // Reload all registers to have an accurate stack trace when we return + // to host mode. This means that the stack should be unwound correctly. + if errno := c.getUserRegisters(&c.dieState.guestRegs); errno != 0 { + throw(c.dieState.message) + } + // If the vCPU is in user mode, we set the stack to the stored stack // value in the vCPU itself. We don't want to unwind the user stack. if guestRegs.RFLAGS&ring0.UserFlagsSet == ring0.UserFlagsSet { diff --git a/pkg/sentry/platform/kvm/bluepill_arm64.s b/pkg/sentry/platform/kvm/bluepill_arm64.s index c61700892..04efa0147 100644 --- a/pkg/sentry/platform/kvm/bluepill_arm64.s +++ b/pkg/sentry/platform/kvm/bluepill_arm64.s @@ -82,6 +82,8 @@ fallback: // dieTrampoline: see bluepill.go, bluepill_arm64_unsafe.go for documentation. TEXT ·dieTrampoline(SB),NOSPLIT,$0 - // TODO(gvisor.dev/issue/1249): dieTrampoline supporting for Arm64. - MOVD R9, 8(RSP) - BL ·dieHandler(SB) + // R0: Fake the old PC as caller + // R1: First argument (vCPU) + MOVD.P R1, 8(RSP) // R1: First argument (vCPU) + MOVD.P R0, 8(RSP) // R0: Fake the old PC as caller + B ·dieHandler(SB) diff --git a/pkg/sentry/platform/kvm/bluepill_arm64_unsafe.go b/pkg/sentry/platform/kvm/bluepill_arm64_unsafe.go index 2f02c03cf..eb5ed574e 100644 --- a/pkg/sentry/platform/kvm/bluepill_arm64_unsafe.go +++ b/pkg/sentry/platform/kvm/bluepill_arm64_unsafe.go @@ -17,10 +17,33 @@ package kvm import ( + "unsafe" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/platform/ring0" ) +// dieArchSetup initialies the state for dieTrampoline. +// +// The arm64 dieTrampoline requires the vCPU to be set in R1, and the last PC +// to be in R0. The trampoline then simulates a call to dieHandler from the +// provided PC. +// //go:nosplit func dieArchSetup(c *vCPU, context *arch.SignalContext64, guestRegs *userRegs) { - // TODO(gvisor.dev/issue/1249): dieTrampoline supporting for Arm64. + // If the vCPU is in user mode, we set the stack to the stored stack + // value in the vCPU itself. We don't want to unwind the user stack. + if guestRegs.Regs.Pstate&ring0.PSR_MODE_MASK == ring0.PSR_MODE_EL0t { + regs := c.CPU.Registers() + context.Regs[0] = regs.Regs[0] + context.Sp = regs.Sp + context.Regs[29] = regs.Regs[29] // stack base address + } else { + context.Regs[0] = guestRegs.Regs.Pc + context.Sp = guestRegs.Regs.Sp + context.Regs[29] = guestRegs.Regs.Regs[29] + context.Pstate = guestRegs.Regs.Pstate + } + context.Regs[1] = uint64(uintptr(unsafe.Pointer(c))) + context.Pc = uint64(dieTrampolineAddr) } diff --git a/pkg/sentry/platform/kvm/kvm.go b/pkg/sentry/platform/kvm/kvm.go index 972ba85c3..a9b4af43e 100644 --- a/pkg/sentry/platform/kvm/kvm.go +++ b/pkg/sentry/platform/kvm/kvm.go @@ -27,6 +27,38 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) +// userMemoryRegion is a region of physical memory. +// +// This mirrors kvm_memory_region. +type userMemoryRegion struct { + slot uint32 + flags uint32 + guestPhysAddr uint64 + memorySize uint64 + userspaceAddr uint64 +} + +// runData is the run structure. This may be mapped for synchronous register +// access (although that doesn't appear to be supported by my kernel at least). +// +// This mirrors kvm_run. +type runData struct { + requestInterruptWindow uint8 + _ [7]uint8 + + exitReason uint32 + readyForInterruptInjection uint8 + ifFlag uint8 + _ [2]uint8 + + cr8 uint64 + apicBase uint64 + + // This is the union data for exits. Interpretation depends entirely on + // the exitReason above (see vCPU code for more information). + data [32]uint64 +} + // KVM represents a lightweight VM context. type KVM struct { platform.NoCPUPreemptionDetection diff --git a/pkg/sentry/platform/kvm/kvm_amd64.go b/pkg/sentry/platform/kvm/kvm_amd64.go index c5a6f9c7d..093497bc4 100644 --- a/pkg/sentry/platform/kvm/kvm_amd64.go +++ b/pkg/sentry/platform/kvm/kvm_amd64.go @@ -21,17 +21,6 @@ import ( "gvisor.dev/gvisor/pkg/sentry/platform/ring0" ) -// userMemoryRegion is a region of physical memory. -// -// This mirrors kvm_memory_region. -type userMemoryRegion struct { - slot uint32 - flags uint32 - guestPhysAddr uint64 - memorySize uint64 - userspaceAddr uint64 -} - // userRegs represents KVM user registers. // // This mirrors kvm_regs. @@ -169,27 +158,6 @@ type modelControlRegisters struct { entries [16]modelControlRegister } -// runData is the run structure. This may be mapped for synchronous register -// access (although that doesn't appear to be supported by my kernel at least). -// -// This mirrors kvm_run. -type runData struct { - requestInterruptWindow uint8 - _ [7]uint8 - - exitReason uint32 - readyForInterruptInjection uint8 - ifFlag uint8 - _ [2]uint8 - - cr8 uint64 - apicBase uint64 - - // This is the union data for exits. Interpretation depends entirely on - // the exitReason above (see vCPU code for more information). - data [32]uint64 -} - // cpuidEntry is a single CPUID entry. // // This mirrors kvm_cpuid_entry2. diff --git a/pkg/sentry/platform/kvm/kvm_arm64.go b/pkg/sentry/platform/kvm/kvm_arm64.go index 2319c86d3..79045651e 100644 --- a/pkg/sentry/platform/kvm/kvm_arm64.go +++ b/pkg/sentry/platform/kvm/kvm_arm64.go @@ -20,17 +20,6 @@ import ( "syscall" ) -// userMemoryRegion is a region of physical memory. -// -// This mirrors kvm_memory_region. -type userMemoryRegion struct { - slot uint32 - flags uint32 - guestPhysAddr uint64 - memorySize uint64 - userspaceAddr uint64 -} - type kvmOneReg struct { id uint64 addr uint64 @@ -53,27 +42,6 @@ type userRegs struct { fpRegs userFpsimdState } -// runData is the run structure. This may be mapped for synchronous register -// access (although that doesn't appear to be supported by my kernel at least). -// -// This mirrors kvm_run. -type runData struct { - requestInterruptWindow uint8 - _ [7]uint8 - - exitReason uint32 - readyForInterruptInjection uint8 - ifFlag uint8 - _ [2]uint8 - - cr8 uint64 - apicBase uint64 - - // This is the union data for exits. Interpretation depends entirely on - // the exitReason above (see vCPU code for more information). - data [32]uint64 -} - // updateGlobalOnce does global initialization. It has to be called only once. func updateGlobalOnce(fd int) error { physicalInit() diff --git a/pkg/sentry/platform/kvm/machine.go b/pkg/sentry/platform/kvm/machine.go index 8076c7529..f1afc74dc 100644 --- a/pkg/sentry/platform/kvm/machine.go +++ b/pkg/sentry/platform/kvm/machine.go @@ -329,10 +329,12 @@ func (m *machine) Destroy() { } // Get gets an available vCPU. +// +// This will return with the OS thread locked. func (m *machine) Get() *vCPU { + m.mu.RLock() runtime.LockOSThread() tid := procid.Current() - m.mu.RLock() // Check for an exact match. if c := m.vCPUs[tid]; c != nil { @@ -343,8 +345,22 @@ func (m *machine) Get() *vCPU { // The happy path failed. We now proceed to acquire an exclusive lock // (because the vCPU map may change), and scan all available vCPUs. + // In this case, we first unlock the OS thread. Otherwise, if mu is + // not available, the current system thread will be parked and a new + // system thread spawned. We avoid this situation by simply refreshing + // tid after relocking the system thread. m.mu.RUnlock() + runtime.UnlockOSThread() m.mu.Lock() + runtime.LockOSThread() + tid = procid.Current() + + // Recheck for an exact match. + if c := m.vCPUs[tid]; c != nil { + c.lock() + m.mu.Unlock() + return c + } for { // Scan for an available vCPU. diff --git a/pkg/sentry/platform/kvm/machine_arm64_unsafe.go b/pkg/sentry/platform/kvm/machine_arm64_unsafe.go index 1c8384e6b..b531f2f85 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 diff --git a/pkg/sentry/platform/ptrace/BUILD b/pkg/sentry/platform/ptrace/BUILD index 95abd321e..30402c2df 100644 --- a/pkg/sentry/platform/ptrace/BUILD +++ b/pkg/sentry/platform/ptrace/BUILD @@ -9,6 +9,7 @@ go_library( "ptrace.go", "ptrace_amd64.go", "ptrace_arm64.go", + "ptrace_arm64_unsafe.go", "ptrace_unsafe.go", "stub_amd64.s", "stub_arm64.s", diff --git a/pkg/sentry/platform/ptrace/ptrace_amd64.go b/pkg/sentry/platform/ptrace/ptrace_amd64.go index db0212538..24fc5dc62 100644 --- a/pkg/sentry/platform/ptrace/ptrace_amd64.go +++ b/pkg/sentry/platform/ptrace/ptrace_amd64.go @@ -31,3 +31,17 @@ func fpRegSet(useXsave bool) uintptr { func stackPointer(r *syscall.PtraceRegs) uintptr { return uintptr(r.Rsp) } + +// x86 use the fs_base register to store the TLS pointer which can be +// get/set in "func (t *thread) get/setRegs(regs *syscall.PtraceRegs)". +// So both of the get/setTLS() operations are noop here. + +// getTLS gets the thread local storage register. +func (t *thread) getTLS(tls *uint64) error { + return nil +} + +// setTLS sets the thread local storage register. +func (t *thread) setTLS(tls *uint64) error { + return nil +} diff --git a/pkg/sentry/platform/ptrace/ptrace_arm64_unsafe.go b/pkg/sentry/platform/ptrace/ptrace_arm64_unsafe.go new file mode 100644 index 000000000..32b8a6be9 --- /dev/null +++ b/pkg/sentry/platform/ptrace/ptrace_arm64_unsafe.go @@ -0,0 +1,62 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// +build arm64 + +package ptrace + +import ( + "syscall" + "unsafe" + + "gvisor.dev/gvisor/pkg/abi/linux" +) + +// getTLS gets the thread local storage register. +func (t *thread) getTLS(tls *uint64) error { + iovec := syscall.Iovec{ + Base: (*byte)(unsafe.Pointer(tls)), + Len: uint64(unsafe.Sizeof(*tls)), + } + _, _, errno := syscall.RawSyscall6( + syscall.SYS_PTRACE, + syscall.PTRACE_GETREGSET, + uintptr(t.tid), + linux.NT_ARM_TLS, + uintptr(unsafe.Pointer(&iovec)), + 0, 0) + if errno != 0 { + return errno + } + return nil +} + +// setTLS sets the thread local storage register. +func (t *thread) setTLS(tls *uint64) error { + iovec := syscall.Iovec{ + Base: (*byte)(unsafe.Pointer(tls)), + Len: uint64(unsafe.Sizeof(*tls)), + } + _, _, errno := syscall.RawSyscall6( + syscall.SYS_PTRACE, + syscall.PTRACE_SETREGSET, + uintptr(t.tid), + linux.NT_ARM_TLS, + uintptr(unsafe.Pointer(&iovec)), + 0, 0) + if errno != 0 { + return errno + } + return nil +} diff --git a/pkg/sentry/platform/ptrace/subprocess.go b/pkg/sentry/platform/ptrace/subprocess.go index 31b7cec53..a644609ef 100644 --- a/pkg/sentry/platform/ptrace/subprocess.go +++ b/pkg/sentry/platform/ptrace/subprocess.go @@ -506,6 +506,9 @@ func (s *subprocess) switchToApp(c *context, ac arch.Context) bool { regs := &ac.StateData().Regs t.resetSysemuRegs(regs) + // Extract TLS register + tls := uint64(ac.TLS()) + // Check for interrupts, and ensure that future interrupts will signal t. if !c.interrupt.Enable(t) { // Pending interrupt; simulate. @@ -526,6 +529,9 @@ func (s *subprocess) switchToApp(c *context, ac arch.Context) bool { if err := t.setFPRegs(fpState, uint64(fpLen), useXsave); err != nil { panic(fmt.Sprintf("ptrace set fpregs (%+v) failed: %v", fpState, err)) } + if err := t.setTLS(&tls); err != nil { + panic(fmt.Sprintf("ptrace set tls (%+v) failed: %v", tls, err)) + } for { // Start running until the next system call. @@ -555,6 +561,12 @@ func (s *subprocess) switchToApp(c *context, ac arch.Context) bool { if err := t.getFPRegs(fpState, uint64(fpLen), useXsave); err != nil { panic(fmt.Sprintf("ptrace get fpregs failed: %v", err)) } + if err := t.getTLS(&tls); err != nil { + panic(fmt.Sprintf("ptrace get tls failed: %v", err)) + } + if !ac.SetTLS(uintptr(tls)) { + panic(fmt.Sprintf("tls value %v is invalid", tls)) + } // Is it a system call? if sig == (syscallEvent | syscall.SIGTRAP) { diff --git a/pkg/sentry/platform/ptrace/subprocess_amd64.go b/pkg/sentry/platform/ptrace/subprocess_amd64.go index e99798c56..cd74945e7 100644 --- a/pkg/sentry/platform/ptrace/subprocess_amd64.go +++ b/pkg/sentry/platform/ptrace/subprocess_amd64.go @@ -21,6 +21,7 @@ import ( "strings" "syscall" + "golang.org/x/sys/unix" "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/seccomp" "gvisor.dev/gvisor/pkg/sentry/arch" @@ -183,13 +184,76 @@ func enableCpuidFault() { // appendArchSeccompRules append architecture specific seccomp rules when creating BPF program. // Ref attachedThread() for more detail. -func appendArchSeccompRules(rules []seccomp.RuleSet) []seccomp.RuleSet { - return append(rules, seccomp.RuleSet{ - Rules: seccomp.SyscallRules{ - syscall.SYS_ARCH_PRCTL: []seccomp.Rule{ - {seccomp.AllowValue(linux.ARCH_SET_CPUID), seccomp.AllowValue(0)}, +func appendArchSeccompRules(rules []seccomp.RuleSet, defaultAction linux.BPFAction) []seccomp.RuleSet { + rules = append(rules, + // Rules for trapping vsyscall access. + seccomp.RuleSet{ + Rules: seccomp.SyscallRules{ + syscall.SYS_GETTIMEOFDAY: {}, + syscall.SYS_TIME: {}, + unix.SYS_GETCPU: {}, // SYS_GETCPU was not defined in package syscall on amd64. }, - }, - Action: linux.SECCOMP_RET_ALLOW, - }) + Action: linux.SECCOMP_RET_TRAP, + Vsyscall: true, + }) + if defaultAction != linux.SECCOMP_RET_ALLOW { + rules = append(rules, + seccomp.RuleSet{ + Rules: seccomp.SyscallRules{ + syscall.SYS_ARCH_PRCTL: []seccomp.Rule{ + {seccomp.AllowValue(linux.ARCH_SET_CPUID), seccomp.AllowValue(0)}, + }, + }, + Action: linux.SECCOMP_RET_ALLOW, + }) + } + return rules +} + +// probeSeccomp returns true iff seccomp is run after ptrace notifications, +// which is generally the case for kernel version >= 4.8. This check is dynamic +// because kernels have be backported behavior. +// +// See createStub for more information. +// +// Precondition: the runtime OS thread must be locked. +func probeSeccomp() bool { + // Create a completely new, destroyable process. + t, err := attachedThread(0, linux.SECCOMP_RET_ERRNO) + if err != nil { + panic(fmt.Sprintf("seccomp probe failed: %v", err)) + } + defer t.destroy() + + // Set registers to the yield system call. This call is not allowed + // by the filters specified in the attachThread function. + regs := createSyscallRegs(&t.initRegs, syscall.SYS_SCHED_YIELD) + if err := t.setRegs(®s); err != nil { + panic(fmt.Sprintf("ptrace set regs failed: %v", err)) + } + + for { + // Attempt an emulation. + if _, _, errno := syscall.RawSyscall6(syscall.SYS_PTRACE, unix.PTRACE_SYSEMU, uintptr(t.tid), 0, 0, 0, 0); errno != 0 { + panic(fmt.Sprintf("ptrace syscall-enter failed: %v", errno)) + } + + sig := t.wait(stopped) + if sig == (syscallEvent | syscall.SIGTRAP) { + // Did the seccomp errno hook already run? This would + // indicate that seccomp is first in line and we're + // less than 4.8. + if err := t.getRegs(®s); err != nil { + panic(fmt.Sprintf("ptrace get-regs failed: %v", err)) + } + if _, err := syscallReturnValue(®s); err == nil { + // The seccomp errno mode ran first, and reset + // the error in the registers. + return false + } + // The seccomp hook did not run yet, and therefore it + // is safe to use RET_KILL mode for dispatched calls. + return true + } + } } diff --git a/pkg/sentry/platform/ptrace/subprocess_arm64.go b/pkg/sentry/platform/ptrace/subprocess_arm64.go index 7b975137f..7f5c393f0 100644 --- a/pkg/sentry/platform/ptrace/subprocess_arm64.go +++ b/pkg/sentry/platform/ptrace/subprocess_arm64.go @@ -160,6 +160,15 @@ func enableCpuidFault() { // appendArchSeccompRules append architecture specific seccomp rules when creating BPF program. // Ref attachedThread() for more detail. -func appendArchSeccompRules(rules []seccomp.RuleSet) []seccomp.RuleSet { +func appendArchSeccompRules(rules []seccomp.RuleSet, defaultAction linux.BPFAction) []seccomp.RuleSet { return rules } + +// probeSeccomp returns true if seccomp is run after ptrace notifications, +// which is generally the case for kernel version >= 4.8. +// +// On arm64, the support of PTRACE_SYSEMU was added in the 5.3 kernel, so +// probeSeccomp can always return true. +func probeSeccomp() bool { + return true +} diff --git a/pkg/sentry/platform/ptrace/subprocess_linux.go b/pkg/sentry/platform/ptrace/subprocess_linux.go index 74968dfdf..2ce528601 100644 --- a/pkg/sentry/platform/ptrace/subprocess_linux.go +++ b/pkg/sentry/platform/ptrace/subprocess_linux.go @@ -20,7 +20,6 @@ import ( "fmt" "syscall" - "golang.org/x/sys/unix" "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/log" "gvisor.dev/gvisor/pkg/procid" @@ -30,54 +29,6 @@ import ( const syscallEvent syscall.Signal = 0x80 -// probeSeccomp returns true iff seccomp is run after ptrace notifications, -// which is generally the case for kernel version >= 4.8. This check is dynamic -// because kernels have be backported behavior. -// -// See createStub for more information. -// -// Precondition: the runtime OS thread must be locked. -func probeSeccomp() bool { - // Create a completely new, destroyable process. - t, err := attachedThread(0, linux.SECCOMP_RET_ERRNO) - if err != nil { - panic(fmt.Sprintf("seccomp probe failed: %v", err)) - } - defer t.destroy() - - // Set registers to the yield system call. This call is not allowed - // by the filters specified in the attachThread function. - regs := createSyscallRegs(&t.initRegs, syscall.SYS_SCHED_YIELD) - if err := t.setRegs(®s); err != nil { - panic(fmt.Sprintf("ptrace set regs failed: %v", err)) - } - - for { - // Attempt an emulation. - if _, _, errno := syscall.RawSyscall6(syscall.SYS_PTRACE, unix.PTRACE_SYSEMU, uintptr(t.tid), 0, 0, 0, 0); errno != 0 { - panic(fmt.Sprintf("ptrace syscall-enter failed: %v", errno)) - } - - sig := t.wait(stopped) - if sig == (syscallEvent | syscall.SIGTRAP) { - // Did the seccomp errno hook already run? This would - // indicate that seccomp is first in line and we're - // less than 4.8. - if err := t.getRegs(®s); err != nil { - panic(fmt.Sprintf("ptrace get-regs failed: %v", err)) - } - if _, err := syscallReturnValue(®s); err == nil { - // The seccomp errno mode ran first, and reset - // the error in the registers. - return false - } - // The seccomp hook did not run yet, and therefore it - // is safe to use RET_KILL mode for dispatched calls. - return true - } - } -} - // createStub creates a fresh stub processes. // // Precondition: the runtime OS thread must be locked. @@ -123,18 +74,7 @@ func attachedThread(flags uintptr, defaultAction linux.BPFAction) (*thread, erro // stub and all its children. This is used to create child stubs // (below), so we must include the ability to fork, but otherwise lock // down available calls only to what is needed. - rules := []seccomp.RuleSet{ - // Rules for trapping vsyscall access. - { - Rules: seccomp.SyscallRules{ - syscall.SYS_GETTIMEOFDAY: {}, - syscall.SYS_TIME: {}, - unix.SYS_GETCPU: {}, // SYS_GETCPU was not defined in package syscall on amd64. - }, - Action: linux.SECCOMP_RET_TRAP, - Vsyscall: true, - }, - } + rules := []seccomp.RuleSet{} if defaultAction != linux.SECCOMP_RET_ALLOW { rules = append(rules, seccomp.RuleSet{ Rules: seccomp.SyscallRules{ @@ -173,9 +113,8 @@ func attachedThread(flags uintptr, defaultAction linux.BPFAction) (*thread, erro }, Action: linux.SECCOMP_RET_ALLOW, }) - - rules = appendArchSeccompRules(rules) } + rules = appendArchSeccompRules(rules, defaultAction) instrs, err := seccomp.BuildProgram(rules, defaultAction) if err != nil { return nil, err diff --git a/pkg/sentry/platform/ring0/aarch64.go b/pkg/sentry/platform/ring0/aarch64.go index f6da41c27..8122ac6e2 100644 --- a/pkg/sentry/platform/ring0/aarch64.go +++ b/pkg/sentry/platform/ring0/aarch64.go @@ -27,26 +27,27 @@ const ( _PTE_PGT_BASE = 0x7000 _PTE_PGT_SIZE = 0x1000 - _PSR_MODE_EL0t = 0x0 - _PSR_MODE_EL1t = 0x4 - _PSR_MODE_EL1h = 0x5 - _PSR_EL_MASK = 0xf - - _PSR_D_BIT = 0x200 - _PSR_A_BIT = 0x100 - _PSR_I_BIT = 0x80 - _PSR_F_BIT = 0x40 + _PSR_D_BIT = 0x00000200 + _PSR_A_BIT = 0x00000100 + _PSR_I_BIT = 0x00000080 + _PSR_F_BIT = 0x00000040 ) const ( + // PSR bits + PSR_MODE_EL0t = 0x00000000 + PSR_MODE_EL1t = 0x00000004 + PSR_MODE_EL1h = 0x00000005 + PSR_MODE_MASK = 0x0000000f + // KernelFlagsSet should always be set in the kernel. - KernelFlagsSet = _PSR_MODE_EL1h + KernelFlagsSet = PSR_MODE_EL1h // UserFlagsSet are always set in userspace. - UserFlagsSet = _PSR_MODE_EL0t + UserFlagsSet = PSR_MODE_EL0t - KernelFlagsClear = _PSR_EL_MASK - UserFlagsClear = _PSR_EL_MASK + KernelFlagsClear = PSR_MODE_MASK + UserFlagsClear = PSR_MODE_MASK PsrDefaultSet = _PSR_D_BIT | _PSR_A_BIT | _PSR_I_BIT | _PSR_F_BIT ) diff --git a/pkg/sentry/platform/ring0/entry_arm64.s b/pkg/sentry/platform/ring0/entry_arm64.s index baa6c4910..d42eda37b 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 @@ -429,6 +467,7 @@ TEXT ·kernelExitToEl0(SB),NOSPLIT,$0 TEXT ·kernelExitToEl1(SB),NOSPLIT,$0 ERET() +// Start is the CPU entrypoint. TEXT ·Start(SB),NOSPLIT,$0 IRQ_DISABLE MOVD R8, RSV_REG @@ -437,18 +476,23 @@ 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 +528,10 @@ 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 +540,9 @@ 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 +554,19 @@ 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 +604,7 @@ 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 +618,7 @@ 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 +663,7 @@ 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 +677,7 @@ TEXT ·El0_fiq_invalid(SB),NOSPLIT,$0 TEXT ·El0_error_invalid(SB),NOSPLIT,$0 B ·Shutdown(SB) +// Vectors implements exception vector table. TEXT ·Vectors(SB),NOSPLIT,$0 B ·El1_sync_invalid(SB) nop31Instructions() diff --git a/pkg/sentry/platform/ring0/kernel_arm64.go b/pkg/sentry/platform/ring0/kernel_arm64.go index c3d341998..ccacaea6b 100644 --- a/pkg/sentry/platform/ring0/kernel_arm64.go +++ b/pkg/sentry/platform/ring0/kernel_arm64.go @@ -16,6 +16,14 @@ package ring0 +// HaltAndResume halts execution and point the pointer to the resume function. +//go:nosplit +func HaltAndResume() + +// HaltEl1SvcAndResume calls Hooks.KernelSyscall and resume. +//go:nosplit +func HaltEl1SvcAndResume() + // init initializes architecture-specific state. func (k *Kernel) init(opts KernelOpts) { // Save the root page tables. diff --git a/pkg/sentry/platform/ring0/pagetables/BUILD b/pkg/sentry/platform/ring0/pagetables/BUILD index 4f2406ce3..581841555 100644 --- a/pkg/sentry/platform/ring0/pagetables/BUILD +++ b/pkg/sentry/platform/ring0/pagetables/BUILD @@ -80,7 +80,7 @@ go_library( "pagetables_amd64.go", "pagetables_arm64.go", "pagetables_x86.go", - "pcids_x86.go", + "pcids.go", "walker_amd64.go", "walker_arm64.go", "walker_empty.go", diff --git a/pkg/sentry/platform/ring0/pagetables/pcids_x86.go b/pkg/sentry/platform/ring0/pagetables/pcids.go index e199bae18..9206030bf 100644 --- a/pkg/sentry/platform/ring0/pagetables/pcids_x86.go +++ b/pkg/sentry/platform/ring0/pagetables/pcids.go @@ -12,8 +12,6 @@ // See the License for the specific language governing permissions and // limitations under the License. -// +build i386 amd64 - package pagetables import ( 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/netfilter/netfilter.go b/pkg/sentry/socket/netfilter/netfilter.go index faa3e892a..b5b9be46f 100644 --- a/pkg/sentry/socket/netfilter/netfilter.go +++ b/pkg/sentry/socket/netfilter/netfilter.go @@ -36,6 +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 @@ -241,6 +246,8 @@ func marshalTarget(target iptables.Target) []byte { return marshalErrorTarget(tg.Name) case iptables.ReturnTarget: return marshalStandardTarget(iptables.RuleReturn) + case iptables.RedirectTarget: + return marshalRedirectTarget() case JumpTarget: return marshalJumpTarget(tg) default: @@ -277,6 +284,19 @@ func marshalErrorTarget(errorName string) []byte { return binary.Marshal(ret, usermem.ByteOrder, target) } +func marshalRedirectTarget() []byte { + // This is a redirect target named redirect + target := linux.XTRedirectTarget{ + Target: linux.XTEntryTarget{ + TargetSize: linux.SizeOfXTRedirectTarget, + }, + } + copy(target.Target.Name[:], redirectTargetName) + + ret := make([]byte, 0, linux.SizeOfXTRedirectTarget) + return binary.Marshal(ret, usermem.ByteOrder, target) +} + func marshalJumpTarget(jt JumpTarget) []byte { nflog("convert to binary: marshalling jump target") @@ -346,6 +366,8 @@ func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { switch replace.Name.String() { case iptables.TablenameFilter: table = iptables.EmptyFilterTable() + case iptables.TablenameNat: + table = iptables.EmptyNatTable() default: nflog("we don't yet support writing to the %q table (gvisor.dev/issue/170)", replace.Name.String()) return syserr.ErrInvalidArgument @@ -405,7 +427,7 @@ 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 @@ -496,10 +518,11 @@ func SetEntries(stack *stack.Stack, optVal []byte) *syserr.Error { } // TODO(gvisor.dev/issue/170): Support other chains. - // Since we only support modifying the INPUT chain right now, make sure - // all other chains point to ACCEPT rules. + // Since we only support modifying the INPUT chain and redirect for + // PREROUTING chain right now, make sure all other chains point to + // ACCEPT rules. for hook, ruleIdx := range table.BuiltinChains { - if hook != iptables.Input { + if hook != iptables.Input && hook != iptables.Prerouting { if _, ok := table.Rules[ruleIdx].Target.(iptables.AcceptTarget); !ok { nflog("hook %d is unsupported.", hook) return syserr.ErrInvalidArgument @@ -571,7 +594,7 @@ 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 iptables.IPHeaderFilter, optVal []byte) (iptables.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)) @@ -621,6 +644,55 @@ func parseTarget(optVal []byte) (iptables.Target, error) { nflog("set entries: user-defined target %q", name) return iptables.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 iptables.RedirectTarget. + var target iptables.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. @@ -643,7 +715,10 @@ func filterFromIPTIP(iptip linux.IPTIP) (iptables.IPHeaderFilter, error) { } func containsUnsupportedFields(iptip linux.IPTIP) bool { - // Currently we check that everything except protocol is zeroed. + // The following features are supported: + // - Protocol + // - Dst and DstMask + // - The inverse destination IP check flag var emptyInetAddr = linux.InetAddr{} var emptyInterface = [linux.IFNAMSIZ]byte{} // Disable any supported inverse flags. diff --git a/pkg/sentry/socket/netstack/netstack.go b/pkg/sentry/socket/netstack/netstack.go index e187276c5..13a9a60b4 100644 --- a/pkg/sentry/socket/netstack/netstack.go +++ b/pkg/sentry/socket/netstack/netstack.go @@ -712,14 +712,44 @@ func (s *SocketOperations) Connect(t *kernel.Task, sockaddr []byte, blocking boo // Bind implements the linux syscall bind(2) for sockets backed by // tcpip.Endpoint. func (s *SocketOperations) Bind(t *kernel.Task, sockaddr []byte) *syserr.Error { - addr, family, err := AddressAndFamily(sockaddr) - if err != nil { - return err + if len(sockaddr) < 2 { + return syserr.ErrInvalidArgument } - if err := s.checkFamily(family, true /* exact */); err != nil { - return err + + family := usermem.ByteOrder.Uint16(sockaddr) + var addr tcpip.FullAddress + + // Bind for AF_PACKET requires only family, protocol and ifindex. + // In function AddressAndFamily, we check the address length which is + // not needed for AF_PACKET bind. + if family == linux.AF_PACKET { + var a linux.SockAddrLink + if len(sockaddr) < sockAddrLinkSize { + return syserr.ErrInvalidArgument + } + binary.Unmarshal(sockaddr[:sockAddrLinkSize], usermem.ByteOrder, &a) + + if a.Protocol != uint16(s.protocol) { + return syserr.ErrInvalidArgument + } + + addr = tcpip.FullAddress{ + NIC: tcpip.NICID(a.InterfaceIndex), + Addr: tcpip.Address(a.HardwareAddr[:header.EthernetAddressSize]), + } + } else { + var err *syserr.Error + addr, family, err = AddressAndFamily(sockaddr) + if err != nil { + return err + } + + if err = s.checkFamily(family, true /* exact */); err != nil { + return err + } + + addr = s.mapFamily(addr, family) } - addr = s.mapFamily(addr, family) // Issue the bind request to the endpoint. return syserr.TranslateNetstackError(s.Endpoint.Bind(addr)) @@ -2637,7 +2667,9 @@ func (s *SocketOperations) Ioctl(ctx context.Context, _ *fs.File, io usermem.IO, } // Add bytes removed from the endpoint but not yet sent to the caller. + s.readMu.Lock() v += len(s.readView) + s.readMu.Unlock() if v > math.MaxInt32 { v = math.MaxInt32 diff --git a/pkg/sentry/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/syscalls/linux/BUILD b/pkg/sentry/syscalls/linux/BUILD index c7883e68e..0d24fd3c4 100644 --- a/pkg/sentry/syscalls/linux/BUILD +++ b/pkg/sentry/syscalls/linux/BUILD @@ -42,6 +42,8 @@ go_library( "sys_socket.go", "sys_splice.go", "sys_stat.go", + "sys_stat_amd64.go", + "sys_stat_arm64.go", "sys_sync.go", "sys_sysinfo.go", "sys_syslog.go", diff --git a/pkg/sentry/syscalls/linux/sys_epoll.go b/pkg/sentry/syscalls/linux/sys_epoll.go index fbef5b376..3ab93fbde 100644 --- a/pkg/sentry/syscalls/linux/sys_epoll.go +++ b/pkg/sentry/syscalls/linux/sys_epoll.go @@ -25,6 +25,8 @@ import ( "gvisor.dev/gvisor/pkg/waiter" ) +// LINT.IfChange + // EpollCreate1 implements the epoll_create1(2) linux syscall. func EpollCreate1(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { flags := args[0].Int() @@ -164,3 +166,5 @@ func EpollPwait(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sy return EpollWait(t, args) } + +// LINT.ThenChange(vfs2/epoll.go) diff --git a/pkg/sentry/syscalls/linux/sys_file.go b/pkg/sentry/syscalls/linux/sys_file.go index 421845ebb..35a98212a 100644 --- a/pkg/sentry/syscalls/linux/sys_file.go +++ b/pkg/sentry/syscalls/linux/sys_file.go @@ -130,6 +130,8 @@ func copyInPath(t *kernel.Task, addr usermem.Addr, allowEmpty bool) (path string return path, dirPath, nil } +// LINT.IfChange + func openAt(t *kernel.Task, dirFD int32, addr usermem.Addr, flags uint) (fd uintptr, err error) { path, dirPath, err := copyInPath(t, addr, false /* allowEmpty */) if err != nil { @@ -512,7 +514,7 @@ func (ac accessContext) Value(key interface{}) interface{} { } } -func accessAt(t *kernel.Task, dirFD int32, addr usermem.Addr, resolve bool, mode uint) error { +func accessAt(t *kernel.Task, dirFD int32, addr usermem.Addr, mode uint) error { const rOK = 4 const wOK = 2 const xOK = 1 @@ -527,7 +529,7 @@ func accessAt(t *kernel.Task, dirFD int32, addr usermem.Addr, resolve bool, mode return syserror.EINVAL } - return fileOpOn(t, dirFD, path, resolve, func(root *fs.Dirent, d *fs.Dirent, _ uint) error { + return fileOpOn(t, dirFD, path, true /* resolve */, func(root *fs.Dirent, d *fs.Dirent, _ uint) error { // access(2) and faccessat(2) check permissions using real // UID/GID, not effective UID/GID. // @@ -562,19 +564,29 @@ func Access(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal addr := args[0].Pointer() mode := args[1].ModeT() - return 0, nil, accessAt(t, linux.AT_FDCWD, addr, true, mode) + return 0, nil, accessAt(t, linux.AT_FDCWD, addr, mode) } // Faccessat implements linux syscall faccessat(2). +// +// Note that the faccessat() system call does not take a flags argument: +// "The raw faccessat() system call takes only the first three arguments. The +// AT_EACCESS and AT_SYMLINK_NOFOLLOW flags are actually implemented within +// the glibc wrapper function for faccessat(). If either of these flags is +// specified, then the wrapper function employs fstatat(2) to determine access +// permissions." - faccessat(2) func Faccessat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { dirFD := args[0].Int() addr := args[1].Pointer() mode := args[2].ModeT() - flags := args[3].Int() - return 0, nil, accessAt(t, dirFD, addr, flags&linux.AT_SYMLINK_NOFOLLOW == 0, mode) + return 0, nil, accessAt(t, dirFD, addr, mode) } +// LINT.ThenChange(vfs2/filesystem.go) + +// LINT.IfChange + // Ioctl implements linux syscall ioctl(2). func Ioctl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { fd := args[0].Int() @@ -650,6 +662,10 @@ func Ioctl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall } } +// LINT.ThenChange(vfs2/ioctl.go) + +// LINT.IfChange + // Getcwd implements the linux syscall getcwd(2). func Getcwd(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { addr := args[0].Pointer() @@ -760,6 +776,10 @@ func Fchdir(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal return 0, nil, nil } +// LINT.ThenChange(vfs2/fscontext.go) + +// LINT.IfChange + // Close implements linux syscall close(2). func Close(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { fd := args[0].Int() @@ -1094,6 +1114,8 @@ func Fcntl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall } } +// LINT.ThenChange(vfs2/fd.go) + const ( _FADV_NORMAL = 0 _FADV_RANDOM = 1 @@ -1141,6 +1163,8 @@ func Fadvise64(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sys return 0, nil, nil } +// LINT.IfChange + func mkdirAt(t *kernel.Task, dirFD int32, addr usermem.Addr, mode linux.FileMode) error { path, _, err := copyInPath(t, addr, false /* allowEmpty */) if err != nil { @@ -1218,7 +1242,7 @@ func rmdirAt(t *kernel.Task, dirFD int32, addr usermem.Addr) error { return syserror.ENOTEMPTY } - if err := fs.MayDelete(t, root, d, name); err != nil { + if err := d.MayDelete(t, root, name); err != nil { return err } @@ -1421,6 +1445,10 @@ func Linkat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal return 0, nil, linkAt(t, oldDirFD, oldAddr, newDirFD, newAddr, resolve, allowEmpty) } +// LINT.ThenChange(vfs2/filesystem.go) + +// LINT.IfChange + func readlinkAt(t *kernel.Task, dirFD int32, addr usermem.Addr, bufAddr usermem.Addr, size uint) (copied uintptr, err error) { path, dirPath, err := copyInPath(t, addr, false /* allowEmpty */) if err != nil { @@ -1480,6 +1508,10 @@ func Readlinkat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sy return n, nil, err } +// LINT.ThenChange(vfs2/stat.go) + +// LINT.IfChange + func unlinkAt(t *kernel.Task, dirFD int32, addr usermem.Addr) error { path, dirPath, err := copyInPath(t, addr, false /* allowEmpty */) if err != nil { @@ -1491,7 +1523,7 @@ func unlinkAt(t *kernel.Task, dirFD int32, addr usermem.Addr) error { return syserror.ENOTDIR } - if err := fs.MayDelete(t, root, d, name); err != nil { + if err := d.MayDelete(t, root, name); err != nil { return err } @@ -1516,6 +1548,10 @@ func Unlinkat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysc return 0, nil, unlinkAt(t, dirFD, addr) } +// LINT.ThenChange(vfs2/filesystem.go) + +// LINT.IfChange + // Truncate implements linux syscall truncate(2). func Truncate(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { addr := args[0].Pointer() @@ -1614,6 +1650,8 @@ func Ftruncate(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sys return 0, nil, nil } +// LINT.ThenChange(vfs2/setstat.go) + // Umask implements linux syscall umask(2). func Umask(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { mask := args[0].ModeT() @@ -1621,6 +1659,8 @@ func Umask(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall return uintptr(mask), nil, nil } +// LINT.IfChange + // Change ownership of a file. // // uid and gid may be -1, in which case they will not be changed. @@ -1987,6 +2027,10 @@ func Futimesat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sys return 0, nil, utimes(t, dirFD, pathnameAddr, ts, true) } +// LINT.ThenChange(vfs2/setstat.go) + +// LINT.IfChange + func renameAt(t *kernel.Task, oldDirFD int32, oldAddr usermem.Addr, newDirFD int32, newAddr usermem.Addr) error { newPath, _, err := copyInPath(t, newAddr, false /* allowEmpty */) if err != nil { @@ -2042,6 +2086,8 @@ func Renameat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysc return 0, nil, renameAt(t, oldDirFD, oldPathAddr, newDirFD, newPathAddr) } +// LINT.ThenChange(vfs2/filesystem.go) + // Fallocate implements linux system call fallocate(2). func Fallocate(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { fd := args[0].Int() diff --git a/pkg/sentry/syscalls/linux/sys_getdents.go b/pkg/sentry/syscalls/linux/sys_getdents.go index f66f4ffde..b126fecc0 100644 --- a/pkg/sentry/syscalls/linux/sys_getdents.go +++ b/pkg/sentry/syscalls/linux/sys_getdents.go @@ -27,6 +27,8 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) +// LINT.IfChange + // Getdents implements linux syscall getdents(2) for 64bit systems. func Getdents(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { fd := args[0].Int() @@ -244,3 +246,5 @@ func (ds *direntSerializer) CopyOut(name string, attr fs.DentAttr) error { func (ds *direntSerializer) Written() int { return ds.written } + +// LINT.ThenChange(vfs2/getdents.go) diff --git a/pkg/sentry/syscalls/linux/sys_lseek.go b/pkg/sentry/syscalls/linux/sys_lseek.go index 297e920c4..3f7691eae 100644 --- a/pkg/sentry/syscalls/linux/sys_lseek.go +++ b/pkg/sentry/syscalls/linux/sys_lseek.go @@ -21,6 +21,8 @@ import ( "gvisor.dev/gvisor/pkg/syserror" ) +// LINT.IfChange + // Lseek implements linux syscall lseek(2). func Lseek(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { fd := args[0].Int() @@ -52,3 +54,5 @@ func Lseek(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall } return uintptr(offset), nil, err } + +// LINT.ThenChange(vfs2/read_write.go) diff --git a/pkg/sentry/syscalls/linux/sys_mmap.go b/pkg/sentry/syscalls/linux/sys_mmap.go index 9959f6e61..91694d374 100644 --- a/pkg/sentry/syscalls/linux/sys_mmap.go +++ b/pkg/sentry/syscalls/linux/sys_mmap.go @@ -35,6 +35,8 @@ func Brk(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallCo return uintptr(addr), nil, nil } +// LINT.IfChange + // Mmap implements linux syscall mmap(2). func Mmap(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { prot := args[2].Int() @@ -104,6 +106,8 @@ func Mmap(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallC return uintptr(rv), nil, err } +// LINT.ThenChange(vfs2/mmap.go) + // Munmap implements linux syscall munmap(2). func Munmap(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { return 0, nil, t.MemoryManager().MUnmap(t, args[0].Pointer(), args[1].Uint64()) diff --git a/pkg/sentry/syscalls/linux/sys_read.go b/pkg/sentry/syscalls/linux/sys_read.go index 227692f06..78a2cb750 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 @@ -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_stat.go b/pkg/sentry/syscalls/linux/sys_stat.go index 11f25e00d..a11a87cd1 100644 --- a/pkg/sentry/syscalls/linux/sys_stat.go +++ b/pkg/sentry/syscalls/linux/sys_stat.go @@ -23,23 +23,7 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) -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.IfChange // Stat implements linux syscall stat(2). func Stat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { @@ -152,7 +136,10 @@ 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 { @@ -297,3 +284,5 @@ 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 new file mode 100644 index 000000000..0a04a6113 --- /dev/null +++ b/pkg/sentry/syscalls/linux/sys_stat_amd64.go @@ -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 amd64 + +package linux + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/kernel" +) + +// 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 new file mode 100644 index 000000000..5a3b1bfad --- /dev/null +++ b/pkg/sentry/syscalls/linux/sys_stat_arm64.go @@ -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 + +package linux + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/kernel" +) + +// 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_write.go b/pkg/sentry/syscalls/linux/sys_write.go index aba892939..506ee54ce 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. // @@ -358,3 +360,5 @@ func pwritev(t *kernel.Task, f *fs.File, src usermem.IOSequence, offset int64) ( return total, err } + +// LINT.ThenChange(vfs2/read_write.go) diff --git a/pkg/sentry/syscalls/linux/sys_xattr.go b/pkg/sentry/syscalls/linux/sys_xattr.go index 9d8140b8a..2de5e3422 100644 --- a/pkg/sentry/syscalls/linux/sys_xattr.go +++ b/pkg/sentry/syscalls/linux/sys_xattr.go @@ -25,6 +25,8 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) +// LINT.IfChange + // GetXattr implements linux syscall getxattr(2). func GetXattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { return getXattrFromPath(t, args, true) @@ -418,3 +420,5 @@ func removeXattr(t *kernel.Task, d *fs.Dirent, nameAddr usermem.Addr) error { return d.Inode.RemoveXattr(t, d, name) } + +// LINT.ThenChange(vfs2/xattr.go) diff --git a/pkg/sentry/syscalls/linux/vfs2/BUILD b/pkg/sentry/syscalls/linux/vfs2/BUILD index 6b8a00b6e..e7695e995 100644 --- a/pkg/sentry/syscalls/linux/vfs2/BUILD +++ b/pkg/sentry/syscalls/linux/vfs2/BUILD @@ -5,18 +5,46 @@ package(licenses = ["notice"]) go_library( name = "vfs2", srcs = [ + "epoll.go", + "epoll_unsafe.go", + "execve.go", + "fd.go", + "filesystem.go", + "fscontext.go", + "getdents.go", + "ioctl.go", "linux64.go", "linux64_override_amd64.go", "linux64_override_arm64.go", - "sys_read.go", + "mmap.go", + "path.go", + "poll.go", + "read_write.go", + "setstat.go", + "stat.go", + "stat_amd64.go", + "stat_arm64.go", + "sync.go", + "xattr.go", ], + marshal = True, visibility = ["//:sandbox"], deps = [ + "//pkg/abi/linux", + "//pkg/fspath", + "//pkg/gohacks", "//pkg/sentry/arch", + "//pkg/sentry/fsbridge", "//pkg/sentry/kernel", + "//pkg/sentry/kernel/auth", + "//pkg/sentry/kernel/time", + "//pkg/sentry/limits", + "//pkg/sentry/loader", + "//pkg/sentry/memmap", "//pkg/sentry/syscalls", "//pkg/sentry/syscalls/linux", "//pkg/sentry/vfs", + "//pkg/sync", "//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..d6cb0e79a --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/epoll.go @@ -0,0 +1,225 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "math" + "time" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" + "gvisor.dev/gvisor/pkg/waiter" +) + +// EpollCreate1 implements Linux syscall epoll_create1(2). +func EpollCreate1(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + flags := args[0].Int() + if flags&^linux.EPOLL_CLOEXEC != 0 { + return 0, nil, syserror.EINVAL + } + + file, err := t.Kernel().VFS().NewEpollInstanceFD() + if err != nil { + return 0, nil, err + } + defer file.DecRef() + + fd, err := t.NewFDFromVFS2(0, file, kernel.FDFlags{ + CloseOnExec: flags&linux.EPOLL_CLOEXEC != 0, + }) + if err != nil { + return 0, nil, err + } + return uintptr(fd), nil, nil +} + +// EpollCreate implements Linux syscall epoll_create(2). +func EpollCreate(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + size := args[0].Int() + + // "Since Linux 2.6.8, the size argument is ignored, but must be greater + // than zero" - epoll_create(2) + if size <= 0 { + return 0, nil, syserror.EINVAL + } + + file, err := t.Kernel().VFS().NewEpollInstanceFD() + if err != nil { + return 0, nil, err + } + defer file.DecRef() + + fd, err := t.NewFDFromVFS2(0, file, kernel.FDFlags{}) + if err != nil { + return 0, nil, err + } + return uintptr(fd), nil, nil +} + +// EpollCtl implements Linux syscall epoll_ctl(2). +func EpollCtl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + epfd := args[0].Int() + op := args[1].Int() + fd := args[2].Int() + eventAddr := args[3].Pointer() + + epfile := t.GetFileVFS2(epfd) + if epfile == nil { + return 0, nil, syserror.EBADF + } + defer epfile.DecRef() + ep, ok := epfile.Impl().(*vfs.EpollInstance) + if !ok { + return 0, nil, syserror.EINVAL + } + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + if epfile == file { + return 0, nil, syserror.EINVAL + } + + var event linux.EpollEvent + switch op { + case linux.EPOLL_CTL_ADD: + if err := event.CopyIn(t, eventAddr); err != nil { + return 0, nil, err + } + return 0, nil, ep.AddInterest(file, fd, event) + case linux.EPOLL_CTL_DEL: + return 0, nil, ep.DeleteInterest(file, fd) + case linux.EPOLL_CTL_MOD: + if err := event.CopyIn(t, eventAddr); err != nil { + return 0, nil, err + } + return 0, nil, ep.ModifyInterest(file, fd, event) + default: + return 0, nil, syserror.EINVAL + } +} + +// EpollWait implements Linux syscall epoll_wait(2). +func EpollWait(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + epfd := args[0].Int() + eventsAddr := args[1].Pointer() + maxEvents := int(args[2].Int()) + timeout := int(args[3].Int()) + + const _EP_MAX_EVENTS = math.MaxInt32 / sizeofEpollEvent // Linux: fs/eventpoll.c:EP_MAX_EVENTS + if maxEvents <= 0 || maxEvents > _EP_MAX_EVENTS { + return 0, nil, syserror.EINVAL + } + + epfile := t.GetFileVFS2(epfd) + if epfile == nil { + return 0, nil, syserror.EBADF + } + defer epfile.DecRef() + ep, ok := epfile.Impl().(*vfs.EpollInstance) + if !ok { + return 0, nil, syserror.EINVAL + } + + // Use a fixed-size buffer in a loop, instead of make([]linux.EpollEvent, + // maxEvents), so that the buffer can be allocated on the stack. + var ( + events [16]linux.EpollEvent + total int + ch chan struct{} + haveDeadline bool + deadline ktime.Time + ) + for { + batchEvents := len(events) + if batchEvents > maxEvents { + batchEvents = maxEvents + } + n := ep.ReadEvents(events[:batchEvents]) + maxEvents -= n + if n != 0 { + // Copy what we read out. + copiedEvents, err := copyOutEvents(t, eventsAddr, events[:n]) + eventsAddr += usermem.Addr(copiedEvents * sizeofEpollEvent) + total += copiedEvents + if err != nil { + if total != 0 { + return uintptr(total), nil, nil + } + return 0, nil, err + } + // If we've filled the application's event buffer, we're done. + if maxEvents == 0 { + return uintptr(total), nil, nil + } + // Loop if we read a full batch, under the expectation that there + // may be more events to read. + if n == batchEvents { + continue + } + } + // We get here if n != batchEvents. If we read any number of events + // (just now, or in a previous iteration of this loop), or if timeout + // is 0 (such that epoll_wait should be non-blocking), return the + // events we've read so far to the application. + if total != 0 || timeout == 0 { + return uintptr(total), nil, nil + } + // In the first iteration of this loop, register with the epoll + // instance for readability events, but then immediately continue the + // loop since we need to retry ReadEvents() before blocking. In all + // subsequent iterations, block until events are available, the timeout + // expires, or an interrupt arrives. + if ch == nil { + var w waiter.Entry + w, ch = waiter.NewChannelEntry(nil) + epfile.EventRegister(&w, waiter.EventIn) + defer epfile.EventUnregister(&w) + } else { + // Set up the timer if a timeout was specified. + if timeout > 0 && !haveDeadline { + timeoutDur := time.Duration(timeout) * time.Millisecond + deadline = t.Kernel().MonotonicClock().Now().Add(timeoutDur) + haveDeadline = true + } + if err := t.BlockWithDeadline(ch, haveDeadline, deadline); err != nil { + if err == syserror.ETIMEDOUT { + err = nil + } + // total must be 0 since otherwise we would have returned + // above. + return 0, nil, err + } + } + } +} + +// EpollPwait implements Linux syscall epoll_pwait(2). +func EpollPwait(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + maskAddr := args[4].Pointer() + maskSize := uint(args[5].Uint()) + + if err := setTempSignalSet(t, maskAddr, maskSize); err != nil { + return 0, nil, err + } + + return EpollWait(t, args) +} diff --git a/pkg/sentry/syscalls/linux/vfs2/epoll_unsafe.go b/pkg/sentry/syscalls/linux/vfs2/epoll_unsafe.go new file mode 100644 index 000000000..825f325bf --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/epoll_unsafe.go @@ -0,0 +1,44 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "reflect" + "runtime" + "unsafe" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/gohacks" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/usermem" +) + +const sizeofEpollEvent = int(unsafe.Sizeof(linux.EpollEvent{})) + +func copyOutEvents(t *kernel.Task, addr usermem.Addr, events []linux.EpollEvent) (int, error) { + if len(events) == 0 { + return 0, nil + } + // Cast events to a byte slice for copying. + var eventBytes []byte + eventBytesHdr := (*reflect.SliceHeader)(unsafe.Pointer(&eventBytes)) + eventBytesHdr.Data = uintptr(gohacks.Noescape(unsafe.Pointer(&events[0]))) + eventBytesHdr.Len = len(events) * sizeofEpollEvent + eventBytesHdr.Cap = len(events) * sizeofEpollEvent + copiedBytes, err := t.CopyOutBytes(addr, eventBytes) + runtime.KeepAlive(events) + copiedEvents := copiedBytes / sizeofEpollEvent // rounded down + return copiedEvents, err +} diff --git a/pkg/sentry/syscalls/linux/vfs2/execve.go b/pkg/sentry/syscalls/linux/vfs2/execve.go new file mode 100644 index 000000000..aef0078a8 --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/execve.go @@ -0,0 +1,137 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/loader" + slinux "gvisor.dev/gvisor/pkg/sentry/syscalls/linux" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +// Execve implements linux syscall execve(2). +func Execve(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + pathnameAddr := args[0].Pointer() + argvAddr := args[1].Pointer() + envvAddr := args[2].Pointer() + return execveat(t, linux.AT_FDCWD, pathnameAddr, argvAddr, envvAddr, 0 /* flags */) +} + +// Execveat implements linux syscall execveat(2). +func Execveat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + dirfd := args[0].Int() + pathnameAddr := args[1].Pointer() + argvAddr := args[2].Pointer() + envvAddr := args[3].Pointer() + flags := args[4].Int() + return execveat(t, dirfd, pathnameAddr, argvAddr, envvAddr, flags) +} + +func execveat(t *kernel.Task, dirfd int32, pathnameAddr, argvAddr, envvAddr usermem.Addr, flags int32) (uintptr, *kernel.SyscallControl, error) { + if flags&^(linux.AT_EMPTY_PATH|linux.AT_SYMLINK_NOFOLLOW) != 0 { + return 0, nil, syserror.EINVAL + } + + pathname, err := t.CopyInString(pathnameAddr, linux.PATH_MAX) + if err != nil { + return 0, nil, err + } + var argv, envv []string + if argvAddr != 0 { + var err error + argv, err = t.CopyInVector(argvAddr, slinux.ExecMaxElemSize, slinux.ExecMaxTotalSize) + if err != nil { + return 0, nil, err + } + } + if envvAddr != 0 { + var err error + envv, err = t.CopyInVector(envvAddr, slinux.ExecMaxElemSize, slinux.ExecMaxTotalSize) + if err != nil { + return 0, nil, err + } + } + + root := t.FSContext().RootDirectoryVFS2() + defer root.DecRef() + var executable fsbridge.File + closeOnExec := false + if path := fspath.Parse(pathname); dirfd != linux.AT_FDCWD && !path.Absolute { + // We must open the executable ourselves since dirfd is used as the + // starting point while resolving path, but the task working directory + // is used as the starting point while resolving interpreters (Linux: + // fs/binfmt_script.c:load_script() => fs/exec.c:open_exec() => + // do_open_execat(fd=AT_FDCWD)), and the loader package is currently + // incapable of handling this correctly. + if !path.HasComponents() && flags&linux.AT_EMPTY_PATH == 0 { + return 0, nil, syserror.ENOENT + } + dirfile, dirfileFlags := t.FDTable().GetVFS2(dirfd) + if dirfile == nil { + return 0, nil, syserror.EBADF + } + start := dirfile.VirtualDentry() + start.IncRef() + dirfile.DecRef() + closeOnExec = dirfileFlags.CloseOnExec + file, err := t.Kernel().VFS().OpenAt(t, t.Credentials(), &vfs.PathOperation{ + Root: root, + Start: start, + Path: path, + FollowFinalSymlink: flags&linux.AT_SYMLINK_NOFOLLOW == 0, + }, &vfs.OpenOptions{ + Flags: linux.O_RDONLY, + FileExec: true, + }) + start.DecRef() + if err != nil { + return 0, nil, err + } + defer file.DecRef() + executable = fsbridge.NewVFSFile(file) + } + + // Load the new TaskContext. + mntns := t.MountNamespaceVFS2() // FIXME(jamieliu): useless refcount change + defer mntns.DecRef() + wd := t.FSContext().WorkingDirectoryVFS2() + defer wd.DecRef() + remainingTraversals := uint(linux.MaxSymlinkTraversals) + loadArgs := loader.LoadArgs{ + Opener: fsbridge.NewVFSLookup(mntns, root, wd), + RemainingTraversals: &remainingTraversals, + ResolveFinal: flags&linux.AT_SYMLINK_NOFOLLOW == 0, + Filename: pathname, + File: executable, + CloseOnExec: closeOnExec, + Argv: argv, + Envv: envv, + Features: t.Arch().FeatureSet(), + } + + tc, se := t.Kernel().LoadTaskImage(t, loadArgs) + if se != nil { + return 0, nil, se.ToError() + } + + ctrl, err := t.Execve(tc) + return 0, ctrl, err +} diff --git a/pkg/sentry/syscalls/linux/vfs2/fd.go b/pkg/sentry/syscalls/linux/vfs2/fd.go new file mode 100644 index 000000000..3afcea665 --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/fd.go @@ -0,0 +1,147 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + slinux "gvisor.dev/gvisor/pkg/sentry/syscalls/linux" + "gvisor.dev/gvisor/pkg/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()) + 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..fc5ceea4c --- /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, + 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, followFinalSymlink) + 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..ddc140b65 --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/getdents.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 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) + 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) + buf[size-1] = 0 // NUL terminator + } else { + // struct linux_dirent { + // unsigned long d_ino; /* Inode number */ + // unsigned long d_off; /* Offset to next linux_dirent */ + // unsigned short d_reclen; /* Length of this linux_dirent */ + // char d_name[]; /* Filename (null-terminated) */ + // /* length is actually (d_reclen - 2 - + // offsetof(struct linux_dirent, d_name)) */ + // /* + // char pad; // Zero padding byte + // char d_type; // File type (only since Linux + // // 2.6.4); offset is (d_reclen - 1) + // */ + // }; + if cb.t.Arch().Width() != 8 { + panic(fmt.Sprintf("unsupported sizeof(unsigned long): %d", cb.t.Arch().Width())) + } + size := 8 + 8 + 2 + 1 + 1 + 1 + len(dirent.Name) + 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) + buf[size-3] = 0 // NUL terminator + buf[size-2] = 0 // zero padding byte + 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/kernel/pipe/buffer_test.go b/pkg/sentry/syscalls/linux/vfs2/ioctl.go index 4d54b8b8f..5a2418da9 100644 --- a/pkg/sentry/kernel/pipe/buffer_test.go +++ b/pkg/sentry/syscalls/linux/vfs2/ioctl.go @@ -1,4 +1,4 @@ -// Copyright 2019 The gVisor Authors. +// Copyright 2020 The gVisor Authors. // // Licensed under the Apache License, Version 2.0 (the "License"); // you may not use this file except in compliance with the License. @@ -12,21 +12,24 @@ // See the License for the specific language governing permissions and // limitations under the License. -package pipe +package vfs2 import ( - "testing" - "unsafe" - - "gvisor.dev/gvisor/pkg/usermem" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/syserror" ) -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") +// Ioctl implements Linux syscall ioctl(2). +func Ioctl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF } + defer file.DecRef() + + ret, err := file.Ioctl(t, t.MemoryManager(), args) + return ret, nil, err } diff --git a/pkg/sentry/syscalls/linux/vfs2/linux64_override_amd64.go b/pkg/sentry/syscalls/linux/vfs2/linux64_override_amd64.go index e0ac32b33..7d220bc20 100644 --- a/pkg/sentry/syscalls/linux/vfs2/linux64_override_amd64.go +++ b/pkg/sentry/syscalls/linux/vfs2/linux64_override_amd64.go @@ -12,6 +12,8 @@ // See the License for the specific language governing permissions and // limitations under the License. +// +build amd64 + package vfs2 import ( @@ -22,110 +24,142 @@ import ( // Override syscall table to add syscalls implementations from this package. func Override(table map[uintptr]kernel.Syscall) { table[0] = syscalls.Supported("read", Read) - - // Remove syscalls that haven't been converted yet. It's better to get ENOSYS - // rather than a SIGSEGV deep in the stack. - delete(table, 1) // write - delete(table, 2) // open - delete(table, 3) // close - delete(table, 4) // stat - delete(table, 5) // fstat - delete(table, 6) // lstat - delete(table, 7) // poll - delete(table, 8) // lseek - delete(table, 9) // mmap - delete(table, 16) // ioctl - delete(table, 17) // pread64 - delete(table, 18) // pwrite64 - delete(table, 19) // readv - delete(table, 20) // writev - delete(table, 21) // access - delete(table, 22) // pipe - delete(table, 32) // dup - delete(table, 33) // dup2 - delete(table, 40) // sendfile - delete(table, 59) // execve - delete(table, 72) // fcntl - delete(table, 73) // flock - delete(table, 74) // fsync - delete(table, 75) // fdatasync - delete(table, 76) // truncate - delete(table, 77) // ftruncate - delete(table, 78) // getdents - delete(table, 79) // getcwd - delete(table, 80) // chdir - delete(table, 81) // fchdir - delete(table, 82) // rename - delete(table, 83) // mkdir - delete(table, 84) // rmdir - delete(table, 85) // creat - delete(table, 86) // link - delete(table, 87) // unlink - delete(table, 88) // symlink - delete(table, 89) // readlink - delete(table, 90) // chmod - delete(table, 91) // fchmod - delete(table, 92) // chown - delete(table, 93) // fchown - delete(table, 94) // lchown - delete(table, 133) // mknod - delete(table, 137) // statfs - delete(table, 138) // fstatfs - delete(table, 161) // chroot - delete(table, 162) // sync + table[1] = syscalls.Supported("write", Write) + table[2] = syscalls.Supported("open", Open) + table[3] = syscalls.Supported("close", Close) + table[4] = syscalls.Supported("stat", Stat) + table[5] = syscalls.Supported("fstat", Fstat) + table[6] = syscalls.Supported("lstat", Lstat) + table[7] = syscalls.Supported("poll", Poll) + table[8] = syscalls.Supported("lseek", Lseek) + table[9] = syscalls.Supported("mmap", Mmap) + table[16] = syscalls.Supported("ioctl", Ioctl) + table[17] = syscalls.Supported("pread64", Pread64) + table[18] = syscalls.Supported("pwrite64", Pwrite64) + table[19] = syscalls.Supported("readv", Readv) + table[20] = syscalls.Supported("writev", Writev) + table[21] = syscalls.Supported("access", Access) + delete(table, 22) // pipe + table[23] = syscalls.Supported("select", Select) + table[32] = syscalls.Supported("dup", Dup) + table[33] = syscalls.Supported("dup2", Dup2) + delete(table, 40) // sendfile + delete(table, 41) // socket + delete(table, 42) // connect + delete(table, 43) // accept + delete(table, 44) // sendto + delete(table, 45) // recvfrom + delete(table, 46) // sendmsg + delete(table, 47) // recvmsg + delete(table, 48) // shutdown + delete(table, 49) // bind + delete(table, 50) // listen + delete(table, 51) // getsockname + delete(table, 52) // getpeername + delete(table, 53) // socketpair + delete(table, 54) // setsockopt + delete(table, 55) // getsockopt + table[59] = syscalls.Supported("execve", Execve) + table[72] = syscalls.Supported("fcntl", Fcntl) + delete(table, 73) // flock + table[74] = syscalls.Supported("fsync", Fsync) + table[75] = syscalls.Supported("fdatasync", Fdatasync) + table[76] = syscalls.Supported("truncate", Truncate) + table[77] = syscalls.Supported("ftruncate", Ftruncate) + table[78] = syscalls.Supported("getdents", Getdents) + table[79] = syscalls.Supported("getcwd", Getcwd) + table[80] = syscalls.Supported("chdir", Chdir) + table[81] = syscalls.Supported("fchdir", Fchdir) + table[82] = syscalls.Supported("rename", Rename) + table[83] = syscalls.Supported("mkdir", Mkdir) + table[84] = syscalls.Supported("rmdir", Rmdir) + table[85] = syscalls.Supported("creat", Creat) + table[86] = syscalls.Supported("link", Link) + table[87] = syscalls.Supported("unlink", Unlink) + table[88] = syscalls.Supported("symlink", Symlink) + table[89] = syscalls.Supported("readlink", Readlink) + table[90] = syscalls.Supported("chmod", Chmod) + table[91] = syscalls.Supported("fchmod", Fchmod) + table[92] = syscalls.Supported("chown", Chown) + table[93] = syscalls.Supported("fchown", Fchown) + table[94] = syscalls.Supported("lchown", Lchown) + table[132] = syscalls.Supported("utime", Utime) + table[133] = syscalls.Supported("mknod", Mknod) + table[137] = syscalls.Supported("statfs", Statfs) + table[138] = syscalls.Supported("fstatfs", Fstatfs) + table[161] = syscalls.Supported("chroot", Chroot) + table[162] = syscalls.Supported("sync", Sync) delete(table, 165) // mount delete(table, 166) // umount2 - delete(table, 172) // iopl - delete(table, 173) // ioperm delete(table, 187) // readahead - delete(table, 188) // setxattr - delete(table, 189) // lsetxattr - delete(table, 190) // fsetxattr - delete(table, 191) // getxattr - delete(table, 192) // lgetxattr - delete(table, 193) // fgetxattr + table[188] = syscalls.Supported("setxattr", Setxattr) + table[189] = syscalls.Supported("lsetxattr", Lsetxattr) + table[190] = syscalls.Supported("fsetxattr", Fsetxattr) + table[191] = syscalls.Supported("getxattr", Getxattr) + table[192] = syscalls.Supported("lgetxattr", Lgetxattr) + table[193] = syscalls.Supported("fgetxattr", Fgetxattr) + table[194] = syscalls.Supported("listxattr", Listxattr) + table[195] = syscalls.Supported("llistxattr", Llistxattr) + table[196] = syscalls.Supported("flistxattr", Flistxattr) + table[197] = syscalls.Supported("removexattr", Removexattr) + table[198] = syscalls.Supported("lremovexattr", Lremovexattr) + table[199] = syscalls.Supported("fremovexattr", Fremovexattr) delete(table, 206) // io_setup delete(table, 207) // io_destroy delete(table, 208) // io_getevents delete(table, 209) // io_submit delete(table, 210) // io_cancel - delete(table, 213) // epoll_create - delete(table, 214) // epoll_ctl_old - delete(table, 215) // epoll_wait_old - delete(table, 216) // remap_file_pages - delete(table, 217) // getdents64 - delete(table, 232) // epoll_wait - delete(table, 233) // epoll_ctl + table[213] = syscalls.Supported("epoll_create", EpollCreate) + table[217] = syscalls.Supported("getdents64", Getdents64) + delete(table, 221) // fdavise64 + table[232] = syscalls.Supported("epoll_wait", EpollWait) + table[233] = syscalls.Supported("epoll_ctl", EpollCtl) + table[235] = syscalls.Supported("utimes", Utimes) delete(table, 253) // inotify_init delete(table, 254) // inotify_add_watch delete(table, 255) // inotify_rm_watch - delete(table, 257) // openat - delete(table, 258) // mkdirat - delete(table, 259) // mknodat - delete(table, 260) // fchownat - delete(table, 261) // futimesat - delete(table, 262) // fstatat - delete(table, 263) // unlinkat - delete(table, 264) // renameat - delete(table, 265) // linkat - delete(table, 266) // symlinkat - delete(table, 267) // readlinkat - delete(table, 268) // fchmodat - delete(table, 269) // faccessat - delete(table, 270) // pselect - delete(table, 271) // ppoll + table[257] = syscalls.Supported("openat", Openat) + table[258] = syscalls.Supported("mkdirat", Mkdirat) + table[259] = syscalls.Supported("mknodat", Mknodat) + table[260] = syscalls.Supported("fchownat", Fchownat) + table[261] = syscalls.Supported("futimens", Futimens) + table[262] = syscalls.Supported("newfstatat", Newfstatat) + table[263] = syscalls.Supported("unlinkat", Unlinkat) + table[264] = syscalls.Supported("renameat", Renameat) + table[265] = syscalls.Supported("linkat", Linkat) + table[266] = syscalls.Supported("symlinkat", Symlinkat) + table[267] = syscalls.Supported("readlinkat", Readlinkat) + table[268] = syscalls.Supported("fchmodat", Fchmodat) + table[269] = syscalls.Supported("faccessat", Faccessat) + table[270] = syscalls.Supported("pselect", Pselect) + table[271] = syscalls.Supported("ppoll", Ppoll) + delete(table, 275) // splice + delete(table, 276) // tee + table[277] = syscalls.Supported("sync_file_range", SyncFileRange) + table[280] = syscalls.Supported("utimensat", Utimensat) + table[281] = syscalls.Supported("epoll_pwait", EpollPwait) + delete(table, 282) // signalfd + delete(table, 283) // timerfd_create + delete(table, 284) // eventfd delete(table, 285) // fallocate - delete(table, 291) // epoll_create1 - delete(table, 292) // dup3 + delete(table, 286) // timerfd_settime + delete(table, 287) // timerfd_gettime + delete(table, 288) // accept4 + delete(table, 289) // signalfd4 + delete(table, 290) // eventfd2 + table[291] = syscalls.Supported("epoll_create1", EpollCreate1) + table[292] = syscalls.Supported("dup3", Dup3) delete(table, 293) // pipe2 delete(table, 294) // inotify_init1 - delete(table, 295) // preadv - delete(table, 296) // pwritev - delete(table, 306) // syncfs - delete(table, 316) // renameat2 + table[295] = syscalls.Supported("preadv", Preadv) + table[296] = syscalls.Supported("pwritev", Pwritev) + delete(table, 299) // recvmmsg + table[306] = syscalls.Supported("syncfs", Syncfs) + delete(table, 307) // sendmmsg + table[316] = syscalls.Supported("renameat2", Renameat2) delete(table, 319) // memfd_create - delete(table, 322) // execveat - delete(table, 327) // preadv2 - delete(table, 328) // pwritev2 - delete(table, 332) // statx + table[322] = syscalls.Supported("execveat", Execveat) + table[327] = syscalls.Supported("preadv2", Preadv2) + table[328] = syscalls.Supported("pwritev2", Pwritev2) + table[332] = syscalls.Supported("statx", Statx) } diff --git a/pkg/sentry/syscalls/linux/vfs2/linux64_override_arm64.go b/pkg/sentry/syscalls/linux/vfs2/linux64_override_arm64.go index 6af5c400f..a6b367468 100644 --- a/pkg/sentry/syscalls/linux/vfs2/linux64_override_arm64.go +++ b/pkg/sentry/syscalls/linux/vfs2/linux64_override_arm64.go @@ -12,6 +12,8 @@ // See the License for the specific language governing permissions and // limitations under the License. +// +build arm64 + package vfs2 import ( diff --git a/pkg/sentry/syscalls/linux/vfs2/mmap.go b/pkg/sentry/syscalls/linux/vfs2/mmap.go new file mode 100644 index 000000000..60a43f0a0 --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/mmap.go @@ -0,0 +1,92 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/memmap" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +// Mmap implements Linux syscall mmap(2). +func Mmap(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + prot := args[2].Int() + flags := args[3].Int() + fd := args[4].Int() + fixed := flags&linux.MAP_FIXED != 0 + private := flags&linux.MAP_PRIVATE != 0 + shared := flags&linux.MAP_SHARED != 0 + anon := flags&linux.MAP_ANONYMOUS != 0 + map32bit := flags&linux.MAP_32BIT != 0 + + // Require exactly one of MAP_PRIVATE and MAP_SHARED. + if private == shared { + return 0, nil, syserror.EINVAL + } + + opts := memmap.MMapOpts{ + Length: args[1].Uint64(), + Offset: args[5].Uint64(), + Addr: args[0].Pointer(), + Fixed: fixed, + Unmap: fixed, + Map32Bit: map32bit, + Private: private, + Perms: usermem.AccessType{ + Read: linux.PROT_READ&prot != 0, + Write: linux.PROT_WRITE&prot != 0, + Execute: linux.PROT_EXEC&prot != 0, + }, + MaxPerms: usermem.AnyAccess, + GrowsDown: linux.MAP_GROWSDOWN&flags != 0, + Precommit: linux.MAP_POPULATE&flags != 0, + } + if linux.MAP_LOCKED&flags != 0 { + opts.MLockMode = memmap.MLockEager + } + defer func() { + if opts.MappingIdentity != nil { + opts.MappingIdentity.DecRef() + } + }() + + if !anon { + // Convert the passed FD to a file reference. + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // mmap unconditionally requires that the FD is readable. + if !file.IsReadable() { + return 0, nil, syserror.EACCES + } + // MAP_SHARED requires that the FD be writable for PROT_WRITE. + if shared && !file.IsWritable() { + opts.MaxPerms.Write = false + } + + if err := file.ConfigureMMap(t, &opts); err != nil { + return 0, nil, err + } + } + + rv, err := t.MemoryManager().MMap(t, opts) + return uintptr(rv), nil, err +} diff --git a/pkg/sentry/syscalls/linux/vfs2/path.go b/pkg/sentry/syscalls/linux/vfs2/path.go new file mode 100644 index 000000000..97da6c647 --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/path.go @@ -0,0 +1,94 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +func copyInPath(t *kernel.Task, addr usermem.Addr) (fspath.Path, error) { + pathname, err := t.CopyInString(addr, linux.PATH_MAX) + if err != nil { + return fspath.Path{}, err + } + return fspath.Parse(pathname), nil +} + +type taskPathOperation struct { + pop vfs.PathOperation + haveStartRef bool +} + +func getTaskPathOperation(t *kernel.Task, dirfd int32, path fspath.Path, shouldAllowEmptyPath shouldAllowEmptyPath, shouldFollowFinalSymlink shouldFollowFinalSymlink) (taskPathOperation, error) { + root := t.FSContext().RootDirectoryVFS2() + start := root + haveStartRef := false + if !path.Absolute { + if !path.HasComponents() && !bool(shouldAllowEmptyPath) { + root.DecRef() + return taskPathOperation{}, syserror.ENOENT + } + if dirfd == linux.AT_FDCWD { + start = t.FSContext().WorkingDirectoryVFS2() + haveStartRef = true + } else { + dirfile := t.GetFileVFS2(dirfd) + if dirfile == nil { + root.DecRef() + return taskPathOperation{}, syserror.EBADF + } + start = dirfile.VirtualDentry() + start.IncRef() + haveStartRef = true + dirfile.DecRef() + } + } + return taskPathOperation{ + pop: vfs.PathOperation{ + Root: root, + Start: start, + Path: path, + FollowFinalSymlink: bool(shouldFollowFinalSymlink), + }, + haveStartRef: haveStartRef, + }, nil +} + +func (tpop *taskPathOperation) Release() { + tpop.pop.Root.DecRef() + if tpop.haveStartRef { + tpop.pop.Start.DecRef() + tpop.haveStartRef = false + } +} + +type shouldAllowEmptyPath bool + +const ( + disallowEmptyPath shouldAllowEmptyPath = false + allowEmptyPath shouldAllowEmptyPath = true +) + +type shouldFollowFinalSymlink bool + +const ( + nofollowFinalSymlink shouldFollowFinalSymlink = false + followFinalSymlink shouldFollowFinalSymlink = true +) diff --git a/pkg/sentry/syscalls/linux/vfs2/poll.go b/pkg/sentry/syscalls/linux/vfs2/poll.go new file mode 100644 index 000000000..dbf4882da --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/poll.go @@ -0,0 +1,584 @@ +// 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()) + return tsRemaining.CopyOut(t, timespecAddr) +} + +// 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()) + return tvRemaining.CopyOut(t, timevalAddr) +} + +// 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..35f6308d6 --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/read_write.go @@ -0,0 +1,511 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs2 + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + slinux "gvisor.dev/gvisor/pkg/sentry/syscalls/linux" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" + "gvisor.dev/gvisor/pkg/waiter" +) + +const ( + eventMaskRead = waiter.EventIn | waiter.EventHUp | waiter.EventErr + eventMaskWrite = waiter.EventOut | waiter.EventHUp | waiter.EventErr +) + +// Read implements Linux syscall read(2). +func Read(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + addr := args[1].Pointer() + size := args[2].SizeT() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Check that the size is legitimate. + si := int(size) + if si < 0 { + return 0, nil, syserror.EINVAL + } + + // Get the destination of the read. + dst, err := t.SingleIOSequence(addr, si, usermem.IOOpts{ + AddressSpaceActive: true, + }) + if err != nil { + return 0, nil, err + } + + n, err := read(t, file, dst, vfs.ReadOptions{}) + t.IOUsage().AccountReadSyscall(n) + return uintptr(n), nil, slinux.HandleIOErrorVFS2(t, n != 0, err, kernel.ERESTARTSYS, "read", file) +} + +// Readv implements Linux syscall readv(2). +func Readv(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + addr := args[1].Pointer() + iovcnt := int(args[2].Int()) + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Get the destination of the read. + dst, err := t.IovecsIOSequence(addr, iovcnt, usermem.IOOpts{ + AddressSpaceActive: true, + }) + if err != nil { + return 0, nil, err + } + + n, err := read(t, file, dst, vfs.ReadOptions{}) + t.IOUsage().AccountReadSyscall(n) + return uintptr(n), nil, slinux.HandleIOErrorVFS2(t, n != 0, err, kernel.ERESTARTSYS, "readv", file) +} + +func read(t *kernel.Task, file *vfs.FileDescription, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { + n, err := file.Read(t, dst, opts) + if err != syserror.ErrWouldBlock || file.StatusFlags()&linux.O_NONBLOCK != 0 { + return n, err + } + + // Register for notifications. + w, ch := waiter.NewChannelEntry(nil) + file.EventRegister(&w, eventMaskRead) + + total := n + for { + // Shorten dst to reflect bytes previously read. + dst = dst.DropFirst(int(n)) + + // Issue the request and break out if it completes with anything other than + // "would block". + n, err := file.Read(t, dst, opts) + total += n + if err != syserror.ErrWouldBlock { + break + } + if err := t.Block(ch); err != nil { + break + } + } + file.EventUnregister(&w) + + return total, err +} + +// Pread64 implements Linux syscall pread64(2). +func Pread64(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + addr := args[1].Pointer() + size := args[2].SizeT() + offset := args[3].Int64() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Check that the offset is legitimate. + if offset < 0 { + return 0, nil, syserror.EINVAL + } + + // Check that the size is legitimate. + si := int(size) + if si < 0 { + return 0, nil, syserror.EINVAL + } + + // Get the destination of the read. + dst, err := t.SingleIOSequence(addr, si, usermem.IOOpts{ + AddressSpaceActive: true, + }) + if err != nil { + return 0, nil, err + } + + n, err := pread(t, file, dst, offset, vfs.ReadOptions{}) + t.IOUsage().AccountReadSyscall(n) + return uintptr(n), nil, slinux.HandleIOErrorVFS2(t, n != 0, err, kernel.ERESTARTSYS, "pread64", file) +} + +// Preadv implements Linux syscall preadv(2). +func Preadv(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + addr := args[1].Pointer() + iovcnt := int(args[2].Int()) + offset := args[3].Int64() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Check that the offset is legitimate. + if offset < 0 { + return 0, nil, syserror.EINVAL + } + + // Get the destination of the read. + dst, err := t.IovecsIOSequence(addr, iovcnt, usermem.IOOpts{ + AddressSpaceActive: true, + }) + if err != nil { + return 0, nil, err + } + + n, err := pread(t, file, dst, offset, vfs.ReadOptions{}) + t.IOUsage().AccountReadSyscall(n) + return uintptr(n), nil, slinux.HandleIOErrorVFS2(t, n != 0, err, kernel.ERESTARTSYS, "preadv", file) +} + +// Preadv2 implements Linux syscall preadv2(2). +func Preadv2(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + // While the glibc signature is + // preadv2(int fd, struct iovec* iov, int iov_cnt, off_t offset, int flags) + // the actual syscall + // (https://elixir.bootlin.com/linux/v5.5/source/fs/read_write.c#L1142) + // splits the offset argument into a high/low value for compatibility with + // 32-bit architectures. The flags argument is the 6th argument (index 5). + fd := args[0].Int() + addr := args[1].Pointer() + iovcnt := int(args[2].Int()) + offset := args[3].Int64() + flags := args[5].Int() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Check that the offset is legitimate. + if offset < -1 { + return 0, nil, syserror.EINVAL + } + + // Get the destination of the read. + dst, err := t.IovecsIOSequence(addr, iovcnt, usermem.IOOpts{ + AddressSpaceActive: true, + }) + if err != nil { + return 0, nil, err + } + + opts := vfs.ReadOptions{ + Flags: uint32(flags), + } + var n int64 + if offset == -1 { + n, err = read(t, file, dst, opts) + } else { + n, err = pread(t, file, dst, offset, opts) + } + t.IOUsage().AccountReadSyscall(n) + return uintptr(n), nil, slinux.HandleIOErrorVFS2(t, n != 0, err, kernel.ERESTARTSYS, "preadv2", file) +} + +func pread(t *kernel.Task, file *vfs.FileDescription, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { + n, err := file.PRead(t, dst, offset, opts) + if err != syserror.ErrWouldBlock || file.StatusFlags()&linux.O_NONBLOCK != 0 { + return n, err + } + + // Register for notifications. + w, ch := waiter.NewChannelEntry(nil) + file.EventRegister(&w, eventMaskRead) + + total := n + for { + // Shorten dst to reflect bytes previously read. + dst = dst.DropFirst(int(n)) + + // Issue the request and break out if it completes with anything other than + // "would block". + n, err := file.PRead(t, dst, offset+total, opts) + total += n + if err != syserror.ErrWouldBlock { + break + } + if err := t.Block(ch); err != nil { + break + } + } + file.EventUnregister(&w) + + return total, err +} + +// Write implements Linux syscall write(2). +func Write(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + addr := args[1].Pointer() + size := args[2].SizeT() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Check that the size is legitimate. + si := int(size) + if si < 0 { + return 0, nil, syserror.EINVAL + } + + // Get the source of the write. + src, err := t.SingleIOSequence(addr, si, usermem.IOOpts{ + AddressSpaceActive: true, + }) + if err != nil { + return 0, nil, err + } + + n, err := write(t, file, src, vfs.WriteOptions{}) + t.IOUsage().AccountWriteSyscall(n) + return uintptr(n), nil, slinux.HandleIOErrorVFS2(t, n != 0, err, kernel.ERESTARTSYS, "write", file) +} + +// Writev implements Linux syscall writev(2). +func Writev(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + addr := args[1].Pointer() + iovcnt := int(args[2].Int()) + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Get the source of the write. + src, err := t.IovecsIOSequence(addr, iovcnt, usermem.IOOpts{ + AddressSpaceActive: true, + }) + if err != nil { + return 0, nil, err + } + + n, err := write(t, file, src, vfs.WriteOptions{}) + t.IOUsage().AccountWriteSyscall(n) + return uintptr(n), nil, slinux.HandleIOErrorVFS2(t, n != 0, err, kernel.ERESTARTSYS, "writev", file) +} + +func write(t *kernel.Task, file *vfs.FileDescription, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) { + n, err := file.Write(t, src, opts) + if err != syserror.ErrWouldBlock || file.StatusFlags()&linux.O_NONBLOCK != 0 { + return n, err + } + + // Register for notifications. + w, ch := waiter.NewChannelEntry(nil) + file.EventRegister(&w, eventMaskWrite) + + total := n + for { + // Shorten src to reflect bytes previously written. + src = src.DropFirst(int(n)) + + // Issue the request and break out if it completes with anything other than + // "would block". + n, err := file.Write(t, src, opts) + total += n + if err != syserror.ErrWouldBlock { + break + } + if err := t.Block(ch); err != nil { + break + } + } + file.EventUnregister(&w) + + return total, err +} + +// Pwrite64 implements Linux syscall pwrite64(2). +func Pwrite64(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + addr := args[1].Pointer() + size := args[2].SizeT() + offset := args[3].Int64() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Check that the offset is legitimate. + if offset < 0 { + return 0, nil, syserror.EINVAL + } + + // Check that the size is legitimate. + si := int(size) + if si < 0 { + return 0, nil, syserror.EINVAL + } + + // Get the source of the write. + src, err := t.SingleIOSequence(addr, si, usermem.IOOpts{ + AddressSpaceActive: true, + }) + if err != nil { + return 0, nil, err + } + + n, err := pwrite(t, file, src, offset, vfs.WriteOptions{}) + t.IOUsage().AccountWriteSyscall(n) + return uintptr(n), nil, slinux.HandleIOErrorVFS2(t, n != 0, err, kernel.ERESTARTSYS, "pwrite64", file) +} + +// Pwritev implements Linux syscall pwritev(2). +func Pwritev(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + addr := args[1].Pointer() + iovcnt := int(args[2].Int()) + offset := args[3].Int64() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Check that the offset is legitimate. + if offset < 0 { + return 0, nil, syserror.EINVAL + } + + // Get the source of the write. + src, err := t.IovecsIOSequence(addr, iovcnt, usermem.IOOpts{ + AddressSpaceActive: true, + }) + if err != nil { + return 0, nil, err + } + + n, err := pwrite(t, file, src, offset, vfs.WriteOptions{}) + t.IOUsage().AccountReadSyscall(n) + return uintptr(n), nil, slinux.HandleIOErrorVFS2(t, n != 0, err, kernel.ERESTARTSYS, "pwritev", file) +} + +// Pwritev2 implements Linux syscall pwritev2(2). +func Pwritev2(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + // While the glibc signature is + // pwritev2(int fd, struct iovec* iov, int iov_cnt, off_t offset, int flags) + // the actual syscall + // (https://elixir.bootlin.com/linux/v5.5/source/fs/read_write.c#L1162) + // splits the offset argument into a high/low value for compatibility with + // 32-bit architectures. The flags argument is the 6th argument (index 5). + fd := args[0].Int() + addr := args[1].Pointer() + iovcnt := int(args[2].Int()) + offset := args[3].Int64() + flags := args[5].Int() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Check that the offset is legitimate. + if offset < -1 { + return 0, nil, syserror.EINVAL + } + + // Get the source of the write. + src, err := t.IovecsIOSequence(addr, iovcnt, usermem.IOOpts{ + AddressSpaceActive: true, + }) + if err != nil { + return 0, nil, err + } + + opts := vfs.WriteOptions{ + Flags: uint32(flags), + } + var n int64 + if offset == -1 { + n, err = write(t, file, src, opts) + } else { + n, err = pwrite(t, file, src, offset, opts) + } + t.IOUsage().AccountWriteSyscall(n) + return uintptr(n), nil, slinux.HandleIOErrorVFS2(t, n != 0, err, kernel.ERESTARTSYS, "pwritev2", file) +} + +func pwrite(t *kernel.Task, file *vfs.FileDescription, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) { + n, err := file.PWrite(t, src, offset, opts) + if err != syserror.ErrWouldBlock || file.StatusFlags()&linux.O_NONBLOCK != 0 { + return n, err + } + + // Register for notifications. + w, ch := waiter.NewChannelEntry(nil) + file.EventRegister(&w, eventMaskWrite) + + total := n + for { + // Shorten src to reflect bytes previously written. + src = src.DropFirst(int(n)) + + // Issue the request and break out if it completes with anything other than + // "would block". + n, err := file.PWrite(t, src, offset+total, opts) + total += n + if err != syserror.ErrWouldBlock { + break + } + if err := t.Block(ch); err != nil { + break + } + } + file.EventUnregister(&w) + + return total, err +} + +// Lseek implements Linux syscall lseek(2). +func Lseek(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + offset := args[1].Int64() + whence := args[2].Int() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + newoff, err := file.Seek(t, offset, whence) + return uintptr(newoff), nil, err +} diff --git a/pkg/sentry/syscalls/linux/vfs2/setstat.go b/pkg/sentry/syscalls/linux/vfs2/setstat.go new file mode 100644 index 000000000..9250659ff --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/setstat.go @@ -0,0 +1,380 @@ +// 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 + } + + return 0, nil, setstatat(t, linux.AT_FDCWD, path, disallowEmptyPath, followFinalSymlink, &vfs.SetStatOptions{ + Stat: linux.Statx{ + Mask: linux.STATX_SIZE, + Size: uint64(length), + }, + }) +} + +// 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() + + return 0, nil, file.SetStat(t, vfs.SetStatOptions{ + Stat: linux.Statx{ + Mask: linux.STATX_SIZE, + Size: uint64(length), + }, + }) +} + +// Utime implements Linux syscall utime(2). +func Utime(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + pathAddr := args[0].Pointer() + timesAddr := args[1].Pointer() + + path, err := copyInPath(t, pathAddr) + if err != nil { + return 0, nil, err + } + + opts := vfs.SetStatOptions{ + Stat: linux.Statx{ + Mask: linux.STATX_ATIME | linux.STATX_MTIME, + }, + } + if timesAddr == 0 { + opts.Stat.Atime.Nsec = linux.UTIME_NOW + opts.Stat.Mtime.Nsec = linux.UTIME_NOW + } else { + var times linux.Utime + if err := times.CopyIn(t, timesAddr); err != nil { + return 0, nil, err + } + opts.Stat.Atime.Sec = times.Actime + opts.Stat.Mtime.Sec = times.Modtime + } + + return 0, nil, setstatat(t, linux.AT_FDCWD, path, disallowEmptyPath, followFinalSymlink, &opts) +} + +// Utimes implements Linux syscall utimes(2). +func Utimes(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + pathAddr := args[0].Pointer() + timesAddr := args[1].Pointer() + + path, err := copyInPath(t, pathAddr) + if err != nil { + return 0, nil, err + } + + opts := vfs.SetStatOptions{ + Stat: linux.Statx{ + Mask: linux.STATX_ATIME | linux.STATX_MTIME, + }, + } + if timesAddr == 0 { + opts.Stat.Atime.Nsec = linux.UTIME_NOW + opts.Stat.Mtime.Nsec = linux.UTIME_NOW + } else { + var times [2]linux.Timeval + if _, err := t.CopyIn(timesAddr, ×); err != nil { + return 0, nil, err + } + opts.Stat.Atime = linux.StatxTimestamp{ + Sec: times[0].Sec, + Nsec: uint32(times[0].Usec * 1000), + } + opts.Stat.Mtime = linux.StatxTimestamp{ + Sec: times[1].Sec, + Nsec: uint32(times[1].Usec * 1000), + } + } + + return 0, nil, setstatat(t, linux.AT_FDCWD, path, disallowEmptyPath, followFinalSymlink, &opts) +} + +// Utimensat implements Linux syscall utimensat(2). +func Utimensat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + dirfd := args[0].Int() + pathAddr := args[1].Pointer() + timesAddr := args[2].Pointer() + flags := args[3].Int() + + if flags&^linux.AT_SYMLINK_NOFOLLOW != 0 { + return 0, nil, syserror.EINVAL + } + + path, err := copyInPath(t, pathAddr) + if err != nil { + return 0, nil, err + } + + var opts vfs.SetStatOptions + if err := populateSetStatOptionsForUtimens(t, timesAddr, &opts); err != nil { + return 0, nil, err + } + + return 0, nil, setstatat(t, dirfd, path, disallowEmptyPath, followFinalSymlink, &opts) +} + +// Futimens implements Linux syscall futimens(2). +func Futimens(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + timesAddr := args[1].Pointer() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + var opts vfs.SetStatOptions + if err := populateSetStatOptionsForUtimens(t, timesAddr, &opts); err != nil { + return 0, nil, err + } + + return 0, nil, file.SetStat(t, opts) +} + +func populateSetStatOptionsForUtimens(t *kernel.Task, timesAddr usermem.Addr, opts *vfs.SetStatOptions) error { + if timesAddr == 0 { + opts.Stat.Mask = linux.STATX_ATIME | linux.STATX_MTIME + opts.Stat.Atime.Nsec = linux.UTIME_NOW + opts.Stat.Mtime.Nsec = linux.UTIME_NOW + return nil + } + var times [2]linux.Timespec + if _, err := t.CopyIn(timesAddr, ×); err != nil { + return err + } + if times[0].Nsec != linux.UTIME_OMIT { + opts.Stat.Mask |= linux.STATX_ATIME + opts.Stat.Atime = linux.StatxTimestamp{ + Sec: times[0].Sec, + Nsec: uint32(times[0].Nsec), + } + } + if times[1].Nsec != linux.UTIME_OMIT { + opts.Stat.Mask |= linux.STATX_MTIME + opts.Stat.Mtime = linux.StatxTimestamp{ + Sec: times[1].Sec, + Nsec: uint32(times[1].Nsec), + } + } + return nil +} + +func setstatat(t *kernel.Task, dirfd int32, path fspath.Path, shouldAllowEmptyPath shouldAllowEmptyPath, shouldFollowFinalSymlink shouldFollowFinalSymlink, opts *vfs.SetStatOptions) error { + root := t.FSContext().RootDirectoryVFS2() + defer root.DecRef() + start := root + if !path.Absolute { + if !path.HasComponents() && !bool(shouldAllowEmptyPath) { + return syserror.ENOENT + } + if dirfd == linux.AT_FDCWD { + start = t.FSContext().WorkingDirectoryVFS2() + defer start.DecRef() + } else { + dirfile := t.GetFileVFS2(dirfd) + if dirfile == nil { + return syserror.EBADF + } + if !path.HasComponents() { + // Use FileDescription.SetStat() instead of + // VirtualFilesystem.SetStatAt(), since the former may be able + // to use opened file state to expedite the SetStat. + err := dirfile.SetStat(t, *opts) + dirfile.DecRef() + return err + } + start = dirfile.VirtualDentry() + start.IncRef() + defer start.DecRef() + dirfile.DecRef() + } + } + return t.Kernel().VFS().SetStatAt(t, t.Credentials(), &vfs.PathOperation{ + Root: root, + Start: start, + Path: path, + FollowFinalSymlink: bool(shouldFollowFinalSymlink), + }, opts) +} diff --git a/pkg/sentry/syscalls/linux/vfs2/stat.go b/pkg/sentry/syscalls/linux/vfs2/stat.go new file mode 100644 index 000000000..97eaedd66 --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/stat.go @@ -0,0 +1,377 @@ +// 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/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) + return stat.CopyOut(t, statAddr) + } + 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) + return stat.CopyOut(t, statAddr) +} + +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) + return 0, nil, stat.CopyOut(t, statAddr) +} + +// 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 + } + + 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) + return 0, nil, statx.CopyOut(t, statxAddr) + } + 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) + return 0, nil, statx.CopyOut(t, statxAddr) +} + +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 + } + + // 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 + } + + return 0, nil, statfs.CopyOut(t, bufAddr) +} + +// 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 + } + + return 0, nil, statfs.CopyOut(t, bufAddr) +} 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/xattr.go b/pkg/sentry/syscalls/linux/vfs2/xattr.go new file mode 100644 index 000000000..89e9ff4d7 --- /dev/null +++ b/pkg/sentry/syscalls/linux/vfs2/xattr.go @@ -0,0 +1,353 @@ +// 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) + 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) + 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, name) + 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, name) + if err != nil { + return 0, nil, err + } + n, err := copyOutXattrValue(t, valueAddr, size, value) + if err != nil { + return 0, nil, err + } + return uintptr(n), nil, nil +} + +// Setxattr implements Linux syscall setxattr(2). +func Setxattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + return 0, nil, setxattr(t, args, followFinalSymlink) +} + +// Lsetxattr implements Linux syscall lsetxattr(2). +func Lsetxattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + return 0, nil, setxattr(t, args, nofollowFinalSymlink) +} + +func setxattr(t *kernel.Task, args arch.SyscallArguments, shouldFollowFinalSymlink shouldFollowFinalSymlink) error { + pathAddr := args[0].Pointer() + nameAddr := args[1].Pointer() + valueAddr := args[2].Pointer() + size := args[3].SizeT() + flags := args[4].Int() + + if flags&^(linux.XATTR_CREATE|linux.XATTR_REPLACE) != 0 { + return syserror.EINVAL + } + + path, err := copyInPath(t, pathAddr) + if err != nil { + return err + } + tpop, err := getTaskPathOperation(t, linux.AT_FDCWD, path, disallowEmptyPath, shouldFollowFinalSymlink) + if err != nil { + return err + } + defer tpop.Release() + + name, err := copyInXattrName(t, nameAddr) + if err != nil { + return err + } + value, err := copyInXattrValue(t, valueAddr, size) + if err != nil { + return err + } + + return t.Kernel().VFS().SetxattrAt(t, t.Credentials(), &tpop.pop, &vfs.SetxattrOptions{ + Name: name, + Value: value, + Flags: uint32(flags), + }) +} + +// Fsetxattr implements Linux syscall fsetxattr(2). +func Fsetxattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + nameAddr := args[1].Pointer() + valueAddr := args[2].Pointer() + size := args[3].SizeT() + flags := args[4].Int() + + if flags&^(linux.XATTR_CREATE|linux.XATTR_REPLACE) != 0 { + return 0, nil, syserror.EINVAL + } + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + name, err := copyInXattrName(t, nameAddr) + if err != nil { + return 0, nil, err + } + value, err := copyInXattrValue(t, valueAddr, size) + if err != nil { + return 0, nil, err + } + + return 0, nil, file.Setxattr(t, vfs.SetxattrOptions{ + Name: name, + Value: value, + Flags: uint32(flags), + }) +} + +// Removexattr implements Linux syscall removexattr(2). +func Removexattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + return 0, nil, removexattr(t, args, followFinalSymlink) +} + +// Lremovexattr implements Linux syscall lremovexattr(2). +func Lremovexattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + return 0, nil, removexattr(t, args, nofollowFinalSymlink) +} + +func removexattr(t *kernel.Task, args arch.SyscallArguments, shouldFollowFinalSymlink shouldFollowFinalSymlink) error { + pathAddr := args[0].Pointer() + nameAddr := args[1].Pointer() + + path, err := copyInPath(t, pathAddr) + if err != nil { + return err + } + tpop, err := getTaskPathOperation(t, linux.AT_FDCWD, path, disallowEmptyPath, shouldFollowFinalSymlink) + if err != nil { + return err + } + defer tpop.Release() + + name, err := copyInXattrName(t, nameAddr) + if err != nil { + return err + } + + return t.Kernel().VFS().RemovexattrAt(t, t.Credentials(), &tpop.pop, name) +} + +// Fremovexattr implements Linux syscall fremovexattr(2). +func Fremovexattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + nameAddr := args[1].Pointer() + + file := t.GetFileVFS2(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + name, err := copyInXattrName(t, nameAddr) + if err != nil { + return 0, nil, err + } + + return 0, nil, file.Removexattr(t, name) +} + +func copyInXattrName(t *kernel.Task, nameAddr usermem.Addr) (string, error) { + name, err := t.CopyInString(nameAddr, linux.XATTR_NAME_MAX+1) + if err != nil { + if err == syserror.ENAMETOOLONG { + return "", syserror.ERANGE + } + return "", err + } + if len(name) == 0 { + return "", syserror.ERANGE + } + return name, nil +} + +func copyOutXattrNameList(t *kernel.Task, listAddr usermem.Addr, size uint, names []string) (int, error) { + if size > linux.XATTR_LIST_MAX { + size = linux.XATTR_LIST_MAX + } + var buf bytes.Buffer + for _, name := range names { + buf.WriteString(name) + buf.WriteByte(0) + } + if size == 0 { + // Return the size that would be required to accomodate the list. + return buf.Len(), nil + } + if buf.Len() > int(size) { + if size >= linux.XATTR_LIST_MAX { + return 0, syserror.E2BIG + } + return 0, syserror.ERANGE + } + return t.CopyOutBytes(listAddr, buf.Bytes()) +} + +func copyInXattrValue(t *kernel.Task, valueAddr usermem.Addr, size uint) (string, error) { + if size > linux.XATTR_SIZE_MAX { + return "", syserror.E2BIG + } + buf := make([]byte, size) + if _, err := t.CopyInBytes(valueAddr, buf); err != nil { + return "", err + } + return gohacks.StringFromImmutableBytes(buf), nil +} + +func copyOutXattrValue(t *kernel.Task, valueAddr usermem.Addr, size uint, value string) (int, error) { + if size > linux.XATTR_SIZE_MAX { + size = linux.XATTR_SIZE_MAX + } + if size == 0 { + // Return the size that would be required to accomodate the value. + return len(value), nil + } + if len(value) > int(size) { + if size >= linux.XATTR_SIZE_MAX { + return 0, syserror.E2BIG + } + return 0, syserror.ERANGE + } + return t.CopyOutBytes(valueAddr, gohacks.ImmutableBytesFromString(value)) +} diff --git a/pkg/sentry/vfs/BUILD b/pkg/sentry/vfs/BUILD index 0b4f18ab5..cb4deb068 100644 --- a/pkg/sentry/vfs/BUILD +++ b/pkg/sentry/vfs/BUILD @@ -42,9 +42,13 @@ go_library( deps = [ "//pkg/abi/linux", "//pkg/context", + "//pkg/fd", "//pkg/fspath", + "//pkg/gohacks", "//pkg/log", + "//pkg/safemem", "//pkg/sentry/arch", + "//pkg/sentry/fs", "//pkg/sentry/fs/lock", "//pkg/sentry/kernel/auth", "//pkg/sentry/memmap", @@ -52,6 +56,7 @@ go_library( "//pkg/syserror", "//pkg/usermem", "//pkg/waiter", + "@org_golang_x_sys//unix:go_default_library", ], ) diff --git a/pkg/sentry/vfs/anonfs.go b/pkg/sentry/vfs/anonfs.go index 2db25be49..925996517 100644 --- a/pkg/sentry/vfs/anonfs.go +++ b/pkg/sentry/vfs/anonfs.go @@ -41,7 +41,14 @@ 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 +) // anonFilesystem is the implementation of FilesystemImpl that backs // VirtualDentries returned by VirtualFilesystem.NewAnonVirtualDentry(). @@ -69,6 +76,16 @@ func (fs *anonFilesystem) Sync(ctx context.Context) error { return nil } +// AccessAt implements vfs.Filesystem.Impl.AccessAt. +// +// TODO(gvisor.dev/issue/1965): Implement access permissions. +func (fs *anonFilesystem) AccessAt(ctx context.Context, rp *ResolvingPath, creds *auth.Credentials, ats AccessTypes) error { + if !rp.Done() { + return syserror.ENOTDIR + } + return GenericCheckPermissions(creds, ats, false /* isDir */, anonFileMode, anonFileUID, anonFileGID) +} + // GetDentryAt implements FilesystemImpl.GetDentryAt. func (fs *anonFilesystem) GetDentryAt(ctx context.Context, rp *ResolvingPath, opts GetDentryOptions) (*Dentry, error) { if !rp.Done() { @@ -167,9 +184,9 @@ func (fs *anonFilesystem) StatAt(ctx context.Context, rp *ResolvingPath, opts St Mask: linux.STATX_TYPE | linux.STATX_MODE | linux.STATX_NLINK | linux.STATX_UID | linux.STATX_GID | linux.STATX_INO | linux.STATX_SIZE | linux.STATX_BLOCKS, Blksize: anonfsBlockSize, Nlink: 1, - UID: uint32(auth.RootKUID), - GID: uint32(auth.RootKGID), - Mode: 0600, // no type is correct + UID: uint32(anonFileUID), + GID: uint32(anonFileGID), + Mode: anonFileMode, Ino: 1, Size: 0, Blocks: 0, diff --git a/pkg/sentry/vfs/epoll.go b/pkg/sentry/vfs/epoll.go index eed41139b..3da45d744 100644 --- a/pkg/sentry/vfs/epoll.go +++ b/pkg/sentry/vfs/epoll.go @@ -202,6 +202,9 @@ func (ep *EpollInstance) AddInterest(file *FileDescription, num int32, event lin // 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() diff --git a/pkg/sentry/vfs/file_description_impl_util.go b/pkg/sentry/vfs/file_description_impl_util.go index c2a52ec1b..45191d1c3 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. diff --git a/pkg/sentry/vfs/filesystem.go b/pkg/sentry/vfs/filesystem.go index 556976d0b..c43dcff3d 100644 --- a/pkg/sentry/vfs/filesystem.go +++ b/pkg/sentry/vfs/filesystem.go @@ -20,6 +20,7 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" ) // A Filesystem is a tree of nodes represented by Dentries, which forms part of @@ -144,6 +145,9 @@ 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. // diff --git a/pkg/sentry/vfs/mount.go b/pkg/sentry/vfs/mount.go index 9912df799..05f6233f9 100644 --- a/pkg/sentry/vfs/mount.go +++ b/pkg/sentry/vfs/mount.go @@ -74,6 +74,10 @@ type Mount struct { // umounted is true. umounted is protected by VirtualFilesystem.mountMu. umounted bool + // flags contains settings as specified for mount(2), e.g. MS_NOEXEC, except + // for MS_RDONLY which is tracked in "writers". + flags MountFlags + // The lower 63 bits of writers is the number of calls to // Mount.CheckBeginWrite() that have not yet been paired with a call to // Mount.EndWrite(). The MSB of writers is set if MS_RDONLY is in effect. @@ -81,6 +85,21 @@ type Mount struct { writers int64 } +func newMount(vfs *VirtualFilesystem, fs *Filesystem, root *Dentry, mntns *MountNamespace, opts *MountOptions) *Mount { + mnt := &Mount{ + vfs: vfs, + fs: fs, + root: root, + flags: opts.Flags, + ns: mntns, + refs: 1, + } + if opts.ReadOnly { + mnt.setReadOnlyLocked(true) + } + return mnt +} + // A MountNamespace is a collection of Mounts. // // MountNamespaces are reference-counted. Unless otherwise specified, all @@ -129,16 +148,22 @@ func (vfs *VirtualFilesystem) NewMountNamespace(ctx context.Context, creds *auth refs: 1, mountpoints: make(map[*Dentry]uint32), } - mntns.root = &Mount{ - vfs: vfs, - fs: fs, - root: root, - ns: mntns, - refs: 1, - } + mntns.root = newMount(vfs, fs, root, mntns, &MountOptions{}) return mntns, nil } +// NewDisconnectedMount returns a Mount representing fs with the given root +// (which may be nil). The new Mount is not associated with any MountNamespace +// and is not connected to any other Mounts. References are taken on fs and +// root. +func (vfs *VirtualFilesystem) NewDisconnectedMount(fs *Filesystem, root *Dentry, opts *MountOptions) (*Mount, error) { + fs.IncRef() + if root != nil { + root.IncRef() + } + return newMount(vfs, fs, root, nil /* mntns */, opts), nil +} + // MountAt creates and mounts a Filesystem configured by the given arguments. func (vfs *VirtualFilesystem) MountAt(ctx context.Context, creds *auth.Credentials, source string, target *PathOperation, fsTypeName string, opts *MountOptions) error { rft := vfs.getFilesystemType(fsTypeName) @@ -201,13 +226,7 @@ func (vfs *VirtualFilesystem) MountAt(ctx context.Context, creds *auth.Credentia // are directories, or neither are, and returns ENOTDIR if this is not the // case. mntns := vd.mount.ns - mnt := &Mount{ - vfs: vfs, - fs: fs, - root: root, - ns: mntns, - refs: 1, - } + mnt := newMount(vfs, fs, root, mntns, opts) vfs.mounts.seq.BeginWrite() vfs.connectLocked(mnt, vd, mntns) vfs.mounts.seq.EndWrite() diff --git a/pkg/sentry/vfs/mount_unsafe.go b/pkg/sentry/vfs/mount_unsafe.go index 1fe766a44..bc7581698 100644 --- a/pkg/sentry/vfs/mount_unsafe.go +++ b/pkg/sentry/vfs/mount_unsafe.go @@ -26,6 +26,7 @@ import ( "sync/atomic" "unsafe" + "gvisor.dev/gvisor/pkg/gohacks" "gvisor.dev/gvisor/pkg/sync" ) @@ -160,7 +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 { @@ -361,12 +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 6af7fdac1..3e90dc4ed 100644 --- a/pkg/sentry/vfs/options.go +++ b/pkg/sentry/vfs/options.go @@ -46,8 +46,21 @@ type MknodOptions struct { DevMinor uint32 } +// MountFlags contains flags as specified for mount(2), e.g. MS_NOEXEC. +// MS_RDONLY is not part of MountFlags because it's tracked in Mount.writers. +type MountFlags struct { + // NoExec is equivalent to MS_NOEXEC. + NoExec bool +} + // MountOptions contains options to VirtualFilesystem.MountAt(). type MountOptions struct { + // Flags contains flags as specified for mount(2), e.g. MS_NOEXEC. + Flags MountFlags + + // ReadOnly is equivalent to MS_RDONLY. + ReadOnly bool + // GetFilesystemOptions contains options to FilesystemType.GetFilesystem(). GetFilesystemOptions GetFilesystemOptions @@ -75,7 +88,8 @@ 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 } diff --git a/pkg/sentry/vfs/resolving_path.go b/pkg/sentry/vfs/resolving_path.go index 8a0b382f6..eb4ebb511 100644 --- a/pkg/sentry/vfs/resolving_path.go +++ b/pkg/sentry/vfs/resolving_path.go @@ -228,7 +228,7 @@ 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] } } diff --git a/pkg/sentry/vfs/vfs.go b/pkg/sentry/vfs/vfs.go index 8f29031b2..2e2880171 100644 --- a/pkg/sentry/vfs/vfs.go +++ b/pkg/sentry/vfs/vfs.go @@ -126,17 +126,23 @@ func (vfs *VirtualFilesystem) Init() error { // Construct vfs.anonMount. anonfsDevMinor, err := vfs.GetAnonBlockDevMinor() if err != nil { - return 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, + 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 nil } @@ -168,6 +174,23 @@ 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) { @@ -382,18 +405,19 @@ func (vfs *VirtualFilesystem) OpenAt(ctx context.Context, creds *auth.Credential // 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 t := linux.FileMode(stat.Mode).FileType(); t != linux.ModeRegular { - ctx.Infof("%q is not a regular file: %v", pop.Path, t) + if stat.Mask&linux.STATX_TYPE == 0 || stat.Mode&linux.S_IFMT != linux.S_IFREG { + fd.DecRef() return nil, syserror.EACCES } } diff --git a/pkg/sentry/watchdog/watchdog.go b/pkg/sentry/watchdog/watchdog.go index bfb2fac26..f7d6009a0 100644 --- a/pkg/sentry/watchdog/watchdog.go +++ b/pkg/sentry/watchdog/watchdog.go @@ -221,7 +221,7 @@ func (w *Watchdog) waitForStart() { return } var buf bytes.Buffer - buf.WriteString("Watchdog.Start() not called within %s:\n") + buf.WriteString(fmt.Sprintf("Watchdog.Start() not called within %s", w.StartupTimeout)) w.doAction(w.StartupTimeoutAction, false, &buf) } @@ -325,7 +325,7 @@ func (w *Watchdog) report(offenders map[*kernel.Task]*offender, newTaskFound boo 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) } @@ -359,7 +359,7 @@ func (w *Watchdog) doAction(action Action, skipStack bool, msg *bytes.Buffer) { case <-metricsEmitted: case <-time.After(1 * time.Second): } - panic(fmt.Sprintf("Stack for running G's are skipped while panicking.\n%s", msg.String())) + panic(fmt.Sprintf("%s\nStack for running G's are skipped while panicking.", msg.String())) default: panic(fmt.Sprintf("Unknown watchdog action %v", action)) |